3DSIG

• Xiao Chen, University of Missouri - Columbia, United States
• Alex Morehead, University of Missouri - Columbia, United States
• Jian Liu, University of Missouri - Columbia, United States
• Jianlin Cheng, University of Missouri - Columbia, United States

Motivation: Proteins interact to form complexes to carry out essential biological functions. Computational methods such as AlphaFold-multimer have been developed to predict the quaternary structures of protein complexes. An important yet largely unsolved challenge in protein complex structure prediction is to accurately estimate the quality of predicted protein complex structures without any knowledge of the corresponding native structures. Such estimations can then be used to select high-quality predicted complex structures to facilitate biomedical research such as protein function analysis and drug discovery.
Results: In this work, we introduce a new gated neighborhood-modulating graph transformer to predict the quality of 3D protein complex structures. It incorporates node and edge gates within a graph transformer framework to control information flow during graph message passing. We trained, evaluated and tested the method (called DProQA) on newly-curated protein complex datasets before the 15th Critical Assessment of Techniques for Protein Structure Prediction (CASP15) and then blindly tested it in the 2022 CASP15 experiment. The method was ranked 3rd among the single-model quality assessment methods in CASP15 in terms of the ranking loss of TM-score on 36 complex targets. The rigorous internal and external experiments demonstrate that DProQA is effective in ranking protein complex structures.
Availability: The source code, data, and pre-trained models are available at https://github.com/jianlin-cheng/DProQA

• Jia-Ning Li, School of Life Sciences, Northwestern Polytechnical University, China
• Guang Yang, School of Life Sciences, Northwestern Polytechnical University, China
• Peng-Cheng Zhao, School of Life Sciences, Northwestern Polytechnical University, China
• Xue-Xin Wei, School of Life Sciences, Northwestern Polytechnical University, China
• Jian-Yu Shi, School of Life Sciences, Northwestern Polytechnical University, China

Motivation: Deep learning-based molecule generation becomes a new paradigm of de novo mol-ecule design since it enables fast and directional exploration in the vast chemical space. However, it is still an open issue to generate molecules, which bind to specific proteins with high binding af-finities while owning desired drug-like physicochemical properties.
Results: To address these issues, we elaborate a novel framework for controllable protein-oriented molecule generation, named CProMG, which contains a 3-D protein embedding module, a dual-view protein encoder, a molecule embedding module, and a novel drug-like molecule decoder. Based on fusing the hierarchical views of proteins, it enhances the representation of protein binding pockets significantly by associating amino acid residues with their comprising atoms. Through joint-ly embedding molecule sequences, their drug-like properties, and binding affinities w.r.t. proteins, it autoregressively generates novel molecules having specific properties in a controllable manner by measuring the proximity of molecule tokens to protein residues and atoms. The comparison with state-of-the-art deep generative methods demonstrates the superiority of our CProMG. Furthermore, the progressive control of properties demonstrates the effectiveness of CProMG when controlling binding affinity and drug-like properties. After that, the ablation studies reveal how its crucial compo-nents contribute to the model respectively, including hierarchical protein views, Laplacian position encoding as well as property control. Last, a case study w.r.t protein illustrates the novelty of CProMG and the ability to capture crucial interactions between protein pockets and molecules. It’s anticipated that this work can boost de novo molecule design.

• Piyumi Amarasinghe, Monash University, Australia
• Lloyd Allison, Monash University, Australia
• Peter Stuckey, Monash University, Australia
• Maria Garcia de la Banda, Monash University, Australia
• Arthur Lesk, Pennsylvania State University, United States
• Arun Konagurthu, Monash University, Australia

The tendency of an amino acid to adopt certain configurations in folded proteins is treated here as a statistical estimation problem. We model the joint distribution of the observed mainchain and sidechain dihedral angles (<φ,ψ,χ1,χ2,...>) of any amino acid by a mixture of a product of von Mises probability distributions. This mixture model maps any vector of dihedral angles to a point on a multi-dimensional torus. The continuous space it uses to specify the dihedral angles provides an alternative to the commonly-used rotamer libraries. These rotamer libraries discretize the space of dihedral angles into coarse angular bins, and cluster combinations of sidechain dihedral angles (<χ1,χ2,...>) as a function of backbone <φ,ψ> conformations. A ‘good‘ model is one that is both concise and explains (compresses) observed data. Competing models can be compared directly and in particular our model is shown to outperform the Dunbrack rotamer library in terms of model-complexity (by three-orders of magnitude) and its fidelity (on average 20% more compression) when losslessly explaining the observed dihedral angle data across experimental resolutions of structures. Our method is unsupervised (with parameters estimated automatically) and uses information-theory to determine the optimal complexity of the statistical model, thus avoiding under/over-fitting, a common pitfall in model-selection problems. Our models are computationally inexpensive to sample from and are geared to support a number of downstream studies, ranging from experimental structure refinement, de novo protein design, and protein structure prediction. We call our collection of mixture models PhiSiCal (φψχal). It is available for download from http://lcb.infotech.monash.edu.au/phisical.

• Md Mahfuzur Rahaman, University of Central Florida, United States
• Nabila Shahnaz Khan, University of Central Florida, United States
• Shaojie Zhang, University of Central Florida, United States

Motivation: The 3D structures of RNA play a critical role in understanding their functionalities. There exist several computational methods to study RNA 3D structures by identifying structural motifs and categorizing them into several motif families based on their structures. Although the number of such motif families is not limited, a few are well-studied. Out of these structural motif families, there exists several families that are visually similar or very close in structure, even with different base interactions. Alternatively, some motif families share a set of base interactions but maintain variation in their 3D formations. These similarities among different motif families, if known, can provide a better insight into the RNA 3D structural motifs as well as their characteristic functions in cell biology.

Results: In this work, we proposed a method, RNAMotifComp, that analyzes the instances of well-known structural motif families and establishes a relational graph among them. We also have designed a method to visualize the relational graph where the families are shown as nodes and their similarity information is represented as edges. We validated our discovered correlations of the motif families using RNAMotifContrast. Additionally, we used a basic Na\""ive Bayes classifier to show the importance of RNAMotifComp. The relational analysis explains the functional analogies of divergent motif families and illustrates the situations where the motifs of disparate families are predicted to be of the same family.

Bio-Ontologies

• Ke Zhang, ShanghaiTech University, China
• Min Wu, I2R, A*STAR, Singapore
• Yong Liu, Nanyang Technological University, Singapore
• Yimiao Feng, ShanghaiTech University, China
• Jie Zheng, ShanghaiTech University, China

Motivation: Synthetic lethality (SL) is a promising strategy for anti-cancer therapy, as inhibiting SL partners of genes with cancer-specific mutations can selectively kill the cancer cells without harming the normal cells. Wet-lab techniques for SL screening have issues like high cost and off-target effect. Computational methods can help address these issues. Previous machine learning methods leverage known SL pairs, and the use of knowledge graph (KG) can significantly enhance the prediction performance. However, the subgraph structures of KG have not been fully explored. Besides, most machine learning methods lack interpretability, which is an obstacle for wide applications of machine learning to SL identification.
Results: We present a model named KR4SL to predict SL partners for a given primary gene. It captures the structural semantics of a KG by efficiently constructing and learning from relational digraphs in the KG. To encode the semantic information of the relational digraphs, we fuse textual semantics of entities into propagated messages and enhance the sequential semantics of paths using recurrent neural network. Moreover, we design an attentive aggregator to identify critical subgraph structures that contributed the most to the SL prediction as explanations. Extensive experiments under different settings show that KR4SL significantly outperforms all baselines. The explanatory subgraphs for the predicted gene pairs can unveil prediction process and mechanisms underlying synthetic lethality. The improved predictive power and interpretability indicate that deep learning is practically useful for SL-based cancer drug target discovery.
Availability: The source code is freely available at https://github.com/JieZheng-ShanghaiTech/KR4SL

CAMDA

• Mohammad Al Olaimat, University of North Texas, United States
• Fahad Saeed, Florida International University, United States
• Serdar Bozdag, University of North Texas, United States
• Jared Martinez, University of North Texas, United States

Alzheimer's disease (AD) is a neurodegenerative disease that affects millions of people worldwide. Mild cognitive impairment (MCI) is an intermediary stage between cognitively normal (CN) state and AD. Not all people who have MCI convert to AD. The diagnosis of AD is made after significant symptoms of dementia such as short-term memory loss are already present. Since AD is currently an irreversible disease, diagnosis at the onset of disease brings a huge burden on patients, their caregivers, and the healthcare sector. Thus, there is a crucial need to develop methods for the early prediction AD for patients who have MCI. Recurrent Neural Networks (RNN) have been successfully used to handle Electronic Health Records (EHR) for predicting conversion from MCI to AD. However, RNN ignores irregular time intervals between successive events which occurs common in EHR data. In this study, we propose two deep learning architectures based on RNN, namely Predicting Progression of Alzheimer’s Disease (PPAD) and PPAD-Autoencoder (PPAD-AE). PPAD and PPAD-AE are designed for early predicting conversion from MCI to AD at the next visit and multiple visits ahead for patients, respectively. To minimize the effect of the irregular time intervals between visits, we propose using age in each visit as an indicator of time change between successive visits. Our experimental results conducted on Alzheimer’s Disease Neuroimaging Initiative (ADNI) and National Alzheimer’s Coordinating Center (NACC) datasets showed that our proposed models outperformed all baseline models for most prediction scenarios in terms of F2 and sensitivity.

Education

• Thomas Cokelaer, Institut Pasteur, France
• Sarah Cohen-Boulakia, Université Paris-Saclay, France
• Frédéric Lemoine, Institut Pasteur, France

Motivation: The reproducibility crisis has highlighted the importance of improving the way bioinformatics data analyses are implemented, executed, and shared. To address this, various tools such as content versioning systems, workflow management systems, and software environment management systems have been developed. While these tools are becoming more widely used, there is still much work to be done to increase their adoption. The most effective way to ensure reproducibility becomes a standard part of most bioinformatics data analysis projects is to integrate it into the curriculum of bioinformatics Master's programs.
Results: In this manuscript, we present the Reprohackathon, a Master's course that we have been running for the last three years at Université Paris-Saclay (France), and that has been attended by a total of 123 students. The course is divided into two parts. The first part includes lessons on the challenges related to reproducibility, content versioning systems, container management, and workflow systems. In the second part, students work on a data analysis project for 3-4 months, reanalyzing data from a previously published study. The Reprohackaton has taught us many valuable lessons, such as the fact that implementing reproducible analyses is a complex and challenging task that requires significant effort. However, providing in-depth teaching of the concepts and the tools during a Master's degree program greatly improves students' understanding and abilities in this area.

EvolCompGen

• Ziyun Guang, Carleton College, United States
• Matthew Smith-Erb, Carleton College, United States
• Layla Oesper, Carleton College, United States

Motivation: The acquisition of somatic mutations by a tumor can be modeled by a type of evolutionary tree. However, it is impossible to observe this tree directly. Instead, numerous algorithms have been developed to infer such a tree from different types of sequencing data. But such methods can produce conflicting trees for the same patient, making it desirable to have approaches that can combine several such tumor trees into a consensus or summary tree. We introduce The Weighted m-Tumor Tree Consensus Problem (W-m-TTCP) to find a consensus tree among multiple plausible tumor evolutionary histories, each assigned a confidence weight, given a specific distance measure between tumor trees. We present an algorithm called TuLiP that is based on integer linear programming (ILP) which solves the W-m-TTCP, and unlike other existing consensus methods, allows the input trees to be weighted differently.

Results: On simulated data we show that TuLiP outperforms two existing methods at correctly identifying the true underlying tree used to create the simulations. We also show that the incorporation of weights can lead to more accurate tree inference. On a Triple-Negative Breast Cancer data set we show that including confidence weights can have important impacts on the consensus tree identified.

• Kirti Biharie, Delft University of Technology, Netherlands
• Lieke Michielsen, Leiden University Medical Center, Netherlands
• Marcel Reinders, Delft University of Technology, Netherlands
• Ahmed Mahfouz, Leiden University Medical Center, Netherlands

Motivation: Knowing the relation between cell types is crucial for translating experimental results from mice to humans. Establishing cell type matches, however, is hindered by the biological differences between the species. A substantial amount of evolutionary information between genes that could be used to align the species is discarded by most of the current methods since they only use one-to-one orthologous genes. Some methods try to retain the information by explicitly including the relation between genes, however, not without caveats.
Results: In this work, we present a model to transfer and align cell types in cross-species analysis (TACTiCS). First, TACTiCS uses a natural language processing model to match genes using their protein sequences. Next, TACTiCS employs a neural network to classify cell types within a species. Afterward, TACTiCS uses transfer learning to propagate cell type labels between species. We applied TACTiCS on scRNA-seq data of the primary motor cortex of human, mouse, and marmoset. Our model can accurately match and align cell types on these datasets. Moreover, our model outperforms Seurat and the state-of-the-art method SAMap. Finally, we show that our gene matching method results in better cell type matches than BLAST in our model.

• Hanqing Zhao, University of Twente, Netherlands
• Matthijs Souilljee, University of Twente, Netherlands
• Pavlos Pavlidis, Foundation for Research and Technology-Hellas, Greece
• Nikolaos Alachiotis, University of Twente, Netherlands

Motivation: Recent methods for selective sweep detection cast the problem as a classification task and use summary statistics as features to capture region characteristics that are indicative of a selective sweep, thereby being sensitive to confounding factors. Furthermore, they are not designed to perform whole-
genome scans or to estimate the extent of the genomic region that was affected by positive selection; both are required for identifying candidate genes and the time and strength of selection.
Results: We present ASDEC (https://github.com/pephco/ASDEC), a neural-network-based framework which can scan whole genomes for selective sweeps. ASDEC achieves similar classification performance to other CNN-based classifiers that rely on summary statistics, but it is trained 10x faster and classifies genomic regions 5x faster by inferring region characteristics from the raw sequence data directly. Deploying
ASDEC for genomic scans achieved up to 15.2x higher sensitivity, 19.4x higher success rates, and 4x higher detection accuracy than state-of-the-art methods. We used ASDEC to scan human chromosome 1 of the Yoruba population (1000Genomes project), identifying 9 known candidate genes.

• Siddhant Grover, Iowa State University, United States
• Alexey Markin, USDA-ARS, United States
• Tavis Anderson, USDA-ARS, United States
• Oliver Eulenstein, Iowa State University, United States

The classic quantitative measure of phylogenetic diversity, PD, has been used to address problems in conservation biology, microbial ecology, and evolutionary biology. PD is the minimum total length of the branches in a phylogeny required to cover a specified set of taxa on the phylogeny. A general goal in the application of PD has been identifying a set of taxa of size k that maximize PD on a given phylogeny; this has been mirrored in active research to develop efficient algorithms for the problem. Other descriptive statistics, such as the minimum PD, average PD, and standard deviation of PD, can provide invaluable insight into the distribution of PD across a phylogeny (relative to a fixed value of k). However, there has been limited or no research on computing these statistics, especially when required for each clade in a phylogeny, enabling direct comparisons of PD between clades. We introduce efficient algorithms for computing PD and the associated descriptive statistics for a given phylogeny and each of its clades. In simulation studies, we demonstrate the ability of our algorithms to analyze large-scale phylogenies with applications in ecology and evolutionary biology.
Availability: The software is available at https://github.com/flu-crew/PD_stats.

• Yasamin Tabatabaee, University of Illinois at Urbana-Champaign, United States
• Chao Zhang, University of California at Berkeley, United States
• Tandy Warnow, University of Illinois at Urbana-Champaign, United States
• Siavash Mirarab, University of California, San Diego, United States

Branch lengths and topology of a species tree are essential in most downstream analyses, including estimation of diversification dates, characterization of selection, understanding adaptation, and comparative genomics. Modern phylogenomic analyses often use methods that account for the heterogeneity of evolutionary histories across the genome due to processes such as incomplete lineage sorting. However, these methods typically do not generate branch lengths in units that are usable by downstream applications, forcing phylogenomic analyses to resort to alternative shortcuts such as estimating branch lengths by concatenating gene alignments into a supermatrix. Yet, concatenation and other available approaches for estimating branch lengths fail to address heterogeneity across the genome. In this paper, we derive expected values of gene tree branch lengths in substitution units under an extension of the multi-species coalescent (MSC) model that allows substitutions with varying rates across the species tree. We present CASTLES, a new technique for estimating branch lengths on the species tree from estimated gene trees that uses these expected values, and our study shows that CASTLES improves on the most accurate prior methods with respect to both speed and accuracy.

Function

• Frimpong Boadu, University of Missouri - Columbia, United States
• Hongyuan Cao, University of Missouri - Columbia, United States
• Jianlin Cheng, University of Missouri - Columbia, United States

Motivation: Millions of protein sequences have been generated by numerous genome and transcriptome sequencing projects. However, experimentally determining the function of the proteins is still a time consuming, low-throughput, and expensive process, leading to a large protein sequence-function gap. Therefore, it is important to develop computational methods to accurately predict protein function to fill the gap. Even though many methods have been developed to use protein sequences as input to predict function, much fewer methods leverage protein structures in protein function prediction because there was lack of accurate protein structures for most proteins until recently.
Results: We developed TransFun - a method using a transformer-based protein language model and 3D-equivariant graph neural networks to distill information from both protein sequences and structures to predict protein function. It extracts feature embeddings from protein sequences using a pre-trained protein language model (ESM) via transfer learning and combines them with 3D structures of proteins predicted by AlphaFold2 through equivariant graph neural networks. Benchmarked on the CAFA3 test dataset and a new test dataset, TransFun outperforms several state-of-the-art methods, indicating the language model and 3D-equivariant graph neural networks are effective methods to leverage protein sequences and structures to improve protein function prediction. Combining TransFun predictions and sequence similarity-based predictions can further increase prediction accuracy.
Availability: The source code of TransFun is available at https://github.com/jianlin-cheng/TransFun .

• Alexandru Dumitrescu, Aalto University, Department of Computer Science, Finland
• Emmi Jokinen, Aalto University, Department of Computer Science, Finland
• Anja Paatero, University of Helsinki, Institute of Biotechnology, Finland
• Juho Kellosalo, University of Helsinki, Institute of Biotechnology, Finland
• Ville Paavilainen, University of Helsinki, Institute of Biotechnology, Finland
• Harri Lähdesmäki, Aalto University, Department of Computer Science, Finland

Motivation: Signal peptides are short amino acid segments present at the N-terminus of newly synthesized proteins that facilitate protein translocation into the lumen of the endoplasmic reticulum, after which they are cleaved off. Specific regions of signal peptides influence the efficiency of protein translocation, and small changes in their primary structure can abolish protein secretion altogether. The lack of conserved motifs across signal peptides, sensitivity to mutations, and variability in the length of the peptides make signal peptide prediction a challenging task that has been extensively pursued over the years.
Results: We introduce TSignal, a deep transformer-based neural network architecture that utilizes BERT language models (LMs) and dot-product attention techniques. TSignal predicts the presence of signal peptides (SPs) and the cleavage site between the SP and the translocated mature protein. We use common benchmark datasets and show competitive accuracy in terms of SP presence prediction and state-of-the-art accuracy in terms of cleavage site prediction for most of the SP types and organism groups. We further illustrate that our fully data-driven trained model identifies useful biological information on heterogeneous test sequences.
Availability: TSignal is available at: https://github.com/Dumitrescu-Alexandru/TSignal.
Contact: alexandru.dumitrescu@aalto.fi, harri.lahdesmaki@aalto.fi

GenCompBio

• Yunpei Xu, Central South University, China
• Hong-Dong Li, Central South University, China
• Cui-Xiang Lin, Central South University, China
• Ruiqing Zheng, Central South University, China
• Yaohang Li, Old Dominion University, China
• Jinhui Xu, State University of New York at Buffalo, United States
• Jianxin Wang, Central South University, China

Motivation: Single cell RNA-sequencing (scRNA-seq) offers a powerful tool to dissect the complexity of biological tissues through cell sub-population identification in combination with clustering approaches. Feature selection is a critical step for improving the accuracy and interpretability of single-cell clustering. Existing feature selection methods do not make full use of the information of cell-type-discriminating ability of genes. We hypothesize that incorporating such information could further boost the performance of single cell clustering.

Results: We develop CellBRF, a feature selection method that considers genes’ relevance to cell types for single-cell clustering. The key idea is to identify genes that are most important for discriminating cell types through random forests guided by predicted cell labels. Moreover, it proposes a class balancing strategy to mitigate the impact of unbalanced cell type distributions on feature importance evaluation. We benchmark CellBRF on thirty-three scRNA-seq datasets representing diverse biological scenarios, and demonstrate that it substantially outperforms state-of-the-art feature selection methods in terms of clustering accuracy and cell neighborhood consistency. Furthermore, we demonstrate the significantly outstanding performance of our selected features with three case studies on cell differentiation stage identification, non-malignant cell subtype identification, and rare cell identification. CellBRF provides a new and effective tool to boost single-cell clustering accuracy.

Availability and implementation: All source codes of CellBRF are freely available at github.com/xuyp-csu/CellBRF.

• Jonas Christian Ditz, University of Tübingen, Germany
• Jacqueline Wistuba-Hamprecht, University of Tübingen, Germany
• Timo Maier, University of Tübingen, Germany
• Rolf Fendel, University of Tübingen, Germany
• Nico Pfeifer, Department of Computer Science, University of Tübingen, Germany
• Bernhard Reuter, Department of Computer Science, University of Tübingen, Germany, Germany

Motivation: Machine learning methods can be used to support scientific discovery in healthcare-related research fields. However, these methods can only be reliably used if they can be trained on high-quality and curated datasets. Currently, no such dataset for the exploration of Plasmodium falciparum protein antigen candidates exists. The parasite Plasmodium falciparum causes the infectious disease malaria. Thus, identifying potential antigens is of utmost importance for the development of antimalarial drugs and vaccines. Since exploring antigen candidates experimentally is an expensive and time-consuming process, applying machine learning methods to support this process has the potential to accelerate the development of drugs and vaccines, which are needed for fighting and controlling malaria.

Results: We developed PlasmoFAB, a curated benchmark that can be used to train machine learning methods for the exploration of Plasmodium falciparum protein antigen candidates. We combined an extensive literature search with domain expertise to create high-quality labels for Plasmodium falciparum specific proteins that distinguish between antigen candidates and intracellular proteins. Additionally, we used our benchmark to compare different well-known prediction models and available protein localization prediction services on the task of identifying protein antigen candidates. We show that available general-purpose services are unable to provide sufficient performance on identifying protein antigen candidates and are outperformed by our models that were trained on this tailored data.

• Leonard Dervishi, Case Western Reserve University, United States
• Wenbiao Li, Case Western Reserve University, United States
• Anisa Halimi, IBM Research Europe, Ireland
• Xiaoqian Jiang, UTHealth at Houston, United States
• Jaideep Vaidya, Rutgers University, United States
• Erman Ayday, Case Western Reserve University, United States

The rapid improvements in genomic sequencing technology have led to the proliferation of locally collected genomic datasets. Given the sensitivity of genomic data, it is crucial to conduct collaborative studies while preserving the privacy of the individuals. However, before starting any collaborative research effort, the quality of the data needs to be assessed. One of the essential steps of the quality control process is population stratification: identifying the presence of genetic difference in individuals due to subpopulations. One of the common methods used to group genomes of individuals based on ethnicity is principal component analysis (PCA). In this paper, we propose a framework to perform population stratification using PCA across multiple collaborators in a privacy-preserving way. In our proposed client-server-based scheme, we initially let the server train a global PCA model on a publicly available genomic dataset which contains individuals from multiple populations. The global PCA model is later used to reduce the dimensionality of the local data by each collaborator (client). After adding noise to achieve local differential privacy (LDP), the collaborators send metadata (in the form of their local PCA outputs) about their research datasets to the server, which then aligns the local PCA results to identify the genetic differences among collaborators' datasets. Our results on real genomic data show that the proposed framework can perform population stratification with high accuracy while preserving the privacy of the research participants.

• Vincent Wagner, University of Stuttgart, Germany
• Nicole Radde, University of Stuttgart, Germany

Motivation: The Chemical Master Equation is a set of linear differential equations that describes the evolution of the probability distribution on all possible configurations of a (bio-)chemical reaction system. Since the number of configurations and therefore the dimension of the CME rapidly increases with the number of molecules, its applicability is restricted to small systems. A widely applied remedy for this challenge are moment-based approaches which consider the evolution of the first few moments of the distribution as summary statistics for the complete distribution.
Here, we investigate the performance of two moment-estimation methods for reaction systems whose equilibrium distributions encounter heavy-tailedness and hence do not possess statistical moments.
Results: We show that estimation via Stochastic Simulation Algorithm trajectories lose consistency over time and estimated moment values span a wide range of values even for large sample sizes. In comparison, the Method of Moments returns smooth moment estimates but is not able to indicate the non-existence of the allegedly predicted moments. We furthermore analyze the negative effect of a CME solution's heavy-tailedness on SSA run times and explain inherent difficulties.
While moment estimation techniques are a commonly applied tool in the simulation of (bio-)chemical reaction networks, we conclude that they should be used with care, as neither the system definition nor the moment estimation techniques themselves reliably indicate the potential heavy-tailedness of the CME's solution.

• Muyu Yang, Carnegie Mellon University, United States
• Jian Ma, Carnegie Mellon University, United States

The spatial positioning of chromosomes relative to functional nuclear bodies is intertwined with genome functions such as transcription, but the sequence patterns and epigenomic features that collectively influence chromatin spatial positioning in a genome-wide manner are not well understood. Here, we develop a new transformer-based deep learning model, called UNADON, that predicts the genome-wide cytological distance to a specific type of nuclear body, as measured by TSA-seq, using both sequence features and epigenomic signals. Evaluations of UNADON in four cell lines (K562, H1, HFFc6, HCT116) show high accuracy in predicting chromatin spatial positioning to nuclear bodies when trained on a single cell line. UNADON also performed well in an unseen cell type. Importantly, we reveal potential sequence and epigenomic factors that affect large-scale chromatin compartmentalization to nuclear bodies. Together, UNADON provides new insights into the principles between sequence features and large-scale chromatin spatial localization, which has important implications for understanding nuclear structure and function.

HiTSeq

• Timofey Prodanov, Heinrich Heine University, Duesseldorf 40225, Germany, Germany
• Vikas Bansal, University of California San Diego, United States

Motivation: Low-copy repeats (LCRs) or segmental duplications are long segments of duplicated DNA that cover > 5% of the human genome. Existing tools for variant calling using short reads exhibit low accuracy in LCRs due to ambiguity in read mapping and extensive copy number variation. Variants in more than 150 genes overlapping LCRs are associated with risk for human diseases.

Methods: We describe a short-read variant calling method, ParascopyVC, that performs variant calling jointly across all repeat copies and utilizes reads independent of mapping quality in LCRs. To identify candidate variants, ParascopyVC aggregates reads mapped to different repeat copies and performs polyploid variant calling. Subsequently, paralogous sequence variants (PSVs) that can differentiate repeat copies are identified using population data and used for estimating the genotype of variants for each repeat copy.

Results: On simulated whole-genome sequence data, ParascopyVC achieved higher precision (0.997) and recall (0.807) than
three state-of-the-art variant callers (best precision = 0.956 for DeepVariant and best recall = 0.738 for GATK) in 167 LCR
regions. Benchmarking of ParascopyVC using the Genome-in-a-Bottle high-confidence variant calls for HG002 genome
showed that it achieved a very high precision of 0.991 and a high recall of 0.909 across LCR regions, significantly better than
FreeBayes (precision = 0.954 and recall = 0.822), GATK (precision = 0.888 and recall = 0.873) and DeepVariant (precision
= 0.983 and recall = 0.861). ParascopyVC demonstrated a consistently higher accuracy (mean F_1 score = 0.947) than other callers
(best F_1 = 0.908) across seven human genomes.

Availability and implementation: ParascopyVC is implemented in Python and is freely available at https://github.com/
tprodanov/ParascopyVC.

• Andrew Mikalsen, University at Buffalo, United States
• Jaroslaw Zola, University at Buffalo, United States

Motivation: The introduction of portable DNA sequencers such as the Oxford Nanopore Technologies MinION has enabled real-time and in the field DNA sequencing. However, in the field sequencing is actionable only when coupled with in the field DNA classification. This poses new challenges for metagenomic software since mobile deployments are typically in remote locations with limited network connectivity and without access to capable computing devices.
Results: We propose new strategies to enable in the field metagenomic classification on mobile devices. We first introduce a programming model for expressing metagenomic classifiers that decomposes the classification process into well-defined and manageable abstractions. The model simplifies resource management in mobile setups and enables rapid prototyping of classification algorithms. Next, we introduce the compact string B-tree, a practical data structure for indexing text in external storage, and we demonstrate its viability as a strategy to deploy massive DNA databases on memory-constrained devices. Finally, we combine both solutions into Coriolis, a metagenomic classifier designed specifically to operate on lightweight mobile devices. Through experiments with actual MinION metagenomic reads and a portable supercomputer-on-a-chip, we show that compared to the state-of-the-art solutions Coriolis offers higher throughput and lower resource consumption without sacrificing quality of classification.
Availability: Source code and test data can be obtained from http://jzola.org/smarten/. Contact: ajmikals@buffalo.edu, jzola@buffalo.edu

• Marjan Hosseini, University of Connecticut, United States
• Aaron Palmer, University of Connecticut, United States
• William Manka, University of Connecticut, United States
• Patrick GS Grady, University of Connecticut, United States
• Venkata Patchigolla, University of Connecticut, United States
• Jinbo Bi, University of Connecticut, United States
• Rachel O'Neill, University of Connecticut, United States
• Zhiyi Chi, University of Connecticut, United States
• Derek Aguiar, University of Connecticut, United States

Motivation: Non-canonical (or non-B) DNA are genomic regions whose three-dimensional conformation deviates from the canonical double helix. Non-B DNA play an important role in basic cellular processes and are associated with genomic instability, gene regulation, and oncogenesis. Experimental methods are low-throughput and can detect only a limited set of non-B DNA structures, while computational methods rely on non-B DNA base motifs, which are necessary but not sufficient indicators of non-B structures. Oxford Nanopore sequencing is an efficient and low-cost platform, but it is currently unknown whether nanopore reads can be used for identifying non-B structures.
Results: We build the first computational pipeline to predict non-B DNA structures from nanopore sequencing. We formalize non-B detection as a novelty detection problem and develop the GoFAE-DND, an autoencoder (AE) that uses goodness-of-fit (GoF) tests as a regularizer. A discriminative loss encourages non-B DNA to be poorly reconstructed and optimizing Gaussian GoF tests allows for the computation of p-values that indicate non-B structures. Based on whole genome nanopore sequencing of NA12878, we show that there exist significant differences between the timing of DNA translocation for non-B DNA bases compared to B-DNA. We demonstrate the efficacy of our approach through comparisons with novelty detection methods using experimental data and data synthesized from a new translocation time simulator. Experimental validations suggest that reliable detection of non-B DNA from nanopore sequencing is achievable.
Availability: Source code is available at https://github.com/bayesomicslab/ONT-nonb-GoFAE-DND.
Contact: {marjan.hosseini, aaron.palmer, derek.aguiar}@uconn.edu
Supplementary information: Supplementary data are available at Bioinformatics online.

• Hartmut Häntze, National Cheng Kung University, Taiwan
• Paul Horton, National Cheng Kung University, Taiwan

Motivation: Alignment-free, k-mer based genotyping methods are a fast alternative to alignment-based methods and are particularly well suited for genotyping larger cohorts. The sensitivity of algorithms, that work with k-mers, can be increased by using spaced seeds, however, the application of spaced seeds in k-mer based genotyping methods has not been researched yet.
Results: We add a spaced seeds functionality to the genotyping software PanGenie and use it to calculate genotypes. This significantly improves sensitivity and F-score when genotyping SNPs, indels and structural variants on reads with low (5x) and high (30x) coverage. Improvements are greater than what could be achieved by just increasing the length of contiguous k-mers. Effect sizes are particularly large for low coverage data. If applications implement effective algorithms for hashing of spaced k-mers, spaced k-mers have thepotential to become an useful technique in k-mer based genotyping.

• Asia Mendelevich, Altius Institute for Biomedical Sciences, United States
• Saumya Gupta, Stem Cell Program, Boston Children’s Hospital; Department of Stem Cell and Regenerative Biology, Harvard University, United States
• Aleksei Pakharev, ---, United States
• Athanasios Teodosiadis, Altius Institute for Biomedical Sciences, United States
• Andrey Mironov, Lomonosov Moscow State University, Institute of Information Transmission Problems, Russia
• Alexander Gimelbrant, Altius Institute for Biomedical Sciences, United States

Analysis of allele-specific expression is strongly affected by the technical noise present in RNA-seq experiments. Previously, we showed that technical replicates can be used for precise estimates of this noise, and we provided a tool for correction of technical noise in allele-specific expression analysis. This approach is very accurate but costly due to the need for two or more replicates of each library. Here, we develop a spike-in approach that is highly accurate at only a small fraction of the cost.
We show that a distinct RNA added as a spike-in before library preparation reflects technical noise of the whole library and can be used in large batches of samples. We experimentally demonstrate the effectiveness of this approach using combinations of RNA from species distinguishable by alignment, namely, mouse, human, and C.elegans. Our new approach, controlFreq, enables highly accurate and computationally efficient analysis of allele-specific expression in (and between) arbitrarily large studies at an overall cost increase of ∼ 5%.

• Alice Lacan, IBISC, University Paris-Saclay (Univ. Evry), France
• Michele Sebag, TAU, CNRS-INRIA-LISN, University Paris-Saclay, France
• Blaise Hanczar, IBISC, University Paris-Saclay (Univ. Evry), France

Motivation: Transcriptomics data is becoming more accessible due to high-throughput and less costly sequencing methods. However, data scarcity prevents exploiting deep learning models’ full predictive power for phenotypes prediction. Artificially enhancing the training sets, namely data augmentation, is suggested as a regularization strategy. Data augmentation corresponds to label-invariant transformations
of the training set (e.g. geometric transformations on images and syntax parsing on text data). Such transformations are, unfortunately, unknown in the transcriptomic field. Therefore, deep generative models such as Generative Adversarial Networks (GANs) have been proposed to generate additional samples.
In this paper, we analyze GAN-based data augmentation strategies with respect to performance indicators and the classification of cancer phenotypes.
Results: This work highlights a significant boost in binary and multiclass classification performances due to augmentation strategies. Without augmentation, training a classifier on only 50 RNA-seq samples yields an accuracy of, respectively, 94% and 70% for binary and tissue classification. In comparison, we achieved 98% and 94% of accuracy when adding 1000 augmented samples. Richer architectures and
more expensive training of the GAN return better augmentation performances and generated data quality overall. Further analysis of the generated data shows that several performance indicators are needed to assess its quality correctly.
Availability: All data used for this research is publicly available and comes from The Cancer Genome Atlas. Reproducible code is available on the GitHub repository: GANs-for-transcriptomics
Contact: alice.lacan@univ-evry.fr
Supplementary information: Supplementary data are available at Bioinformatics online.

• Giulio Ermanno Pibiri, Ca' Foscari University of Venice, Italy
• Yoshihiro Shibuya, University Gustave Eiffel, France
• Antoine Limasset, CNRS, France

Motivation: Minimal perfect hashing is the problem of mapping a static set of n distinct keys into the address space {1,...,n} bijectively. It is well-known that n log2(e) bits are necessary to specify a minimal perfect hash function (MPHF) f, when no additional knowledge of the input keys is to be used. However, it is often the case in practice that the input keys have intrinsic relationships that we can exploit to lower the bit complexity of f. For example, consider a string and the set of all its distinct k-mers as input keys: since two consecutive k-mers share an overlap of k−1 symbols, it seems possible to beat the classic log2(e) bits/key barrier in this case. Moreover, we would like f to map consecutive k-mers to consecutive addresses, as to preserve as much as possible their relationships also in the codomain. This is a useful feature in practice as it guarantees a certain degree of locality of reference for f, resulting in a better evaluation time when querying consecutive k-mers.
Results: Motivated by these premises, we initiate the study of a new type of locality-preserving MPHF designed for k-mers extracted consecutively from a collection of strings. We design a construction whose space usage decreases for growing k and discuss experiments with a practical implementation of the method: in practice, the functions built with our method can be several times smaller and even faster to query than the most efficient MPHFs in the literature.

• Can Firtina, ETH Zurich, Switzerland
• Nika Mansouri Ghiasi, ETH Zurich, Switzerland
• Joel Lindegger, ETH Zurich, Switzerland
• Gagandeep Singh, ETH Zurich, Switzerland
• Meryem Banu Cavlak, ETH Zurich, Switzerland
• Haiyu Mao, ETH Zurich, Switzerland
• Onur Mutlu, ETH Zurich, Switzerland

Nanopore sequencers generate electrical raw signals in real-time while sequencing long genomic strands. These raw signals can be analyzed as they are generated, providing an opportunity for real-time genome analysis. An important feature of nanopore sequencing, Read Until, can eject strands from sequencers without fully sequencing them, which provides opportunities to computationally reduce the sequencing time and cost. However, existing works utilizing Read Until either 1) require powerful computational resources that may not be available for portable sequencers or 2) lack scalability for large genomes, rendering them inaccurate or ineffective.

We propose RawHash, the first mechanism that can accurately and efficiently perform real-time analysis of nanopore raw signals for large genomes using a hash-based similarity search. To enable this, RawHash ensures the signals corresponding to the same DNA content lead to the same hash value, regardless of the slight variations in these signals. RawHash achieves an accurate hash-based similarity search via an effective quantization of the raw signals such that signals corresponding to the same DNA content have the same quantized value and, subsequently, the same hash value.

We evaluate RawHash on three applications: 1) read mapping, 2) relative abundance estimation, and 3) contamination analysis. Our evaluations show that RawHash is the only tool that can provide high accuracy and high throughput for analyzing large genomes in real-time. When compared to the state-of-the-art techniques, UNCALLED and Sigmap, RawHash provides 1) 25.8x and 3.4x better average throughput and 2) significantly better accuracy for large genomes, respectively. Source code is available at https://github.com/CMU-SAFARI/RawHash.

• Camille Marchet, CNRS, France
• Antoine Limasset, CNRS, France

The Sequence Read Archive public database has reached 45 Peta-bytes of raw sequences and doubles its nucleotide
content every two years. Although BLAST-like methods can routinely search for a sequence in a small collection of
genomes, making accessible immense public resources accessible is beyond the reach of alignment-based strategies.
In recent years, abundant literature tackled the task of finding a sequence in extensive sequence collections using
k-mer-based strategies. At present, the most scalable methods are approximate membership query data structures
that conjugate the ability to query small signatures or variants while being scalable to collections up to 10,000
eukaryotic samples. Here, we present PAC, a novel approximate membership query data structure for querying
collections of sequence datasets. PAC index construction works in a streaming fashion without any disk footprint
besides the index itself. It shows a 3 to 6 fold improvement in construction time compared to other compressed
methods for comparable index size. A PAC query can need single random access and be performed in constant
time in favorable instances. Using limited computation resources, we built PAC for very large collections. They
include 32,000 human RNA-seq samples in five days, the entire Genbank bacterial genome collection in a single
day for an index size of 3.5TB. The latter is to our knowledge the largest sequence collection ever indexed using
an approximate membership query structure. We also showed that PAC’s ability to query 500,000 transcript
sequences in less than an hour. PAC’s open-source software is available at https://github.com/Malfoy/PAC.

• Xiang Li, Department of Computer Science and Engineering, The Pennsylvania State University, United States
• Qian Shi, Department of Computer Science and Engineering, The Pennsylvania State University, United States
• Ke Chen, Department of Computer Science and Engineering, The Pennsylvania State University, United States
• Mingfu Shao, Department of Computer Science and Engineering, The Pennsylvania State University, United States

Modern methods for computation-intensive tasks in sequence analysis (e.g., read mapping, sequence alignment, genome assembly, etc.) often first transform each sequence into a list of short, regular-length seeds so that compact data structures and efficient algorithms can be employed to handle the ever-growing large-scale data. Seeding methods using kmers have gained tremendous success in processing sequencing data with low mutation/error rates. However, they are much less effective for sequencing data with high error rates as kmers cannot tolerate errors. We propose SubseqHash, a strategy that uses subsequences, rather than substrings, as seeds. Formally, SubseqHash maps a string of length n to its smallest subsequence of length k, k < n, according to a given order over all length-k strings. Finding the smallest subsequence of a string by enumeration is impractical as the number of subsequences grows exponentially. To overcome this barrier, we propose a novel algorithmic framework that consists of a specifically designed order (termed ABC order) and an algorithm that computes the minimized subsequence under an ABC order in polynomial time. We first show that the ABC order exhibits the desired property and the probability of hash collision using the ABC order is close to a theoretical upper bound. We then show that SubseqHash overwhelmingly outperforms the substring-based seeding methods in producing high-quality seeds for three critical applications: read mapping, sequence alignment, and overlap detection. SubseqHash presents a major algorithmic breakthrough for tackling the high error rates and we expect it to be widely adapted for long-reads analysis.

• Sandra Romain, INRIA, France
• Claire Lemaitre, INRIA, France

Motivation: Structural variation (SV) is a class of genetic diversity whose importance is increasingly revealed by genome re-sequencing, especially with long-read technologies. One crucial problem when analyzing and comparing SVs in several individuals is their accurate genotyping, that is determining whether a described SV is present or absent in one sequenced individual, and if present, in how many copies. There are only a few methods dedicated to SV genotyping with long read data, and all either suffer of a biais towards the reference allele by not representing equally all alleles, or have difficulties genotyping close or overlapping SVs due to a linear representation of the alleles. Results: We present, SVJedi-graph, a novel method for SV genotyping that relies on a variation graph to represent in a single data structure all alleles of a set of SVs. The long reads are then mapped on the variation graph and the resulting alignments that cover allele-specific edges in the graph are used to estimate the most likely genotype for each SV. Running SVJedi-graph on simulated sets of close and overlapping deletions showed that this graph model prevents the biais towards the reference alleles and allows maintaining high genotyping accuracy whatever the SV proximity, contrary to other state-of-the-art genotypers. On the human gold standard HG002 dataset, SVJedi-graph obtained the best performances, genotyping 99.5 % of the high confidence SV callset with an accuracy of 95% in less than 30 minutes.

• Jarno Alanko, University of Helsinki, Finland
• Simon Puglisi, University of Helsinki, Finland
• Tommi Mäklin, University of Helsinki, Finland
• Jaakko Vuohtoniemi, University of Helsinki, Finland

Motivation: Huge data sets containing whole-genome sequences of bacterial strains are now commonplace and represent a rich and important resource for modern genomic epidemiology and metagenomics. In order to efficiently make use of these data sets, efficient indexing data structures - that are both scalable and provide rapid query throughput - are paramount.

Results: Here, we present Themisto, a scalable colored k-mer index designed for large collections of microbial reference genomes, that works for both short and long read data. Themisto indexes 179 thousand Salmonella enterica genomes in 9 hours. The resulting index takes 142 gigabytes. In comparison, the best competing tools Metagraph and Bifrost were only able to index 11 thousand genomes in the same time. In pseudoalignment, these other tools were either an order of magnitude slower than Themisto, or used an order of magnitude more memory. Themisto also offers superior pseudoalignment quality, achieving a higher recall than previous methods on Nanopore read sets.

Availability and implementation: Themisto is available and documented as a C++ package at https://github.com/algbio/themisto available under the GPLv2 license.

iRNA

• Alexander Jürgen Petri, Department of Mathematics, Stockholm University, Sweden
• Kristoffer Sahlin, Department of Mathematics, Stockholm University, Sweden

With advances in long-read transcriptome sequencing, we can now fully sequence transcripts, which greatly improves our ability to study transcription processes. A popular long-read transcriptome sequencing technique is Oxford Nanopore Technologies (ONT), which through its cost-effective sequencing and high throughput, has the potential to characterize the transcriptome in a cell. However, due to transcript variability and sequencing errors, long cDNA reads need substantial bioinformatic processing to produce a set of isoform predictions from the reads. Several genome and annotation-based methods exist to produce transcript predictions. However, such methods require high-quality genomes and annotations and are limited by the accuracy of long-read splice aligners. In addition, gene families with high heterogeneity may not be well represented by a reference genome and would benefit from reference-free analysis. Reference-free methods to predict transcripts from ONT, such as RATTLE, exist, but their sensitivity is not comparable to reference-based approaches.

We present isONform, a high-sensitivity algorithm to construct isoforms from ONT cDNA sequencing data. The algorithm is based on iterative bubble popping on gene graphs built from fuzzy seeds from the reads. Using simulated, synthetic, and biological ONT cDNA data, we show that isONform has substantially higher sensitivity than RATTLE albeit with some loss in precision. On biological data, we show that isONform's predictions have much higher consistency with the annotation-based method StringTie2 compared to RATTLE. We believe isONform can be used both for isoform construction for organisms without well-annotated genomes and as an orthogonal method to verify predictions of reference-based methods.

• Tianshuo Zhou, Oregon State University, United States
• Ning Dai, Oregon State University, United States
• Sizhen Li, Oregon State University, United States
• Max Ward, University of Western Australia, Australia
• David H. Mathews, University of Rochester Medical Center, United States
• Liang Huang, Oregon State University, United States

Motivation: RNA design is the selection of a sequence or set of sequences that will fold to desired structure, i.e. the inverse problem of RNA folding. However, the sequences designed by existing algorithms often suffer from low ensemble stability, and such problem is even worse for long sequence design. Additionally, for many methods only a small number of sequences satisfying MFE criterion can be found by each design. Those drawbacks limit their use in both applications and advanced research.
Results: We propose an innovative optimization paradigm SAMFEO, which optimizes ensemble objectives (equilibrium probability or ensemble defect) and yields successfully designed RNA sequences as byproducts. We develop a search method which leverages structure level and ensemble level information at different stages of the optimization: initialization, sampling, mutation and updating. Our work, though less complicated than others, is the first algorithm that is able to design thousands of RNA sequences for the puzzles from the Eterna100 benchmark. In addition, our proposal solves the most Eterna100 puzzles among all the optimization based methods in our study. The only baseline solving more puzzles than our work is dependent on human rules. Surprisingly, our approach shows great superiority on designing long sequences when using structures adapted from the database 16S Ribosomal RNAs.

Microbiome

• Mihir Mongia, Carnegie Mellon, United States
• Romel Baral, Carnegie Mellon, United States
• Abhinav Adduri, Carnegie Mellon, United States
• Donghui Yan, Carnegie Mellon, United States
• Yudong Liu, Carnegie Mellon, United States
• Yuying Bian, Carnegie Mellong, United States
• Paul Kim, Carnegie Mellon, United States
• Bahar Behsaz, Carnegie Mellon, United States
• Hosein Mohimani, Carnegie Mellon, United States

Microbial natural products represent a major source of bioactive compounds for drug discovery. Among these molecules, Non-Ribosomal Peptides (NRPs) represent a diverse class that include antibiotics, immunosuppressants, anticancer agents, toxins, siderophores, pigments, and cytostatics. The discovery of novel NRPs remains a laborious process because many NRPs consist of non-standard amino acids that are assembled by Non-Ribosomal Peptide Synthetases (NRPSs). Adenylation domains (A-domains) in NRPSs are responsible for selection and activation of monomers appearing in NRPs. During the past decade, several support vector machine-based algorithms have been developed for predicting the specificity of the monomers present in NRPs. These algorithms utilize physiochemical features of the amino acids present in the A-domains of NRPSs. In this paper, we benchmarked the performance of various machine learning algorithms and features for predicting specificities of NRPSs and we showed that the extra trees model paired with one hot encoding features outperforms the existing approaches. Moreover, we show that unsupervised clustering of 453,560 A-domains reveals many clusters that correspond to potentially novel amino acids. While it is challenging to predict the chemical structure of these amino acids, we developed novel techniques to predict their various properties, including polarity, hydrophobicity, charge, and presence of aromatic rings, and carboxyl, and hydroxyl groups.

• Qi Wang, Systems, Synthetic, and Physical Biology (SSPB) Graduate Program, Rice University, Houston, Texas, USA, United States
• Michael Nute, Anvil Diagnostics, Southborough, MA, USA, United States
• Todd Treangen, Department of Computer Science, Rice University, Houston, TX, USA, United States

Motivation: Interactions among microbes within microbial communities have been shown to play crucial roles in human health. In spite of recent progress, low-level knowledge of bacteria driving microbial interactions within microbiomes remains unknown, limiting our ability to fully decipher and control microbial communities.

Results: We present a novel approach for identifying species driving interactions within microbiomes. Bakdrive infers ecological networks of given metagenomic sequencing samples and identifies minimum sets of driver species (MDS) using control theory. Bakdrive has three key innovations in this space: (i) it leverages inherent information from metagenomic sequencing samples to identify driver species, (ii) it explicitly takes host-specific variation into consideration, and (iii) it does not require a known ecological network. In extensive simulated data, we demonstrate identifying driver species identified from healthy donor samples and introducing them to the disease samples, we can restore the gut microbiome in recurrent Clostridioides difficile (rCDI) infection patients to a healthy state. We also applied Bakdrive to two real datasets, rCDI and Crohn's disease patients, uncovering driver species consistent with previous work. Bakdrive represents a novel approach for capturing microbial interactions.

Availability: Bakdrive is open-source and available at: https://gitlab.com/treangenlab/bakdrive

• Wei Wei, the Pennsylvania State University, United States
• Andrew Millward, the Pennsylvania State University, United States
• David Koslicki, Penn State University, United States

Metagenomic samples have high spatiotemporal variability. Hence, it is useful to summarize and characterize the microbial makeup of a given environment in a way that is biologically reasonable and interpretable. The UniFrac metric has been a robust and widely-used metric for measuring the variability between metagenomic samples. We propose that the characterization of metagenomic environments can be achieved by finding the average, a.k.a. the barycenter, among the samples with respect to the UniFrac distance. However, it is possible that such a UniFrac-average includes negative entries, making it no longer a valid representation of a metagenomic community. To overcome this intrinsic issue, we propose a special version of the UniFrac metric, termed L2UniFrac, which inherits the phylogenetic nature of the traditional UniFrac and with respect to which one can easily compute the average, producing biologically meaningful environment-specific “representative samples”. We demonstrate the usefulness of such representative samples as well as the extended usage of L2UniFrac in efficient clustering of metagenomic samples, and provide mathematical characterizations and proofs to the desired properties of L2UniFrac. A prototype implementation is provided at: https://github.com/ KoslickiLab/L2-UniFrac.git.

• Jiayu Shang, Department of Electrical Engineering, City University of Hong Kong, Hong Kong (SAR), China, Hong Kong
• Cheng Peng, Department of Electrical Engineering, City University of Hong Kong, Hong Kong (SAR), China, Hong Kong
• Xubo Tang, Department of Electrical Engineering, City University of Hong Kong, Hong Kong (SAR), China, Hong Kong
• Yanni Sun, Department of Electrical Engineering, City University of Hong Kong, Hong Kong (SAR), China, Hong Kong

Motivation: As viruses that mainly infect bacteria, phages are key players across a wide range of ecosystems. Analyzing phage proteins is indispensable for understanding phages' functions and roles in microbiomes. High-throughput sequencing enables us to obtain phages in different microbiomes with low cost. However, compared to the fast accumulation of newly identified phages, phage protein classification remains difficult. In particular, a fundamental need is to annotate virion proteins, the structural proteins such as major tail, baseplate, etc. Although there are experimental methods for virion protein identification, they are too expensive or time-consuming, leaving a large number of proteins unclassified. Thus, there is a great demand to develop a computational method for fast and accurate phage virion protein classification.
Results: In this work, we adapted the state-of-the-art image classification model, Vision Transformer, to conduct virion protein classification. By encoding protein sequences into unique images using chaos game representation, we can leverage Vision Transformer to learn both local and global features from sequence ``images''. Our method, PhaVIP, has two main functions: classifying PVP and non-PVP sequences and annotating the types of PVP, such as capsid and tail. We tested PhaVIP on several datasets with increasing difficulty and benchmarked it against alternative tools. The experimental results show that PhaVIP has superior performance. After validating the performance of PhaVIP, we investigated two applications that can use the output of PhaVIP: phage taxonomy classification and phage host prediction. The results showed the benefit of using classified proteins over all proteins.

• Aniket Mane, Simon Fraser University, Canada
• Mahsa Faizrahnemoon, Simon Fraser University, Canada
• Tomas Vinar, Comenius University, Slovakia
• Brona Brejova, Comenius University, Slovakia
• Cedric Chauve, Simon Fraser University, Canada

The analysis of bacterial isolates to detect plasmids is important due to their role in the propagation of antimicrobial resistance. In short-read sequence assemblies, both plasmids and bacterial chromosomes are typically split into several contigs of various lengths, making identification of plasmids a challenging problem. In plasmid contig binning, the goal is to distinguish short-read assembly contigs based on their origin into plasmid and chromosomal contigs and subsequently sort plasmid contigs into bins, each bin corresponding to a single plasmid. Previous works on this problem consist of de novo approaches and reference-based approaches. De novo methods rely on contig features such as length, circularity, read coverage, or GC content. Reference-based approaches compare contigs to databases of known plasmids or plasmid markers from finished bacterial genomes.
Recent developments suggest that leveraging information contained in the assembly graph improves the accuracy of plasmid binning. We present PlasBin-flow, a hybrid method that defines contig bins as subgraphs of the assembly graph. PlasBin-flow identifies such plasmid subgraphs through a mixed integer linear programming model that relies on the concept of network flow to account for sequencing coverage, while also accounting for the presence of plasmid genes and the GC content that often distinguishes plasmids from chromosomes. We demonstrate the performance of PlasBin-flow on a real data set of bacterial samples.

• Shaojun Pan, Fudan University, China
• Xing-Ming Zhao, Fudan University, China
• Luis Pedro Coelho, Fudan University, China

Motivation: Metagenomic binning methods to reconstruct metagenome-assembled genomes (MAGs) from environmental samples have been widely used in large-scale metagenomic studies. The recently proposed semi-supervised binning method, SemiBin, achieved state-of-the-art binning results in several environments. However, this required annotating contigs, a computationally costly and potentially biased process.

Results: We propose SemiBin2, which uses self-supervised learning to learn feature embeddings from the contigs. In simulated and real datasets, we show that self-supervised learning achieves better results than the semi-supervised learning
used in SemiBin1 and that SemiBin2 outperforms other state-of-the-art binners. Compared to SemiBin1, SemiBin2 can reconstruct 8.3%–21.5% more high-quality bins and requires only 25% of the running time and 11% of peak memory usage
in real short-read sequencing samples. To extend SemiBin2 to long-read data, we also propose ensemble-based DBSCAN clustering algorithm, resulting in 13.1%–26.3% more high-quality genomes than the second best binner for long-read data.

Availability and implementation: SemiBin2 is available as open source software at https://github.com/BigDataBiology/-
SemiBin/ and the analysis script used in the study can be found at https://github.com/BigDataBiology/SemiBin2_benchmark.

Contact: Correspondence should be addressed to xmzhao@fudan.edu.cn and luispedro@big-data-biology.org.

Supplementary information: Supplementary data are available online.

MLCSB

• Mogan Gim, Korea University, South Korea
• Junseok Choe, Korea University, South Korea
• Seungheun Baek, Korea University, South Korea
• Jueon Park, Korea University, South Korea
• Chaeeun Lee, Korea University, South Korea
• Minjae Ju, LG CNS, AI Research Center, South Korea
• Sumin Lee, LG AI Research, South Korea
• Jaewoo Kang, Korea University, South Korea

Protein-ligand binding affinity prediction is an important task in drug design and development. Cross-modal attention mechanism has become a core component of deep learning models due to the significance of model explainability. Non-covalent interactions, one of the key chemical aspects of this task, should be incorporated in protein-ligand attention mechanism. We propose ArkDTA, a novel deep neural architecture for explainable binding affinity prediction guided by non-covalent interactions. Experimental results show that ArkDTA achieves predictive performance comparable to current state-of-the-art models while significantly improving model explainability. Qualitative investigation into our novel attention mechanism reveals that ArkDTA can identify potential regions for non-covalent interactions between candidate drug compounds and target proteins, as well as guiding internal operations of the model in a more interpretable and domain-aware manner. ArkDTA is available at https://github.com/dmis-lab/ArkDTA

• Aurélien Beaude, Université Paris Saclay, France
• Milad Rafiee Vahid, Sanofi R&D Data and Data Science, United States
• Franck Augé, Sanofi R&D Data and Data Science, France
• Farida Zehraoui, Université d'Évry, France
• Blaise Hanczar, Université d'Évry, France

The increasing availability of high-throughput omics data allows for considering a new medicine centered on individual patients. Precision medicine relies on exploiting these high-throughput data with machine-learning models, especially the ones based on deep-learning approaches, to improve diagnosis. Due to the high-dimensional small-sample nature of omics data, current deep-learning models end up with many parameters and have to be fitted with a limited training set. Furthermore, interactions between molecular entities inside an omics profile are not patient-specific but are the same for all patients.

In this article, we propose AttOmics, a new deep-learning architecture based on the self-attention mechanism. First, we decompose each omics profile into a set of groups, where each group contains related features. Then, by applying the self-attention mechanism to the set of groups, we can capture the different interactions specific to a patient. The results of different experiments carried out in this paper show that our model can accurately predict the phenotype of a patient with fewer parameters than deep neural networks. Visualizing the attention maps can provide new insights into the essential groups for a particular phenotype.

• Jonas Christian Ditz, University of Tübingen, Germany
• Bernhard Reuter, University of Tübingen, Germany
• Nico Pfeifer, University of Tübingen, Germany

Motivation: The size of available omics datasets is steadily increasing with technological advancement in recent years. While this increase in sample size can be used to improve the performance of relevant prediction tasks in healthcare, models that are optimized for large datasets usually operate as black boxes. In high stakes scenarios, like healthcare, using a black-box model poses safety and security issues. Without an explanation about molecular factors and phenotypes that affected the prediction, healthcare providers are left with no choice but to blindly trust the models. We propose a new type of artificial neural network, named Convolutional Omics Kernel Network (COmic). By combining convolutional kernel networks with pathway-induced kernels, our method enables robust and interpretable end-to-end learning on omics datasets ranging in size from a few hundred to several hundreds of thousands of samples. Furthermore, COmic can be easily adapted to utilize multi-omics data.

Results: We evaluated the performance capabilities of COmic on six different breast cancer cohorts. Additionally, we trained COmic models on multi-omics data using the METABRIC cohort. Our models performed either better or similar to competitors on both tasks. We show how the use of pathway-induced Laplacian kernels opens the black-box nature of neural networks and results in intrinsically interpretable models that eliminate the need for post-hoc explanation models.

• Zhijian Huang, Central South University, China
• Pan Zhang, Central South University, China
• Lei Deng, Central South University, China

Motivation: The coronavirus disease 2019 (COVID-19) remains a global public health emergency. Although people, especially those with underlying health conditions, could benefit from several approved COVID-19 therapeutics, the development of effective antiviral COVID-19 drugs is still a very urgent problem. Accurate and robust drug response prediction to a new chemical compound is critical for discovering safe and effective COVID-19 therapeutics.

Results: In this study, we propose DeepCoVDR, a novel COVID-19 drug response prediction method based on deep transfer learning with graph transformer and cross-attention. First, we adopt a graph transformer and feed-forward neural network to mine the drug and cell line information. Then, we use a cross-attention module that calculates the interaction between the drug and cell line. After that, DeepCoVDR combines drug and cell line representation and their interaction features to predict drug response. To solve the problem of SARS-CoV-2 data scarcity, we apply transfer learning and use the SARS-CoV-2 dataset to fine-tune the model pre-trained on the cancer dataset. The experiments of regression and classification show that DeepCoVDR outperforms baseline methods. We also evaluate DeepCoVDR on the cancer dataset, and the results indicate that our approach has high performance compared with other state-of-the-art methods. Moreover, we use DeepCoVDR to predict COVID-19 drugs from FDA-approved drugs and demonstrate the effectiveness of DeepCoVDR in identifying novel COVID-19 drugs.

• Yueqi Sheng, School of Engineering and Applied Sciences, Harvard University, Boston, MA, 02134, USA, United States
• Boaz Barak, School of Engineering and Applied Sciences, Harvard University, Boston, MA, 02134, USA, United States
• Mor Nitzan, The Hebrew University of Jerusalem, Jerusalem, 9190401, Israel, Israel

Single-cell RNA-sequencing technologies have greatly enhanced our understanding of
heterogeneous cell populations and underlying regulatory processes. However, structural (spatial or
temporal) relations between cells are lost during cell dissociation. These relations are crucial for identifying
associated biological processes. Many existing tissue-reconstruction algorithms use prior information
about subsets of genes that are informative with respect to the structure or process to be reconstructed.
When such information is not available, and in the general case when the input genes code for
multiple processes, including being susceptible to noise, biological reconstruction is often computationally
challenging.
We propose an algorithm that iteratively identifies manifold-informative genes using existing
reconstruction algorithms for single-cell RNA-seq data as subroutine. We show that our algorithm improves
the quality of tissue reconstruction for diverse synthetic and real scRNA-seq data, including data from the
mammalian intestinal epithelium and liver lobules.

• Eric Lee, Department of Molecular Oncology, BC Cancer Agency, Vancouver, BC, Canada, Canada
• Kevin Chern, Department of Statistics, University of British Columbia, Vancouver, British Columbia, Canada, Canada
• Michael Nissen, Terry Fox Laboratory, British Columbia Cancer Research Centre, Vancouver, British Columbia, Canada, Canada
• Xuehai Wang, Terry Fox Laboratory, British Columbia Cancer Research Centre, Vancouver, British Columbia, Canada, Canada
• Imaxt Consortium, CRUK IMAXT Grand Challenge Consortium, Cambridge, UK, United Kingdom
• Chris Huang, Translational Medicine Hematology, Bristol Myers Squibb, Summit NJ, USA, United States
• Anita K. Gandhi, Translational Medicine Hematology, Bristol Myers Squibb, Summit NJ, USA, United States
• Alexandre Bouchard-Côté, Department of Statistics, University of British Columbia, Vancouver, British Columbia, Canada, Canada
• Andrew P. Weng, Terry Fox Laboratory, British Columbia Cancer Research Centre, Vancouver, British Columbia, Canada, Canada
• Andrew Roth, Department of Molecular Oncology, BC Cancer Agency, Vancouver, BC, Canada, Canada

Motivation: Recent advances in spatial proteomics technologies have enabled the profiling of dozens of proteins in thousands of single cells in situ. This has created the opportunity to move beyond quantifying the composition of cell types in tissue, and instead probe the spatial relationships between cells. However, current methods for clustering data from these assays only consider the expression values of cells and ignore the spatial context. Furthermore, existing approaches do not account for prior information about the expected cell populations in a sample.
Results: To address these shortcomings, we developed SpatialSort, a spatially aware Bayesian clustering approach that allows for the incorporation of prior biological knowledge. Our method is able to account for the affinities of cells of different types to neighbor in space, and by incorporating prior information about expected cell populations, it is able to simultaneously improve clustering accuracy and perform automated annotation of clusters. Using synthetic and real data, we show that by using spatial and prior information SpatialSort improves clustering accuracy. We also demonstrate how SpatialSort can perform label transfer between spatial and non-spatial modalities through the analysis of a real world diffuse large B-cell lymphoma dataset.

• Haoting Zhang, University of Cambridge, United Kingdom
• Carl Henrik Ek, University of Cambridge, United Kingdom
• Magnus Rattray, University of Manchester, United Kingdom
• Marta Milo, Oncology R&D AstraZeneca, United Kingdom

There exists a range of different quantification frameworks to estimate the synergistic effect of drug combinations. The diversity and disagreement in estimates make it challenging to determine which combinations from a large drug screening should be proceeded with. Furthermore, the lack of accurate uncertainty quantification for those estimates precludes the choice of optimal drug combinations based on the most favourable synergistic effect. In this work, we propose SynBa, a flexible Bayesian approach to estimate the uncertainty of the synergistic efficacy and potency of drug combinations, so that actionable decisions can be derived from the model outputs. The actionability is enabled by incorporating the Hill equation into SynBa, so that the parameters representing the potency and the efficacy can be preserved. Existing knowledge may be conveniently inserted due to the flexibility of the prior, as shown by the empirical Beta prior defined for the normalised maximal inhibition. Through experiments on large combination screenings and comparison against benchmark methods, we show that SynBa provides improved accuracy of dose-response predictions and better-calibrated uncertainty estimation for the parameters and the predictions.

• Alperen Dalkiran, Middle East Technical University, Turkey
• Ahmet Atakan, Middle East Technical University, Turkey
• Ahmet Süreyya Rifaioğlu, Heidelberg University, Germany
• Maria Martin, EMBL-EBI, United Kingdom
• Rengul Atalay, University of Chicago, United States
• Aybar Acar, Middle East Technical University, Turkey
• Tunca Dogan, Hacettepe University, Turkey
• Volkan Atalay, Middle East Technical University, Turkey

Utilizing AI-driven approaches for DTI prediction require large volumes of training data which are not available for the majority of target proteins. In this study, we investigate the use of deep transfer learning for the prediction of interactions between drug candidate compounds and understudied target proteins with scarce training data. The idea here is to first train a deep neural network classifier with a generalized source training dataset of large size and then reuse this pre-trained neural network as an initial configuration for re-training/fine-tuning purposes with a small-sized specialized target training dataset. To explore this idea, we selected six protein families that have critical importance in biomedicine: kinases, G-protein-coupled receptors (GPCRs), ion channels, nuclear receptors, proteases, and transporters. The protein families of transporters and nuclear receptors were individually set as the target datasets, while the other five families were used as the source datasets. Several size-based target family training datasets were formed in a controlled manner. Here we present a disciplined evaluation by pre-training a feed-forward neural network with source training datasets and applying different modes of transfer learning from the pre-trained source network to a target dataset. The performance of deep transfer learning is evaluated and compared with that of training the same deep neural network from scratch. We found that when the training dataset is smaller than 100 compounds, transfer learning yields significantly better performance compared to training the system from scratch, suggesting an advantage to using transfer learning to predict binders to under-studied targets.

NetBio

• Enio Gjerga, University Hospital Heidelberg, Germany
• Isabel S. Naarmann-de Vries, University Hospital Heidelberg, Germany
• Christoph Dieterich, University Hospital Heidelberg, Germany

Alternative RNA splicing plays a crucial role in defining protein function. However, despite its relevance, there is a lack of tools that characterise effects of splicing on protein interaction networks in a mechanistic manner (i.e. presence or absence of protein-protein interactions due to RNA splicing). To fill this gap, we present LINDA (Linear Integer programming for Network reconstruction using transcriptomics and Differential splicing data Analysis) as a method that integrates resources of protein-protein and domain-domain interaction, transcription factor targets and differential splicing/transcript analysis to infer splicing-dependent effects on cellular pathways and regulatory networks. We have applied LINDA to a panel of 54 shRNA depletion experiments in HepG2 and K562 cells from the ENCORE initiative. Through computational benchmarking, we could show that the integration of splicing effects with LINDA can identify pathway mechanisms contributing to known bioprocesses better than other state of the art methods, which do not account for splicing. Additionally, we have experimentally validated some of the predicted splicing effects that the depletion of HNRNPK in K562 cells has on signalling. LINDA has been implemented as an R-package and it is available online in: https://dieterich-lab.github.io/LINDA/ along with results and tutorials.

• Addie Woicik, University of Washington, United States
• Mingxin Zhang, University of Washington, United States
• Hanwen Xu, University of Washington, United States
• Sara Mostafavi, University of Washington, United States
• Sheng Wang, University of Washington, United States

The exponential growth of genomic sequencing data has created ever-expanding repositories of gene networks. Unsupervised network integration methods are critical to learn informative representations for each gene, which are later used as features for downstream applications. However, these network integration methods must be scalable to account for the increasing number of networks and robust to an uneven distribution of network types within hundreds of gene networks. To address these needs, we present Gemini, a novel network integration method that uses memory-efficient high-order pooling to represent and weight each network according to its uniqueness. Gemini then mitigates the uneven distribution through mixing up existing networks to create many new networks. We find that Gemini leads to more than a 10% improvement in F1 score, 14% improvement in micro-AUPRC, and 71% improvement in macro-AURPC for protein function prediction by integrating hundreds of networks from BioGRID, and that Gemini's performance significantly improves when more networks are added to the input network collection, while the comparison approach's performance deteriorates. Gemini thereby enables memory-efficient and informative network integration for large gene networks, and can be used to massively integrate and analyze networks in other domains. Gemini can be accessed at: https://github.com/MinxZ/Gemini.

• Paolo Pellizzoni, ETH Zurich, Switzerland
• Giulia Muzio, ETH Zurich, Switzerland
• Karsten Borgwardt, ETH Zurich, Switzerland

Motivation: Complex phenotypes, such as many common diseases and morphological traits, are controlled by multiple genetic factors, namely genetic mutations and genes, and are influenced by environmental conditions. Deciphering the genetics underlying such traits requires a systemic approach, where many different genetic factors and their interactions are considered simultaneously. Many association mapping techniques available nowadays follow this reasoning, but have some severe limitations. In particular, they require binary encodings for the genetic markers, forcing the user to decide beforehand whether to use, for example, a recessive or a dominant encoding. Moreover, most methods cannot include any biological prior or are limited to testing only lower-order interaction among genes for association with the phenotype, potentially missing a large number of marker combinations.
Results: We propose HOGImine, a novel algorithm that expands the class of discoverable genetic meta-markers by considering higher-order interactions of genes and by allowing multiple encodings for the genetic variants. Our experimental evaluation shows that the algorithm has a substantially higher statistical power compared to previous methods, allowing it to discover genetic mutations statistically associated with the phenotype at hand that could not be found before. Our method can exploit prior biological knowledge on gene interactions, such as protein-protein interaction networks, genetic pathways and protein complexes, to restrict its search space. Since computing higher-order gene interactions poses a high computational burden, we also develop novel algorithmic techniques to make our approach applicable in practice, leading to substantial runtime improvements compared to state-of-the-art methods.

• Sara Mohammad-Taheri, Northeastern Univerusity, United States
• Vartika Tewari, Northeastern Univerusity, United States
• Rohan Kapre, Northeastern Univerusity, United States
• Ehsan Rahiminasab, Google Inc, United States
• Karen Sachs, Next Generation Analytics, United States
• Charles Tapley Hoyt, Laboratory of Systems Pharmacology, Harvard Medical School, United States
• Jeremy Zucker, Pacific Northwest National Laboratory, United States
• Olga Vitek, Northeastern Univerusity, United States

Causal query estimation commonly selects a {\it valid adjustment set}, i.e. a subset of covariates in a model that eliminates the bias of the estimator.
The same query may have multiple valid adjustment sets, each with a different variance.
When networks are partially observed, current methods use graph-based criteria to find an adjustment set that minimizes asymptotic variance.
Many models share the same graph topology, and thus share the same functional dependencies, but may differ in the specific functions that generate the observational data.
In these cases, the topology-based criteria fail to distinguish the variances of the adjustment sets.
This deficiency can lead to sub-optimal adjustment sets, and to miss-characterization of the effect of the intervention.
We propose an approach for deriving {\it optimal adjustment sets}
that take into account the nature of the data generation process, bias, finite-sample variance, and cost.
It empirically learns the data generating processes from historical experimental data, and characterizes the properties of the estimators by simulation.
We demonstrate the utility of the proposed approach in four biomolecular case studies with different topologies and different data generation processes.
The implementation and reproducible case studies are at
\url{https://anonymous.4open.science/r/OptimalAdjustmentSet-E543}

• Kerr Ding, Georgia Institute of Technology, United States
• Sheng Wang, University of Washington, United States
• Yunan Luo, Georgia Institute of Technology, United States

Despite the advances in sequencing technology, massive proteins with known sequences remain functionally unannotated. Biological network alignment (NA), which aims to find the node correspondence between species' protein-protein interaction (PPI) networks, has been a popular strategy to uncover missing annotations by transferring functional knowledge across species. Traditional NA methods assumed that topologically similar proteins in PPIs are functionally similar. However, it was recently reported that functionally unrelated proteins can be as topologically similar as functionally related pairs, and a new data-driven or supervised NA paradigm has been proposed, which uses protein function data to discern which topological features correspond to functional relatedness. Here, we propose GraNA, a deep learning framework for the supervised NA paradigm. Employing graph neural networks, GraNA utilizes within-network interactions and across-network anchor links for learning protein representations and predicting functional correspondence between across-species proteins. A major strength of GraNA is its flexibility to integrate multi-faceted non-functional relationship data, such as sequence similarity and ortholog relationships, as anchor links to guide the mapping of functionally related proteins across species. Evaluating GraNA on a benchmark dataset composed of several NA tasks between different pairs of species, we observed that GraNA accurately predicted the functional relatedness of proteins and robustly transferred functional annotations across species, outperforming a number of existing NA methods. When applied to a case study on a humanized yeast network, GraNA also successfully discovered functionally replaceable human-yeast protein pairs that were documented in previous studies.

• Van-Giang Trinh, Aix-Marseille University, France
• Belaid Benhamou, Aix-Marseille University, France
• Thomas Henzinger, Institute of Science and Technology Austria, Austria
• Samuel Pastva, Institute of Science and Technology Austria, Austria

Boolean networks are simple but efficient mathematical formalism for modeling complex biological systems. However, having only two levels of activation is sometimes not enough to fully capture the dynamics of real-world biological systems. Hence the need for multi-valued networks, a generalization of Boolean networks. Despite the importance of multi-valued networks for modeling biological systems, only limited progress has been made on developing theories, analysis methods, and tools that can support them. In particular, the recent use of trap spaces in Boolean networks made a great impact on the field of systems biology, but there has been no similar concept defined and studied for multi-valued networks to date.

In this work, we generalize the concept of trap spaces in Boolean networks to that in multi-valued networks. We then develop the theory and the analysis methods for trap spaces in multi-valued networks. In particular, we implement all proposed methods in a Python package called trapmvn. Not only showing the applicability of our approach via a realistic case study, we also evaluate the time efficiency of the method on a large collection of real-world models. The experimental results confirm the time efficiency, which we believe enables more accurate analysis on larger and more complex multi-valued models.

RegSys

• Ali Tugrul Balci, University of Pittsburgh, United States
• Mark Maher Ebeid, University of Pittsburgh, United States
• Panayiotis Benos, University of Florida, United States
• Dennis Kostka, University of Pittsburegh, United States
• Maria Chikina, University of Pittsburgh, United States

Motivation: Sequence-based deep learning approaches have been shown to predict a multitude of functional genomic readouts, including regions of open chromatin and RNA expression of genes. However, a major limitation of current methods is that model interpretation relies on computationally demanding post-hoc analyses, and even then we often cannot explain the internal mechanics of highly parameterized models. Here, we introduce a deep learning architecture called tiSFM (totally interpretable sequence to function model). tiSFM improves upon the performance of standard multi-layer convolutional models while using fewer parameters. Additionally, while tiSFM is itself technically a multi-layer neural network, internal model parameters are intrinsically interpretable in terms of relevant sequence motifs. Results: tiSFM’s model architecture makes use of convolutions with a fixed set of kernel weights representing known transcription factor (TF) binding site motifs. Analyzing published open chromatin measurements across hematopoietic lineage cell types we demonstrate that tiSFM outperforms a state-of-the-art convolutional neural network model custom-tailored to this dataset. We also show that it correctly identifies context specific activities of transcription factors with known roles in hematopoietic differentiation, including Pax5 and Ebf1 for B-cells, and Rorc for innate lymphoid cells. tiSFM’s model parameters have biologically meaningful interpretations, and we show the utility of our approach on a complex task of predicting the change in epigenetic state as a function developmental transition.

• Christopher Hill, NIH, United States
• Sanjarbek Hudaiberdiev, NIH, United States
• Ivan Ovcharenko, NIH, United States

Motivation: Predicting the regulatory function of non-coding DNA using only the DNA sequence continues to be a major challenge in genomics. With the advent of improved optimization algorithms, faster GPU speeds, and more intricate machine learning libraries, hybrid convolutional and recurrent neural network architectures can be constructed and applied to extract crucial information from non-coding DNA.
Results: Using a comparative analysis of the performance of thousands of Deep Learning (DL) architectures, we developed ChromDL, a neural network architecture combining bidirectional gated recurrent units (BiGRU), convolutional neural networks (CNNs), and bidirectional long short-term memory units (BiLSTM), which significantly improves upon a range of prediction metrics compared to its predecessors in transcription factor binding site (TFBS), histone modification (HM), and DNase-I hypersensitive site (DHS) detection. Combined with a secondary model, it can be utilized for accurate classification of gene regulatory elements. The model can also detect weak transcription factor (TF) binding with higher accuracy as compared to previously developed methods and has the potential to accurately delineate TF binding motif specificities. Availability: The ChromDL source code can be found at https://github.com/chrishil1/ChromDL.

• Mihir Bafna, Georgia Institute of Technology, United States
• Hechen Li, Georgia Institute of Technology, United States
• Xiuwei Zhang, Georgia Institute of Technology, United States

Motivation: Gene regulatory networks (GRNs) in a cell provide the tight feedback needed to synchronize cell actions. However, genes in a cell also take input from, and provide signals to, other, neighboring cells. These cell-cell interactions (CCIs) and the GRNs deeply influence each other. Many computational methods have been developed for GRN inference in cells. More recently, methods were proposed to infer CCIs using single cell gene expression data with or without cell spatial location information. However, in reality, the two processes do not exist in isolation and are subject to spatial constraints. Despite this rationale, no methods currently exist to infer GRNs and CCIs using the same model.

Results: We propose CLARIFY, a tool that takes GRNs as input, uses them to predict CCIs, and simultaneously, uses the CCIs to refine and output cell-specific GRNs. CLARIFY uses a novel multi-level graph neural network, which mimics cellular networks at a higher level and cell specific GRNs at a deeper level. We applied CLARIFY to two real spatial transcriptomic datasets, one using SeqFISH and the other using MERFISH, and also tested on simulated datasets from scMultiSim. We compared the quality of predicted GRNs and CCIs with state-of-the-art baseline methods that either inferred only GRNs or only CCIs. The results show that CLARIFY consistently outperforms the baseline in terms of commonly used evaluation metrics. Our results point to the importance of co-inference of CCIs and GRNs and to the use of layered graph neural networks as an inference tool for biological networks.

• Yanlin Zhang, McGill University, Canada
• Mathieu Blanchette, McGill University, Canada

Motivation: Accurately assessing contacts between DNA fragments inside the nucleus with Hi-C experiment is crucial for understanding the role of 3D genome organization in gene regulation. This challenging task is due in part to the high sequencing depth of Hi-C libraries required to support high resolution analyses. Most existing Hi-C data are collected with limited sequencing coverage, leading to poor chromatin interaction frequency estimation. Current computational approaches to enhance Hi-C signals focus on the analysis of individual Hi-C data sets of interest, without taking advantage of the facts that (i) several hundred Hi-C contact maps are publicly available, and (ii) the vast majority of local spatial organizations are conserved across multiple cell types.

Results: Here, we present RefHiC-SR, an attention-based deep learning framework that uses a reference panel of Hi-C datasets to facilitate the enhancement of Hi-C data resolution of a given study sample. We compare RefHiC-SR against tools that do not use reference samples and find that RefHiC-SR outperforms other programs across different cell types, and sequencing depths. It also enables high accuracy mapping of structures such as loops and topologically associating domains.

Availability: https://github.com/BlanchetteLab/RefHiC

Contact: blanchem@cs.mcgill.ca

• Cassandra Burdziak, Memorial Sloan Kettering Cancer Center, United States
• Chujun Zhao, Memorial Sloan Kettering Cancer Center/Columbia University, United States
• Doron Haviv, Memorial Sloan Kettering Cancer Center/Weill Cornell Medicine, United States
• Direna Alonso-Curbelo, Memorial Sloan Kettering Cancer Center/The Barcelona Institute of Science and Technology, Spain
• Scott Lowe, Memorial Sloan Kettering Cancer Center/Howard Hughes Medical Institute, United States
• Dana Pe'Er, Memorial Sloan Kettering Cancer Center/Howard Hughes Medical Institute, United States

Motivation: Transcriptional dynamics governed by the action of regulatory proteins are fundamental to systems ranging from normal development to disease progression. There has been substantial progress in deriving mechanistic insight into regulators of static populations with single-cell transcriptomic data, yet prevalent methods tracking phenotypic dynamics are naive to the regulatory drivers of gene expression variability through time.
Results: We introduce scKINETICS (Key regulatory Interaction NETwork for Inferring Cell Speed), a dynamical model of gene expression change which is fit with the simultaneous learning of per-cell transcriptional velocities and a governing gene regulatory network. This is accomplished through an expectation-maximization approach derived to learn the impact of each regulator on its target genes, leveraging biologically-motivated priors from epigenetic data, gene-gene co-expression, and constraints on cells’ future states imposed by the phenotypic manifold. Applying this approach to an acute pancreatitis dataset recapitulates a well-studied axis of acinar-to-ductal trans-differentiation whilst proposing novel regulators of this process, including factors with previously-appreciated roles in driving pancreatic tumorigenesis. In benchmarking experiments, we show that scKINETICS successfully extends and improves existing velocity approaches to generate interpretable, mechanistic models of gene regulatory dynamics.

TransMed

• Monica Dayao, Carnegie Mellon University, United States
• Alexandro Trevino, Enable Medicine, United States
• Honesty Kim, Enable Medicine, United States
• Matthew Ruffalo, Carnegie Mellon University, United States
• H. Blaize D'Angio, Enable Medicine, United States
• Ryan Preska, Enable Medicine, United States
• Umamaheswar Duvvuri, University of Pittsburgh, United States
• Aaron Mayer, Enable Medicine, United States
• Ziv Bar-Joseph, Carnegie Mellon University, United States

Spatial proteomics data has been used to map cell states and improve our understanding of tissue organization. More recently, these methods have been extended to study the impact of such organization on disease progression and patient survival. However, to date, the majority of supervised learning methods utilizing these data types did not take full advantage of the spatial information, impacting their performance and utilization. Taking inspiration from ecology and epidemiology, we developed novel spatial feature extraction methods for use with spatial proteomics data. We used these features to learn prediction models for cancer patient survival. As we show, using the spatial features led to consistent improvement over prior methods that used the spatial proteomics data for the same task. In addition, feature importance analysis revealed new insights about the cell interactions that contribute to patient survival.

VarI

• Yasser Mohseni Behbahani, Sorbonne Université, France
• Elodie Laine, Sorbonne Université - Laboratory of Computational and Quantitative Biology (LCQB, CNRS-SU), France
• Alessandra Carbone, Sorbonne Universite, France

The spectacular recent advances in protein and protein complex structure prediction hold promise for reconstructing interactomes at large scale and residue resolution. Beyond determining the 3D arrangement of interacting partners, modeling approaches should be able to unravel the impact of sequence variations on the strength of the association. In this work, we report on Deep Local Analysis (DLA), a novel and efficient deep learning framework that relies on a strikingly simple deconstruction of protein interfaces into small locally oriented residue-centered cubes and on 3D convolutions recognizing patterns within cubes. Merely based on the two cubes associated with the wild-type and the mutant residues, DLA accurately estimates the binding affinity change for the associated complexes. It achieves a Pearson correlation coefficient of 0.735 on about 400 mutations on unseen complexes. Its generalization capability on blind datasets of complexes is higher than the state-of-the-art methods. We show that taking into account the evolutionary constraints on residues contributes to predictions. We also discuss the influence of conformational variability on performance. Beyond the predictive power on the effects of mutations, DLA is a general framework for transferring the knowledge gained from the available non-redundant set of complex protein structures to various tasks. For instance, given a single partially masked cube, it recovers the identity and physico-chemical class of the central residue. Given an ensemble of cubes representing an interface, it predicts the function of the complex. Source code and models are available at http://gitlab.lcqb.upmc.fr/DLA/DLA.git.