• James Zou

This talk will explore how we can develop and use generative AI to help researchers. I will first discuss how generative AI can act as research co-advisors. Then I will first discuss how we use genAI to expand researchers' creativity by designing and experimentally validating new drugs. Finally, I will explore the role of language as the foundational data modality for biomedicine.

• Eric Sun, Eric Sun, Stanford University
• Rong Ma, Rong Ma, Harvard T.H. Chan School of Public Health
• James Zou, James Zou, Stanford University

Spatially resolved single-cell transcriptomics have provided unprecedented insights into gene expression {\it in situ}, particularly in the context of cell interactions or organization of tissues. However, current technologies for profiling spatial gene expression at single-cell resolution are generally limited to the measurement of a small number of genes. To address this limitation, several algorithms have been developed to impute or predict the expression of additional genes that were not present in the measured gene panel. Current algorithms do not leverage the rich spatial and gene relational information in spatial transcriptomics. To improve spatial gene expression predictions, we introduce SPRITE (Spatial Propagation and Reinforcement of Imputed Transcript Expression) as a meta-algorithm that processes predictions obtained from existing methods by propagating information across gene correlation networks and spatial neighborhood graphs. SPRITE improves spatial gene expression predictions across multiple spatial transcriptomics datasets. Furthermore, SPRITE predicted spatial gene expression leads to improved clustering, visualization, and classification of cells. SPRITE is available as a software package and can be used in spatial transcriptomics data analysis to improve inferences based on predicted gene expression.

• Charles Broadbent, Charles Broadbent, University of Minnesota Twin Cities
• Tianci Song, Tianci Song, University of Minnesota Twin Cities
• Rui Kuang, Rui Kuang, University of Minnesota Twin Cities

Spatial transcripome (ST) profiling can reveal cells’ structural organizations and functional roles in tissues. However, deciphering the spatial context of gene expressions in ST data is a challenge—the high-order structure hiding in whole transcriptome space over 2D/3D spatial coordinates requires modeling and detection of interpretable high-order elements and components for further functional analysis and interpretation. This paper presents a new method GraphTucker—-graph-regularized Tucker tensor decomposition for learning high-order factorization in ST data. GraphTucker is based on a non-negative Tucker decomposition algorithm regularized by a high-order graph that captures spatial relation among spots and functional relation among genes. In the experiments on several Visium and Stereo-seq datasets, the novelty and advantage of modeling multi-way multilinear relationships among the components in Tucker decomposition are demonstrated as opposed to the Canonical Polyadic Decomposition (CPD) and conventional matrix factorization models by evaluation of detecting spatial components of gene modules, clustering spatial coefficients for tissue segmentation and imputing complete spatial transcriptomes. The results of visualization show strong evidences that GraphTucker detect more interpretable spatial components in the context of the spatial domains in the tissues. Availability: https://github.com/kuanglab/GraphTucker.

• Sokratia Georgaka, Sokratia Georgaka, University of Manchester
• William Geraint Morgans, William Geraint Morgans, University of Manchester
• Qian Zhao, Qian Zhao, University of Manchester
• Diego Sanchez, Diego Sanchez, Cancer Research UK Manchester Institute
• Mohamed Ghafoor, Mohamed Ghafoor, Division of Informatics
• Syed Murtuza Baker, Syed Murtuza Baker, University of Manchester
• Mudassar Iqbal, Mudassar Iqbal, The University of Manchester
• Magnus Rattray, Magnus Rattray, The University of Manchester

Spatially resolved transcriptomics has revolutionised the study of the gene expression within tissues, allowing researchers to maintain the spatial context. Accompanying these spatial transcriptomics datasets are often histology images, providing rich information on tissue architecture, organisation and pathology, complementing the spatial gene expression. However, in traditional pipelines, histological information is typically discarded during tasks such as dimensionality reduction of the spatial transcriptomics data.
To address this limitation, we propose Cellpie, a novel approach based on fast, joint non-negative matrix factorisation (NMF). Cellpie simultaneously decomposes spatial gene expression and histology image features into interpretable components. Through joint NMF, CellPie generates non-negative factor matrices representing parts-based representation (factors) of the data, facilitating the identification of biologically relevant patterns of variation. In addition, CellPie extracts the corresponding leading genes and image features that are strongly associated with each factor. These genes and features serve as marker genes and morphological characteristics, respectively, providing insights into the biological processes underlying the observed patterns in the spatial gene expression data. Furthermore, they enable the discoverer of links between molecular signalling and tissue morphology.
We demonstrated CellPie on two distinct tissue types, showcasing its improved accuracy in downstream analysis tasks compared to published dimensionality reduction methods.

• Euxhen Hasanaj, Euxhen Hasanaj, Carnegie Mellon University
• Sachin Mathur, Sachin Mathur, Sanofi
• Ziv Bar-Joseph, Ziv Bar-Joseph, Carnegie Mellon University

Motivation: Analysis of time series transcriptomics data from clinical trials is challenging. Such studies usually profile very few time points from several individuals with varying response patterns and dynamics. Current methods for these datasets are mainly based on linear, global orderings using visit times which do not account for the varying response rates and subgroups within a patient cohort.

Results We developed a new method that utilizes multi-commodity flow algorithms for trajectory inference in large scale clinical studies. Recovered trajectories satisfy individual-based timing restrictions while integrating data from multiple patients. Testing the method on multiple drug datasets demonstrated an improved performance compared to prior approaches suggested for this task, while identifying novel endotypes that correspond to heterogeneous patient response patterns.

Availability: The source code and instructions to download the data have been deposited on GitHub at https://github.com/euxhenh/Truffle

• David van Dijk, David van Dijk, Yale University

In this talk, I will showcase the work of my lab in revolutionizing biomedical data analysis through foundation models and large language models (LLMs). First, we introduce CINEMA-OT, a causal-inference-based approach using optimal transport for single-cell perturbation analysis. CINEMA-OT allows individual treatment-effect analysis, response clustering, and synergy analysis, revealing potential mechanisms in airway antiviral response and immune cell recruitment. Next, we present CaLMFlow, combining flow matching with integral equations and causal language models. By fine-tuning LLMs on flow matching and conditioning on natural language prompts, CaLMFlow predicts single-cell perturbation responses and performs protein backbone generation. We then explore "Cell2Sentence" (C2S), a technique translating single-cell transcriptomics into a language for LLMs. C2S automates the generation of natural language insights directly from biological data and generates cells based on textual prompts, enhancing data interpretation and synthesis. Additionally, I will discuss "BrainLM," the first fMRI foundation model to decode brain activity, predict clinical variables, and improve our understanding of brain function and disease. Finally, I will present some of our efforts to integrate foundation models with graphs with the aim to leverage pre-trained textual and non-textual foundation models for graph-based tasks.

• Andrzej Mizera, Andrzej Mizera, IDEAS NCBR & University of Warsaw
• Jakub Zarzycki, Jakub Zarzycki, IDEAS NCBR & University of Warsaw

Cellular reprogramming can be used for both the prevention and cure of different diseases. However, the efficiency of discovering reprogramming strategies with classical wet-lab experiments is hindered by lengthy time commitments and high costs. In this study, we develop a novel computational framework based on deep reinforcement learning that facilitates the identification of reprogramming strategies. For this aim, we formulate a control problem in the context of cellular reprogramming for the frameworks of BNs and PBNs under the asynchronous update mode. Furthermore, we introduce the notion of a pseudo-attractor and a procedure for identification of pseudo-attractor state during training. Finally, we devise a computational framework for solving the control problem, which we test on a number of different models.

• Ding Bai, Ding Bai, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
• Caleb Ellington, Caleb Ellington, Carnegie Mellon University
• Shentong Mo, Shentong Mo, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
• Le Song, Le Song, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)
• Eric Xing, Eric Xing, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) & Carnegie Mellon University

Genetic perturbations (e.g. knockouts, variants) have laid the foundation for our understanding of many diseases, implicating pathogenic mechanisms and indicating therapeutic targets. However, experimental assays are fundamentally limited by the number of measurable perturbations. Computational methods can fill this gap by predicting perturbation effects under novel conditions, but accurately predicting the transcriptional responses of cells to unseen perturbations remains a significant challenge. We address this by developing a novel attention-based neural network, AttentionPert, which accurately predicts gene expression under multiplexed perturbations and generalizes to unseen conditions. AttentionPert integrates global and local effects in a multi-scale model, representing both the non-uniform system-wide impact of the genetic perturbation and the localized disturbance in a network of gene-gene similarities, enhancing its ability to predict nuanced transcriptional responses to both single and multi-gene perturbations. In comprehensive experiments, AttentionPert demonstrates superior performance across multiple datasets outperforming the state-of-the-art method in predicting differential gene expressions and revealing novel gene regulations. AttentionPert marks a significant improvement over current methods, particularly in handling the diversity of gene perturbations and in predicting out-of-distribution scenarios.

• Wei Huang, Wei Huang, College of Computer and Information Engineering
• Hui Liu, Hui Liu, College of Computer and Information Engineering

Phenotype-based screening has emerged as a powerful approach for identifying compounds that actively interact with cells. Transcriptional and proteomic profiling of cell population and single cell provide insights into the cellular changes that occur at the molecular level in response to external perturbations, such as drugs or genetic manipulations. In this paper, we propose cycleCDR, a novel deep learning framework to predict cellular response to drugs or gene perturbations. We leverage the power of autoencoders to maps the unperturbed cellular states to a latent space, in which we postulate the effects of drug perturbations on cellular states follow a linear additive model. Next, we introduce the cycle consistency constraints to ensure that unperturbed cellular state subjected to drug perturbation in the latent space would produce the perturbed cellular state through the decoder. Conversely, removal of perturbations from the perturbed cellular states could restore the unperturbed cellular state. The cycle consistency constraints and linear modeling in latent space enable to learn transferable representations of external perturbations, so that our model can generalize well to unseen drugs. We validate our model on four different types of datasets, including bulk transcriptional responses, bulk proteomic responses, and single-cell transcriptional responses to drug/gene perturbations. The experimental results demonstrate that our model consistently outperforms existing state-of-the-art methods, indicating our method is highly versatile and applicable to a wide range of scenarios.

• Asa Ben-Hur, Asa Ben-Hur, Colorado State University
• Ahmed Daoud, Ahmed Daoud, Colorado State University

Complex deep learning models trained on very large datasets have become key enabling tools for current research in natural language processing and computer vision.
By providing pre-trained models that can be fine-tuned for specific applications, they enable researchers to create accurate models with minimal effort and computational resources.
Large scale genomics deep learning models come in two flavors: the first are large language models of DNA sequences trained in a self-supervised fashion, similar to the corresponding natural language models; the second are supervised learning models that leverage large scale genomics datasets from ENCODE and other sources.
We argue that these models are the equivalent of foundation models in natural language processing in their utility, as they encode within them chromatin state in its different aspects, providing useful representations that allow quick deployment of accurate models of gene regulation.
We demonstrate this premise by leveraging the recently created Sei model to develop simple, interpretable models of intron retention, and demonstrate their advantage over models based on the DNA langauage model DNABERT-2.
Our work also demonstrates the impact of chromatin state on the regulation of intron retention.
Using representations learned by Sei, our model is able to discover the involvement of transcription factors and chromatin marks in regulating intron retention, providing better accuracy than a recently published model trained from scratch for this purpose.

• Anupama Jha, Anupama Jha, University of Washington
• Borislav Hristov, Borislav Hristov, University of Washington
• Xiao Wang, Xiao Wang, University of Washington
• Sheng Wang, Sheng Wang, University of Washington
• William Greenleaf, William Greenleaf, Stanford University
• Anshul Kundaje, Anshul Kundaje, Stanford University
• Erez Lieberman Aiden, Erez Lieberman Aiden, Baylor College of Medicine
• William Stafford Noble, William Stafford Noble, University of Washington

The 3D nuclear DNA architecture is composed of intrachromosomal and interchromosomal contacts. Despite the functional relevance of interchromosomal contacts, existing predictive models for 3D genome folding have focused on modeling intrachromosomal contacts from nucleotide sequences, mainly ignoring the contributions of interchromosomal contacts. To remedy this, we propose TwinC, an interpretable convolutional neural network model that uses a paired sequence design to model Hi-C interchromosomal contacts from replicate Hi-C experiments. TwinC accepts two 100~kb nucleotide sequences as input and predicts interchromosomal contacts between them. We use Hi-C experiments from 20 human donor heart samples from the ENCODE project to show that TwinC achieves high predictive accuracy (AUROC=0.80) on a cross-chromosomal test set. Furthermore, despite TwinC's computational simplicity and faster training time, it performs at par with the state-of-the-art orca model. Subsequently, we show that TwinC learns the importance of local chromatin accessibility features in the formation of interchromosomal contacts and identifies transcription factories located on different chromosomes that cluster in the nucleus. Our results suggest that by leveraging pooled contacts from multiple donors and employing a twin sequence design, TwinC can learn to accurately predict interchromosomal contacts and identify sequence signatures relevant to their 3D structure in the nucleus.

• Mogan Gim, Mogan Gim, Korea University
• Jueon Park, Jueon Park, Korea University
• Soyon Park, Soyon Park, Korea University
• Sanghoon Lee, Sanghoon Lee, Korea University
• Seungheun Baek, Seungheun Baek, Korea University
• Junhyun Lee, Junhyun Lee, Korea University
• Ngoc-Quang Nguyen, Ngoc-Quang Nguyen, Korea University
• Jaewoo Kang, Jaewoo Kang, Korea University

Motivation: Molecular core structures and R-groups are essential concepts especially in compound analysis and lead optimization. Integration of these concepts with conventional graph pre-training approaches can promote deeper understanding in both local and global properties of molecules. We propose MolPLA, a dual molecular pre-training framework that promotes understanding in a molecule's core structure with peripheral R-groups and extends it with the ability to help chemists find replaceable R-groups in lead optimization scenarios.
Results: Experimental results on molecular property prediction show that MolPLA exhibits predictability comparable to current state-of-the-art models. Qualitative analysis implicate that MolPLA is capable of distinguishing core and R-group sub-structures, identifying decomposable regions in molecules and contributing to lead optimization scenarios by rationally suggesting R-group replacements given various query core templates.

• Di Wu, Di Wu, University of Pennsylvania
• Natalie Maus, Natalie Maus, University of Pennsylvania
• Anupama Jha, Anupama Jha, University of Washington
• San Jewell, San Jewell, University of Pennsylvania
• Jacob Gardner, Jacob Gardner, University of Pennsylvania
• Yoseph Barash, Yoseph Barash, University of Pennsylvania

Alternative splicing (AS) of pre-mRNA is a highly regulated process with significant splicing changes occurring across human tissues. The tissue-specific changes in splicing, combined with the fact splicing defects are related to numerous disease made the ability to predict or manipulate AS a long-term goal, with applications ranging from identifying novel regulatory mechanisms to designing therapeutic targets. Here, we take advantage of generative model architectures to address the prediction and design of tissue-specific RNA splicing outcomes. First, we construct a predictive model, TrASPr, which combines multiple localized transformers to predict splicing in a tissue-specific manner. Then, we exploit TrASPr as an Oracle to produce labeled data for a Bayesian Optimization (BO) algorithm with a custom loss function for RNA splicing outcome design. We demonstrate TrASPr significantly outperforms recently published models and identifies relevant regulatory features also captured by the BO generative process.

• Daniel Olson, Daniel Olson, Department of Computer Science
• Daphne Demekas, Daphne Demekas, R. Ken Coit College of Pharmacy
• Thomas Colligan, Thomas Colligan, R. Ken Coit College of Pharmacy
• Travis Wheeler, Travis Wheeler, R. Ken Coit College of Pharmacy

The homology search tool HMMER is extremely sensitive and can identify homologous protein pairs even when there is very little %id between them. Search tools such as MMSeqs2 and DIAMOND are extremely efficient and capable of rapidly searching extremely large protein databases like TrEMBL, but are less sensitive than HMMER. The varying performance (speed and accuracy) of these tools is largely influenced by the choice of filtering strategies that are used to eliminate candidate alignments before running more expensive alignment algorithms. Motivated by a desire to retain full HMM sensitivity with greater speed, we have developed a new filtering method, called NEAR (Neural Embeddings for Amino acid Relationships). NEAR is a method based on representation learning that is designed to rapidly identify good sequence alignment candidates from a large protein database. NEAR's neural embedding model computes per-residue embeddings for target and query protein sequences and identifies alignment candidates with a pipeline consisting of k-NN search, filtration, and neighbor aggregation.
NEAR’s ResNet embedding model is trained using an N-pairs loss function guided by sequence alignments generated by the widely used HMMER3 tool.
Benchmarking results reveal improved performance relative to state-of-the-art neural embedding models specifically developed for protein sequences, as well as enhanced speed relative to the alignment-based filtering strategy used in HMMER3’s sensitive alignment pipeline. We present NEAR as a standalone filter, but have plans to integrate NEAR into our search tool NAIL.

• Jiri Reinis, Jiri Reinis, CeMM
• Andreas Reicher, Andreas Reicher, CeMM
• Monika Malik, Monika Malik, CeMM
• Pavel Ruzicka, Pavel Ruzicka, CeMM
• Maria Ciobanu, Maria Ciobanu, CeMM
• Stefan Kubicek, Stefan Kubicek, CeMM
• Andre Rendeiro, Andre Rendeiro, CeMM
• Victoria Kartysh, Victoria Kartysh, CeMM
• Tatjana Tomek, Tatjana Tomek, CeMM
• Marton Siklos, Marton Siklos, CeMM

Imaging-based methods are widely used for studying the subcellular localization of proteins in living cells. While routine for individual proteins, global monitoring of protein dynamics following chemical or genetic perturbation typically relies on arrayed panels of fluorescently tagged cell lines, limiting throughput and scalability. Here, we describe a strategy that combines high-throughput microscopy, computer vision, and machine learning to detect perturbation-induced changes in multicolor tagged visual proteomics cell (vpCell) pools.

We use genome-wide and cancer-focused intron-targeting sgRNA libraries to generate vpCell pools and a large arrayed collection of clones (4,576 clones, 1,158 unique fluorescently tagged proteins). Each vpCell clone expresses two different endogenously tagged fluorescent proteins. Individual clones can be identified in the pool by image analysis alone, training a machine learning model on localization patterns and expression levels of the tagged proteins. This enables simultaneous live-cell monitoring of large sets of proteins.

To demonstrate broad applicability and scale, we test the effects of antiproliferative compounds on a pool with cancer-related proteins, on which we identify widespread protein localization changes and novel inhibitors of the nuclear import/export machinery. The time-resolved characterization of changes in subcellular localization and abundance of proteins upon perturbation in pooled format highlights the power of the vpCell approach for drug discovery and mechanism of action studies.

Finally, we present an interactive online web atlas of 1,158 fluorescently labeled proteins in clonal cell lines, available at https://vpcells.cemm.at.

• Zhangzhi Peng, Zhangzhi Peng, Duke Unversity
• Benjamin Schussheim, Benjamin Schussheim, Duke Unversity
• Pranam Chatterjee, Pranam Chatterjee, Duke Unversity

Proteins serve as the workhorses of living organisms, orchestrating a wide array of vital functions. Post-translational modifications (PTMs) of their amino acids greatly influence the structural and functional diversity of different protein types and uphold proteostasis, allowing cells to swiftly respond to environmental changes and intricately regulate complex biological processes. To this point, efforts to model the complex features of proteins have involved the training of large and expressive protein language models (pLMs) such as ESM-2 and ProtT5, which accurately encode structural, functional, and physicochemical properties of input protein sequences. However, the over 200 million sequences that these pLMs were trained on merely scratch the surface of proteomic diversity, as they neither input nor account for the effects of PTMs. In this work, we fill this major gap in protein sequence modeling by introducing PTM tokens into the pLM training regime. We then leverage recent advancements in structured state space models (SSMs), specifically Mamba, which utilizes efficient hardware-aware primitives to overcome the quadratic time complexities of Transformers. After adding a comprehensive set of PTM tokens to the model vocabulary, we train bidirectional Mamba blocks whose outputs are fused with state-of-the-art ESM-2 embeddings via a novel gating mechanism. We demonstrate that our resultant PTM-aware pLM, PTM-Mamba, improves upon ESM-2's performance on various PTM-specific tasks. PTM-Mamba is the first and only pLM that can uniquely input and represent both wild-type and PTM sequences, motivating downstream modeling and design applications specific to post-translationally modified proteins.