• TBD

• Fritz Sedlazeck

TBD

• Sumedha Saxena, School of Biomedical Science, Li Ka Shing Faculty of
  Medicine, The University of Hong Kong, Hong Kong
• Edmond S. K. Ma, Division of Clinical Pathology & Molecular Pathology, Hong
  Kong Sanatorium Hospital, Hong Kong
• Aya El Helali, Department of Clinical Oncology, Li Ka Shing Faculty of
  Medicine, The University of Hong Kong, Hong Kong
• David J. H. Shih, School of Biomedical Science, Li Ka Shing Faculty of
  Medicine, The University of Hong Kong, Hong Kong

Advanced tumor sequencing has enabled precision oncology,
yet matching
patients to appropriate clinical trials remains challenging
due to unstructured
eligibility criteria and heterogeneous patient data. We
developed a suite of
AI-enabled tools to improve the matching of cancer patients
in Hong Kong to
relevant trials. Using the open-source MatchMiner platform
from the DanaFarber Cancer Institute, we customized its
MatchEngine to include additional
clinical criteria (HER2, ER, PR and PD-L1 status) and the
latest OncoTree terminology for diagnoses. We created two
applications, matchminer-patient and
nct2ctml, both leveraging a locally hosted large language
model (DeepSeekR1-Distill-Qwen-32B-quantized.w4a16) to
convert unstructured text and image inputs into structured
data on diagnoses, biomarkers, and genomic alterations.
Matchminer-patient provides a web interface for clinicians
to enter a
free text diagnosis / description and upload multiple
tumor-sequencing screenshots; SuryaOCR extracts text from
each image and passes it to the LLM.
Nct2ctml harmonizes clinical trial records from local and
international registries
into Clinical Trial Markup Language (CTML). We converted
124 cancer trials
and integrated patient data from multiple sources; AI
assistance reduced manual curation to 11% for trials and
30% for patient data. Structured datasets
uploaded to MatchMiner generated an average of 7–10 trial
matches per patient,
reviewed at molecular tumor board meetings to inform
personalized treatment
planning.

• Kezia Irene, Kalbe Farma Tbk, Indonesia
• Belinda Mutiara, Kalbe Farma Tbk, Indonesia
• Jocelyn Siswanto, Kalbe Farma Tbk, Indonesia
• Jonathan Susanto, Universitas Multimedia Nusantara, Indonesia
• Restu Kresnadi, Kalbe Farma Tbk, Indonesia

Polygenic risk scores (PRS) for type 2 diabetes (T2D) often lose accuracy when applied outside the populations in which they were developed. Partitioned PRS (PPRS) mitigate this by decomposing T2D risk into biologically interpretable pathways (e.g., obesity, body fat), but have not been adapted to Indonesians. We present PPRS-ID, an Indonesian-adjusted PPRS that integrates a locally derived obesity PRS with a population-specific linkage disequilibrium resource. We analyzed 2,936 Indonesian participants to construct an Indonesian obesity PRS. To localize the T2D partitions, we harmonized SNPs with the published T2D PPRS and addressed limited direct overlap via LD proxy mapping. For this, we built a merged GRCh38 LD reference by liftover of West Javanese whole-genome sequences (n=227), performing per-chromosome QC and imputation against 1000 Genomes, and then merging the imputed West Java genomes with 1000 Genomes to form a combined panel. Signed LD correlations (r) were computed within 1 Mb of PPRS loci, enabling projection of T2D effects onto Indonesian obesity SNPs. Partition-specific hybrid weights were then formed by blending projected T2D betas with Indonesian obesity betas using biologically informed α parameters. An ancestry analysis confirmed that Indonesian samples cluster distinctly from other 1000 Genomes groups, supporting the need for population-aware LD reference. Direct overlap between T2D PPRS and our obesity PRS comprised a single SNP; LD mapping recovered 10 additional proxies with r² > 0.4. The evaluation on the UK Biobank Asian subset, PPRS-ID achieved an AUC of 0.633 for T2D discrimination. In a head-to-head test of the obesity pathway within Indonesians, the Indonesian obesity PRS outperformed the original obesity partition (AUC 0.594 vs. AUC 0.465). PPRS-ID demonstrates a feasible path to population-tailored, pathway-aware T2D risk prediction in Indonesians. Ongoing work focuses on larger Southeast Asian validation, refined partition weighting, and assessment of clinical utility.

• Halleluyah Oludele, College of Medicine, University of Ibadan, Nigeria
• Abdul Latif Koney Shardow, Kumasi Centre for Collaborative Research in Tropical
  Medicine, Ghana
• Maame Apaflo, Noguchi Memorial Institute for Medical Research, Ghana
• Jonas Ibekwe, College of Medicine, Department of Medicine and Surgery,
  University of Ibadan, Nigeria
• Phazha Baeti, Department of Medical Laboratory Science, University of
  Bostwana, Gaborone, Botswana, Botswana
• Julien Nguinkal, Bernhard-Nocht-Institut für Tropenmedizin, Germany
• Olaitan Awe, African Society for Bioinoformatics and Computational
  Biology, South Africa

Background: The widespread rollout of antiretroviral
therapy (ART) in Africa has significantly reduced HIV
mortality but has also driven the emergence of
drug-resistant HIV-1 strains. Genomic surveillance is vital
to guide treatment policies in high-burden regions.

Objective: We characterize HIV-1 drug resistance mutations
(DRMs) in the pol gene across Africa (1990–2024), examining
temporal trends, ART impact, regional subtype dynamics, and
cross-resistance patterns. Our findings aim to inform
resistance screening and optimize therapeutic strategies.

Methods: From 217,227 sequences in the Los Alamos HIV
database, we analyzed 105,970 African pol sequences,
filtering for quality and completeness. Drug resistance
profiles were computed using the Stanford HIVDB’s SierraPy
tool, extracting DRM data via custom JSON parsing. Analyses
included: (1) subtype distribution and diversity (Shannon
index); (2) DRM prevalence, frequency, and regional
co-occurrence; and (3) time series analysis of resistance
scores using moving averages, clustering, and
correlation-based cross-resistance modeling.

Results: Subtype C dominated (46.8%) but declined over
time; A1 and CRF02_AG increased. DRMs were present in 34.3%
of sequences, with NNRTI resistance (EFV/NVP ~27%) most
common. Key mutations included M184V (NRTI), K103N and
Y181C (NNRTI). Resistance surged post-2000, aligned with
ART scale-up. Strong within-class cross-resistance was
observed for PIs and NRTIs (r > 0.90), while INSTI
resistance emerged post-2016. EFV/NVP resistance patterns
were nearly identical (r = 0.98); newer agents showed
distinct profiles.

Conclusion: Our analysis highlights evolving HIV-1
resistance driven by ART pressure. The high prevalence of
NNRTI resistance and regional variability underscore the
need for localized, data-informed treatment protocols and
continuous genomic surveillance.

• Elizabeth Amelia, Department of Bioengineering - Imperial College London, United Kingdom
• Chayanit Piyawajanusorn, Department of Bioengineering - Imperial College London, HRH
  Princess Chulabhorn College of Medical Sciences, Thailand
• Pedro Ballester, Department of Bioengineering - Imperial College London,
  Royal Society Wolfson Fellow, United Kingdom

Background: Transcriptomic biomarkers for predicting response to Anti-PD1 immunotherapy rarely generalise beyond their discovery cohort, limiting clinical utility. This failure, often due to dataset shift, remains poorly dissected.

Methods: We developed 20 gene-panel biomarkers using a consensus pipeline (BioAdapt) on four melanoma and renal cell carcinoma cohorts (N=360) and validated them across five independent datasets, yielding 75 cross-cohort experiments. Our framework comprised three stages: (1) Diagnosis of dataset dissimilarity using the Maximum Mean Discrepancy test, (2) Technical Correction using label-free normalisation, and (3) Biological Stratification of patients by similarity to training data using UMAP.

Results: Direct model transfer produced near-random performance (mean MCC ~0). Diagnostic testing confirmed severe dataset shift (mean Maximum Mean Discrepancy = 0.584). Correction modestly improved results but remained poor. In contrast, biological stratification uncovered subgroups with meaningful gains, nearly doubling corrected performance (best-case MCC: 0.41).

Conclusion & Significance: Our large-scale analysis shows that biomarker generalisation failure stems less from batch effects than from model brittleness: fitted weights are highly local to training data. While technical correction is necessary, stratifying patients by biological similarity can recover predictive power. This work quantifies the limits of current biomarkers and outlines a path toward robust, subgroup-specific clinical translation.

• Hao Jiang

Single-cell transcriptomics has revolutionized the characterization of cell states, with multi-omics profiling further enabling deep functional insights. However, effectively mining cellular heterogeneity from noisy, sparse, and complex single-cell data remains challenging. We propose two novel ensemble learning frameworks that construct and optimize the consensus matrix in an adaptive and robust manner. First, we introduce scEWE, a self-representative ensemble learning method that employs a high-order element-wise weighting strategy to assign cell-specific weights to base clusterings. Second, we present scRECL, a contrastive ensemble learning approach that leverages siamese neural networks trained on multiple K-nearest neighbor partitions to derive low- dimensional embeddings. By integrating multiplex graphs for representative cell selection, scRECL effectively reduces noise and redundancy while enhancing latent feature representation. Both method capture nonlinear cellular relationships and significantly improve the accuracy of cellular similarity representation. Extensive experiments demonstrate that our approaches outperform existing state-of-the-art techniques, underscoring their effectiveness and potential for robust single-cell data analysis.

• Dongjun Guo, University College London, United Kingdom
• Deborah Dunn-Walters, University of Surrey, United Kingdom
• Franca Fraternali, University College London, United Kingdom
• Joseph Ng, University College London, United Kingdom

The development of stable antibodies formed by compatible
heavy (H) and light (L) chain pairs is crucial in both the
in vivo maturation of antibody-producing cells and the ex
vivo designs of therapeutic antibodies. We present here a
novel machine learning framework, ImmunoMatch, for
deciphering the molecular rules governing the pairing of
antibody chains. Fine-tuned on an antibody-specific
language model, ImmunoMatch learns from paired H and L
sequences from single human B cells to distinguish cognate
H-L pairs and randomly paired sequences. We find that the
predictive performance of ImmunoMatch can be augmented by
training separate models on the two types of antibody L
chains in humans, kappa and lambda, in line with the in
vivo mechanism of B cell development in the bone marrow.
ImmunoMatch can be applied to facilitate pairing of heavy
and light immunoglobulin chains in Spatial VDJ sequencing
data, and understand factors affecting pairing specificity.
Using ImmunoMatch, we illustrate that refinement of H-L
chain pairing is a hallmark of B cell maturation in both
health and disease. We find further that ImmunoMatch is
sensitive to sequence differences at the H-L interface.
ImmunoMatch focuses on H-L chain pairing as a specific,
under-explored problem in antibody developability, and
facilitates the computational assessment and modelling of
stably assembled immunoglobulins towards large-scale
optimisation of efficacious antibody therapeutics.

• Rintaro Shiroshima, Kyushu Institute of Technology, Japan
• Kazuhiro Takemoto, Kyushu Institute of Technology, Japan

Vision-language models (VLMs) are increasingly integrated into medical workflows, yet their security vulnerabilities remain largely unexplored. While recent studies demonstrated susceptibility of proprietary VLMs to prompt injection attacks, comparative security of medical-specific models like Google’s MedGemma against multiple attack vectors has not been systematically evaluated. This study comprehensively assesses vulnerability of medical-domain and proprietary VLMs to both prompt injection and adversarial perturbation attacks. Using medical images with confirmed malignant lesions across six modalities, we evaluated attack success rates across MedGemma, GPT-5, Claude 4 Sonnet, and Claude 4.1 Opus. White-box attacks utilized full model access, while black-box attacks employed transfer-based methods using surrogate models. Results reveal complex vulnerability patterns: MedGemma exhibited strongest resistance to prompt injection (38% ASR) but, as an open-source model, showed expected high vulnerability to white-box adversarial attacks (>80% ASR). Even in black-box deployment, MedGemma demonstrated moderate vulnerability (37% ASR). Claude 4 Sonnet appeared most robust against transfer attacks (6% ASR), however this reflected a high baseline lesion miss rate (48%) rather than genuine robustness. GPT-5 and Claude 4.1 Opus showed intermediate vulnerabilities. No single model provided universal robustness. These findings emphasize that robust medical AI security requires multi-layered defenses targeting both text-based and image-based attack vectors.

• Shumin Li, The University of Hong Kong, Hong Kong
• Ruibang Luo, The University of Hong Kong, Hong Kong
• Yuanhua Huang, The University of Hong Kong, Hong Kong

Large-scale genomic language models (gLMs) hold promise for
modeling gene regulation, yet their ability to predict
personal gene expression remains largely unexplored. We
developed xDecoder, an individual gene expression decoder
framework to assess the representational power of gLMs
(Evo2, Nucleotide Transformer, Caduceus) and
sequence-to-function (S2F) models (Enformer, Borzoi,
AlphaGenome) using paired personal genome–transcriptome
data. We find that training on personal genomes improves
prediction for seen genes in new individuals, suggesting
potential applications in few-shot contexts such as rare
variant interpretation. However, both gLM- and S2F-derived
embeddings struggle to capture inter-individual variability
for unseen genes, highlighting a persistent challenge in
transferring cis-regulatory grammar. To address this, we
incorporated personal chromatin accessibility (ATAC-seq)
data, which increased performance for unseen-gene
prediction moderately but significantly. These results
highlight the limitations of DNA–RNA models alone and point
to the importance of integrating individual-level
multi-omic data as a promising direction for improving
personal gene expression prediction.

• Priya Ramarao-Milne, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Anubhav Kaphle, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Anne Klein, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Michael Kuiper, Data 61, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Roc Reguant, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Brendan Hosking, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Julika Wenzel, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Hawlader Al-Mamun, Data 61, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Lujain Elazab, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Qing Zhong, ProCan®, Children's Medical Research Institute, Australia, Australia
• Roger Reddel, ProCan®, Children's Medical Research Institute, Australia, Australia
• Laurence Wilson, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Yatish Jain, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Natalie Twine, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia
• Denis Bauer, AeHRC, Commonwealth Scientific and Industrial Research
  Organisation, Australia, Australia

Background: Resistance to targeted cancer therapies remains
a major barrier to effective treatment, with both intrinsic
and acquired mechanisms contributing to poor clinical
outcomes. While large-scale molecular profiling has
advanced our understanding of cancer biology, the molecular
determinants of treatment resistance remain incompletely
understood. This underscores the need for scalable,
interpretable approaches to identify robust biomarkers and
interaction networks influencing drug response. Recent
advances in proteomics, including the release of the
world’s largest pan-cancer proteomic dataset, offer
unprecedented opportunities to explore protein-level
determinants of drug susceptibility.
Methods: We present synerOmics, a machine learning
framework designed to identify putative synergistic protein
interactions by leveraging parent-child co-occurrences in
random forest regression trees, effectively capturing
non-linear, context-specific interactions that may underlie
drug susceptibility. We validated synerOmics using
simulated data and two independent cancer proteomic
datasets, identifying reproducible interaction signatures.
Results: synerOmics identified molecular pathways
associated with drug resistance; notably, interactions
enriched for endoplasmic reticulum stress and unfolded
protein response were consistently identified across
datasets. The framework rediscovered known sensitivity
markers for tyrosine kinase inhibitors, including EGFR and
ERBB2 in the context of lapatinib response, demonstrating
its ability to capture established biology while revealing
novel interaction networks. Interestingly, synerOmics
uncovered previously uncharacterised hub proteins
associated with lapatinib resistance, demonstrating
predictive and prognostic value across 11 tumour types in
publicly available TCGA datasets. Lastly, to investigate
the structural basis of how our candidate biomarker may
influence lapatinib activity and contribute to resistance,
we applied AlphaFold3 for structure prediction of
ligand-bound protein complexes, enabling us to assess
conformational dynamics and potential alterations in
drug–protein interactions that may underpin resistance.
Conclusion: By focusing on interpretability of interaction
patterns, synerOmics provides a scalable and generalisable
tool for biomarker discovery. This work highlights the
potential of machine learning-guided proteomic integration
to guide experimental validation in precision oncology.

• Jerry Xu, MIT PRIMES, United States
• Shaojun Pei, Brigham and Women's Hospital, United States
• Gil Alterovitz, Harvard Medical School, United States

Structural comparison between proteins is key to many
research tasks, including evolutionary analysis,
peptidomimetics, and functional annotation. Traditional
structure alignment tools based on three-dimensional
protein structures, such as TM-Align, DALI, or ProBiS, are
accurate, but they are computationally expensive and
impractical at scale. Existing protein language models
(PLMs), such as TM-Vec, improve computational efficiency
but only capture global structural similarity, overlooking
important motif-level structural details. In this paper, we
propose a novel PLM consisting of a Siamese neural network,
enabling efficient embedding-based structural comparison
while also capturing both global and local structural
similarity. Our model was trained on a dual loss function
combining TM-score, a global similarity metric, and a
variation of lDDT scores, a per-residue similarity metric.
We tested against two datasets: a varied TM-score dataset
from TM-Vec, and a high TM-score mutant dataset from VIPUR.
Against these sets, our model achieved a TM-score MAE of
0.0741 and 0.0583, respectively, and a lDDT-score MAE of
0.0788 and 0.0038, respectively. Our model fulfills two key
roles: first, it rapidly detects global structural
differences. Second, it supports fine-grained structural
assessments, improving sensitivity to subtle but
functionally important structural changes.

• Noam Shomron, Steven Brenner

• Noam Shomron, Steven Brenner

• Stephen Kwok-Wing Tsui

Accurate protein classification has long been a fundamental challenge in bioinformatics and molecular biology. With the advent of large language models (LLMs), such as the GPT series, new opportunities have arisen for improving protein classification accuracy. These models, originally designed for natural language processing, have demonstrated remarkable capabilities in understanding and generating text. Leveraging these capabilities, researchers have started exploring their potential in protein classification tasks. In this talk, I will introduce two projects: allergen prediction and antibody characterization. The allergen prediction project applies protein LLMs and outperforms previous studies. The antibody characterization project utilizes 18 popular LLMs across 5 families in 3 specific antibody characterization scenarios. Both projects demonstrate the superior performance of applying LLMs in protein classification problems. We anticipate that the rapid development of LLMs will further revolutionize this specific area and the general bioinformatics field.

• Shujun Wang, The Hong Kong Polytechnic University, Hong Kong
• Ye Du, The Hong Kong Polytechnic University, Hong Kong
• Qian Zhao, The Hong Kong Polytechnic University, Hong Kong

De novo peptide sequencing is a fundamental computational
technique for ascertaining amino acid sequences of peptides
directly from tandem mass spectrometry data, eliminating
the need for reference databases. Cutting-edge models
usually encode the observed mass spectra into latent
representations from which peptides are predicted
autoregressively.

However, the issue of missing fragmentation, attributable
to factors such as suboptimal fragmentation efficiency and
instrumental constraints, presents a formidable challenge
in practical applications. To tackle this obstacle, we
propose a novel computational paradigm called Latent
Imputation before Prediction (LIPNovo). LIPNovo is devised
to compensate for missing fragmentation information within
observed spectra before executing the final peptide
prediction. Rather than generating raw missing data,
LIPNovo performs imputation in the latent space, guided by
the theoretical peak profile of the target peptide
sequence. The imputation process is conceptualized as a
set-prediction problem, utilizing a set of learnable peak
queries to reason about the relationships among observed
peaks and directly generate the latent representations of
theoretical peaks through optimal bipartite matching. In
this way, LIPNovo manages to supplement missing information
during inference and thus boosts performance. Despite its
simplicity, experiments on three benchmark datasets
demonstrate that LIPNovo outperforms state-of-the-art
methods by large margins. Code is available at
https://github.com/usr922/LIPNovo.

• Jingcheng Zhang, The University of Hong Kong, Hong Kong
• Zhenxian Zheng, The University of Hong Kong, Hong Kong
• Ruibang Luo, The University of Hong Kong, Hong Kong

DNA has emerged as a compelling archival storage medium,
offering unprecedented information density and
millennia-scale durability. Despite its promise, DNA-based
data storage faces critical challenges due to error-prone
processes during DNA synthesis, storage, and sequencing. In
this study, we introduce Gungnir, a codec system using the
proof-of-work idea to address substitution, insertion, and
deletion errors in a sequence. With a hash signature for
each data fragment, Gungnir corrects the errors by testing
the educated guesses until the hash signature is matched.
For practicality, especially when sequenced with nanopore
long-read, Gungnir also considers biochemical constraints
including GC-content, homopolymers, and error-prone motifs
during encoding. In silico benchmarking demonstrates its
outperforming error resilience capacity against the
state-of-art methods and achieving complete binary data
recovery from a single sequence copy containing 20%
erroneous bases. Gungnir requires neither keeping many
redundant sequence copies to address storage degradation,
nor high-coverage sequencing to address sequencing error,
reducing the overall cost of using DNA for storage.

• Jiaxuan Li, Southern University of Science and Technology, China
• Shimin Shuai, Southern University of Science and Technology, China

Viral infections pose major economic and public health
challenges worldwide, with vaccines as a critical tool for
prevention. Synonymous recoding of viral genes through
codon and codon-pair deoptimization offers a promising
approach to design live attenuated vaccines (LAVs) by
reducing viral fitness without altering protein sequences.
This strategy has been successfully applied to develop
vaccines for a range of pathogens affecting human and
livestock. To support this approach, we developed CocoVax,
the first web server dedicated to codon and codon-pair
deoptimization for LAV design. CocoVax features four
modules: Virus Database, Gene Recoder, Sequence Evaluator,
and Reference Library, guiding users through the entire
vaccine development process. With its intuitive interface,
CocoVax enables rapid generation of vaccine candidates
using only a pathogen’s gene sequence, providing a valuable
resource for researchers in virology and vaccine
development. CocoVax is freely accessible at
https://comics.med.sustech.edu.cn/cocovax with no login
required.

• Xinyue Chen, Zhejiang University–University of Edinburgh Institute(ZJU-UoE Institute), Zhejiang University, China, China
• Fenglan Pan, Zhejiang University–University of Edinburgh Institute(ZJU-UoE Institute), Zhejiang University, China, China
• Jiale Qin, Department of Human Genetics and Women's Hospital, ZhejiangUniversity School of Medicine, Hangzhou, 310058, China, China
• Zhiang Chen, Department of Human Genetics and Women's Hospital, ZhejiangUniversity School of Medicine, Hangzhou, 310058, China, China
• Xiaodong Wu, Department of Human Genetics and Women's Hospital, ZhejiangUniversity School of Medicine, Hangzhou, 310058, China, China
• Chen Li, Department of Human Genetics and Women's Hospital, ZhejiangUniversity School of Medicine, Hangzhou, 310058, China, China

Ovarian cancer is a highly lethal gynecological malignancy,
often diagnosed at advanced stages. Standard therapy
combines cytoreductive surgery with paclitaxel-carboplatin,
yet responses differ markedly between patients. Capturing
how exposure and tumor response contribute to these
divergent outcomes is challenging, as conventional
experimental systems are limited in scale and duration, and
existing computational models rarely simulate full
treatment courses. New approaches are needed to study
tumor-drug interactions over clinically relevant timescales
and within sizeable 3D tumor spheroids.

We developed OVTwin, a GPU-accelerated digital twin
framework, that models drug diffusion, cellular uptake, and
cell fate dynamics in a three-dimensional ovarian tumor
spheroid, and performs massively parallel simulations of
the standard paclitaxel-carboplatin regimen over three
chemotherapy cycles (63 days). The framework employs a
fully GPU-parallelized architecture to run large-scale
simulations and support exploration of multiple treatment
settings. OVTwin simulated spheroids containing up to
500,000 cells within one hour, accurately capturing the
time-course dynamics of cell proliferation and apoptosis
under the standard paclitaxel-carboplatin regimen.

By capturing the dynamics of tumor-drug interactions at
cellular resolution and enabling rapid, large-scale
simulations, OVTwin provides a framework for exploring
patient-specific chemotherapy responses. This study
highlights the potential of GPU-accelerated digital twins
to advance precision oncology in cancer treatment.

• Feng Yan, Walter and Eliza Hall Institute of Medical Research, Australia
• Pedro L. Baldoni, Walter and Eliza Hall Institute of Medical Research, Australia
• James Lancaster, Walter and Eliza Hall Institute of Medical Research, Australia
• Matthew E. Ritchie, Walter and Eliza Hall Institute of Medical Research, Australia
• Mathew G. Lewsey, La Trobe Institute for Molecular Sciences, La Trobe
  University, Australia
• Quentin Gouil, Walter and Eliza Hall Institute of Medical Research, Australia
• Nadia M. Davidson, Walter and Eliza Hall Institute of Medical Research, Australia

Long-read RNA sequencing has advanced transcriptomics by
enabling the full length of transcripts to be profiled, yet
most analyses still rely on a high-quality reference genome
and gene annotation. Recently, long-read de novo
transcriptomic assembly methods have been developed;
however, benchmarking of long-read transcriptome assembly
and established protocols for downstream analyses remain
limited.

Here, we comprehensively evaluate three long-read de novo
transcriptome assembly tools, RATTLE, RNA-Bloom2 and
isONform, and compare to the short-read assembler Trinity.
We assess assembly quality, computational efficiency,
expression quantification, and critically the impact on
downstream differential transcript and gene analysis. The
datasets used include simulated data and sequin spike-in
(truth known), and human cell line and pea (Pisum sativum)
samples (truth unknown but defined using the
reference-guided method Bambu). To represent contemporary
analysis scenarios, we analysed data ranging from 6 million
to 60 million reads, Nanopore cDNA and direct RNA
sequencing protocols.

Our results show that long reads generate longer assembled
contigs, but a more redundant transcriptome. RNA-Bloom2,
coupled with Corset for transcript clustering, was the best
performing in both accuracy and computational efficiency.
Our findings offer guidance when selecting the most
effective strategy for long-read differential expression
analysis, when a high-quality reference genome is
unavailable.

• Changde Cheng, The University of Alabama at Birmingham, United States
• Sajesan Aryal, The University of Alabama at Birmingham, United States
• Brittany Curtiss, The University of Alabama at Birmingham, United States
• Xinyue Zhou, The University of Alabama at Birmingham, United States
• Rui Lu, The University of Alabama at Birmingham, United States

Resistance to single-agent therapy in acute myeloid
leukemia (AML) often arises from compensatory epigenetic
circuits. We present a causal inference framework that
treats therapy as an intervention (do-operator) and defines
resistance as the deviation of dual perturbations from the
additive expectation of monotherapies. By modeling the
causal structure of the leukemia stem-cell epigenetic
landscape that maintains identity and fitness, the
framework shows that synergistic and antagonistic
interactions can be inferred directly from single
perturbations, suggesting interaction inference may be
simpler than previously assumed. We applied this model to
Perturb-seq of 16 epigenetic regulators in KMT2A-rearranged
AML (>31,000 single-cell transcriptomes). From
single-knockout profiles, the framework correctly predicted
synergistic pairs (Menin+KAT6A, Menin+DOT1L) that enhanced
differentiation and induced cell death, as well as an
antagonistic pair (DOT1L+PCGF1) that conferred resistance.
Predictions were validated by pharmacologic inhibition and
bulk RNA-seq, confirming the model’s predictive accuracy.
This work presents a mechanism-guided approach for rational
prioritization of epigenetic drug combinations in AML and
other cancers.

• Myeongbin Oh, Myongji University, South Korea
• Minsik Oh, Myongji University, South Korea

Single-cell RNA-seq profiles cells at high resolution but
suffers from high-dimensional, sparse, and dropout data.
Large-scale models capture global gene patterns yet miss
condition-specific signals, while graph methods recover
local structure without broad context. To bridge this gap,
we introduce a dual-encoder with cross-attention that
integrates global gene representations with structure from
protein–protein interaction and cell-similarity graphs.
Jointly encoding genes and cells captures both global and
local context, producing robust, interpretable
representations across disease datasets.

• Bairui Du, The University of Tokyo, Japan
• Hisanori Kiryu, The University of Tokyo, Japan

Understanding the tumor microenvironment (TME) requires
models that capture not only single-cell expression but
also the higher-order gene–gene relationships that shape
local niches. We introduce CoSMIM (Covariance of Spatial
Multi-scale Integrated Microenvironment), a novel
representation that encodes gene–gene covariance structures
across cellular, niche, and tissue levels. Unlike
conventional expression matrices, CoSMIM leverages
multi-scale covariance to preserve context-dependent
regulatory patterns that emerge in spatial transcriptomics
and imaging-derived data.
We applied CoSMIM to breast cancer invasive ductal
carcinoma (IDC) datasets, where the spatial heterogeneity
of immune infiltration and stromal remodeling poses a major
challenge for interpretation. By integrating CoSMIM with a
multi-decoder conditional variational autoencoder (CVAE),
we learned a unified latent representation that
reconstructs niche-specific covariance features across
scales. This approach revealed distinct TME states in IDC,
including immune-suppressed epithelial niches,
fibroblast-driven stromal clusters, and regions of
coordinated immune activation.
These findings demonstrate that CoSMIM provides a scalable
framework for niche-level analysis of the TME, with
immediate relevance to breast cancer. Beyond IDC, the
method generalizes to other tumor types and can integrate
multi-omics spatial platforms, offering a new pathway to
dissect microenvironmental heterogeneity at multiple scales.

• Florian Schmidt, ImmunoScape Pte Ltd, Singapore
• Kan Xing Wu, ImmunoScape Pte Ltd, Singapore
• Yovita Ida Purwanti, ImmunoScape Pte Ltd, Singapore
• Nicholas Tan, ImmunoScape Pte Ltd, Singapore
• Daniel Carbajo, ImmunoScape Pte Ltd, Singapore
• Bok Ke Xin, National Cancer Center Singapore, Singapore
• Andreas Wilm, ImmunoScape Pte Ltd, Singapore
• Michael Fehlings, ImmunoScape Pte Ltd, Singapore
• Daniel MacLeod, ImmunoScape Pte Ltd, United States
• Alessandra Nardin, ImmunoScape Pte Ltd, Singapore
• Daniel Tan, National Cancer Center Singapore, Singapore
• Katja Fink, ImmunoScape Pte Ltd, Singapore

Immune checkpoint inhibitors (ICI) are becoming the
standard of care in multiple cancer indications.
Yet, previously identified biomarkers of response have
relatively low accuracy, a patient’s response to ICI
treatment is virtually unpredictable even with confirmed
expression of the relevant targets such as PD-1 or PD-L1.
Here, we comprehensively phenotyped peripheral blood CD8+ T
cells from patients with non-small cell lung cancer by
analyzing surface markers, transcriptome and T-cell
receptor (TCR) repertoire in single cell resolution.
The cohort was comprised of patients that a) responded to
anti-PD(L)1 treatment for a prolonged period of time, b)
were new-on-treatment responders and c) were
new-on-treatment non-responders.
We identified response specific signals in cell type/state
proportions, in TCR repertoire diversity and TCR inter
donor similarity. Furthermore, using machine learning, we
identified cell type specific signatures that predicted the
ICI response with an accuracy between 66\% and 93\% at
single cell and up to 94\% at patient level. CD44, GIMAP4,
CD69 and CCL4L2 were among the most relevant markers
defining the predictive signature of effector memory
T-cells on lung cancer samples.
Our findings suggest that CD8+ T cell subset-specific
models reach an accuracy that possesses the potential to
inform treatment decisions in a clinical setting.

• Menghan Wang, Beijing Normal-Hong Kong Baptist University, China
• Junya Yang, Beijing Normal-Hong Kong Baptist University, China
• Jiaxing Chen, Beijing Normal-Hong Kong Baptist University, China

Cell–cell communication (CCC) plays a central role in
tissue homeostasis, development, and disease progression.
Although single-cell and spatial transcriptomics have
greatly advanced the study of ligand–receptor interactions,
existing methods remain limited in modeling competition,
capacity constraints, and multistep relay cascades. We
present CellFlow, a novel framework that formulates CCC as
a bounded minimum-cost flow problem. By incorporating
receptor abundance constraints, spatial distance penalties,
and molecular affinity weighting at the single-cell level,
CellFlow precisely characterizes competition and allocation
within signaling pathways. Benchmarking with partial
differential equation–based simulations demonstrated that
CellFlow significantly outperforms CellChat and COMMOT in
reconstructing realistic communication fields. Applied to
human lymph node spatial transcriptomics, CellFlow revealed
the critical role of CCR7 in T-zone organization and
delineated receptor-specific as well as cluster-level
communication patterns. Overall, CellFlow provides a
biologically grounded and computationally scalable tool for
inferring communication from single-cell and spatial
transcriptomic data, opening new avenues for systematically
dissecting tissue signaling networks and exploring
therapeutic targets.

• Hibiki Sugiyama, The Universiy of Osaka, Japan
• Hironori Shigeta, The Universiy of Osaka, Japan
• Shunji Umetani, The Universiy of Osaka, Japan
• Shigeto Seno, The Universiy of Osaka, Japan

Spatial transcriptomics enables the measurement of gene expression while preserving spatial context. A key step in such analyses is mapping transcripts to their corresponding cells, since downstream interpretation relies on accurate cell-level profiles. Conventional approaches typically segment stained images to define cell boundaries, but image-based methods often suffer from low quality, labor-intensive preparation, and poor performance in dense or noisy regions. We propose an image-free method that simultaneously estimates cell types and locations directly from transcript point clouds. The central idea is to use the distinctive expression profiles of cell types to guide a hierarchical partitioning, which resolves coarse cell-type domains into individual cells. First, we perform quadtree-based spatial partitioning to approximate coarse domains for each cell type. Within each domain, a probabilistic clustering model integrates spatial coordinates and gene identities, where each mixture component corresponds to an individual cell. We evaluated the method on both synthetic and real datasets, including a human lung cancer sample acquired with the Xenium platform. The results show that our approach successfully assigns transcripts to individual cells without relying on images, offering a robust foundation for downstream spatial analyses.

• Lana Garmire

Drug repurposing is a cost-effective strategy for accelerating therapeutic discovery, yet existing single-cell RNA-seq (scRNA-seq)-based methods often overlook the spatial context critical for capturing tissue-specific drug responses. We introduce STDrug (Spatial Transcriptomics aided Drug repurposing), a personalized computational framework that leverages spatial transcriptomics data to improve drug repurposing. STDrug identifies paired spatial domains between diseased and control tissues using graph convolutional networks and coherent point drift alignment. It then prioritizes candidate drugs by integrating tumor-reversible gene signatures, perturbation-based reversal scores, and GPT-4o-assisted gene weighting through an XGBoost model. By incorporating spatial domain interactions, drug efficacy, and toxicity, STDrug computes robust patient-level drug scores. Applied to hepatocellular carcinoma and prostate cancer, STDrug outperformed scRNA-seq-based approaches, achieving higher AUCs and consistent predictions across patients. Real-world validation using large-scale electronic health records and in vitro validation support the clinical relevance of top-ranked candidates, establishing STDrug as a spatially informed platform for personalized drug repurposing.

• Yuki Matsukiyo, Department of Complex Systems Science, Graduate School of
  Informatics, Nagoya University, Japan
• Chikashige Yamanaka, Department of Complex Systems Science, Graduate School of
  Informatics, Nagoya University, Japan
• Yoshihiro Yamanishi, Department of Complex Systems Science, Graduate School of
  Informatics, Nagoya University, Japan

Drug development is a complex and resource-intensive, making efficient hit identification crucial. Recently, de novo drug design using deep generative models has been proposed as a promising approach for more efficient hit identification. However, existing methods have focused primarily on generating chemically valid structures and often neglect the biological context, particularly the cellular responses to drug candidates. In this study, we propose a novel computational method that integrates transcriptome data to generate drug candidate compounds for therapeutic targets. The proposed method consists of two steps. First, we selected source molecules based on a transcriptome profile analysis. Then, we generated new molecules by fusing substructures of multiple source molecules using a Transformer-based variational autoencoder and Bayesian optimization. We applied the proposed method to the inhibitor or activator design for 10 therapeutic target proteins. In performance comparison with previous methods, our proposed method generated molecules more closely resembling known ligands for seven out of ten target proteins, indicating its enhanced capability to produce more ligand-like molecules compared to the previous methods. This method paves the way for integrating biological context into molecular design, potentially accelerating the discovery of drug candidate compounds.

• Sanjana Fatema Chowdhury, Bangladesh Council of Scientific and Industrial Research
  (BCSIR), Bangladesh
• Murshed Hasan Sarkar, Bangladesh Council of Scientific and Industrial Research
  (BCSIR), Bangladesh
• Syed Muktadir Al Sium, Bangladesh Council of Scientific and Industrial Research
  (BCSIR), Bangladesh
• Md. Salim Khan, Bangladesh Council of Scientific and Industrial Research
  (BCSIR), Bangladesh

Coronavirus disease 2019 (COVID-19), once a global
pandemic, continues to pose challenges to human health
through its long-term consequences and evolving variants.
Emerging evidence highlights the role of the human
microbiome in influencing disease severity, yet studies
exploring host–pathogen interactions at the transcriptomic
level remain limited. Here, we analyzed metatranscriptomic
profiles of forty nasopharyngeal samples from COVID-19
patients across Bangladeshi cohorts. Sequencing data were
processed to assess taxonomic composition, microbial
diversity, and antimicrobial resistance gene (ARG)
expression. COVID-19 positive and asymptomatic patients
exhibited a higher abundance of pathogenic and
multidrug-resistant bacteria, whereas healthy and recovered
individuals showed greater fungal diversity. Functional
enrichment further indicated active ARG expression in
positive cases, linking the respiratory microbiome to
immune modulation. Differential expression analysis
revealed upregulation of immune-related genes, including
pro-inflammatory cytokines, in positive cases. Notably,
asymptomatic patients showed reduced TLR4 expression,
suggesting reduced innate immune activation that may
underline their clinical outcomes. These findings provide
insights into host–microbiome interactions underlying
COVID-19 severity and underscore the need for validation in
larger, ethnically diverse cohorts with comprehensive
clinical metadata.

• Tong Wang, Shandong University, China
• Lingyu Li, The University of Hong Kong, China
• Zhi-Ping Liu, Shandong University, China

Complex diseases remains difficult to detect early because conventional diagnostics rely on static biomarkers that emerge late. We present getDNB, a computational framework that identifies dynamic network biomarker (DNB) from temporally evolving gene regulatory networks via graph embeddings and anomaly detection. Briefly, getDNB has three steps: (i) construct stage-specific regulatory networks to capture molecular dynamics during disease progression; (ii) use graph convolutional networks to derive topology-preserving low-dimensional embeddings; and (iii) quantify gene-level abnormalities via K-means clustering and outlier scores, then refine candidates using minimum dominating set and shortest-path criteria to ensure connectivity and reduce redundancy. Moreover, we define the Dynamic Network Index (DNI) to quantify temporal disorder and flag critical transition states. Applied to a real-world hepatocellular carcinoma dataset, getDNB identified 33 robust DNBs and their interaction network, achieving high predictive accuracy (AUROC = 0.929). Notably, the DNI showed a pronounced increase at the pre-disease stage, consistent with complex systems transition theory. Functional enrichment associated these DNBs with oncogenic pathways, including hepatocellular carcinoma, hepatitis B infection, and cell cycle regulation. In conclusion, getDNB offers a mechanism-informed approach to dynamic biomarker discovery, enabling sensitive detection of early-warning signals in HCC with potential translational value. getDNB has been published in Bioinformatics: https://doi.org/10.1093/bioinformatics/btaf518.

• Tsz Pan Tong, University of Luxembourg, Luxembourg
• Jun Pang, University of Luxembourg, Luxembourg

Gene regulatory networks (GRNs) are essential in studying cell differentiation and development. However, existing GRN inference models often disregard the network nature of GRNs, undermining the potential for GRN inference from scRNA-seq data. Popular GRN models such as PIDC and GENIE3 only consider at most triplet interactions among genes, significantly underestimating the regulation complexity in real biological systems. To showcase the importance of the network structure in GRNs, we proposed a novel GRN aggregation method, MuxGRN, which leverages multiplex link prediction to aggregate GRNs inferred from various state-of-the-art models and extracts the information propagated within different layers of GRNs. We tested MuxGRN on 14 real scRNA-seq datasets aggregated from 9 GRN models and compared it with 2 baseline edge-level aggregation methods. Our evaluation shows that MuxGRN achieves a top-tier averaged AUPRC among individual GRN models and discovers up to 26.55\% new true positive edges, significantly outperforming baselines. This highlights the potential of link prediction in GRN aggregation for mining weak signals from inferred GRNs, and suggests that future development of GRN models should emphasize the network nature of GRNs.

• Kai Ye

Genomic sequences are inherently linear, and researchers have traditionally favored simplified modeling and computation directly at the sequence level. Following a reductionist approach, these sequences are often further abstracted into random strings for analysis. However, this simplification overlooks the intrinsic non-randomness of genomic sequences. As a result, when faced with diverse species, complex genetic backgrounds, or various disease contexts, researchers are frequently forced to introduce ad hoc modifications. This gradually transforms initially elegant mathematical models into overly complex, patchwork systems that are difficult to maintain and extend. To address these limitations, we previously explored solving complex genomic problems—such as complex structural variant detection and genome annotation—by transforming one-dimensional sequences into two-dimensional image representations. This approach has yielded promising results. By projecting linear sequences into image space, we enable hybrid encoding of both data and domain knowledge. The external (shallow) features of sequences—such as linear and nonlinear base arrangements and repetitive patterns—can be effectively captured through image-based representations. Simultaneously, the internal (deep) structural, functional, and biological regularities of the genome can be encoded using the multi-channel spatial advantages of images. Overall, this image-based strategy enhances both the dimensionality and depth of genomic sequence representation, allowing us to uncover complex patterns and relationships that remain hidden in one-dimensional sequence space.

• TBD

TBD

• Xuegong Zhang

Building computational virtual models of cells that can simulate the biochemical processes of cells has been a long-standing goal for many scientists in the fields of molecular and cell biology, computational biology, and systems biology. Recently, the unprecedented accumulation of massive single-cell transcriptomics and other omics data, together with the revolutionary breakthrough in AI foundation models, have aroused a new wave for building AI models of virtual cells or AIVCs. As a common phenomenon in a rapidly developing field, people often use the same word to mean different things and also use different words to mean the same thing. This talk will provide an overview of the different views and approaches for building virtual cells, propose a multi-level construct of AIVC which academic labs of smaller scales can work on, and share our practices in building prototypic AIVC models with AI foundation models and biology-aware interpretable AI models.

• Xing Yi

TBD

• Hafida Hamdache, CRCT inserm COMPO Inria, France
• Vera Pancaldi, CRCT inserm, France
• Sebastien Benzekry, COMPO Inria, France

The clinical outcomes can vary from spontaneous regression
to high metastatic disease. This extracranial tumour arises
from a neural crest-derived cell and can harbor different
phenotypes. This leads downstream to disruption of
homeostasis and a metabolic shift in response to the tumour
needs. In this study we aim to characterize the
neuroblastoma tumour microenvironment (TME) and metabolic
profiles, integrating features in a machine learning
framework to predict survival in High Risk patients. We
have used Multideconv to retrieve cellular composition of
the TME. CellTFusion allowed the identification of specific
cell niches with the underlying transcription regulatory
network (TRN). Metaflux was used to characterize metabolic
profiles of patients and ml.tidy to predict survival in
high risk patients. We have identified 6 TME clusters that
stratify patient survival and 126 cell subgroups, among
which a subset predicts NB risk categories with high
accuracy. Different cell niches and TRN associated with
immune activation or proliferation were found to be related
to clinical outcomes. Combining cell niches with altered
metabolic pathways such as beta oxidation linked to fatty
acid metabolism and hydroxylysine phosphorylation related
to cell proliferation and signaling could predict HRNB
survival with a C-index of 0.71.

• Louisse Mirabueno, EMBL's European Bioinformatics Institute (EMBL-EBI), United Kingdom
• Jorge Batista da Rocha, EMBL's European Bioinformatics Institute (EMBL-EBI), United Kingdom
• Aleena Mushtaq, EMBL's European Bioinformatics Institute (EMBL-EBI), United Kingdom

The Ensembl Genome Browser has been an indispensable tool in the field of genomics, aiding scientists around the world in their research for over 25 years. The Ensembl database provides visualisation and comprehensive analyses of integrated genomic data, including genes, variants, comparative genomics and gene regulation, for over 5,000 eukaryotic and over 30,000 prokaryotic genomes. With the rapid increase of genomic data in the past decade, a new Ensembl data platform (https://beta.ensembl.org/) has been designed to provide efficient access to more genomes than ever before. The new platform offers enhanced features, improved data visualisation, and an intuitive user interface. It currently presents genomic features for over 3,000 species with a new user interface, including visualisation of genes and variants on the Human Pangenome Reference Consortium (HPRC) and the Darwin Tree of Life assemblies. Regular updates will expand this coverage, with thousands of genomes to be added over time to match and exceed the breadth of the legacy Ensembl resource. Attendees will learn about the range of data and tools available through the new platform, and see the latest features it offers for the retrieval and interpretation of genomic data.

• Jiabao Li, The Hong Kong University of Science and Technology, Hong Kong
• Jiguang Wang, The Hong Kong University of Science and Technology, Hong Kong

Advances in single-cell and spatial transcriptomics have
transformed our understanding of tumor cell states and
lineage dynamics, enabling high-resolution investigation of
cellular heterogeneity and evolution.

Building on Schiffman et al.’s Markovian framework for
inferring cell state transitions from single-cell lineage
tracing, we introduce StateSim, a spatial-temporal model
that simulates tumor cell state transitions within tissue,
incorporating spatial growth and mutation accumulation.
StateSim assumes that daughter cells are spatial neighbors,
shaping the spatial organization of cell states, and models
two cell states with distinct birth and death rates. To
quantify spatial patterns, we developed StateMap, a
network-based approach that maps tumor cell states and
their spatial relationships, summarizing transition
patterns through network statistics.

Applying StateSim under various transition scenarios, we
demonstrate that distinct transition modes yield unique
spatial statistics, robustly correlated with transition
rates (Pearson r = 0.9938, P = 6.87e-10). Using spatial
transcriptomics data from glioma samples, StateMap
identified subgroups with distinct transition patterns,
further validated by spatial statistics such as bivariate
Moran’s I. Approximate Bayesian computation with StateSim
enabled inference of posterior probabilities for cell state
transition rates and phylogenetic trajectories. Our
framework enables in-silico lineage tracing from spatial
data, advancing studies of tumor heterogeneity and
evolution.

• Hyun Uk Kim

The utilization of bio big data through computational models can aid in predicting drug targets for various diseases as well as drug responses. In cancer research, for example, a substantial amount of bio big data has been accumulated, including patient-specific omics data (e.g., RNA-seq) and medical data (e.g., survival data). In this talk, I will first elaborate on computational approaches that leverage omics data, particularly RNA-seq, to predict drug targets. Genome-scale metabolic models (GEMs), which simulate a target cell’s metabolic phenotypes by accounting for all metabolic reactions in the cell, play a key role here. These computational approaches aim to predict oncometabolites and drug targets for drug-resistant cancer cells and high-risk cancer patients. Oncometabolites, which accumulate abnormally in cancer cells due to gene mutations, exhibit pro-oncogenic functions and may also serve as biomarkers. Furthermore, I will discuss AI-driven approaches for predicting drug responses, where medical data play a crucial role. Ongoing efforts to generate and apply meaningful bio big data, alongside the effective use of computational models, are set to revolutionize our approach to addressing medical challenges.

• TBD