• Jatin Joshi

• Sung Wing Kin

Deletions and tandem duplications (commonly called CNVs) represent the majority of structural variations in a human genome. They can be identified using short reads, but because they frequently occur in repetitive regions, existing methods fail to detect most of them.

To improve the detection rate, we introduce a new CNV short reads-based caller named SurVIndel2. SurVindel2 builds on statistical techniques we previously developed, but also employs a novel type of evidence, hidden split reads, that can uncover many CNVs missed by existing algorithms. Comparing with other popular callers, SurVIndel2 outperforms them both on human and non-human datasets. We demonstrate the practical utility of the method by generating a catalogue of CNVs for the 1000 Genomes Project that contains hundreds of thousands of CNVs missing from the most recent public catalogue.

Furthermore, we also study the calling of indels of size <50bp. We showed that SurVIndel2 is able to complement small indels predicted by Google DeepVariant, and the two software used in tandem produce a remarkably complete catalogue of variants in an individual.

• Xiaotong Gu, School of Chemistry and Molecular Biosciences, The
  University of Queensland, Australia
• Ashar Malik, School of Chemistry and Molecular Biosciences, The
  University of Queensland, Australia
• David Ascher, School of Chemistry and Molecular Biosciences, The
  University of Queensland, Australia

The functional essentiality of protein glycosylation across
the human population has not been systematically quantified
and it is a challenge for interpreting missense variants
within the glycoproteome. To address this, we integrated
multi-modal data from UniProt, ClinVar, and gnomAD,
leveraging AlphaFold structures to systematically analyse
the genetic and structural context of variants relative to
glycosylation sites. Our analysis revealed that pathogenic
variants are significantly enriched in structurally
constrained regions and exhibit strong signatures of
negative selection. Using a deep neural network that
combines sequence and structure-aware embeddings, we
successfully predicted variant pathogenicity, achieving
high discriminative performance (Accuracy =84.8%, ROC AUC
=0.855) on an independent test set. Complementing this
genomic perspective, we introduce SGGly, a highly accurate
computational tool for N-linked glycosylation site
prediction. SGGly moves beyond limited sequence windows by
integrating full protein sequence ProtBERT embeddings with
structural features such as solvent accessibility. On a
low-redundancy test set, our model achieved exceptional
predictive performance (MCC =0.856, AUC =0.985),
outperforming existing tools and confirming that integrated
sequence and structural context is essential for accurate
site determination. Variant-level pathogenicity and
detailed proteomic and genomic context are available in a
3D interactive visualisation, while SGGly is accessible at
https://biosig.lab.uq.edu.au/sggly/.

• Chuyun Zhang, Department of Biomedical Sciences, College of Biomedicine, City University of Hong Kong, Hong Kong
• Kei Hang Katie Chan, Department of Biomedical Sciences, College of Biomedicine, City University of Hong Kong; Department of Electrical Engineering, City University of Hong Kong; Department of Epidemiology and Biostatistics, Center for Global Cardiometabolic Health and Nutrition, University of California, Irvine School of Medicine, Hong Kong

Background: Alzheimer's disease (AD) involves a complex
interplay of distinct cellular states and dysfunctional
cross-talk. While single-cell omics can chart this
heterogeneity, a major analytical gap persists. Current
approaches rely on sequential, disconnected analyses that
cannot resolve how genetic risk converges with specific
cell states and communication networks. This critical gap
impedes a systems-level understanding of the AD
microenvironment, creating a pressing need for an
integrative analytical framework.

Methodology: We developed a novel computational framework
that systematically integrates five analytical dimensions:
(1) compositional perturbation analysis using
quasi-binomial and linear regression models with adaptive
dispersion to handle biological heterogeneity, (2)
single-cell decorrelated module networks (scDemon) for
multi-resolution gene module discovery, (3) cell-cell
communication inference using LIANA, (4) trajectory
analysis with scFates, and (5) single-cell disease-relevance
scoring (scDRS). We validated this framework on 12
integrated snRNA-seq datasets from the SSREAD resource,
encompassing human entorhinal and prefrontal cortex.

Results: We identified three AD-enriched cell types:
microglia (OR=6.43, p_adj=0.0001), vasculature, and
oligodendrocyte, with microglia most significant.ScDemon
revealed novel microglial functional programs—including
filopodia dynamics (MYO10/PARVG) and proliferation (CDKN1A)
modules—that form distinct transcriptional states along a
PAGA trajectory. LIANA prioritized pathological RTN4-LINGO1
signaling that inhibits neuronal repair, while scDRS mapped Alzheimer's disease genetic risk to microglial cells.

Conclusion: Our work establishes a unified multi-omics
framework that bridges gene expression, cell-cell
communication, and genetic data layers to resolve complex
disease mechanisms. By systematically linking polygenic
risk to specific cellular states and their communication
networks, this integrative approach advances systems
pathology and provides a generalizable blueprint for
multi-omics discovery across neurodegenerative and complex
diseases.

• Syed Muktadir Al Sium, Bangladesh Council of Scientific and Industrial Research, Bangladesh
• Mahafujul Islam Quadery Tonmoy, Bangladesh Council of Scientific and Industrial Research, Bangladesh
• Sanjana Fatema Chowdhury, Bangladesh Council of Scientific and Industrial Research, Bangladesh
• Fatema Tuz Zohra, Bangladesh Council of Scientific and Industrial Research, Bangladesh
• Jean-Christophe Nebel, Kingston University London, United Kingdom
• Farzana Rahman, Kingston University London, United Kingdom

Background; Long non-coding RNAs (lncRNAs) are crucial regulators in cancer, yet their epigenetic control in Lung Adenocarcinoma (LUAD) remains underexplored. Aim; This study aimed to integrate multi-omics data to identify novel prognostic biomarkers and elucidate their potential mechanism. Methods; We analyzed RNA-Seq and Illumina 450k methylation data from 473 LUAD and 32 normal The Cancer Genome Atlas (TCGA) samples. We performed differential expression and methylation analysis to identify candidate lncRNAs whose expression was negatively correlated with promoter methylation. The prognostic value was evaluated using Kaplan-Meier curves and multivariate Cox regression. A lncRNA-miRNA-mRNA regulatory network was constructed to investigate potential mechanisms. For validation, FAM83A-AS2 expression was quantified in a preliminary set of clinical blood samples using RT-qPCR. Results; Our multi-omics analysis identified two lncRNAs, FAM83A-AS2 and AC012213.1, whose high expression was driven by significant promoter hypomethylation and correlated with significantly lower patient survival. Differential analysis and correlation revealed promoter hypomethylation of specific CpG sites (cg19924352 for FAM83A-AS2; cg16648062 and cg20129213 for AC012213.1) drives their upregulation. Multivariate Cox regression confirmed their status as independent prognostic markers after adjusting for clinical covariates (HR=1.55, p<0.01 for FAM83A-AS2; HR=1.30, p<0.05 for AC012213.1), with strong diagnostic potential (AUC=0.72). A regulatory network analysis implicated these lncRNAs in modulating key LUAD-associated genes like RALGPS2, HOXA13 via miRNA MIR126 and MIR34C. Gene set enrichment analysis further linked these lncRNAs to fundamental molecular processes like chromatin modification and DNA methylation. Importantly, the pilot wet-lab validation on an initial set of clinical blood samples supported these findings, demonstrating a marked upregulation of FAM83A-AS2 in patients compared to healthy controls. Conclusion; This study presents FAM83A-AS2 and AC012213.1 as promising biomarkers for risk stratification and potential therapeutic targets in LUAD.

• Dimple Khatri, UMBC, United States
• Sanjan Tp Gupta, IIT Madras, India

Medical imaging workflows integrate radiology images with
their corresponding free-text reports. Large language
models (LLMs) and large vision–language models (LVLMs)
achieve strong results but face deployment barriers in
hospitals due to computational demands, privacy risks, and
infrastructure needs. Small language models (SLMs) and
small vision–language models (SVLMs), typically under 10B
parameters, provide a more efficient and auditable
alternative for on-premise, privacy-preserving applications
in radiology. Recent advancements, including CheXzero,
MedCLIP, XrayGPT, LLaVA-Med, MedFILIP, and MedBridge, show
that smaller multimodal models support classification,
retrieval, and report generation. Complementary baselines
from lightweight SLMs such as DistilBERT, TinyBERT,
BioClinicalBERT, and T5-Small highlight opportunities for
radiology report understanding.

Building on these efforts, we propose a reproducible
evaluation framework anchored on MIMIC-CXR, with potential
extensions to CT, MRI, and ophthalmology datasets. Our
framework integrates task metrics such as AUROC, ROUGE, and
F1 score, together with efficiency measures including VRAM
usage, latency, and model size, alongside trust dimensions
like factuality, calibration, and robustness. We also
conduct ablation studies on model architecture, tokenizers,
and parameter-efficient fine-tuning (e.g., qLoRA), while
analyzing trade-offs between accuracy, efficiency, and
stability. This work establishes reproducible baselines and
guidance for deploying radiology AI.

• Ashitha Washington, Computational Biology and Bioinformatics Lab, National
  Institute of Technology Calicut, India
• Ravindra Kumar, Computational Biology and Bioinformatics Lab, National
  Institute of Technology Calicut, India

Polycystic ovary syndrome (PCOS) represents a multifaceted
endocrine condition marked by genetic, molecular, and
phenotypic variability. To uncover consistent
transcriptomic biomarkers and prognostic gene networks
linked to PCOS, we performed an integrative analysis of
RNA-Seq data compiled from publicly available Gene
Expression Omnibus datasets, comprising 65 PCOS cases and
61 healthy controls across diverse cell types. Data
preprocessing involved normalization followed by
differential expression analysis. Feature selection was
then performed via Elastic Net regression, effectively
managing multicollinearity and refining the feature set to
83 candidate genes for subsequent modeling.
Multiple machine learning classifiers were trained and
validated using a 60:20:20 data split, with hyperparameter
optimization to enhance predictive performance. Among
these, the Support Vector Machine (SVM) model exhibited the
highest classification capability, achieving 92.31%
accuracy on the internal validation set and an impressive
AUC of 0.98. Model explainability was strengthened using
SHAP and LIME analyses, pinpointing the most influential
genes driving model predictions. Logistic regression based
on the key gene clusters produced a prognostic framework
with an AUC of 0.82 and precision of 0.8, suggesting their
robustness as biomarkers despite PCOS heterogeneity.
Functional enrichment results revealed that these genes are
predominantly involved in RNA-binding processes, ribosomal
machinery, and immune regulation. Overall, this integrative
multi-cohort analysis coupled with advanced machine
learning provides a powerful strategy for identifying
clinically actionable biomarkers and prognostic signatures
in PCOS, offering new avenues for molecular diagnosis and
therapeutic development.

• Michael Oluwasola, Universiti Putra Malaysia, Malaysia
• Noor Dina Muhd Noor, Universiti Putra Malaysia, Malaysia
• Thean Chor Leow, Universiti Putra Malaysia, Malaysia

The generation of small molecules capable of simultaneously
modulating multiple biological targets is an emerging
frontier in polypharmacology, offering new opportunities
for treating complex and multifactorial diseases. However,
existing generative frameworks often rely on autoregressive
or fragment-based methods that inadequately capture the
intricate spatial correlations, stereochemical
dependencies, and multi-target constraints inherent in real
molecular systems. To overcome these limitations, we
introduce GEMINI-Mol, a unified deep generative framework
that combines SE(3)-equivariant transformers, diffusion
probabilistic modeling, and graph neural networks for fully
3D-aware multi-target molecular design. By enforcing
rotational and translational equivariance, GEMINI-Mol
preserves molecular geometry and stereochemistry while
modeling long-range atomic interactions underlying
multi-site binding. The architecture integrates built-in
chemical validity filters and uncertainty quantification
mechanisms to ensure the generation of physically plausible
and interpretable molecules. A progressive training
strategy enables efficient co-optimization of diffusion,
transformer, and graph modules, supported by gradient
checkpointing and optimized attention for large-scale
molecular modeling. Across comprehensive benchmarks,
GEMINI-Mol demonstrates superior performance in generating
structurally diverse, chemically sound, and drug-like
molecules with favorable predicted affinities toward
multiple protein targets. Collectively, GEMINI-Mol
establishes a next-generation generative paradigm that
unites geometric deep learning and probabilistic modeling
for rational, multitarget drug discovery.

• Somenath Dutta, Pusan National University, South Korea
• Sudipta Sardar, Pusan National University, South Korea

MicroRNAs (miRNAs) are critical 22-23 nucleotide regulatory
molecules that control gene expression through the
miRNA-induced silencing complex (miRISC), influencing
diverse cellular processes including development,
differentiation, and disease progression. Despite two
decades of research, the molecular mechanisms governing
miRNA-target interactions remain incompletely understood,
with functional targeting occurring through both canonical
seed-region pairing and non-canonical mechanisms involving
complex structural interactions. Current computational
approaches for miRNA target prediction have evolved from
early heuristic methods to machine learning and deep
learning frameworks, yet most rely on one-dimensional
sequence representations that may fail to capture the
intricate spatial relationships and long-range dependencies
critical for target recognition.
We present miTarCGR, a revolutionary deep learning
framework that transforms miRNA target prediction into a
computer vision problem using Frequency Chaos Game
Representation (FCGR). This innovative approach converts
linear miRNA and target sequences into two-dimensional
graphical representations that preserve both local and
global sequence characteristics, enabling convolutional
neural networks to identify complex patterns extending
beyond simple linear complementarity. By representing both
miRNA sequences and candidate target sites as FCGR images,
miTarCGR captures compositional features, structural
motifs, and discontinuous base pairing patterns that are
difficult to detect using conventional sequence-based
methods.
Through comprehensive evaluation on multiple benchmark
datasets, miTarCGR demonstrates superior performance
compared to state-of-the-art methods in both site-level and
gene-level target prediction tasks. The framework
incorporates advanced explainability techniques to provide
interpretable insights into learned features, crucial for
advancing biological understanding of miRNA targeting
mechanisms and building confidence in computational
predictions. Our results suggest that two-dimensional
sequence representation provides a more comprehensive view
of miRNA-target interactions, potentially leading to novel
targeting mechanism discovery and improved therapeutic
target identification. This work represents a paradigm
shift in miRNA target prediction, offering a powerful tool
for understanding post-transcriptional regulation and
advancing precision medicine applications.

• Xuewei Cao, Center for Statistical Genetics, The Gertrude H. Sergievsky
  Center, Columbia University, New York, NY, USA, United States
• Haochen Sun, Computational and Systems Biology, Sloan Kettering
  Institute, Memorial Sloan Kettering Cancer Center, New
  York, NY, USA, United States
• Ru Feng, Center for Statistical Genetics, The Gertrude H. Sergievsky
  Center, Columbia University, New York, NY, USA, United States
• Rahul Mazumder, Operations Research Center, Massachusetts Institute of
  Technology, Cambridge, MA, USA, United States
• Carlos F Buen Abad Najar, Department of Human Genetics, University of Chicago,
  Chicago, IL, USA, United States
• Yang Li, Department of Human Genetics, University of Chicago,
  Chicago, IL, USA, United States
• Philip L. de Jager, Department of Neurology, Columbia University, New York, NY,
  USA, United States
• David Bennett, Rush Alzheimer’s Disease Center and Department of
  Neurological Sciences, Rush University Medical Center,
  Chicago, IL, United States
• Kushal Dey, Computational and Systems Biology, Sloan Kettering
  Institute, Memorial Sloan Kettering Cancer Center, New
  York, NY, USA, United States
• Gao Wang, Center for Statistical Genetics, The Gertrude H. Sergievsky
  Center, Columbia University, New York, NY, USA, United States

Background: Multi-trait QTL (xQTL) colocalization has shown
great promises in identifying causal variants with shared
genetic etiology across multiple molecular modalities,
contexts, and complex diseases. However, the lack of
scalable and efficient methods to integrate large-scale
multi-omics data limits deeper insights into xQTL
regulation.
Methodology: We propose ColocBoost, a multi-task learning
colocalization method that can scale to hundreds of traits,
while accounting for multiple causal variants within a
genomic region of interest. ColocBoost employs a
specialized gradient boosting framework that can adaptively
couple colocalized traits while performing causal variant
selection, thereby enhancing the detection of weaker shared
signals compared to existing pairwise and multi-trait
colocalization methods.
Results: We applied ColocBoost genome-wide to 17 gene-level
single-nucleus and bulk xQTL data from the aging brain
cortex of ROSMAP individuals (average N=595), encompassing
6 cell types, 3 brain regions and 3 molecular modalities
(expression, splicing, and protein abundance). Across
molecular xQTLs, ColocBoost identified 16,503 distinct
colocalization events, exhibiting 10.7(±0.74)-fold
enrichment for heritability across 57 complex
diseases/traits and showing strong concordance with
element-gene pairs validated by CRISPR screening assays.
When colocalized against Alzheimer’s disease (AD) GWAS,
ColocBoost identified up to 2.5-fold more distinct
colocalized loci, explaining twice the AD disease
heritability compared to fine-mapping without xQTL
integration. This improvement is largely attributable to
ColocBoost’s enhanced sensitivity in detecting gene-distal
colocalizations, as supported by strong concordance with
known enhancer-gene links, highlighting its ability to
identify biologically plausible AD susceptibility loci with
underlying regulatory mechanisms. Notably, several genes
including BLNK and CTSH showed sub-threshold associations
in GWAS, but were identified through multi-omics
colocalizations which provide new functional support for
their involvement in AD pathogenesis.
Conclusions: Overall, ColocBoost provides a novel framework
to identify colocalized disease-critical functional signals
for varying number of phenotypes. R package colocboost is
freely available on CRAN
(https://CRAN.R-project.org/package=colocboost).

• Shivani Singh, Sharda University, India
• Sunita Sharma, Sharda University, India

Pyruvylation is a process in which a pyruvate group is
transferred to the sugar moiety. It is primarily found in
the majority of pathogenic and non-pathogenic species
across a vast class of bacteria, fungi, and yeast. Our
study explores the evolutionary relationship between
different classes of pyruvyltransferases across diverse
species, focusing on structural and sequence conservation
in bacteria. The structural similarity strongly resembles
the GT-A and GT-B classes of glycosyltransferases. E. coli
demonstrates a GT-B class with a higher number of
positively charged residues such as histidine, lysine, and
arginine in the binding site. We performed PCA and REMD
(Replica Exchange Molecular Dynamics) simulations, each
with 500 ns and two replicates. This approach helps
investigate the catalytic site and acceptor substrate
specificity along with the global conformational space.
Bayesian inference was implemented on a characterized set
of 59 pyruvyltransferases to construct a phylogenetic tree,
and MEME was used to study conserved motifs and residues.
By correlating structural folds, donor binding motifs, and
phylogenetic relationships, we aim to understand the
evolutionary origin of ketal pyruvyltransferases and
identify structural features that could be exploited for
inhibitor design or drug repurposing against bacterial
pathogens such as E. coli, A. baumannii, M. tuberculosis,
and B. fragilis.

• Lana Garmire

• Joshua Wing Kei Ho

• R. Sowdhamini

Our laboratory has been interested in relationships amongst protein domains where evolutionary
divergences may be large, but there are strong similarities in the overall shape and biological
function. We proposed structural motifs as constraints to drive computational searches in sequence
space. We next built phylogenetic trees to describe these evolutionary variations. This helped us to
ask various questions like – is there a common evolutionary ancestor for protein domains that
cluster together? Can we extrapolate common biological function whenever protein domains co-
cluster? Where do functionally important residues reside? In the process, we chose to study
medicinal plants where abundant information is present on evolutionary divergence of enzymatic
domains. We chose plants like Tulsi to perform genome assembly and I will share our experiences.
Finally, I will also present our more recent study of transcriptome of Aparajitha or Sankapushpi
where different biosynthetic enzymes and their tissue localisations could be analysed.

• Rishik Kondadadi, Eastview High school, United States

Drug repurposing provides an accelerated path to
therapeutic discovery in epilepsy, where nearly 30% of
patients remain resistant to available medications.
Computational approaches leveraging transcriptomic
signatures have shown potential, but critical limitations
persist. Connectivity mapping methods such as tau scoring
achieve only moderate accuracy, while recent deep learning
architectures (e.g., Lv et al., 2024) improve performance
but rely on sequential pipelines that lose biological
context between stages. Graph attention networks represent
a further advance for modeling drug–gene interactions, yet
they remain fundamentally correlational, leaving
predictions susceptible to confounding and spurious
associations.

We introduce the Causal Temporal Diffusion Network (CTDN),
a novel framework designed to move beyond correlation-based
learning by integrating mechanistic reasoning into drug
repurposing. CTDN introduces five key innovations: (1)
causal discovery modules that identify true drug–gene
relationships, (2) diffusion-based propagation to model the
spread of biological effects, (3) temporal dynamics to
capture disease progression, (4) few-shot meta-learning to
address the scarcity of positive training examples, and (5)
neural-process–based uncertainty quantification to improve
interpretability and robustness.

Evaluated on a benchmark of 600 test samples spanning 26
unique antiepileptic drugs (AEDs), CTDN achieves an AUROC
of 0.7574, substantially outperforming traditional machine
learning methods. Moreover, CTDN delivers 65.4% recall@100
and 40% precision@10, demonstrating strong practical
utility for candidate ranking. These results highlight that
integrating causal reasoning and temporal biological
dynamics enables more accurate and mechanistically grounded
therapeutic predictions.

By addressing the shortcomings of correlation-driven
approaches, CTDN establishes a new paradigm for
computational drug repurposing in epilepsy and provides a
foundation for broader application in treatment-resistant
neurological disorders.

• Hsin-Ying Chang, Bioinformatics program, Taiwan International Graduate
  Program, Academia Sinica and National Taiwan University,
  Taipei, Taiwan
• Hsi-Ching Yen, Institute of Plant and Microbial Biology, Academia Sinica,
  Taipei, Taiwan
• Hsiu-An Chu, Institute of Plant and Microbial Biology, Academia Sinica,
  Taipei, Taiwan
• Chih-Horng Kuo, Institute of Plant and Microbial Biology, Academia Sinica,
  Taipei, Taiwan

Cyanobacteria are ecologically important phototrophs with
diverse biotechnological potential. To better characterize
the diversity of thermophilic cyanobacteria in Taiwan, we
conducted environmental sampling across 12 non-acidic hot
springs. we sampled 12 non-acidic hot springs and isolated
27 novel strains of Thermosynechococcus taiwanensis.
Together with previously studied isolates, 32 genomes from
11 springs were analyzed, revealing compact genomes
(2.64–2.70 Mb; ~2,537 genes) and a closed pan-genome of
~3,030 genes. Core-genome phylogeny and gene flow analyses
divided strains into two populations, shaped partly by
isolation by distance. To investigate divergence and
potential adaptations, we identified genomic regions of
reduced nucleotide diversity, revealing 149 and 289 genes
in populations A and B, suggesting selective sweeps within
each population. Only 16 genes were common to both
populations, indicating that selective sweeps primarily
targeted distinct sets of genes in each population. Key
genes involved in photosynthesis, motility, and ion
transport were highlighted. Overall, this study provides a
population genomics view of a hot spring cyanobacterial
species in Taiwan. In addition to shedding light on
microbial genomics and evolutionary processes, the genomes
and strains obtained here serve as important resources with
potential applications in future biotechnology.

• 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, Zhejiang University–University of Edinburgh Institute(ZJU-UoE Institute), Zhejiang University, China, China

Background: 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.

Methodology: We developed OVTwin (Ovarian Cancer Digital
Twin), 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 efficiently run
large-scale simulations and support exploration of multiple
treatment settings.

Result: 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. Additional
simulations of paclitaxel-only, carboplatin-only, and
untreated controls further confirmed that the combination
regimen achieved the most effective tumor suppression, in
line with clinical practice.

Conclusion: 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.

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

Background
Transcriptomic biomarkers for predicting immunotherapy response often fail to generalise across patient cohorts, limiting their clinical utility. This failure is frequently attributed to dataset shift, but the underlying mechanisms remain poorly dissected. Understanding and addressing this brittleness is critical for translating biomarkers into robust tools for precision oncology.

Methodology
We developed 20 gene-panel biomarkers using a consensus pipeline (BioAdapt) across four melanoma and renal cell carcinoma cohorts (N=360). To rigorously test generalisability, each cohort served as both training and independent test data, producing 75 cross-cohort validation experiments. Our framework comprised three stages: (1) diagnosis of dataset dissimilarity using the Maximum Mean Discrepancy test, (2) technical correction via label-free normalisation, and (3) biological stratification of patients by embedding similarity to training cohorts using UMAP.

Results
Direct transfer of models across cohorts yielded near-random performance (mean Matthews Correlation Coefficient ≈ 0). Statistical testing confirmed severe dataset shift (mean Maximum Mean Discrepancy = 0.584). Technical correction modestly improved performance but remained weak. In contrast, biological stratification identified patient subgroups where predictive accuracy nearly doubled compared to corrected models alone, with best-case MCC = 0.41.

Conclusions
Our results reveal that biomarker failure is driven less by batch effects than by inherent model brittleness: fitted weights are highly local to the discovery cohort. While technical correction is necessary, stratifying patients by biological similarity enables meaningful recovery of predictive signal. This study establishes a systematic framework for quantifying biomarker limitations and highlights adaptive stratification as a promising route toward subgroup-specific translation in cancer immunotherapy.

• Melissa Jane Fullwood

The integration of artificial intelligence (AI) with large-scale biological datasets is reshaping the landscape of cancer research. In this talk, I will present AI4Loop, a deep learning framework developed to predict promoter-promoter (P–P) chromatin interactions directly from RNA-Seq data. This model enables genome-wide inference of 3D chromatin architecture at scale, overcoming the technical and financial limitations of experimental mapping techniques such as Hi-C.
Applied to over 12,000 patient samples across 32 cancer types from The Cancer Genome Atlas (TCGA), AI4Loop reveals pervasive gains in P–P interactions across most malignancies, implicating chromatin interaction rewiring as a common oncogenic mechanism. Furthermore, by integrating transcriptomic profiles from over 50,000 drug-perturbation experiments, we identify small molecules capable of selectively disrupting cancer-specific chromatin interactions—findings supported by orthogonal validation using Hi-C.
In addition to this study, I will highlight broader applications of big data analysis and AI in understanding cancer. This work exemplifies a paradigm in which we can use AI and big data analysis strategies as tools to not only analyze complex biological data but also generates actionable hypotheses and identify novel therapeutic strategies—thereby advancing cancer biology and medicine.

• Jatin Joshi