Multi-Omics for Drug Response and Combination Therapy Prediction

Precision oncology and pharmacology are shifting from single-biomarker thinking toward integrated, patient- and context-specific prediction of drug response. Multi-omics profiling—spanning genomics, transcriptomics, epigenomics, proteomics, metabolomics, microbiome signals, and single-cell or spatial data—offers a rich view of disease mechanisms, but it also creates a core challenge: how do we fuse heterogeneous, high-dimensional data into models that produce actionable and trustworthy therapy predictions, especially for drug combinations?

This topic focuses on concrete, modern computational approaches that use large language models (LLMs) and sequence-based transformer architectures to predict (1) drug sensitivity and resistance, (2) synergy/antagonism for combination therapy, and (3) biomarkers and mechanisms that explain model outputs. We welcome studies that move beyond “black-box multi-omics” toward structured, model-driven integration, including transformer-based fusion of omics modalities, representation learning on biological sequences (DNA/RNA/protein), and LLM-driven extraction and reasoning over biomedical knowledge.

A central goal is to bridge three information layers:
1. molecular state (multi-omics, single-cell/spatial),

2. perturbation space (drug target profiles, chemical structures, dose–response), and

3. knowledge space (pathways, networks, literature, clinical evidence). Transformers and LLMs offer practical routes to unify these layers—via self-supervised pretraining, multimodal attention mechanisms, retrieval-augmented generation (RAG), and graph/sequence hybrid architectures—supporting both prediction and interpretation.

Scope and Themes (examples, not limited to):
o Transformer-based multi-omics fusion (early/late fusion, cross-attention, mixture-of-experts, modality dropout).

o Sequence-based transformers for regulatory and functional prediction (e.g., encoding variants, promoters/enhancers, splicing, protein domains) linked to drug response.

o LLM-enabled knowledge integration: RAG pipelines that ground predictions in curated databases and literature; automated hypothesis generation for combinations.

o Drug and target representations using transformers (SMILES/graph transformers, protein sequence embeddings, target/pathway-aware attention).

o Combination therapy prediction: synergy scoring, multi-drug dose–response modeling, and mechanism-aware combination design.

o Causal and counterfactual learning for treatment effect estimation, resistance mechanisms, and patient stratification.

o Interpretability and trust: attention audits, concept bottlenecks, pathway attributions, uncertainty calibration, and robust evaluation across cohorts.

o Benchmarks and reproducibility: standardized datasets, leakage-resistant splits, external validation, and clinically meaningful metrics.

This topic invites contributions ranging from new model architectures and pretraining strategies to clinically grounded validation studies and open benchmarks. Submissions should aim to show clear utility for therapy selection, rational combination design, and mechanistic insight—making transformer and LLM-based approaches not just fashionable, but operational for drug response prediction.