Multimodal Foundation Models for Target ID and Lead Optimization

Drug discovery is entering a new phase where foundation models can learn transferable representations across biology and chemistry, then adapt to specific tasks with limited labeled data. This Research Topic will capture that momentum by focusing on multimodal foundation models that fuse protein and nucleic-acid sequences, 3D structures, small-molecule chemistry, omics profiles, phenotypes, and clinical/real-world signals to improve how we identify therapeutic targets and optimize leads.

We welcome work that advances models which can jointly predict target–disease links, target engagement, binding affinity, selectivity, polypharmacology, resistance risks, and downstream developability (e.g., ADMET, PK/PD proxies, safety flags, manufacturability). A central goal is to move beyond “single-modality winners” and toward rigorous comparisons and ablations across modality combinations—for example, when structure+ligand signals outperform (or fail to outperform) omics+phenotype, and how these tradeoffs shift by therapeutic area, data regime, and assay type.

A strong emphasis will be placed on generalization and real-world usefulness: prospective tests, out-of-distribution evaluation (new targets, new scaffolds, rare diseases, novel modalities), uncertainty estimation, and robust benchmarking that avoids leakage and overly optimistic splits. We also encourage contributions that address practical deployment, including scalable training, data governance, interpretability, human-in-the-loop design, and integration into existing discovery workflows (hit finding, multi-parameter optimization, and decision-making under uncertainty).

Suitable subtopics include (but are not limited to)
- Multimodal pretraining strategies (contrastive learning, masked modeling, diffusion, graph+3D hybrids, retrieval-augmented methods)
- Cross-modal alignment between sequence ↔ structure ↔ ligand ↔ omics/phenotype and task transfer
- Target identification from multi-omics, perturbation screens, single-cell, spatial data, EHR-derived phenotypes, and literature signals
- Binding, selectivity, and mechanism modeling (allostery, induced fit, kinetics, off-target profiling, pathway-level effects)
- Lead optimization objectives: potency + selectivity + ADMET + safety + developability as multi-objective learning
- Benchmarks and evaluation protocols for multimodal models (OOD splits, scaffold/target splits, time splits, prospective validation)
- Deployment and reproducibility: model monitoring, calibration, uncertainty, compute efficiency, and workflow integration
- Case studies showing measurable impact on target nomination, hit triage, or lead series advancement

This Topic aims to define what “good” looks like for multimodal foundation models in drug discovery: credible generalization, clear modality value, and deployable performance that helps teams make better choices, faster.