OmniMedical: Multi-Agent Explainable Drug Discovery Platform

An end-to-end, multi-agent and explainable AI platform for small-molecule drug discovery and intelligent delivery

Project Overview

OmniMedical is an end-to-end multi-agent explainable drug discovery platform that covers the full pipeline from target-centric molecular generation and virtual screening to lead optimization and intelligent delivery design.
The system is designed as a rigorous AI-for-drug-discovery research platform rather than a demo or competition project.

Demo Video

If the embedded player cannot load properly, you can open / download the video file directly.

System Architecture Figure

OmniMedical System Architecture

Research Motivation

Current small-molecule drug discovery pipelines suffer from high experimental cost, long development cycles and fragmented tooling. In particular:

  • Conventional virtual screening pipelines are expensive to scale and difficult to interpret.
  • Existing AI models often focus on isolated stages (e.g., generation or docking) and rarely provide an end-to-end view from target to candidate.
  • Delivery design and late-stage optimization are typically decoupled from early-stage molecular design.

OmniMedical aims to provide a single, modular research platform that integrates these stages into an explainable, multi-agent workflow.

System Design

OmniMedical is organized into three conceptual layers:

  1. Data & Knowledge Layer
    • Curated disease–target databases, molecular libraries and biomedical knowledge graphs.
  2. Algorithm & Agent Layer
    • Task-specific agents for molecular generation, drug repurposing, virtual screening, evaluation and delivery design.
    • Coordination mechanisms that combine large language models with graph neural networks and physics-inspired models.
  3. Application Layer
    • Web-based interface for interactive analysis, visualization and experiment management.

Core Modules

1. Target-aware Multimodal Molecular Generation

  • Uses both 3D target structures and protein sequences as conditioning signals.
  • Structure branch: TamGen-style structure-guided generator for pocket-specific ligand design.
  • Sequence branch: reinforcement learning–enhanced large language model for SMILES/graph generation.
  • Cross-consistency constraints encourage candidates that are simultaneously structure-compatible and chemically feasible.

2. Knowledge-graph-driven Drug Repurposing

  • Large-scale biomedical knowledge graph connecting diseases, targets, drugs and pathways.
  • Multi-agent reasoning framework:
    • Graph reasoning agent proposes candidate disease–target–drug triples.
    • Mechanism analysis agent aggregates literature and pathway evidence.
    • Safety agent filters candidates using side-effect and pharmacokinetic profiles.

3. Three-stage Virtual Screening Pipeline

  1. GNN-based Primary Screening
    • Uses DumplingGNN to score large molecular libraries based on physicochemical properties, ADMET profiles and toxicity surrogates.
  2. Deep-learning-based Docking and Scoring
    • Applies a Boltz2-style deep docking model to predict binding poses and affinity for top-ranked molecules.
  3. Pose Refinement and BioScore Aggregation
    • Refines poses and computes a BioScore-derived composite score, yielding an interpretable ranking.

4. Multi-dimensional Evaluation and Explainability

  • Multi-criteria scoring across potency, ADMET, toxicity and developability.
  • Attention and attribution techniques provide substructure-level explanations for model decisions.
  • Full traceability from final score back to individual module contributions.

5. Lead Optimization and Delivery Design

  • Retrieval-augmented reinforcement learning loop for iterative lead optimization.
  • Automatic proposal of delivery strategies (e.g., formulation and route hints) with structured reports for downstream experimental design.

Experimental Evaluation

  • Benchmarked on multiple public ADMET and property-prediction datasets, where individual components of OmniMedical achieve competitive or state-of-the-art performance.
  • End-to-end studies show that the integrated pipeline outperforms conventional docking-centric workflows in both screening efficiency and hit quality, while providing richer interpretability.

Relation to Other Work

OmniMedical is conceptually related to recent multimodal and cross-architecture frameworks (e.g., DrugCLIP-like systems), but emphasizes:

  • Explicit multi-agent decomposition of the drug discovery pipeline.
  • Tight integration of GNNs, deep docking models and LLM-based agents.
  • End-to-end explainability across generation, screening and delivery stages.