Ensemble Multi-Modal Agentic AI Pipeline for Parkinson’s Disease Drug Discovery

Tuan Cao

Keywords: AI for drug discovery, Parkinson’s disease, Multi-modal models, Large language models, Generative chemistry, Mechanism-of-action modeling, Knowledge graphs

TL;DR: We present an ensemble, multi-modal agentic AI pipeline that fuses LLM reasoning with predictive screening to generate and triage PD candidates—narrowing 1,106 designs to five and surfacing an ambroxol-like lead with stronger GCase binding

Abstract: Parkinson’s disease (PD) remains without a cure, but recent advances in multi-modal AI offer new avenues for discovering disease-modifying treatments. We present a novel ensemble AI pipeline that integrates multiple state-of-the-art AI platforms – to identify and evaluate drug candidates for PD. Our system combines each platform’s outputs using ensemble learning to overcome the limitations of any single model. Focusing on the hypothesis of enhancing glucocerebrosidase (GCase) activity , the pipeline discovered five novel small molecules. Each candidate was evaluated across mechanism of action, blood-brain barrier permeability, ADMET properties, toxicity, manufacturability, and patent novelty.These results underscore the potential of combining multi-modal foundation models and LLM agents to accelerate drug discovery

Submission Number: 38