How AI is Revolutionizing Chemistry and Drug Discovery: From Data to Breakthroughs

Introduction

Li Yu, researcher at the Guangdong‑Hong Kong‑Macao Greater Bay Area Digital Economy Research Institute (IDEA), introduces the emerging field of AI for Science, describing the shift to a fourth paradigm of data‑intensive scientific discovery.

Challenges in Chemistry and Pharma

The traditional drug‑development pipeline suffers from three major pain points: long cycles (over 10 years), high costs (exceeding $1 billion), and low success rates (over 90% of candidates fail in clinical trials). Domestic data show 32% of target drugs rely on imports and 54% of domestic drugs need stability improvements, highlighting an urgent need for innovation.

AI for Science offers a historic opportunity by coupling data and models to provide accurate predictions and dramatically shorten research cycles, exemplified by AlphaFold’s rapid protein‑structure prediction.

AI Solutions: Model Building

LigUnity : a contrast‑learning model for precise protein‑ligand affinity prediction, achieving superior virtual screening speed (40 days → 40 hours) and iterative hit‑to‑lead optimization.

AbLingua : a language model for antibodies that expands the amino‑acid token vocabulary to 8 000, delivering performance comparable to much larger models.

idealFold : an antibody‑structure predictor that combines AbLingua embeddings with a dedicated structure module, achieving industry‑leading accuracy.

InstructMol : a multimodal molecular assistant aligning molecular graph embeddings with large language models (e.g., Llama) to perform tasks such as molecular description generation and reaction prediction.

Presto : an extension of InstructMol capable of handling arbitrary numbers of molecular inputs/outputs, excelling in yield prediction and reaction‑condition selection.

Data as the Foundation

High‑quality, large‑scale chemical data are essential but fragmented across patents, literature, and databases. Extraction challenges include complex content (structures, tables, multimodal data), diverse layouts, and multilingual documents.

IDEA’s solution combines "big models" (GPT‑4, Gemini) for multilingual understanding and reasoning with specialized "small models" for tasks such as layout‑recognition, table parsing (96% accuracy), optical chemical structure recognition (OCSR, >95% accuracy), and entity‑relation linking (94% accuracy). This pipeline reduces the time to parse a thousand‑page patent from weeks to under an hour.

Future Outlook

Goals include building an end‑to‑end AI system that links literature understanding, target discovery, molecular design, property prediction, and synthesis planning via multi‑agent collaboration; creating a dry‑wet experimental closed loop by integrating AI predictions with automated lab platforms; and expanding expertise from drug discovery to new materials and energy domains.

The ultimate vision is for AI to break human cognitive limits, enabling discoveries such as novel chemical reactions that could merit Nobel‑level impact.