DrugGen 2: A disease-aware language model for enhancing drug discovery

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology · Depth: Expert, extended

Summary

DrugGen-2 is a novel generative model designed to enhance drug discovery by creating small molecules conditioned on both disease ontology (MeSH) and target protein sequences. Developed by fine-tuning a pre-trained GPT-2 model, DrugGen-2 utilizes a two-step strategy involving supervised fine-tuning and reinforcement learning via group relative policy optimization (GRPO). This process was guided by reward functions optimizing for chemical validity, novelty, diversity, and high predicted binding affinity. When evaluated on five protein targets relevant to diabetic nephropathy, DrugGen-2 significantly outperformed baseline models like DrugGPT and DrugGen. It generated a higher number of unique molecules (409-444 vs. 50-219), demonstrated greater structural similarity to approved drugs (0.70 vs. 0.30-0.64), and achieved superior predicted binding affinities (median 9.26-9.97 vs. 5.86-8.49). Molecular docking analyses further identified candidate ligands with predicted affinities, such as -9.917 for ACE, exceeding reference drugs like enalapril (-8.283).

Key takeaway

For AI Scientists and Research Scientists focused on drug discovery, DrugGen-2 provides a powerful framework for de novo design and drug repurposing. By integrating disease ontology with target protein sequences, your generative models can produce molecules with significantly higher predicted binding affinities and greater relevance to specific disease contexts. You should explore incorporating similar disease-aware conditioning into your computational drug design pipelines to accelerate the identification of potent and clinically relevant drug candidates.

Key insights

DrugGen-2 is a disease-aware generative AI model that significantly improves de novo drug design by integrating disease ontology.

Principles

• Disease context is crucial for effective target-based drug design.
• Reinforcement learning with tailored rewards improves generative molecular models.
• Hybrid SFT and RL training enhances generative model performance.

Method

Fine-tune a pre-trained GPT-2 on a curated disease-target-drug dataset via supervised fine-tuning, followed by Group Relative Policy Optimization (GRPO) guided by rewards for binding affinity, diversity, and novelty.

In practice

• Design novel small molecules for specific disease-target pairs.
• Repurpose existing drugs by generating context-aware ligands.

Topics

• DrugGen-2
• Generative Models
• Drug Discovery
• Disease-Aware AI
• Reinforcement Learning
• Molecular Design

Code references

• alimotahharynia/DrugGen-2
• huggingface/trl
• rdkit/rdkit

Best for: AI Scientist, Research Scientist

Related on AIssential

• DrugGen 2: A disease-aware language model for enhancing drug discovery — DrugGen-2 integrates disease context and target sequences to generate novel small molecules with superior binding affinities.
• 3 Questions: Using AI to accelerate the discovery and design of therapeutic drugs — AI, particularly deep learning and generative models, significantly accelerates novel antibiotic discovery and design.
• Generative Artificial Intelligence in Bioinformatics: A Systematic Review of Models, Applications, and Methodological Advances — Generative AI, particularly specialized models, is transforming bioinformatics by enabling de novo design and superior analytical performance.
• Bridging the phenotype-target gap for molecular generation via multi-objective reinforcement learning — ExMolRL combines phenotype-guided generation with target-aware reinforcement learning for de novo drug discovery.
• SPOKE: A massive biomedical knowledge graph for precision health and drug discovery — SPOKE integrates vast biomedical data into a knowledge graph for drug discovery and precision medicine.
• HEDGEHOG: Hierarchical Evaluation of Drug Generators Through Rigorous Filtration — Molecular generators often fail to produce compounds simultaneously satisfying all practical drug discovery criteria.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.