Customized Amazon Nova models improve molecular-property prediction in drug discovery

2026-04-15 · Source: Amazon Science homepage · Field: Science & Research — Health & Medical Research, Life Sciences & Biology, Research Methodology & Innovation · Depth: Advanced, medium

Summary

Amazon's Generative AI Innovation Center, in collaboration with Nimbus Therapeutics, has developed a customized Amazon Nova large language model (LLM) that significantly improves molecular-property prediction in drug discovery. This single, fine-tuned LLM unifies the prediction of 11 critical molecular properties across lipophilicity, permeability, and clearance, a task traditionally requiring multiple specialized graph neural networks (GNNs). The approach utilizes supervised fine tuning (SFT) and reinforcement fine tuning (RFT) on Nova 2 Lite, achieving accuracy comparable to or exceeding GNNs on 7 of 11 properties, with an average RMSE only 5% higher than baseline GNNs. This solution streamlines the drug discovery workflow, reduces operational complexity, and enables conversational AI capabilities for medicinal chemists, potentially accelerating drug development which currently takes 10-15 years and costs over $2 billion per drug.

Key takeaway

For medicinal chemists and biotech teams evaluating molecular properties, this advancement means you can consolidate multiple GNN-based prediction tasks into a single, interactive LLM. This not only simplifies your workflow and reduces infrastructure overhead but also enables conversational interaction for reasoning and molecular modification suggestions, accelerating early-stage drug design and increasing viable candidate throughput.

Key insights

Fine-tuned LLMs can unify complex molecular property prediction, outperforming or matching specialized GNNs.

Principles

• Foundational knowledge precedes performance optimization.
• Huber loss provides stable, effective RFT rewards.

Method

Customize a general-purpose LLM (Nova 2 Lite) using supervised fine tuning (SFT) on 55,000+ molecules, followed by reinforcement fine tuning (RFT) with Huber loss-based rewards.

In practice

• Use Nova Forge for LLM customization on SageMaker.
• Employ SFT for domain-specific knowledge acquisition.
• Apply RFT for predictive judgment and error minimization.

Topics

• Customized LLMs
• Drug Discovery
• Molecular Property Prediction
• Graph Neural Networks
• Supervised Fine Tuning

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Graph-Theoretic Models for the Prediction of Molecular Measurements — Enhanced classical graph-theoretic models can rival deep learning for molecular property prediction with significantly lower computational cost.
• SMDD-Bench: Can LLMs Solve Real-World Small Molecule Drug Design Tasks? — SMDD-Bench reveals current LLMs solve only 40.2% of complex drug design tasks, underscoring significant challenges in autonomous scientific discovery.
• Pretraining a foundation model for small-molecule natural products — NaFM is a foundation model for natural products, outperforming existing methods in drug discovery tasks.
• Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design — Probing molecular edit responses guides LLM agents to jointly optimize binding affinity and druggability in drug design.
• A physics-informed graph neural network to approximate docking-based binding affinity for DYRK2 in Alzheimer’s drug repurposing — Integrating physical interaction terms into GNNs can create efficient, interpretable surrogates for computationally intensive molecular docking.
• Knowledge Graphs and Reasoning LLMs for Finding Simple Yet Effective Transcriptomic Perturbation Predictors — Simple K-nearest neighbour models leveraging knowledge graphs, enhanced by RL-trained LLMs, offer competitive transcriptomic perturbation prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Amazon Science homepage.