Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization

2026-07-14 · Source: Machine Learning · Field: Science & Research — Health & Medical Research, Life Sciences & Biology, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

conDitar-dev is a novel conditional diffusion-based framework for structure-based drug design (SBDD) that generates molecules with high binding affinities and improved ADMET properties. It comprises three modules: msPRL for multi-scale pocket representation learning, conDitar as a pocket-conditioned diffusion model, and paOPT for optimizing ligand developability during generation. This framework addresses limitations in existing diffusion SBDD methods by decoupling pocket and ligand learning and considering developability beyond just binding affinity. On a new human disease targets benchmark, conDitar achieved an average binding score of -8.85 kcal/mol, outperforming other SBDD baselines. Furthermore, conDitar-dev improved ADMET property performance by up to 73% compared to conDitar alone. Experimental validation on PD-L1 yielded molecules with SPR-derived K_D values of 3.49 and 3.75 μM, while CSF1R inhibitors showed IC_50 values as low as 200 nM, also revealing drug repositioning potential.

Key takeaway

For research scientists focused on de novo drug design, conDitar-dev offers a robust approach to generate molecules with both strong binding affinity and favorable ADMET profiles. You should consider integrating multi-scale pocket representations and property-aware optimization into your SBDD workflows. This method can significantly improve the developability of your drug candidates, potentially reducing downstream experimental validation costs and accelerating drug discovery timelines.

Key insights

conDitar-dev integrates multi-scale pocket learning, conditional diffusion, and property optimization to generate developable drug candidates with high binding affinity.

Principles

• Decouple pocket and ligand representation learning.
• Optimize for ADMET properties alongside binding affinity.
• Multi-scale pocket representations enhance design.

Method

The conDitar-dev framework employs msPRL for pocket representations, conDitar for pocket-conditioned diffusion, and paOPT for generation-time ADMET property optimization.

In practice

• Design ligands for PD-L1 with K_D values of 3.49 and 3.75 μM.
• Identify CSF1R inhibitors with IC_50 values as low as 200 nM.
• Uncover drug repositioning opportunities.

Topics

• Drug Discovery
• De Novo Drug Design
• Structure-Based Drug Design
• Diffusion Models
• ADMET Properties
• Ligand Generation

Best for: AI Scientist, Research Scientist

Related on AIssential

• Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design — ConGLUDe unifies structure- and ligand-based drug design via contrastive geometric learning, enabling diverse drug discovery tasks.
• DrugGen 2: A disease-aware language model for enhancing drug discovery — DrugGen-2 is a disease-aware generative AI model that significantly improves de novo drug design by integrating disease ontology.
• 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.
• Sample-efficient generative molecular design using memory manipulation — A novel framework named Saturn, which applies the Mamba architecture to generative molecular design, significantly improves sample efficiency for drug discovery.
• MolHIT: Advancing Molecular-Graph Generation with Hierarchical Discrete Diffusion Models — MolHIT uses hierarchical discrete diffusion and decoupled atom encoding to generate valid molecular graphs.
• How a Google DeepMind Spinoff Hunts Hidden Drug Targets — IsoDDE enables discovery of novel, cryptic protein binding pockets, expanding druggable targets beyond AlphaFold's capabilities.

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 Machine Learning.