RL-guided diffusion docks unseen pockets at 11.5%; XAI spots key ligand atoms

Today's Overview

RL-guided diffusion generates semi-flexible ligands that dock to unseen pockets with 11.5% success RL steering of a diffusion denoising process lets the network sample ligand flexibility while remaining in drug-like regions.
XAI pinpoints ligand atoms that touch the kinase binding pocket Consensus XAI attributions recover up to 76 % of ligand atoms within 2 Å of the kinase pocket without 3D input.

Featured

01RL-guided diffusion generates semi-flexible ligands that dock to unseen pockets with 11.5% success

Structure-based design usually treats ligands as static objects, yet real binding involves conformational adaptation. The authors reformulate pocket-conditioned diffusion as a Markov decision process and let a reinforcement-learning agent iteratively explore torsion and atomic positions while staying in drug-like chemical space. After self-supervised pre-training on 1.3 M target-free and 0.3 M target-specific molecules, the model yields poses with a mean Vina score of –7.23 kcal mol⁻¹ on the test set and succeeds in 11.53% of cases, outperforming the previous best generator (8.9%). A fast sampler reduces generation time 20-fold. All validation is in silico; no experimental binding or selectivity data are reported, and performance on very shallow or highly charged pockets was not examined.

RL steering of a diffusion denoising process lets the network sample ligand flexibility while remaining in drug-like regions.The pipeline reaches –7.23 kcal mol⁻¹ Vina score and 11.53% success rate on unseen pockets, beating the prior top method.A 20× faster sampler is provided, but all metrics derive from docking simulations without in-vitro or in-vivo confirmation.

Source: Steering semi-flexible molecular diffusion model for structure-based drug design with reinforcement learning.

02XAI pinpoints ligand atoms that touch the kinase binding pocket

Graph neural networks predict drug–target interactions but remain chemically opaque, slowing adoption in lead optimisation. The authors ask whether post-hoc explainability can flag the specific ligand atoms that physically contact a protein pocket, focusing on two clinically relevant families: kinases and GPCRs.

Four attribution techniques were applied to trained GNNs; although inter-method consistency was modest, a consensus attribution recovered 76 % of atoms lying within 2 Å of the kinase binding site and frequently contacted the regulatory DFG motif. Validation was purely in silico—mapping to 3D crystal structures—so performance may degrade for flexible binding sites or novel chemotypes not represented in the training set.

Consensus XAI attributions recover up to 76 % of ligand atoms within 2 Å of the kinase pocket without 3D input.Attributed atoms often contact functionally important residues such as the DFG motif, confirming biological relevance.All validations are in silico; accuracy on flexible or unseen binding sites remains untested.

Source: Explainability Methods from Machine Learning Detect Important Drugs' Atoms in Drug-Target Interactions.

Also Worth Noting

BOAT: Multi-objective Bayesian optimization for antibody lead optimizationAntibody & Biologics

BOAT uses uncertainty-aware surrogate modeling coupled with a genetic algorithm to jointly optimize predicted antibody properties, outperforming both genetic and newer generative baselines in systematic benchmarks while revealing practical limits tied to sequence dimensionality and oracle cost. link

DualGPT-AB: Dual-stage GPT for multi-property CDRH3 optimizationAntibody & Biologics

DualGPT-AB, a conditional GPT plus RL framework that generates CDRH3 sequences meeting multiple design constraints, yielded 8 of 100 candidates with strong HER2 binding and superior tumoricidal activity to Herceptin in vitro. link

PRiMeFlowOmics & Biomarkers

PRiMeFlow uses flow matching in gene-expression space to predict single-cell perturbation responses and won the Generalist Prize in the ARC Virtual Cell Challenge. link

Today's Observation

Today’s pair shows how reinforcement learning and explainability are being layered onto generative docking models to push beyond static-pocket limits. Paper 1 frames the core chemistry problem as generating conformer-ensembles that remain drug-like while adapting to pockets never seen during training. Their RL-guided diffusion sampler achieves a –7.23 kcal mol⁻¹ average Vina score and an 11.5% top-ranked success rate on the flexible PDBBind test set, outperforming the previous best method and offering a 20× speed-up. Practitioners should note, however, that all metrics are in-silico docking simulations; no biochemical or cellular read-outs are provided, so synthetic follow-up will be needed to confirm affinity.

Paper 2 tackles the downstream question of which generated atoms actually matter. Using only 2D SMILES, consensus XAI attributions recover 76% of ligand atoms that end up within 2 Å of the kinase binding site and frequently highlight atoms contacting the DFG motif, suggesting the explanations capture functionally relevant interactions. Yet the analysis is confined to kinases with relatively rigid ATP sites; accuracy on flexible or truly novel pockets remains unproven. Together, the papers outline a plausible in-silico pipeline—RL diffusion for conformer sampling followed by attribution-guided medicinal-chemistry tweaks—but both steps still rest on docking scores and retrospective kinase data.

The above is personal commentary for reference only. Refer to the original papers for authoritative content.