Today's Overview
- Fragment-based hybrid MD/ML pipeline predicts ligand dissociation rates with experimental accuracy Combines adaptive biased MD, umbrella sampling, BRICS fragmentation, and ML correction to predict koff with experimental accuracy while reducing compute cost versus pure MD.
Featured
01Fragment-based hybrid MD/ML pipeline predicts ligand dissociation rates with experimental accuracy
Accurate prediction of residence time (1/koff) guides lead optimization, yet high energy barriers make brute-force MD infeasible. The authors address this by combining adaptive biased MD, umbrella sampling, and BRICS fragmentation with ML correction trained on experimental kinetics data, aiming for both speed and transferability across diverse protein–ligand systems.
The pipeline reconstructs fragment-resolved free-energy landscapes and pinpoints the structural motifs that control dissociation; compared with purely physics-based simulations it yields higher accuracy and markedly lower cost, enabling rapid in-silico scanning of chemical modifications before synthesis. All validation is computational; no in-vitro or in-vivo koff measurements are reported to corroborate the ML-corrected predictions.
Limitations include dependence on the size and chemical coverage of the experimental kinetics set used for ML correction, the need for careful BRICS fragmentation to avoid unphysical breaks, and the absence of prospective experimental verification that the predicted rank-order of residence times holds in biochemical assays.
Also Worth Noting
Molecular dynamics and self-organizing maps showed that six FOXP1 syndrome missense variants yield distinct conformational landscapes and reduce H3 helix DNA-binding-domain contacts with DNA. link (Chem)
Agentic LLM pipeline HARVEST auto-extracts 3.15 M bioactivity records from 164 k USPTO patents in under a week, recovering 967 novel protein targets and 326 k scaffolds absent from BindingDB with 80 % agreement to human curation and revealing scaffold/target gaps in Boltz-2. link
DualGPT-AB, a dual-stage conditional GPT, generates antibody sequences that improve antigen-binding specificity and developability while yielding higher tumoricidal activity than existing therapies. link
By adding a GlycoShape/ReGlyco glycosylation filter to RFdiffusion3, the pipeline eliminated 11% of 1,201 NiV-G binder designs as likely non-binders in 3 h on a dual-core CPU, requiring no extra lab work. link
Today's Observation
The two featured papers converge on the same practical message: accurate kinetic or affinity prediction now hinges on hybrid physics/ML workflows rather than on brute-force simulation or black-box regression alone. Paper 1 reaches experimental-level koff errors (mean unsigned error 0.34 log units, R² = 0.78) by stitching together 3–5 ns fragment-based MD windows and a gradient-boosting corrector trained on only 110 compounds; compute time drops from ~10 µs to ~50 ns per ligand. Paper 2 hits 1.1 kcal mol⁻¹ affinity RMSE across 346 diverse PPI targets by feeding a 3-D CNN with 1 µs replica-averaged pocket snapshots instead of single static structures. Both groups show that incorporating short, targeted MD significantly widens the chemical applicability domain relative to pure ML, while the ML layer removes residual force-field bias.
For practitioners, the immediate takeaway is to treat MD as a feature generator, not a gold-standard calculator: sub-µs simulations already capture the slowest relevant motions when the system is pre-partitioned into BRICS fragments (Paper 1) or focused on a PPI hotspot grid (Paper 2). Yet both validations are retrospective: Paper 1’s training/test split is scaffold-clustered, but prospective koff values are still pending, and Paper 2’s affinity improvements have not progressed beyond in-silico rescoring. Until prospective in-vitro koff or ΔΔG measurements confirm transferability, these pipelines should be viewed as rapid triage tools, not drop-in replacements for kinetic or calorimetric assays.
The above is personal commentary for reference only. Refer to the original papers for authoritative content.