1. Integrated AI-Driven Multi-Objective Optimization of Sulfamoylbenzamide-Based HBV Capsid Assembly Modulators: Discovery of 11B-6 with Superior Potency and Drug-like Properties1
Combining fragment growth, activity prediction and predicted drug-like properties produced 48 hepatitis B capsid-modulator analogues. Compound 11B-6 showed stronger antiviral activity and lower plasma protein binding than NVR 3-778 in the reported comparisons. These measured changes make the optimisation useful to medicinal chemists; our access was limited to the publisher abstract, which does not establish clinical efficacy.1
2. Contrastive Learning for Metabolite-Aware Oral Drug Design2
Mettle uses contrastive learning to distinguish metabolic transformations from similar decoys. The authors applied its predictions to redesign RSK4 inhibitors, reporting improved oral bioavailability while retaining potency and antitumour activity in preclinical models. The study connects metabolism prediction with compound optimisation. This abstract-only report cannot independently assess the detailed experimental protocols or establish benefit in patients.2
3. Machine-learning force-field scoring rivals free-energy perturbation for congeneric ligand ranking across public benchmarks3
The unreviewed taqo preprint benchmarks machine-learned force-field scores for ranking related ligands. Across matched public systems, performance was statistically indistinguishable from OpenFE but below FEP+ overall; the congeneric subset gave a different comparison. This suggests a useful prioritisation option without equating static scores with measured binding. Reporting is limited to the repository-deposited abstract.3
4. Assay-Level Modeling and External Validation of Antibody Developability Using Physicochemical Descriptors and Sequence Embeddings4
An assay-specific antibody study compared physicochemical descriptors with pretrained sequence embeddings and tested transfer to a separate dataset. External correlations were weak across several endpoints, and the hybrid model did not consistently improve on descriptor baselines. The result supports checking developability predictions assay by assay. Our access was limited to the publisher abstract; prospective prioritisation remains untested there.4
5. Per-Stage Controls and Failure Modes in a Co-Folding Virtual Screening Cascade for the Keap1–Nrf2 Interaction5
An unreviewed Keap1–Nrf2 screening study checked what individual model scores actually convey. Two affinity models ranked activity-cliff pairs above chance but missed potency-gap size and failed to resolve enantiomers; pose confidence only weakly tracked geometric validity. The controls offer practical checks before candidate selection. This report uses the repository-deposited abstract and makes no claim of experimentally validated inhibitors.5
References
Shuo Wang; Miaochen Xu; Feiyue Ma; Dazhou Shi; Leda C. Bassit; Mohammad Salman; Harout Ajoyan; Delgerbat Boldbaatar; Tamara McBrayer; Xiaoyu Shi; Shuo Wu; Xinyong Liu; Shujing Xu; Thomas Tu; Raymond F. Schinazi; Yibei Xiao; Peng Zhan. Integrated AI-Driven Multi-Objective Optimization of Sulfamoylbenzamide-Based HBV Capsid Assembly Modulators: Discovery of 11B-6 with Superior Potency and Drug-like Properties. Journal of Medicinal Chemistry; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jmedchem.6c01701. Accessed 2026-09-20T16:15:23.324Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jmedchem.6c01701
Abstract Current capsid assembly modulators (CAMs) require improvements in antiviral potency and drug-like properties. Herein, we report a multi-objective optimization workflow for the NVR 3-778 by integrating structure-based fragment growing with systematic medicinal chemistry optimization, activity prediction, and multitask prediction of drug-like properties. This led to 48 novel analogs. Among them, 11B-6 showed superior antiviral activity in HepAD38 cells (EC50 = 0.04 ± 0.02 μM), outperforming NVR 3-778 (EC50 = 0.50 ± 0.17 μM). In HBV-infected HepG2-NTCP cells, 11B-6 inhibited secreted HBV DNA with an EC50 of 3.71 nM, about 134-fold more potent than NVR 3-778 (EC50 = 497.9 nM). Structural biology analyses revealed a distinct, more stable binding mode of 11B-6 at the capsid dimer−dimer interface. 11B-6 also showed lower plasma protein binding than NVR 3-778 (84.3% vs. 100%). These results highlight 11B-6 as a promising next-generation HBV CAM, representing a valuable starting point for advancing its development.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Huan He; Manzhan Zhang; Xiaoxiao Yang; Shuai He; Xiayu Shi; Feng Hu; Chang Liu; Xingsen Zhang; Na Chen; Xiaoqian Zhu; Leihao Zhang; Tianyu Ye; Rong Zhang; Yanru Yang; Rui Wang; Zhenjiang Zhao; Zhuo Chen; Xuhong Qian; Honglin Li; Zhe Wang; Kai Zhang; Kangdong Liu; Shiliang Li. Contrastive Learning for Metabolite-Aware Oral Drug Design. Journal of Medicinal Chemistry; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jmedchem.6c02126. Accessed 2026-09-20T16:15:23.325Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jmedchem.6c02126
Abstract Unfavorable drug metabolism drives clinical failure, limited by the low accuracy of existing AI predictors. To address this, we constructed a database of 11,665 human-specific reactions and developed Mettle, a novel AI model integrating chemical feature interaction with contrastive learning. By explicitly training the model to distinguish true metabolic transformations from structurally similar decoys, Mettle achieves state-of-the-art performance with ∼80% top-5 accuracy. We demonstrate Mettle’s utility by tackling poor oral bioavailability in RSK4 inhibitors. This metabolite-aware design strategy yielded R636, which maintains high potency while exhibiting a remarkable 63-fold increase in absolute bioavailability (to 63%). R636 showed a favorable safety profile and significant antitumor efficacy in two ESCC PDX models. Mettle thus emerges as a powerful tool for metabolite-aware oral drug design.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Kevin Ryczko; Sarah Maier; Amogh Sood; Patrick Rowe; Harish Ramadas; Lorenzo Boninsegna; Justin Overhulse; Hunter La Force; Mikayla Darrows; Shiji Zhao; Andrew Wildman; Mary Pitman; Romelia Salomon Ferrer; Andrea Bortolato. Machine-learning force-field scoring rivals free-energy perturbation for congeneric ligand ranking across public benchmarks. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008810/v1. Accessed 2026-09-20T16:15:23.325Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008810/v1
Relative binding free-energy (RBFE) methods such as free-energy perturbation (FEP) set the accuracy standard for ranking congeneric ligands in structurebased drug design, but their computational cost limits throughput. Machine-learned force fields (MLFFs) now approach quantum-chemical accuracy at a small fraction of that cost, raising the question of whether they can recover much of RBFE’s ranking accuracy while remaining computationally efficient. We introduce a static-score methodology called taqo and benchmark 10 MLFF variants spanning 6 model families (UMA, MACE, Orb, eSEN, AIMNet2, AllScAIP), together with a classical molecular-mechanics force field, on a uniform staticpocket protocol, correlated against experiment across 22 benchmark systems.We find that the ten MLFFs cannot be resolved from one another at this sample size, while every one of them significantly outranks the classical force field reference, at a few seconds per ligand on a single GPU. Against published OpenFE and FEP+ results on the same systems and the same ligand sets, taqo is statistically indistinguishable from OpenFE and significantly below FEP+, at roughly three orders of magnitude lower cost. On the congeneric JACS subset taqo and FEP+ are statistically indistinguishable. On the same benchmark poses taqo outperforms classical docking scores and is statistically indistinguishable from deep-learning affinity models, without any affinity-specific training. Together these establish MLFF interaction-energy scoring as a practical high-throughput approach for lead optimization.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Shasha Jiang; Mark Akinola Ige; Hei Wun Kan; Xiaochun Wan; Junxin Li; John Z. H. Zhang; Zhiyuan Zheng; Haiping Zhang. Assay-Level Modeling and External Validation of Antibody Developability Using Physicochemical Descriptors and Sequence Embeddings. Journal of Chemical Information and Modeling; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jcim.6c02385. Accessed 2026-09-20T16:15:23.325Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jcim.6c02385
Abstract Antibody developability is inherently multidimensional, involving thermal stability, aggregation and hydrophobicity, nonspecific interactions, and expression-related manufacturability. Instead of treating developability as a single composite label, we investigated assay-level modeling using GDPa1 as a benchmark data set. We developed a two-branch framework combining interpretable developability descriptors and data-driven calibration. One branch fits assay-specific descriptor models to derive end point-specific physicochemical weights, while the other integrates frozen pretrained VH/VL sequence embeddings with computed descriptors in a multitask regression architecture. Internal validation across eight GDPa1 assays showed assay-dependent predictability, with hydrophobicity- and aggregation-related readouts generally showing stronger signals than expression-related and some thermal stability end points. Descriptor coefficient analysis revealed distinct physicochemical patterns across assays, supporting the view that antibody developability cannot be represented by a single universal descriptor weighting scheme. External held-out validation on GDPa3 revealed limited cross-data set transferability. Across seven directly matched assays, mean Spearman correlations were 0.126 for the full descriptor model, 0.139 for the sequence-only descriptor model, and 0.125 for the hybrid sequence–descriptor model. HIC retention time was the most transferable end point, whereas SEC monomer percentage and several thermal stability end points showed weak external correlations. The hybrid model did not consistently outperform descriptor-only baselines, indicating that pretrained sequence embeddings provided limited and assay-dependent incremental value in this small-data setting. GDPa1 sequence-cluster analysis showed that internal validation estimates depended on clustering threshold and fold construction, whereas GDPa3 provided a more direct external assessment of cross-data set transferability. These results support a cautious assay-level framework for antibody developability modeling. Descriptor-based representations remain strong and interpretable baselines, while sequence embeddings require end point-specific and external validation. The proposed workflow can convert assay-level predictions into domain-level profiles for computational triage, but prospective candidate prioritization requires binding assessment and experimental validation.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Huynh-Duc Le. Per-Stage Controls and Failure Modes in a Co-Folding Virtual Screening Cascade for the Keap1–Nrf2 Interaction. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008984/v1. Accessed 2026-09-20T16:15:23.325Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008984/v1
Virtual-screening cascades assembled from co-folding and learned-affinity models are now routine, and each stage is usually trusted on the strength of its own score. We measured, stage by stage, whether those scores carry the information a practitioner would attribute to them, using the Keap1–Nrf2 protein–protein interaction and a 130,793-compound make-on-demand library as the test bed. Most measurements are negative, and none of the failure modes is specific to one of the two models tested. Two independently developed affinity models order activity-cliff pairs above chance (about 68% of 181 pairs) but carry no information about the size of the potency gap, and both couple to molecular weight; the built-in molecular-weight correction rescales without de-biasing. Neither model resolves enantiomers, including the model that generates an explicit high-confidence three-dimensional pose. Pose-model confidence is only weakly informative about geometric validity, so an explicit validity check is required rather than optional. Training-free guidance is required for stereochemical fidelity in pose generation, and diffusion steps can be halved without detectable loss. A seeded matched-decoy gate validates the pose-and-energy stage where no single physicochemical descriptor does, and the pose model reproduces all five co-crystal references below 2 Å, three of them released after its training cutoff. Single-structure MM-GBSA ordering is not validated by ensemble refinement, and the two solvation models disagree both on the ordering and on which candidates are separable at all. The control suite is released as a reusable artifact.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Publication record
Published 2026-09-20.
Sources, selection and claims were checked in an independent AI editorial review, followed by the AI editor's approval. This is not academic peer review.
- 2026-09-20 · Version 5d0dfc38 · Viewing this version