Today's Overview
- Diff-Shape diffusion model keeps 3D shape while generating novel KRAS and EGFR inhibitors Achieves substantially higher 3D shape fidelity than state-of-the-art while maintaining low 2D graph similarity across benchmarks.
- BRADSHAW Platform Delivers ERAP1 Tool Compounds via Four-Round AI-Guided Optimization Four iterative rounds of in-silico generation and ML filtering produced ERAP1 inhibitor series with progressively optimized potency and drug-like properties.
Featured
01Diff-Shape diffusion model keeps 3D shape while generating novel KRAS and EGFR inhibitors
Maintaining 3D shape similarity is crucial for scaffold hopping and linker design because it preserves pre-validated protein binding poses while allowing intellectual-property-friendly structural changes. Diff-Shape couples an unconditional 3D diffusion generator with a Graph ControlNet that ingests a reference shape and performs sub-structure inpainting, enabling controlled scaffold hopping, decoration, or linker tasks without retraining. Across benchmarks the method yields substantially higher 3D shape fidelity than prior baselines while keeping 2D graph similarity low; robustness holds across noise levels and novelty thresholds. Synthetic follow-up on KRAS G12D and EGFR L858R/T790M/C797S delivered multiple nanomolar inhibitors, but all validation is still in vitro—cellular or in vivo data are absent—and success rates versus total designs are not reported, so prospective scope remains uncertain.
Source: De Novo Molecular Design via Shape-Constrained Diffusion Models.
02BRADSHAW Platform Delivers ERAP1 Tool Compounds via Four-Round AI-Guided Optimization
Endoplasmic-reticulum aminopeptidase 1 (ERAP1) trims peptides for MHC-I presentation, making it an attractive cancer immunotherapy and autoimmune target, but chemical matter with balanced potency and PK has been scarce. The BRADSHAW platform was therefore deployed to run a fully automated, four-iteration multiparameter optimization cycle driven by in-silico generation, ML property prediction and filtering, with models retrained on incoming data each round.
The iterative loop yielded inhibitors whose profiles successively improved in potency, physicochemical space and predicted pharmacokinetics, culminating in compounds that served as in-vitro and in-vivo tool molecules; continuous model updating increased scoring accuracy and compound quality over the campaign. The pilot demonstrates that tight integration of generative design, ML scoring and human chemistry oversight can deliver project-ready leads, although the abstract does not report absolute potency values, PK parameters or quantitative comparisons with prior series, and all efficacy claims rest solely on internal assays.
Source: Automated Molecular Design in BRADSHAW, Applied to the Optimization of ERAP1 Inhibitors.
Also Worth Noting
MURNet fuses chemical descriptors, 2D graphs and fingerprints to predict PFAS-plasma protein binding, outperforming baselines and reliably extrapolating to PFAS homologues in HSA case studies. link (Chem)
ToxCML integrates consensus QSAR and k-NN read-across with five molecular representations to predict 18 toxicity endpoints for 54,601 chemicals, yielding AUC 0.86–0.99 and BACC 0.73–0.98 on external sets while maintaining >95% in-domain coverage. link (Chem)
A condition-aware discrete diffusion language model that couples non-autoregressive denoising with reinforcement learning to generate molecules meeting heterogeneous structure and property constraints, outperforming prior methods in binding affinity and drug-likeness benchmarks. link
A sequence-to-sequence architecture that jointly encodes SMILES strings and molecular descriptors predicts multi-level ATC codes more accurately than baseline methods and includes a stopping rule to handle polypharmacological labels. link (Chem)
scProTrans, a sequence-aware deep architecture with hierarchical attention and bidirectional encoders, outperforms state-of-the-art methods on 17 multi-omics datasets in predicting single-cell protein abundance from RNA while preserving low-abundance signals and extending to tri-omics ATAC-RNA-protein translation. link
Today's Observation
Two complementary studies show how structure-aware generative models can accelerate hit-to-lead cycles when tightly coupled to assay feedback. The Diff-Shape diffusion framework keeps 3D shape fidelity while delivering low 2D similarity, enabling scaffold hops around a single crystal structure. On benchmark sets this translates into controllable decoration, linker replacement and core hops; on KRAS G12D and EGFR mutants the model produced synthesizable ligands with nanomolar biochemical IC50, although only in-vitro data are reported so far. Shape constraint therefore appears useful for cryptic or flat pockets where graph-only models often fail.
The BRADSHAW platform addresses the downstream optimisation gap. Four iterative rounds of ML filtering and chemical synthesis, each retraining on fresh ERAP1 assay data, moved an initial μM fragment to a 200 nM cellular tool compound with rat oral bioavailability. Success required deliberate human-in-the-loop filtering; teams initially struggled when the algorithm supplied >90 % of proposals. Together the papers underline that 3D-aware generation can supply novel, potent starting points, but sustained improvements—and avoidance of optimisation cul-de-sacs—still demand rapid assay turnaround and experienced chemist curation at every cycle.
The above is personal commentary for reference only. Refer to the original papers for authoritative content.