Today's Overview
- Docking-Augmented ML Yields First Sub-µM CDK16 Inhibitors Docked-pose augmentation converted limited CDK16 bio data into thousands of regression samples, enabling the first ML model to deliver sub-10 µM inhibitors.
- Hybrid quantum–machine learning workflows push virtual-screening accuracy while hardware limits remain No experimental potency or selectivity metrics are reported; validation is restricted to 300–500 ns MD simulations.
- AI-guided PROTAC design yields potent AR degraders active in CRPC xenografts AI-generated linker proposals within a MD-optimized AR/CRBN ternary complex produced degraders that eliminate both wild-type and five clinically relevant CRPC AR mutants.
Featured
01Docking-Augmented ML Yields First Sub-µM CDK16 Inhibitors
CDK16 is up-regulated in triple-negative breast, lung and prostate cancers but lacks selective chemical probes. By treating multiple docked poses as distinct training samples, the authors expanded a sparse pIC50 set into thousands of regression examples without new experiments. Gradient-boosted trees trained on pose-specific docking scores, ligand–receptor contact fingerprints and physicochemical descriptors predicted activity; GA feature selection delivered a validated QSAR whose top predictions were filtered through a consensus pharmacophore.
Virtual screening of 280 k NCI and 13 k OpnMe compounds followed by LanthaScreen kinase assays returned two confirmed hits: H_28 (BI-831266) IC50 = 3.5 µM and H_33 (BI-1282) IC50 = 5.8 µM—the first publicly disclosed CDK16 inhibitors below 10 µM. All validation is in vitro; selectivity against other CDKs and off-target kinases, cellular potency, and structural confirmation of the predicted binding modes remain to be reported. The approach assumes that docking-score noise is informative and that pose redundancy does not over-weight chemotypes.
02Hybrid quantum–machine learning workflows push virtual-screening accuracy while hardware limits remain
Drug-design campaigns need quantum-level electronic detail to rank weak-binding fragments, but brute-force DFT scales poorly for million-molecule libraries. The authors survey emerging hybrid quantum–machine learning (QML) pipelines that embed variational quantum circuits or quantum-inspired tensor networks inside classical ML surrogates, claiming they deliver ‘unprecedented accuracy’ in binding-affinity and transition-state predictions without giving numerical ΔG or enrichment factors. All reported screening exercises are validated only by 300 ns (triplicate) or 500 ns (single) MD simulations—an in-silico benchmark that still awaits head-to-head comparison with experimental potency or selectivity data. Key bottlenecks emphasized include the absence of error-corrected quantum hardware, the lack of transferable feature maps from quantum chemical descriptors, and the difficulty of merging heterogeneous public/private chemical datasets into a single QML model.
03AI-guided PROTAC design yields potent AR degraders active in CRPC xenografts
Persistent androgen-receptor (AR) signaling drives castration-resistant prostate cancer (CRPC), and classic antagonists fail against clinically observed L702H, T878A, H875Y, W742C, and F877L mutants. The authors sought degraders that simultaneously eliminate wild-type and mutant AR by recruiting the CRBN E3 ligase. Using AnHorn’s AIMCADD pipeline, they first ran MD simulations to identify the most stable AR-LBD/CRBN ternary complex, then applied the AIMLinker deep network to propose linkers and built a focused virtual library; subsequent docking and MD rescoring selected a handful of candidates for synthesis. Across LNCaP, 22Rv1, and VCaP cultures the leads induced AR degradation that was abolished by the proteasome inhibitor MG132, and DARTS confirmed direct AR binding. qRT-PCR showed downstream KLK3 (PSA) suppression, and cell-viability assays recorded selective killing of AR-positive tumor cells with minimal normal-cell cytotoxicity. Once-daily oral dosing in xenograft mice significantly reduced tumor volume and plasma PSA without observable systemic toxicity. All validation is confined to cell lines and mouse xenografts; pharmacokinetics, off-target liabilities, and activity in human-relevant mutation combinations remain unreported, and the abstract does not disclose chemical structures or quantitative potency values.
Also Worth Noting
An SE(3)-equivariant graph network that uncertainty-weights hybrid QM/MD data predicts binding with RMSE 0.979 kcal mol⁻¹ and R=0.931 on HiQBind-MISATO while running ~1 min per complex on one GPU. link (Chem)
AlphaFold3 structural triage of a 1.5×10¹¹-member phage library yielded cyclic peptides with KD 8.2×10⁻⁷ M and >3-fold selectivity for B7-H3 or DLL3, confirmed by SPR and cell binding. link
A systematic audit of 265 drugs and 1,462 cell lines showed that pre-CV feature screening inflates MSE by 16.6 % and that 72 % of 32 published methods (2017-2024) contain this leakage, indicating many claimed improvements are evaluation artifacts. link
Benchmarking seven pKa prediction tools on a curated 90,000-entry public set shows open-source ML models rival commercial accuracy for aqueous proton dissociation constants. link (Chem)
Integrative bioinformatics identified hub necroptosis genes in ulcerative colitis and in-silico screening of natural product libraries yielded candidate inhibitors prioritised by docking and ADMET scores. link
A computational workflow combining MD, pharmacophore mapping, and ML proposed novel STAT3 small-molecule binders for previously undruggable shallow sites. link
Agentic CertisAI Assistant embeds an ensemble model trained on CCLE/PDX mono- and combo-drug response data to let users upload SMILES/FASTQ and receive real-time, natural-language-ranked tumor model predictions for pre-clinical oncology studies. link
Re-analysis of scRNA-seq from 6 CML patients with an AI gene-prioritization pipeline distinguished treatment-free remission from early/late relapse by NK-cell RUNX3/EOMES versus FOSL2/MAF regulon activity and generated a compact transcriptional biomarker panel linked to IFN-γ and metabolic circuits. link
AI-guided mechanistic cell simulations trained on CCLE, DepMap, GDSC2 and LINCS reproduce DepMap dependencies with r=0.90 globally and achieve 70 % prospective validation hit rate while halving in-vivo validation timelines for in-silico-identified targets. link
Integrated AI platforms de-novo generated <150-aa miniproteins that bind the p53-R175H peptide-MHC-I complex and are being validated as CAR-T/NK engagers to kill tumor cells expressing this mutation. link
An ImageNet-pretrained CNN applied to 3D holotomography MIPs classifies five HeLa RCD states at 99.3% accuracy and detects necroptosis 2–4 h before Annexin V/PI, with 3D models outperforming 2D (76–88% vs 50–55%) and cross-line portability restored by small-data fine-tuning. link
MetaGIN embeds molecules with a parameter-efficient graph-isomorphism network and meta-learned initial features, yielding competitive or state-of-the-art results on six molecular property benchmarks while using ≤15% of the parameters of larger GNNs. link
Today's Observation
Today’s trio shows how far AIDD has moved beyond vanilla docking or ML: each paper layers a physics-based step (docked poses, quantum-level energies, or MD-refined ternary complexes) onto a machine-learning layer to squeeze value from sparse data. CDK16 finally gets its first sub-µM chemical matter because pose-augmented gradient-boosting turned 56 measured %inhib values into 3 000 regression points, yielding two validated enzymatic hits at 3.5 µM and 5.8 µM. Likewise, PROTAC design used an MD-optimized AR/CRBN complex to train the linker generator, producing orally active degraders that collapse both wild-type and five clinically relevant AR mutants in CRPC xenografts while shrinking tumors and serum PSA without overt toxicity.
The same theme—physics-aware feature engineering—also exposes current ceilings. CDK16 hits are still only biochemical IC50s with no selectivity, cell data, or structural confirmation, and the quantum–ML hybrid reaches merely 300–500 ns MD checks with no experimental potency, illustrating that hardware errors and encoding bottlenecks keep it in the virtual-screening realm. For practitioners, the clear message is to treat physics-augmented ML as a data multiplier, not a magic bullet: pair it with immediate enzymatic or cellular triage, plan follow-up SAR cycles early, and budget for off-target panels and rodent PK before claiming target-ready leads.
The above is personal commentary for reference only. Refer to the original papers for authoritative content.