First selective CDK16 inhibitors via docking-ML regression

Today's Overview

Docking-augmented ML regression yields first selective CDK16 inhibitors Multiple docked poses of each ligand expanded a 38-compound CDK16 data set into a feature-rich set that enabled reliable regression modelling.

Featured

01Docking-augmented ML regression yields first selective CDK16 inhibitors

CDK16 promotes growth of triple-negative breast, lung and prostate tumours, yet no potent, selective tool compounds exist to validate it as a drug target. Starting from only 38 public actives, the authors address data scarcity by augmenting the training set with multiple docked poses of each ligand, extracting pose-specific scores, contact fingerprints and physicochemical descriptors for regression modelling.

Gradient-boosted trees trained on this augmented set outperformed seven other algorithms; genetic feature selection further improved generalisability. The resulting model built a pharmacophore that, when screened against 300 k NCI/OpnMe compounds, prioritised two confirmed hits: BI-831266 (IC50 3.5 µM) and BI-1282 (IC50 5.8 µM) in LanthaScreen kinase assays—potencies not previously reported for CDK16. All validation is in vitro; selectivity versus other CDKs and cellular activity remain to be tested, and the initial 38-compound training set limits the chemical space sampled.

Multiple docked poses of each ligand expanded a 38-compound CDK16 data set into a feature-rich set that enabled reliable regression modelling.Virtual screening driven by the docking-augmented ML model produced two micromolar CDK16 inhibitors with IC50 values of 3.5 µM and 5.8 µM.The approach is limited to in vitro confirmation; selectivity profiling and cell-based validation of the hits have not yet been reported.

Source: Computer-aided discovery of CDK16 inhibitors: a docking-augmented machine learning regression modelling approach.

Also Worth Noting

DUW-MGCA: QM/MD coattention for protein-ligand affinityDocking & Binding

An SE(3)-equivariant graph network that uses heteroscedastic uncertainty weighting to fuse QM and MD trajectories attains RMSE 0.979 kcal mol⁻¹ and R 0.931 on HiQBind-MISATO while keeping GPU inference ≈1 min per complex. link (Chem)

Today's Observation

Expanding a sparse 38-molecule SAR into a regression-ready descriptor matrix via multi-pose docking let a simple ML model rank million-compound libraries for CDK16. The tactic—treating each docked pose as an independent “pseudo-experiment”—delivered two low-micromolar inhibitors (3.5 µM and 5.8 µM) without new biochemical data, showing that pose redundancy can substitute for library size when targets have uncertain binding modes.

Practitioners should note the approach is still in vitro only; selectivity and cell data are missing, and success hinged on a crystal structure accurate enough to generate meaningful poses. The same trick may falter for floppy pockets or when docking noise outweighs true signal, so reserve it for kinases (or other rigid sites) where pose ensembles are chemically plausible.

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