This Week In Cheminformatics: Issue #010

Novel NP Inspired Chemotypes for Kinase, Virtual Screening with GPU using UniDock-Pro, Benchmarking Count Fingerprint Performance and a Long List of Papers from the Last Week of February.

Mar 01, 2026

Highlights

Natural Product-like Fragments Unlock Novel Chemotypes for a Kinase Target─Exploring Options beyond the Flatland

Santura et al. caught my eye because it brings "escape from flatland" idea within fragment-based (kinase) drug discovery. They screened an 87-membered natural product-like fragment library against protein kinase A and achieved a surprisingly high (41%) crystallographic hit rate. What I found most compelling from a cheminformatics perspective is that these hits yielded 32 novel (Bemis-Murcko) scaffolds. These scaffolds populate an underexplored, three-dimensional chemical space, which the authors demonstrated through significantly higher Fsp3, FC_Stereo, and nPBF metrics compared to established ChEMBL and PDB kinase ligands. It is a highly grounded, data-rich read that provides a clear structural rationale for integrating spatially complex, sp3-enriched building blocks into targeted screening libraries without sacrificing hit rates.

Count your bits: fingerprint benchmarking to assess broad chemical space representation

We often default to standard folded binary fingerprints without second-guessing the parameterization, but Huber and Pollmann’s recent benchmarking preprint strongly challenges this habit when dealing with large, heterogeneous chemical spaces. They empirically demonstrate that folding high-occupancy fingerprints specifically RDKit path-based and MAP4 into fixed vectors introduces severe bit-collision artifacts that systematically distort Tanimoto similarities and artificially inflate scores for larger molecules. I found their evaluation against graph-based MCES baselines particularly compelling; it shows that using count or log-count representations alongside unfolded sparse arrays consistently mitigates these collision penalties and improves structural specificity. To make implementing this practical rather than purely theoretical, they released chemap, an open-source Python library that standardizes the generation of these unfolded, folded, and frequency-scaled fingerprint variants. If you are building predictive ML pipelines or running similarity searches at scale, you should definitely review their quantitative breakdown on exactly when standard vector folding falls apart. Good read !

UniDock-Pro: A Unified GPU-Accelerated Platform for High-Throughput Structure-Based, Ligand-Based, and Synergistic Hybrid Virtual Screening

Boyang Ni and Douglas R. Houston introduce UniDock-Pro, which brings inter-ligand batch parallelism to a unified GPU framework for structure-based, ligand-based, and hybrid virtual screening. What stands out is they’ve chosen to use a continuously differentiable energy landscape optimally conditioned for their Monte Carlo and BFGS local optimization pipeline. This targeted adjustment yields ~2.5-fold boost in early enrichment on the DUDE-Z benchmark. Furthermore, their Hybrid mode merges receptor and ligand grid maps on-the-fly, supported by a newly introduced “Force Field Complementarity Analysis” that spatially quantifies where these two force fields cooperate or conflict during the conformational search. Code is here.