This Week In Cheminformatics: Issue #028

Here, Huber & Pollmann benchmark molecular fingerprints on virtual screening retrieval across specificity, compound-size dependence, ranking agreement, correlation with MCES scores, and chemical class prediction on large, heterogeneous natural product/metabolomics datasets. The key findings are count fingerprints are almost always better than binary for this task, folding-induced bit collisions are far worse than commonly assumed for high-occupancy fingerprints like RDKit and MAP4 (folded MAP4 showed essentially zero correlation with rascalMCES on their data, while unfolded MAP4 hit ~0.6), and larger Morgan/FCFP radii (r=9) consistently performed well across tasks. They also introduce a “frequency-folding” alternative that keeps a fixed vector size while avoiding collisions by selecting the most occupied bits from an unfolded representation. This works well for Morgan/FCFP but not for RDKit, which they attribute to information loss given RDKit’s very high bit occupancy. The code is here. Good read!

Chen et al. evaluate how different curve-fitting functions and drug concentration windows impact cross-database reproducibility of large pharmacogenomics screens like GDSC, CTRP, etc. Interestingly, the authors demonstrate that calculating the AUC derived from a sigmoidal fit over a strictly filtered / truncated data with shared dose range (0.03 µM to 10.00 µM) outperforms full-range AUC, IC50, and EC50 metrics in cross-platform agreement. This “TruncAUC” protocol maximizes pairwise reproducibility and improves the identification of sensitive cell lines. Code can be found here.

van Hoorn et al. propose a method for automatically generating initial transition state (TS) guesses directly from reactant and product SMILES. They do this by constructing and interpolating between reactant and product force fields to approximate the potential energy surface along the reaction coordinate. The method produces reliable starting geometries for single-ended TS optimizations at a fraction of the usual cost. Benchmarks across 121 diverse reactions demonstrated convergence for 115 transition states, requiring an average of just 107 gradient evaluations per reaction. It is implemented natively in VeloxChem.

• FragScan: A Quantitative Fragment Scanning Strategy for Rational Drug Discovery

• GeoPep: A Geometry-Aware Masked Language Model for Protein-Peptide Binding Site Prediction

• Tree-Guided Graph Neural Networks with Multilevel Optimization for Protein–Protein Interaction Prediction

• tmGNN-XAI: An Explainable Graph Neural Network Tool for Predicting Electronic Properties of Transition Metal Complexes from SMILES

• Derisking Affinity Optimization for Macrocycles and Cyclic Peptides: High-Precision Free Energy Simulations across Five Diverse Targets

• When Chemistry is Too Colourful: Gamut Clipping in 8-bit sRGB Risks Misinterpretation of Camera-Based Chemical Analysis

• Cluster-based virtual reaction generation with reaction site-centered buffer zone

• An End-User Audit of Reproducibility, Data Leakage, and Overfitting of the Top-Ranked ADMET Prediction Models in TDC Leaderboards

• PFASGroups: An Open-Source Framework for Automated Identification, Structural Classification, and Prioritization of Per- and Polyfluoroalkyl Substances

• SPPIDER-seq: sequence-based partner-aware predictor of protein-protein interaction sites

• A survey of models composed of graph neural networks and large language models for molecular science

• MetagenomicKG: a knowledge graph for metagenomic applications

• Computational tool choice impacts CRISPR spacer-protospacer detection

• QUAD: a composite risk framework integrating uncertainty, applicability domain, and model disagreement for reliable QSAR predictions

• Integrating evolutionary and compositional features with ML and DL for robust and interpretable druggable protein prediction

• RGTBind: RBF-gate graph transformer with spatially biased attention for protein–DNA binding-site prediction

• DeepKbhb: Context-Aware Prediction of Human Lysine β-Hydroxybutyrylation Sites

• Unified heterogeneity-aware benchmark of drug synergy prediction: a cross-study analysis of traditional machine learning and graph deep learning models

• HyperDC: A Non-Uniform Hypergraph Framework for Dual- and Higher-Order Drug Combination Recommendation Across Diverse Complex Diseases

• Molecular Mechanisms of KIT Receptor Dimerization and Oncogenic Activation Revealed by Multiscale Simulations

• Simulation Guided Design of a Potentially Hyperactive Ice Nucleating Protein

• Structural and Thermodynamic Discrimination between Agonists and Antagonists of Retinoic Acid Receptor γ and the Vitamin D Receptor

• MolPy: A Large Language Model-Friendly Toolkit for Reactive Topology Editing in Polymer Simulations

• Setting the Bases of the Photogenotoxicity of p-Aminobenzoic Acid

• SMASH: Screening Molecules Accurately on Small Hardware. Fast, user-friendly, enhanced with a machine learning virtual screening tool

• Multimodal feature fusion for molecular property classification

• P2MAT: A machine learning (ML) driven software for Property Prediction of MATerial

• Topological data analysis for antibody-drug conjugate payload discovery: a computational framework for mechanistic classification and target validation

• PES2MP: A Python Application for Automating Collisional Dynamics of Linear Rigid-Rotors

• RANQSAR: a standalone open-source application for reproducible machine learning-based QSAR analysis

• OpenStats: how to combine statistics and research data management (RDM) to leverage efficient scientific data analysis by guided statistics

• KCFtools: rapid alignment-free method for introgression screening and GWAS using k-mer profiles

• PACEff Builder: An Efficient Platform for Constructing PACE Hybrid-Resolution Models for Molecular Dynamics Simulations of Aqueous Protein, Peptide Assembly, and Membrane Protein Systems

• TransKla: A Local-Global Cross-Attention Based Transformer Approach for Prediction of Lysine Lactylation Sites

• FoldDoF: Utilizing the Primary Degrees of Freedom of Protein Backbone for Geometric Modeling and Generation

• QSAR in the Browser: An Interactive Cheminformatics Web Application

• BFEE-Docking: A User-Friendly and Customizable End-to-End Tool from High-Throughput Virtual Screening to Binding Free-Energy Calculations

• Efficient Coupled-Cluster Python Frameworks for Next-Generation GPUs: A Comparative Study of CuPy and PyTorch on the Hopper and Grace Hopper Architecture

• Knowledge Distillation of a Protein Language Model Yields a Foundational Implicit Solvent Model

• Intricate Role of Cholesterol in Membrane Fusion

• Decoding Galectin–Glycan Recognition with 19F-Tagged Lectins: from Simple Glycans to the Cellular Glycocalyx

• Navigating Misfolded Terrain: ER-Associated Degradation of Membrane Proteins

• One Face, Three Solutions: Structural Convergence in PD-L1 Inhibition across Antibodies, Macrocycles, and Small Molecules

• Rational Development of Activatable Prodrugs of the GSTP1 Inhibitor NBDHEX: Turn-On NIR Fluorogenic Drug Delivery with Selective Anticancer Activity

• Recent Highlights in the Discovery and Design of Antibody–Drug Conjugates

• Late-Stage Carbonyl Removal via Sequential Double Carbon–Carbon Bond Cleavage

• Open- and Closed-Shell Roles of Sensitizer and Annihilator in Pseudo-Single Component Mixtures for Upconversion

• To Biotic or Abiotic: Biohybrid Systems for Artificial Photosynthesis

Rhys@RhysSullivan

coworker told me we need to run some "smoke tests"

Rhys @RhysSullivan

Horrifying to be listening to a human being speak and realize they’re just a vessel for Claude

6:54 PM · Jul 3, 2026 · 170K Views

56 Replies · 90 Reposts · 3.97K Likes

himanshu@himanshustwts

LMAO

ali @waterloo_intern

we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B - 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance turns out the student is not bounded by the teacher. it also converged on one universal truth. we open-sourced the

5:34 AM · Jul 4, 2026 · 696K Views

50 Replies · 247 Reposts · 8.65K Likes

Dmitrii Kovanikov@ChShersh

@lauriewired Unexpected Lisp content from Laurie

4:33 PM · Jul 2, 2026 · 8.16K Views

2 Replies · 5 Reposts · 205 Likes

Akshat@star_stufff

probably the most based linkedin profile

6:03 PM · Jul 2, 2026 · 12.8K Views

6 Replies · 24 Reposts · 398 Likes

Arnav Gupta@championswimmer

John Carmack is so based he can explain how Task Manager works to the guy who created Task Manager.

12:32 AM · Jul 3, 2026 · 446K Views

98 Replies · 297 Reposts · 7.88K Likes

pedro@pedropimlico

1:09 AM · Jul 3, 2026 · 1M Views

42 Replies · 12.7K Reposts · 114K Likes

Ran Blekhman@blekhman

Claude Science is incredible. I gave it some sequencing data, and in 8 hours it did a full analysis, generated figures, wrote a paper, submitted it for publication, got rejected, revised and resubmitted, got rejected again, it is now applying for positions in industry

1:25 AM · Jul 3, 2026 · 474K Views

101 Replies · 680 Reposts · 9.09K Likes

Mykola Harmash@mykolaharmash

ok vscode, so copy-paste is now an async operation and needs a spinner, got you! 🚮

8:51 AM · Jun 30, 2026 · 171K Views

56 Replies · 31 Reposts · 2.3K Likes

pushkar /ˈpʊʃkər/@thepushkarp

pushkar /ˈpʊʃkər/ @thepushkarp

why do they have to make the word "kernel" mean so many things?

5:43 AM · Jul 5, 2026 · 1.59K Views

2 Replies · 25 Likes

Ash Jogalekar@curiouswavefn

Jefferson took time out from his duties four times on that day to record the temperature. Also, when Jefferson came to Washington as vice president, he carried with him several big bones. Jefferson believed they belonged to a gigantic extinct predator, and he presented a

1776 Live @250YearsAgoLive

For some unknown reason, Thomas Jefferson is preoccupied this morning with recording the temperature. He writes that upon awaking at 6 AM, the temperature is a pleasant 68.0 degrees Fahrenheit, but that it is creeping up toward 70. On his way to Congress, Jefferson purchases a

5:56 AM · Jul 5, 2026 · 2.82K Views

2 Reposts · 25 Likes

Jacob X. Li ✈️ ICML@jacobli99

If I see this on ur poster I might just walk away

11:33 PM · Jul 4, 2026 · 48.3K Views

22 Replies · 3 Reposts · 556 Likes

Math Lady Hazel 🇦🇷@mathladyhazel

Today I learned that the FBI spent decades tracking mathematician Paul Erdős, only to conclude that the guy was just really into math.

muckrock.com

The FBI spent decades tracking mathematician Paul Erdős, only to conclude that the guy was just really into math

12:04 AM · Jul 5, 2026 · 4.88K Views

9 Replies · 4 Reposts · 101 Likes

𝕡𝕨𝕟𝕚𝕖@0xpwnie

12:11 PM · Jul 4, 2026 · 25.5K Views

17 Replies · 69 Reposts · 764 Likes

Jake Wintermute 🧬/acc@SynBio1

Biologists using fable like

11:11 PM · Jul 3, 2026 · 8.22K Views

4 Replies · 11 Reposts · 334 Likes

Best,
Manas