This Week In Cheminformatics: Issue #031

Eric Masson published Equilibrist as an open source streamlit web app for fitting equilibrium and rate constants from concentration using NMR, and UV-vis/fluorescence titration data via a plain-text reaction scripting language. The implementation includes 2D RMSE profiling to visualize pairwise parameter correlation, a sloppy-spectrum eigendecomposition to show which parameter combinations the data actually constrain, and many more interesting checks. There are also in built uncertainty estimates using bootstrap, jackknife and monte carlo. Pretty useful alternative to SupraFit. The code is available on Github.

Ahmed et al. benchmark an ML docking pipeline (DiffDock-Pocket for pose generation, rescored with the Pafnucy 3D-CNN affinity model) against CDOCKER and Surflex-Dock on DUD-E and LIT-PCBA, using AUC and EF1 (enrichment at the top 1% of the ranked list vs. random). On DUD-E, DiffDock-Pocket/Pafnucy and Surflex-Dock performance is similar (AUC 0.70/0.70, EF1 7.64/7.40). CDOCKER wasn't run on LIT-PCBA (too slow), but while comparing DiffDock-Pocket/Pafnucy with Surflex-Dock, they observed both drop in performance (AUC 0.59/0.48, EF1 1.94/1.73). ML workflow's performance margin over Surflex is proportionally larger on LIT-PCBA than on DUD-E. They also show DiffDock-Pocket beats CDOCKER on pose reproduction (79.8% vs. 73.9% of poses within 2Å RMSD, using 40 samples vs. CDOCKER's 500 !), but when both sets of poses are rescored with the same FACTS solvation function, CDOCKER has better AUC. This shows pose accuracy and scoring performance are not correlated. It is a good reminder that DUD-E performance numbers for ML docking methods are mostly upper bounds.

PPARγ has been targeted for type 2 diabetes for quite some time now. This is usually done via full agonists like rosiglitazone, pioglitazone, and troglitazone or partial agonists like TZDs, halofenate, and amorfrutins. It’s been shown that the safety profiles of full agonists is much worse than partial agonists. So, this begs the question what makes a partial agonist, a partial agonist. In this paper, Lin et al. try to answer this question and show that their “Ligand Drift Rate Descriptor“ correlates with efficacy classes. From earlier works using NMR, MD, etc. studies it’s shown that these partial agonists might bind beyond the canonical binding site and might sample multiple conformations and binding sites within the pocket. The authors took this inspiration and docked each ligand 100 times, took the top 20 poses from each and counted how many of them fail to stay within a H-bond distance of some key residues that define the canonical binding region. This descriptor is what they called Ligand Drift Rate Descriptor. With some parameter tweaking, using GOLD’s GA, pose diversity of ~0.5 Å RMSD, and receptor flexibility enabled, at a 10 Å H-bond displacement threshold, the Ligand Drift Rate was able to separate full agonists from partial agonists in both training and ChEMBL validation set (N=1079). They show that partial agonists consistently drifted more than full agonists across essentially all tested configurations, supporting their two-state Boltzmann-like model of ligand occupancy. They did note however that choice in docking algorithm, pocket structure, H-bond displacement threshold, retained pose count (20 vs more) matters a lot and changing it might yield unexpected results.

• Exploration of Minimum Energy Paths Through Feature-Vector Force-Driven Reaction Coordinate Exploration Approach

• Computational mapping of productive POI–E3 ligase conformations to guide de novo degrader design: application to WEE1 and PKMYT1 PROTACs

• NNP/CG-MM: Embedding of All-Atom Neural Network Potentials into a Coarse-Grained Molecular Mechanics Environment

• Assessing the Stability of Molecular Glues with Weighted Ensemble Simulations

• Relation-Aware Pretraining and Reaction Center Modeling for Chemical Reaction Graph Representation Learning

• MolT5-Linker: A Transformer-Based Sequence- and Putative Site-Guided Generative Framework for Antibody-Drug Conjugate Linkers

• DeepHeptox: An Interpretable Deep Learning Model for Multi-Endpoint Hepatotoxicity Prediction of Chemical Compounds

• CYPMol: A Single Model Framework Integrating Functional Residues with Protein Features and Molecule Embeddings to Predict CYP Substrates, Inhibitors, and Metabolism Sites

• Self-Supervised Learning for Molecular Property Prediction: Methods, Multimodal Insights, and Benchmark Comparisons

• lmp2gro: A Python Tool for Converting LAMMPS Data Files into GROMACS Topologies

• A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic Simulations

• Chemical Information Modeling for Component-Level Review Prioritization of Adverse-Reaction Signals in Multi-Component Medicines

• Mapping the Glycan Recognition Landscape of Galectin-3 through Enhanced Sampling Simulations

• FGSGT-DDI: An LLM-Enhanced Functional Group Semantic Graph Transformer for Drug–Drug Interaction Prediction

• ADMET-Driven Chemical Navigability Rules for Early-Stage Virtual Compound Prioritization

• Epistatic Modulation of Sec/Cys Catalysis in GPX6 Revealed by EVB Free-Energy Landscapes

• MolGuidance: Advanced Guidance Strategies for Conditional Molecular Generation with Flow Matching

• Estimating Permeabilities from Biased Molecular Dynamics with a Green–Kubo Formalism for Transport across the Blood-Brain Barrier

• Structural Thermodynamics of Membrane Electroporation: Voltage-Driven Interdigitation Induces Electro-Topological Disparity

• Deep Learning-Driven Anticancer Drug Discovery: Emodepside as a Potential Therapeutic Candidate for Triple-Negative Breast Cancer

• TextDTI: A Multimodal Context Representation Learning Framework for Drug–Target Interaction Prediction

• Energetically Anchored Machine-Learning Interatomic Potential Embeddings for Reliable Hydrogen Evolution Electrocatalyst Prediction

• DF-S4: Disentangled FiLM-Conditioned Molecular Generation Using the S4 Architecture

• In Silico Isomerization Produces Apt Negative Data for VHTS Validation

• Identifying Cryptic Binding Sites with Mixed Solvent MD Simulation and SiteMap

• A Multimodal Semi-Supervised Learning Framework for Pharmaceutical Cocrystals Prediction

• Unveiling the Structural Determinants of PFAS Toxicity: A Graph Neural Network and Interpretability Analysis

• DisoPatho: A Cross-View Feature-Adaptive Interaction Encoding Framework for Predicting Disease-Associated Variants in Intrinsically Disordered Regions

• MetaboT: an LLM-based multi-agent framework for interactive analysis of mass spectrometry metabolomics knowledge graphs

• Artificial intelligence in impurity prediction: current landscape, challenges, and future directions

• Implicit solvent effects on the binding interactions of amines with CO2

• How Structural Subtleties Modulate Excited-State Lifetimes in Cyclometalated Cobalt(III) Complexes: Nonadiabatic Molecular Dynamics Study

• Rapid and energy-efficient ultra-large library screening for drug discovery on a SpiNNaker2 neuromorphic chip

• AlphaDTA: integrating AlphaFold3 embeddings and 3D complex structures for drug–target binding affinity prediction

• ComplexMatrixComb: predicting anticancer drug combination dosages for target growth inhibition via complex-number matrix factorization

• Some Recent Observations on the Role of Carbon Tetrel Bonding in Drug Design

• From Structure to Selectivity: Integrating CADD and AIDD in Rational Medicinal Chemistry Strategies for DYRK1A Inhibitor Discovery

• Generative AI-Assisted Discovery of HPK1 Inhibitors

• Automated DNA-Encoded Library Synthesis and Activity-Based Screening at the Attomole Scale

• Predicting Ligand Binding Modes by Scaffold-Guided Structure Refinement

• Membrane Protein Clustering as a Key Determinant of Endocytosis: Natural Mechanisms, Nonintentional Transporter Internalization, and Engineered Multivalent Nanocarriers

• On the Role of Hydrogen Bond Strength and Charge Transfer in an On-Water Diels–Alder Reaction: Semiempirical and Free Energy Calculations

• Genetically Encoded Pentafluorophenylalanine Enables Quantitative Probing of Local Protein Malleability by 19F NMR

Did anyone notice how horrible the typesetting (on web) has become since the latest ACS website migration. Also what’s up with all these migrations first RCS now ACS. Reminds me of this XKCD.

Alex Kontorovich@AlexKontorovich

Terry’s notes on one of Jean Bourgain’s papers. (Look close… 😂 Jean was one of my longest term collaborators, until his untimely passing a few years ago.) #ICM2026

11:48 PM · Jul 24, 2026 · 38.1K Views

10 Replies · 37 Reposts · 465 Likes

Ash Jogalekar@curiouswavefn

In chemistry this is usually a good sign that you need to switch to an alternative protocol.

6:23 PM · Jul 22, 2026 · 1.19K Views

2 Replies · 11 Likes

Navid@navidmafi

can we stop polluting `$HOME` everyone? how about we put these clearly configuration files in .... ~/.config 😳

11:45 PM · Jul 22, 2026 · 150K Views

107 Replies · 157 Reposts · 3.92K Likes

frankie@FrankieIsLost

"What is the charge? Using an open source model? A succulent Chinese open source model?"

12:33 AM · Jul 23, 2026 · 235K Views

60 Replies · 727 Reposts · 7.11K Likes

sharon✨coming to venice🎬@iamsharonle

watching The Odyssey the way Christopher Nolan intended

8:36 PM · Jul 22, 2026 · 3.62M Views

354 Replies · 28.1K Reposts · 430K Likes

Kai Fell@kai_fell

hey @BetterSoftwareC when are you uploading this guy's talk?

6:12 PM · Jul 22, 2026 · 17.3K Views

7 Replies · 11 Reposts · 388 Likes

Hugging Models@HuggingModels

OpenAI: “We’re launching an investigation to find whoever hacked Hugging Face.” opens internal logs “…” “Guys.”

OpenAI @OpenAI

We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation. Sharing preliminary findings to help defenders understand emerging risks:

2:32 AM · Jul 22, 2026 · 911K Views

84 Replies · 603 Reposts · 11.2K Likes

frankie@FrankieIsLost

the counterexamples will continue until morale improves

Dmitry Rybin @DmitryRybin1

Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years. The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60. Chat with GPT 5.6 Pro where this was found: https://t.co/Oi2PQoab2h

5:23 PM · Jul 22, 2026 · 78.8K Views

8 Replies · 99 Reposts · 2.17K Likes

Charles Rosenbauer@bzogrammer

Hear me out, I have an idea on how to expand C++'s syntax in a relatively clean way.

♡ mari/cohe ♡ @noinconsistency

C++26 reflection syntax looks abhorrent

6:55 AM · Jul 26, 2026 · 54.1K Views

19 Replies · 66 Reposts · 1.37K Likes

Evgeny Gazdovsky@gazdovsky

Just hold my beer

Hassan Abedi 📚🌿🦉🪬 @Hassan_Abedi

@evisdrenova

1:27 AM · Jul 26, 2026 · 10.9K Views

1 Reply · 7 Reposts · 345 Likes

River@rivdb

this is cool when youre 16 and have no real work to do

7:31 PM · Jul 26, 2026 · 164K Views

57 Replies · 210 Reposts · 5.05K Likes

François Fleuret@francoisfleuret

So @AnthropicAI and @DarioAmodei are all over the place lobbying against open-source AI, and their foot soldiers are making snarky comments about companies who support the said open-source AI? Nice.

Julian Schrittwieser @Mononofu

I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!

8:25 AM · Jul 26, 2026 · 61.7K Views

30 Replies · 41 Reposts · 997 Likes

special thanks to Wim and Andreas for the banger playlists that kept me writing in the flight.

Best,
Manas