Highlights
Do humans and large language models agree on the quality of synthesis plans?
Partially. The best ”agreeing” model, Gemini 3.1 Pro, reached MCC of ~0.4 against human expert consensus, while Claude Opus 4.8 (0.2), GPT-5.5 (0.35), and Llama 3.1 70B (0.07) showed weaker and less reliable agreement. Betteridge’s Law of Headlines almost holds.
Chemically Interpretable Explanations for Molecular Property Prediction via Fragment-Level Shapley Values
SHAP (SHapley Additive exPlanations) assigns each input feature a number showing how much it changes a specific prediction, using Shapley values from game theory. Shapley values do this by looking at every possible subset of features, checking what the model predicts with and without a given feature added to that subset, and averaging that difference across all subsets. This average (marginal) contribution is the feature’s Shapley value. This means with N features, there are 2N possible subsets i.e. for a 2048-bit fingerprint, there are 22048 possible subsets. That’s way too many calculations. So, there are approximations to it such as kernalSHAP, treeSHAP, etc. It’s obviously not perfect with many approximations such as assuming feature independence, etc. read more here. In this paper, Roth redefined the “players” in the Shapley value game to be resulting BRICS fragments instead of fingerprint bits or graph nodes, which makes exact Shapley value computation feasible by exhaustively evaluating the model on all “fragment subsets”. They applied this to RF and GCN models trained on aqueous solubility, Ames mutagenicity, and SARS-CoV-2 Mpro potency and compared the resulting fragment-level attributions against standard SHAP and GNNExplainer. They show Fragment-level Shapley values recovered known chemical trends across all three tasks. Compared to SHAP on bits and GNNExplainer, the fragment-based attributions were seemingly more chemically meaningful, though the two models (RF vs. GCN) differed somewhat in exact fragment rankings.
Targeting BCL‑2 through Deep Learning-Based Drug Repurposing: A Multimodal Approach Combining Diffusion-Based Generative Modeling, Neural Relational Inference, and In Vitro Validation
Sayyah et al. screened FDA-approved drugs against BCL-2 and found relugolix as a hit. I did a report on BCL-2 and venetoclax as the BH3 mimetic for my medchem class last year, so I remember that venetoclax is no ordinary drug from a design, synthesis and structural perspective and if you don’t know the FBDD story behind it, you need to check out the amazing https://fbdd-lit.blogspot.com and https://practicalfragments.blogspot.com and papers like this one. So, when I see a repurposing claim it’s very interesting, and even better for the nerds, this paper does in vitro studies to back up the claim. The authors used NeuralPlexer to predict BCL-2–ligand complex structures for 3094 FDA-approved drugs, then filtered candidates through docking refinement, MD simulations, and more to compare each candidate’s dynamic interaction pattern to the reference inhibitor, Venetoclax. This identified relugolix (oral drug primarily used to treat advanced prostate cancer or heavy menstrual bleeding due to uterine fibroids) as a promising BCL-2 hit, which they then confirmed experimentally via TR-FRET binding and cancer cell viability assays. Pretty cool !
Long List
Cheminformatics
Q2 performance: Toward Unbiased Selection of Machine-Learning Regression Models
Robust generative transition-state models for unseen chemistry
Computational techniques to study breast cancer scaffolds for antiangiogenesis: a review
Sequencing saturation does not uniquely determine molecular recovery in UMI transcriptomics
Phylogenetic inference under the balanced minimum evolution criterion via semidefinite programming
Assessing the influence of different alignment tools on the accuracy of a forensic epigenetic clock
MMGNN: Multilevel, Multicolor Graph Neural Networks for Molecular Property Prediction
Extending the Python RESP Framework for Extra Point Charge Fitting
Methodological Considerations in Small-Sample Multitask QSPR Modeling for PAMPA Permeability
MMGNN: Multilevel, Multicolor Graph Neural Networks for Molecular Property Prediction
Differentiable Thermodynamic Phase-Equilibria for Machine Learning
Ser715 Phosphorylation Induces β-Hairpin Unfolding and Destabilization of FOXM1 Autoinhibition
Dual-Attention Multimodal Framework for Molecular Property Prediction
FastRet: Fast and Simple Retention Time Prediction in Liquid Chromatography
Accurate Identification of Covalently Ligandable Cysteines Using CCSite
MedChem
Chemical Stabilization and Transformation of Schiff Base Scaffold: A
Path toward Superior Platinum-Based Anti-Cancer Stem Cell Drugs
Structure and Dynamics of the HIV‑1 Envelope Protein on the Virion
How Polymer Chains Remember Their Crystalline Past: Simulations
Other
Palate Cleanser
Best,
Manas