This Week In Cheminformatics: Issue #032

Sultan et al. present a rigorous pre-training scaling study for molecular transformers: BERT-style models trained on 0/30/60/100% of GuacaMol (~1.3M molecules), evaluated across seven ADME datasets with Butina splits. The standout result is that pre-training gains plateau at 30-60% of the data (~400-800K molecules), with more data not helping and sometimes hurting.
However, an extra step before fine-tuning where they keep (unsupervised) training the model on the small (~ thousands of molecules) downstream dataset, to predict RDKit physicochemical descriptors. This step itself gives significant gains across the seven ADME datasets they tested, and lets a model pre-trained on far less data match MolBERT and perform better than MolFormer. Note that RDKit descriptors with random forest work just as well as every transformer on most datasets, and R² stays in the 0.2-0.4 range everywhere. So, the takeaway isn’t that transformers are solution for everything but if you’re going to use a transformer, this cheap domain-adaptation trick might just work better than simply scaling up pre-training.

Kim et al. present DeepBioisostere, a fragment-based generative model that picks the fragment to remove, selects a replacement from a learned fragment library, and predicts attachment orientation, all conditioned on target property deltas. The core idea is a shared atom / fragment embedding space (hierarchical message passing, following Modof) that scores candidate insertion fragments by comparing them directly against the removal-site embedding + context. They show insertion choice is context-sensitive (same fragment removed from two molecules gives different QED-improving substitutions), and in a hit-to-lead benchmark against outputs from Pocket2Mol / DeepICL / TargetDiff / DecompDiff, it clearly beats random / frequency / MMPA baselines on joint QED + SAscore improvement while keeping docking scores within ~1 kcal/mol. Interesting paper.

Wan et al. benchmark Boltz-2 on two large compound libraries (16,780 3CLPro, 21,702 TNKS2 ligands), comparing its predicted structures and binding affinities against docking and their own ESMACS free-energy calculations. Structurally, Boltz-2 frequently places ligands in a different pocket than the 3CLPro crystal structure, and a subset of ligands show wrong hybridization relative to their SMILES. Energetically, correlation with ESMACS is weak to moderate across the full data sets (r = 0.24 for 3CLPro, r = 0.45 for TNKS2), and drops to near zero when restricted to the top 100 ranked compounds per target! They also examine the BindingDB comparison that often gets cited in Boltz-2’s favor. Boltz-2 correlates well with BindingDB experimental values (r = 0.77 for TNKS2), but the authors show BindingDB’s own internal experimental reproducibility is only around r = 0.56, and the post-training cutoff test compounds overlap closely in chemical space with the pre-cutoff training data. So the high correlation is more consistent with memorization of similar scaffolds than genuine affinity prediction. Good read.

• Exhaustive molecular string enumeration for data augmentation and structure exploration

• In silico identification of selective cyclodecapeptide inhibitors targeting Plasmodium falciparum Grp78 chaperone

• Scalable and Generalizable Analog Design via Learning Medicinal Chemistry Intuition from Matched Molecular Pair Transformations

• Expanding accessible chemical space for fragment-based enumeration by orders of magnitude through optimization of the Chemically Reasonable Mutations (CReM) framework

• Supervised Restricted Data Fusion With Common, Local, and Distinct Components

• How well can off-the-shelf LLMs elucidate molecular structures from mass spectra using chain-of-thought reasoning?

• Gaussian Mixture Modelling and DBSCAN for Reclassifying Legacy pXRF Compositional Data: A Chemometric Workflow Applied to Archaeological Ceramics

• Activity-cliff awareness enables robust graph learning for molecular property prediction

• Transformer-based molecular fragment prediction using SMILES and DeepSMILES representations in a fragment-based drug discovery pipeline

• Design of Thermotropic Liquid Crystal Molecules With a Wide Temperature Range for the Liquid Crystal Phase Using Machine Learning

• Mapping Trofinetide Polypharmacology in Rett Syndrome: A Multi-Stage Computational Analysis

• Design of a metabolically-stable peptide therapeutic with triple-hormone-receptor agonist activity

• First principles exploration of high hydrogen storage capacity of lithium-based perovskite hydrides

• Local Asymmetric Least Squares (LAsLS) for Noisy and Complex Baseline Correction

• Prediction of solubility of hydrogen in chemicals using QSPR-based machine learning approach: a comparative study

• EPSpec: An Evidence-Guided, Prior-Retrieval Agent for Near-Infrared Spectral Band Selection

• MolDualNet as a multi-modal architecture for small-data analog-space molecular property prediction

• Data-driven catalyst design for direct catalytic N2O decomposition

• BRIDGE: an interactive application for multi-omics data analysis, visualization and integration

• Linkerability of Protein Ligands: Insights From Cocrystal Structures and Implications for DNA-Encoded Libraries

• Snake Optimization Algorithm for Variable Selection in Spectral Quantitative Analysis of Complex Samples

• Unlocking a shortcut: acceleration of non-aqueous electrolyte development by active machine learning for sodium-based batteries

• Range-aware Bayesian optimization for discovering diverse designs within target property windows

• Evaluating BioEmu-Generated Kinase Ensembles Reveals Structure Selection as the Virtual Screening Bottleneck

• Quantifying the Energetics of Protein-Mediated Stabilization of

  Membrane Deformation

• Automated Reaction Network Enumeration and Profiling Platform

  for Mechanism-Guided Catalyst Design and Reactivity Optimization

• MDTransport: A Modular, Open-Source, Extensible Python Tool for

  Computing Transport Properties from GROMACS and LAMMPS

  Simulations

• Gaps in AI-Driven Pharmacokinetic Property Prediction for Early

  Drug Development: A Scoping Review

• Predicting Enzyme pH Optima from Structure Using Equivariant

  Graph Neural Networks

• HSAF-DDI: Heterogeneous Semantic-Aware Drug−Drug Interaction

  Prediction with Hierarchical Adaptive Fusion

• Local Geometry Recovers but Cooperative Structure Does Not:

  Residue-Resolved Limits of All-Atom Reconstruction from SingleBead Coarse-Grained Disordered Protein Ensembles

• From Ternary Modeling to Predictive PROTAC Design: A

  Computational Perspective

• N‑Terminal Chirality and Sequence Variations Modulate the

  Conformational Landscape of Amyloid-Beta 42

• Native Contact Ratio as a Topological Metric for Machine Learning

  Based Molecular Docking

• Short-Term Memory Active Learning for Drug Development

• From Dynamics to Diagnosis and Therapy: A Multiscale

  Computational Framework for MALT1-Targeted Cancer Theranostics

• Soft Non-diagonality Penalty Enables Latent Space-Level

  Interpretability of Parameter-Efficient Peptide LM at No

  Performance Cost

• Property-Biased Covalent DNA-Encoded Library Screening Enabled the Discovery of AM-8719, A Structurally Novel, CNS-Penetrant KRAS G12C Inhibitor

• Discovery of a Novel P2X2/3 Allosteric Antagonist as an Efficient Bitter Taste Blocker

• Homolysis of Alkoxyamines Triggered by Carbohydrate-Active Enzyme: A Step toward a General Platform to Generate Toxic Alkyl Radicals

• Structure-Based Discovery of JN210 as a Potent Dual DCN1/HDAC Inhibitor for the Treatment of Nonsmall Cell Lung Cancer

• Toxicological Knowledge−Guided Graph Learning for Interpretable and Generalizable Molecular Toxicity Prediction

• Umpolung Character of Styrenes: Mechanism and Stereoselective Studies of α, β-Substituted Amino Acids Using Chiral Ir-Phosphine Complexes

Ash Jogalekar@curiouswavefn

First they came for the mathematicians, And I did not speak up because I was not a mathematician and felt smug in my experimental verification barrier

5:09 PM · Aug 1, 2026 · 22.4K Views

11 Replies · 34 Reposts · 423 Likes

Alex Hirsch@_AlexHirsch

@sama What if you just talked to your children

7:06 PM · Jul 31, 2026 · 2.95M Views

400 Replies · 17.3K Reposts · 200K Likes

Security Trybe@SecurityTrybe

8:24 PM · Aug 1, 2026 · 264K Views

30 Replies · 1.16K Reposts · 9.85K Likes

Parsa@ParsaIsBack

11:24 PM · Jul 31, 2026 · 50.4K Views

8 Replies · 69 Reposts · 1.41K Likes

François Fleuret@francoisfleuret

People are joking with "Gaussian processes" as if it was some kind of ancient arcane ml magic. I wrote this 17y ago. fleuret.org/public/EN_note…

2:01 PM · Aug 1, 2026 · 61K Views

26 Replies · 65 Reposts · 900 Likes

Christian Lück@another_clue

Valid PHP. I'm going to hell. #phptips

8:54 AM · Jul 31, 2026 · 245K Views

97 Replies · 120 Reposts · 2.83K Likes

Anshul Kundaje@anshulkundaje

Wow! Just incredible. This kind of lineage tracing will be so important for learning causal basis of cell state transitions. (Happening in academia btw .. at universities)

Jay Shendure @JShendure

Thrilled to post thread re: new single-cell lineage of mouse embryo reconstructed w/ DNA Typewriter. One animal, zygote to late organogenesis (E13.5). Tree has 1,340,794 transcriptionally profiled, annotated tips (cells), 1,142,588 dated internal nodes, rooted at zygote 1/n

7:32 PM · Jul 31, 2026 · 22.1K Views

3 Replies · 21 Reposts · 185 Likes

Ash Jogalekar@curiouswavefn

“I belonged to a small minority of boys who were lacking in physical strength and athletic prowess and were squeezed between the twin oppressions of a vicious headmaster and bullying boys. We found our refuge in a territory that was equally inaccessible to our Latin-obsessed

4:37 AM · Jul 31, 2026 · 4.95K Views

18 Reposts · 100 Likes

Sebastian Kozuch Comp Chem@KozuchSebastian

Will this be a new slide in my homogeneous catalysis course?

Mark Gandelman @MarkGandelman

Happy to share our latest Angewandte Chemie paper, in which we report a new class of nitrogen-based ligands that exhibit L-, X-, and for the first time, Z-type coordination, enabling an interesting concept of redox-by-coordination. https://t.co/bJSfnkq91V

12:04 PM · Jul 30, 2026 · 217 Views

2 Likes

NOT Potato Bolshevik@NotPotBol

1:58 AM · Jul 31, 2026 · 2.04M Views

93 Replies · 22.3K Reposts · 237K Likes

owl@owl_posting

its Real

8:14 PM · Jul 30, 2026 · 1.03M Views

78 Replies · 840 Reposts · 27.3K Likes

Jahir Sheikh@jahirsheikh8

Me trying to defend my O(n³) solution during the coding interview

9:58 AM · Jul 30, 2026 · 444K Views

104 Replies · 605 Reposts · 8.6K Likes

Cyber City Circuits@MakeAugusta

3:15 PM · Aug 3, 2026 · 26.9K Views

11 Replies · 78 Reposts · 1.15K Likes

Lexer@LexerLux

rationalists: "von neumann was the smartest man of ALL TIME! THE GOAT!" von neumann:

2:38 AM · Aug 3, 2026 · 249K Views

75 Replies · 150 Reposts · 5.47K Likes

Best,
Manas