This Week In Cheminformatics: Issue #032
Domain adaptation for Transformers, MPO guided by Bioisostere replacement, Boltz‑2 Evaluation, and a long list of papers
Highlights
Transformers for molecular property prediction: domain adaptation efficiently improves performance
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.
Autonomous bioisosteric replacement for multi-property optimization in drug design
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.
Reliability of AI Methods in Drug Discovery: Evaluation of Boltz‑2 for Structure and Binding Affinity Prediction
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.
Long List
Cheminformatics
Exhaustive molecular string enumeration for data augmentation and structure exploration
Supervised Restricted Data Fusion With Common, Local, and Distinct Components
Activity-cliff awareness enables robust graph learning for molecular property prediction
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
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
Range-aware Bayesian optimization for discovering diverse designs within target property windows
Quantifying the Energetics of Protein-Mediated Stabilization of
Automated Reaction Network Enumeration and Profiling Platform
for Mechanism-Guided Catalyst Design and Reactivity Optimization
MDTransport: A Modular, Open-Source, Extensible Python Tool for
Gaps in AI-Driven Pharmacokinetic Property Prediction for Early
Predicting Enzyme pH Optima from Structure Using Equivariant
HSAF-DDI: Heterogeneous Semantic-Aware Drug−Drug Interaction
Native Contact Ratio as a Topological Metric for Machine Learning
From Dynamics to Diagnosis and Therapy: A Multiscale
Computational Framework for MALT1-Targeted Cancer Theranostics
Soft Non-diagonality Penalty Enables Latent Space-Level
MedChem
Other
Palate Cleanser
Best,
Manas