This Week In Cheminformatics: Issue #029

M-JEPA, reaction network reduction, MolRank, how to find interesting molecules, automated parallel synthesis, and a long list of papers

Jul 13, 2026

Highlights

Automated Parallel Synthesis Accelerates Virtual Screening Hit Discovery

McKenna, Šícho, and colleagues built COMBINAUT, a repurposed peptide synthesizer running SPS workflows (amide/sulfonamide coupling, reductive amination, SNAr, nitro reduction, heterocyclization) across 715 in-house building blocks, enumerating a 22.9M compound virtual library and physically producing hits within 32h instead of the 4-6 week make-on-demand turnaround !! They report that not all but 78/100 selected VS hits actually got synthesized to spec, and only 9 of those were validated in radioligand binding against the CCR2 allosteric pocket (12% hit rate, pretty good !). They also try to justify these numbers in the paper carefully. They then used the same platform for rapid SAR on two hits, getting 4-6x potency improvements in a few hours of parallel synthesis per round, ending with two cell-active antagonists (IC50 256 nM and 1.3 µM in β-arrestin assays). It’s a modest library by ultra-large VS standards (22.9M vs billions), but the closed-loop docking followed by synthesis, assay, re-docking cycle all in hours rather than weeks is the actual contribution here, and it’s pretty amazing if you think about it.

M-JEPA: Predictive Self-Supervised Learning for Molecular Graphs with Scaffold-Shift Evaluation on Tox21

M-JEPA ports the JEPA recipe (context encoder, EMA teacher, latent-space prediction instead of contrastive InfoNCE) to molecular graphs, using BFS-grown connected-subgraph masks as the target region. Under matched compute, JEPA beats a matched InfoNCE baseline on ESOL proxy RMSE and trains faster, and hybrid fine-tuning from the JEPA checkpoint lifts Tox21 mean ROC-AUC only modestly (0.561 to 0.609) but drops mean ECE substantially (0.279 to 0.064) which is a useful reminder that ranking and calibration under scaffold “shift” are separable properties and gains in one don’t guarantee the other. Their connected vs noncontiguous masking ablation was statistically non-signigicant on both proxy RMSE and motif-level coherence, so the “chemically coherent” argument for BFS masking seems basically aesthetic.

Graph-Based Generation and Reduction of Complex Chemical Reaction Networks

Here, Fite & Gross add a neat methodological contribution to the reaction network reduction problem. Most reduction algorithms are validated on a handful of published mechanisms, which tells you little about generality. The authors build a preferential attachment based generator fitted to ammonia, methane, and hydrogen combustion mechanisms, producing 2,613 artificial networks with sampled thermodynamics and Arrhenius kinetics, to get a large, unbiased test set. On top of that they introduce MolRank, a PageRank style topological algorithm that estimates upper bounds on maximal species concentration and reaction rate using only thermodynamic data. MolRank is markedly better at identifying redundant species than redundant reactions. They also report Eyring-Polanyi based scores give better bounds than true rate constants but don’t preserve ranking quality.

Strategies for Identifying Molecules of Interest in Large Chemical Spaces

Klein et al. (BioSolveIT + Pfizer) benchmark SpaceMACS, SpaceLight, and FTrees across 2,917 CHEMBL queries on REAL Space, eXplore, and PfizerVL, and the headline result is that the three methods barely overlap under 1.2% shared hits across all three for a given query. MCS similarity, fingerprint (ECFP4/fCSFP) similarity, and FTrees pharmacophore similarity are sampling quite different neighborhoods of a chemical space, so which one and at what thresholds should you use them becomes the real question. This comparison was done with similarity scored on synthons rather than fully enumerated products. New bond formed at assembly for instance are invisible to standard ECFP4, which is why fCSFP was built with smaller, non-circular features to survive the synthon boundary better (recovering 90-100% of close analogs vs 70-95% for ECFP4). This paper gives you calibrated per-algorithm recommended thresholds for very similar, similar, and dissimilar compounds from search results (check Figure 16). Good read.