This Week In Cheminformatics: Issue #004

Random Forest Baseline Strikes Again, Better Atropisomer Prediction & Long List of Papers

Jan 19, 2026

Highlights

Prediction of Atropisomerism for Drug-like Molecules

In this new JCIM study, Balduf and colleagues demonstrate that standard force fields are systematically blind to this “Class 2” atropisomer, which presents barriers between 20-28 kcal/mol that allow isolation on the bench but risk racemization in vivo. Standard force fields like OPLS4 systematically overestimate rotational barriers above 30 kcal/mol. This misclassifies “Class 2” atropisomers as “Class 3” (stable enantiomers). OPLS4 correctly identified only 13 of 25 Class 2 molecules in this benchmark which is barely better than a coin flip. This suggests you cannot trust a force field to locate the transition state in all cases. Their proposed workflow uses OPLS4 solely for rough filtration, then geometry optimization and TS search is done by QRNN-TB (a semi-empirical ML potential). The results show locating the TS on the ML-corrected surface before the final DFT energy score improves accuracy drastically in particular RMSE drops from 5.4 kcal/mol (OPLS4) to 2.0 kcal/mol (QRNN-TB/DFT) and Accuracy jumps from 69% to 91%, correctly flagging 20 out of 25 Class 2 risks.

Kinetic predictions for SN2 reactions using the BERT architecture: comparison and interpretation

In this Digital Discovery study, Wilson et al. benchmark a transformer-based BERT model against a Random Forest (RF) baseline for predicting SN2 reaction kinetics, revealing that both machine learning approaches significantly outperform traditional quantum mechanics. The models achieved an RMSE of approximately 1.1 log k on external test data, surpassing high-level CCSD(T) calculations (RMSE 1.9 log k) while reducing prediction costs from hours to milliseconds. Although the RF model was vastly faster to train (256 seconds versus BERT’s 53 hours), the transformer architecture demonstrated superior chemical reasoning in specific domains, successfully learning solvent effects purely from text tokens and extrapolating reliably to reaction rates outside the training distribution. However, the study clearly highlights blind spots in the BERT model as it failed to recognize the rate-enhancing allylic effects of aromatic groups that the simpler Random Forest model correctly identified.