AI Chem., Volume 1, Issue 2 (June 2026) – 4 articles

Contrastive learning-based models such as DrugCLIP have recently emerged as scalable tools for structure-based virtual screening by embedding protein structures and small molecules into a shared representation space. While these approaches demonstrate high throughput and competitive screening performance in ligand retrieval tasks, their ability to correctly identify biologically relevant ligand-binding pockets has not been systematically evaluated. Here, we construct a benchmarking dataset comprising 42 pharmacologically diverse human protein targets with experimentally validated drug-bound structures spanning multiple target families. Using this dataset, we evaluate the pocket recognition capability of DrugCLIP and compare its performance with a traditional structure-based workflow (Fpocket combined with ESSA) and a machine learning-based method (P2Rank). DrugCLIP shows robust performance for well-characterized target classes, including kinases (10/10) and nuclear receptors (5/5), but exhibits markedly reduced accuracy for ion channels (1/4), GPCRs (3/5), and transporters (3/5). Notably, pocket prediction accuracy does not strongly correlate with structural data availability, suggesting that intrinsic pocket characteristics rather than training data abundance primarily affect model performance. Across the benchmark, DrugCLIP achieves an overall success rate of 71% (95% CI: 56–83%), compared with 79% (95% CI: 64–88%) for Fpocket+ESSA, and 93% (95% CI: 81–98%) for P2Rank. McNemar’s test showed no significant difference between DrugCLIP and Fpocket+ESSA (p = 0.508), whereas P2Rank significantly outperformed DrugCLIP (p = 0.012). Together, these results provide a quantitative evaluation of pocket recognition by contrastive learning-based models and highlight key limitations of embedding-based approaches for pocket localization. Full article

We designed a supervised machine learning framework to predict standard Gibbs free energies, ΔG°, of formal hydrogen atom transfer (f-HAT) for phenolic antioxidants across different radicals and media, enabling rapid and chemically interpretable screening. We curated a DFT dataset of 71 molecules (phenolic compounds and anthocyanidins), with 207 reaction sites, 10 radical reactive oxygen/sulfur species, and three environments (leading to a total of 6210 ΔG° values). The models amass 106 numerical RDKit descriptors, augmented with one-hot encodings of medium, site, radical, and structural class, and were evaluated through a leave-one-molecule-out protocol. Among the tested regression algorithms, the random forest regressor provides the best balance of accuracy and robustness with both R² test (≈0.94) and MAE (2.74 kcal mol⁻¹; RMSE (≈5.0 kcal mol⁻¹)), close to DFT chemical accuracy. The feature-importance analysis revealed that “electronic” and “experimental” (site/group) descriptors primarily drive predictions, with the radical’s maximum absolute partial charge being the most important descriptor in the prediction of a radical’s ΔG°. These results suggest that descriptor-driven RF (Random Forest) models can generalize across chemical space to provide interpretable ΔG° predictions, providing a path for chemists towards a scalable route to prioritize antioxidant candidates for broader molecular families. Full article

Ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco) efficiency constrains carbon fixation, making it a high-value target in biotechnology. The core task of this work is a supervised regression and ranking problem on Rubisco: given a numerical representation of a protein sequence (a PLM embedding), we predict continuous phenotypic scores such as an enzyme kinetic proxy or fitness value. The predictions then guide which variants to test next. Engineering Rubisco is a point of focus but remains challenging due to selection forces in vivo and the combinatorial space of potential mutants for ex vivo uses. We combine protein language model (PLM) embeddings with tabular learning to model Rubisco variant landscapes in two regimes. First, we analyze deep mutational scanning data providing inferred kinetic proxies, including K_m for CO₂ and V_max. Second, we model a cyanobacterial screening dataset measuring mutant fitness under differing oxygen and nitrogen regimes, enabling an oxygen tolerance objective. Across tasks, a tabular foundation (TabPFN-2.5) model outperforms gradient-boosted trees on rank-based criteria for variant prioritization, including Spearman correlation and top 5% hit recovery. We then simulate active-learning campaigns initialized with 200 measured variants and iteratively acquiring batches of 48. Model-guided selection recovers more top-performing mutants than random sampling at fixed experimental budgets, even with a conservative XGBoost surrogate. We also demonstrate that Rubisco large-subunit embeddings predict cyanobacterial doubling time and cross-species kinetic parameters, suggesting that Rubisco representation remains meaningful across organisms even with multi-objective cellular constraints. Together, these results support a practical, data-efficient workflow for enzyme engineering and motivate objective-aware design strategies that complement directed evolution. Full article

Infrared (IR) spectroscopy is a powerful tool for characterizing molecular structures and chemical groups, offering advantages such as low cost, rapid analysis, and non-destructive testing. When analyzing heterogeneous objects, spectra are typically measured from different regions to capture the local variations, presenting a multi-instance learning (MIL) problem. However, existing methods primarily rely on multi-instance assumptions or explicit bag representations, often failing to fully capture the intrinsic information from implicit representations. We introduce a bag dissimilarity regularized MIL framework for analyzing IR spectra of heterogeneous objects, which integrates both explicit and implicit representations to effectively learn the MIL bags. Specifically, a bag dissimilarity regularization term is utilized to extract implicit representations, which subsequently guide the classifier based on explicit representations to enhance generalization performance. The proposed method was validated on two heterogeneous detection tasks: polydimethylsiloxane (PDMS) block assessment and polyethylene terephthalate (PET) fiber inspection. Experimental results demonstrate that our approach significantly outperforms existing methods on both datasets with a considerable margin. Full article