Breast cancer is a highly prevalent and molecularly heterogeneous malignancy, where therapeutic resistance and off‐target toxicity remain major challenges. Here, we introduce BreastCAR, a multi‐target quantitative structure–activity relationship framework trained on ChEMBL bioactivity data covering 53 cancer‐relevant targets and 233975 compounds. The model integrates 2048‐bit Morgan fingerprints with physicochemical descriptors and applies random forest, XGBoost, LightGBM (LGB), support vector machine with RBF kernel (SVM‐RBF) and multilayer perceptron (MLP) for both regression (pIC 50 ) and three‐class activity classification. The framework achieves a mean R 2 of 0.723 and root mean squared error of 0.661, with a peak R 2 of 0.894 for NOTC1. Classification performance shows a mean F1 score of 0.826, area under the curve–receiver operating characteristic of 0.919, Matthews correlation coefficient of 0.679 and balanced accuracy of 0.759. Y‐randomisation confirmed non‐trivial model learning (permuted R 2 = −0.167 vs. real 0.723; permuted F1 = 0.467 vs. real 0.826). Applicability domain analysis (Tanimoto k‐NN, 95th‐percentile threshold) demonstrated that 94.8% of test compounds fall within the training AD. To experimentally validate the top‐ranked BreastCAR predictions, a panel of 13 benzylidene‐scaffold compounds (R26‐01 to R26‐13) was evaluated in vitro against MCF‐7 human breast cancer cells using the MTT antiproliferative assay ( n = 3, IC 50 determined by non‐linear regression). Two compounds—R26‐06 (4‐bromobenzylidene, IC 50 = 5.84 ± 0.24 μM, SI = 9.2) and R26‐07 (3‐chlorobenzylidene, IC 50 = 6.72 ± 0.28 μM, SI = 8.4)—surpassed tamoxifen (IC 50 = 14.23 ± 0.58 μM) with statistical significance ( p < 0.001), demonstrating that BreastCAR‐guided prioritisation translates computational predictions into experimentally confirmed antiproliferative activity. A FastAPI‐based web interface enables real‐time SMILES input, structural visualisation and activity prediction. The complete code and models are provided to ensure reproducibility.
Designing selective kinase inhibitors is a multiobjective optimization problem constrained by the structural conservation of the human kinome, along with the lack of high‐quality interaction data. Traditional deep learning methods often train complex neural networks directly on small datasets and fail to generalize to new chemical scaffolds. To this end, we present Geometric Manifold Alignment , a novel linear approach that shifts away from architectural complexity. We introduce Multi‐View Linear Manifold ( MVLM ), which aligns the latent spaces of ChemBERTa (small molecules) and ESM‐2 (protein sequences) into a shared geometric interaction space. We pretrain MVLM on the Universal Foundation Set of 473 923 high‐confidence interactions, where binding affinity is represented as geometric proximity between drug and target embeddings. Achieving an AUROC of 0.936 on KIBA and 0.928 on Davis, MVLM establishes the new state‐of‐the‐art on industry‐standard benchmarks, outperforming existing methods ( p < 0.001). In this work, interpretability is geometric where interaction scores are cosine similarities in a shared manifold, and the learned structure recovers kinase phylogenetic relationships without supervision. We also demonstrate the model’s utility as a transparent discovery engine by retrieving known clinical targets of FDA‐approved drugs. Our findings suggest that aligning pretrained chemical and biological representations provides a scalable and interpretable approach to advancing computational drug discovery.
The histamine H 3 receptor (H 3 R) is an attractive central nervous system target due to its regulatory role in histaminergic neurotransmission and its ability to modulate the release of several other neurotransmitters implicated in cognition, arousal, and neurodegenerative disorders. In this study, an integrated chemometric workflow was developed to identify novel H 3 R ligands by combining structure‐based and ligand‐based pharmacophore modeling, QSAR analysis, molecular docking, and experimental validation. A structure‐based pharmacophore model derived from the H 3 R crystal structure and a ligand‐based pharmacophore model generated from known H 3 R inhibitors were validated by ROC analysis. The top‐performing pharmacophore models showed excellent discrimination between actives and decoys. The QSAR model demonstrated good predictive performance ( r 2 = 0.75, q 2 = 0.78, predicted r 2 = 0.67), supporting its application in hit prioritization. The validated pharmacophore models were subsequently used to virtually screen the ChemDiv CNS BBB library of 22 790 compounds. Docking, consensus scoring, and QSAR‐guided filtering identified 20 top‐ranked hits for experimental evaluation in a [ 3 H]‐ N α ‐methylhistamine displacement assay. One hit, D466‐0370 ( 60 ), exhibited submicromolar affinity for H 3 R ( K i = 641 nM), while seven additional compounds showed low micromolar binding affinities. In silico ADMET profiling for the most active hits further supported their compatibility with CNS‐relevant chemical space. Overall, this integrated workflow efficiently identified structurally novel H3R ligands with promising biological activities.
Melting point prediction plays a pivotal role in chemistry, materials science, and pharmaceuticals, where it informs compound identification, purity assessment, and design of thermally robust materials. Despite advances in computational modeling, current methods face challenges in accuracy, interpretability, and computational efficiency, largely due to limited experimental datasets and the nonlinear, discontinuous relationship between molecular structure and melting behavior. In this study, we curate a harmonized dataset of 117 752 unique organic crystals by integrating and standardizing multiple public sources, emphasizing reproducibility and molecular uniqueness. We systematically evaluate classical machine learning models using molecular embeddings derived from SMILES strings, including ChemBERTa, MoLFormer, Morgan fingerprints, and Uni-Mol, and compare them with advanced graph neural network (GNN) architectures such as our custom implementation and the widely used Chemprop framework. Our analysis demonstrates that gradient-boosted tree models offer notable computational efficiency, while GNNs and transformer-based models like MoLFormer provide superior predictive accuracy, particularly when enhanced with attention mechanisms. By integrating ChemBERTa and MoLFormer embeddings in the TabM model, we achieve a mean absolute error (MAE) of 26.75 K and an RMSE of 35.36 K, surpassing individual model performance. These results highlight the benefits of hybrid modeling strategies that balance interpretability, scalability, and accuracy, offering a robust framework for accelerating materials discovery and pharmaceutical development through improved melting point prediction. Nevertheless, Chemprop exhibits considerably higher accuracy but demands approximately five times greater resources during training.
While anti-human epidermal growth factor receptor 2 (HER2) drugs, including trastuzumab improve the prognosis of patients with HER2-positive cancers, acquired resistance by cancer cells remains a significant clinical challenge. We had previously reported that YES1 is overexpressed in trastuzumab-resistant breast and lung cancer cells, and that its inhibition restores trastuzumab sensitivity. However, no approved drug specifically targets YES1. In this study, as a drug repositioning approach aimed at rapid clinical application, we constructed a machine learning model capable of learning the interactions between compounds and target proteins and conducted a screening of approved drugs. The strategy identified chlorpromazine and raloxifene as hit compounds. In vitro validation demonstrated that they effectively inhibited YES1 activation at 15 mu M after 48 h. The findings suggested that the identified approved drugs might represent promising therapeutic options for overcoming anti-HER2 drug resistance.
