Automatic Prediction\nof Molecular Properties Using\nSubstructure Vector Embeddings within a Feature Selection Workflow

Machine learning\n(ML) methods provide a pathway to accurately\npredict\nmolecular properties, leveraging patterns derived from structure–property\nrelationships within materials databases. This approach holds significant\nimportance in drug discovery and materials design, where the rapid,\nefficient screening of molecules can accelerate the development of\nnew pharmaceuticals and chemical materials for highly specialized\ntarget application. Unsupervised and self-supervised learning methods\napplied to graph-based or geometric models have garnered considerable\ntraction. More recently, transformer-based language models have emerged\nas powerful tools. Nevertheless, their application entails considerable\ncomputational resources, owing to the need for an extensive pretraining\nprocess on a vast corpus of unlabeled chemical data sets. To this\nend, we present a semisupervised strategy that harnesses substructure\nvector embeddings in conjunction with a ML-based feature selection\nworkflow to predict various molecular and drug properties. We evaluate\nthe efficacy of our modeling methodology across a diverse range of\ndata sets, encompassing both regression and classification tasks.\nOur findings demonstrate superior performance compared to most existing\nstate-of-the-art algorithms, while offering advantages in terms of\nbalancing model accuracy with computational requirements. Moreover,\nour approach provides deeper insights into feature interactions that\nare essential for model interpretability. A case study is conducted\nto predict the lipophilicity of chemical molecules, exemplifying the\nrobustness of our strategy. The result underscores the importance\nof meticulous feature analysis and selection over a mere reliance\non predictive modeling with a high degree of algorithmic complexity.