Molecular Property Prediction
Overview
Molecular property prediction studies how to infer quantities such as solubility, toxicity, binding affinity, and quantum energies from molecules, mixtures, or molecular environments. Common approaches use fingerprints and physicochemical descriptors, graph neural networks, molecular language models, and 3D equivariant or multimodal models; some also encode physical laws or learn from experimental and simulated data together. Models are assessed on benchmarks such as MoleculeNet and TDC, with scaffold, temporal, and other out-of-distribution splits used to test transfer to unfamiliar chemistry. A central challenge is that strong in-distribution scores and larger pretrained models do not guarantee reliable extrapolation, while careful classical baselines, data quality, uncertainty estimates, and evaluation design can strongly affect results.
Latest papers
204 papers29 papers in the last four weeks, level with the four weeks before. 0.2% of all new papers.