MolSight: Molecular Property Prediction with Images

MolSight introduces a vision-based approach to molecular property prediction (MPP), utilizing 2 million molecule images and evaluating 10 vision architectures across various tasks, including physical-property regression and drug-discovery classification. The study employs a chemistry-informed curriculum that categorizes molecules by structural complexity, leading to significant performance improvements, with the best configuration achieving top results on 5 out of 10 benchmarks while requiring 80 times fewer FLOPs compared to leading multi-modal models. This research highlights the potential of using simple 2D skeletal diagrams for efficient and effective MPP, offering practitioners a scalable alternative to traditional methods.