Molecular property prediction in the ultra-low data regime

Data scarcity remains a major obstacle to effective machine learning in molecular property prediction and design, affecting diverse domains such as pharmaceuticals, solvents, polymers, and energy carriers. Although multi-task learning (MTL) can leverage correlations among properties to improve predictive performance, imbalanced training datasets often degrade its efficacy through negative transfer. Here, we present adaptive checkpointing with specialization (ACS), a training scheme for multi-task graph neural networks that mitigates detrimental inter-task interference while preserving the benefits of MTL. We validate ACS on multiple molecular property benchmarks, where it consistently surpasses or matches the performance of recent supervised methods. To illustrate its practical utility, we deploy ACS in a real-world scenario of predicting sustainable aviation fuel properties, showing that it can learn accurate models with as few as 29 labeled samples. By enabling reliable property prediction in low-data regimes, ACS broadens the scope and accelerates the pace of artificial intelligence-driven materials discovery and design.