Exploring data augmentation: Multi-task methods for molecular property prediction

The effectiveness of machine learning (ML) for molecular property prediction is often limited by scarce and incomplete experimental datasets. A particular promising approach to facilitate training ML models in low-data regimes is multi-task learning. We investigate how additional molecular data – even potentially sparse or weakly related – can be augmented through multi-task learning to enhance prediction quality. Through controlled experiments on progressively larger subsets of the QM9 dataset [Ruddigkeit et al. (2012), J. Chem. Inf. Model; Ramakrishnan et al. (2014), Sci. Data], we evaluate the conditions under which multi-task learning outperforms single-task models. We extend these insights to a practical real-world dataset of fuel ignition properties that is small and inherently sparse, offering recommendations for augmenting auxiliary data to improve predictive accuracy. This work provides a systematic framework for data augmentation in molecular property prediction, with implications for data-constrained applications.