Generated Molecules as Pretraining Data When Reference Data Are Limited

Molecular foundation models are becoming increasingly important in AI-driven drug discovery, where pretrained molecular representations support a broad range of downstream applications, including molecular property prediction and virtual screening. However, their effectiveness depends critically on the diversity and coverage of pretraining data, while the continued expansion of high-quality reference molecular corpora can be constrained by data availability and acquisition cost. Recent molecular generative models offer a promising means of alleviating this limitation by producing chemically plausible and structurally diverse molecules, yet whether such generated molecules provide useful learning signals for downstream molecular models remains insufficiently understood. Here, we investigate whether synthetic molecules can augment molecular pretraining when reference data are limited. Using GraphMVP as a representative self-supervised molecular representation learning framework, we compare three controlled pretraining conditions: 25,000 GEOM reference molecules (Ref25K), the same 25,000 reference molecules supplemented with 25,000 MolDiff-generated molecules (Mix50K), and a full corpus of 50,000 GEOM reference molecules (Ref50K). All conditions follow the same pretraining protocol with the same optimization budget and are evaluated using the same downstream fine-tuning procedure across eight molecular property prediction tasks. Mix50K achieves a macro-average ROC-AUC of 0.7232, compared with 0.7060 for Ref25K and 0.7222 for Ref50K. Thus, Mix50K improves over Ref25K while achieving overall performance comparable to Ref50K. It also obtains a higher mean ROC-AUC than Ref50K on four of the eight downstream tasks. These results indicate that generated molecules can provide useful pretraining signals and suggest synthetic molecular augmentation as a promising strategy for expanding molecular foundation-model pretraining when reference data are limited.