Evaluating World Models for Molecular Drug Discovery

More sophisticated AI representations do not automatically outperform established molecular features in drug discovery. Deepsense.ai tested a JEPA-inspired world-model approach using graph-based molecular representations, with self-supervised pre-training on roughly three million ChEMBL molecules and evaluation on antimicrobial and HIV datasets. Fine-tuning removed much of the pre-training advantage, while combining learned representations with classical molecular fingerprints performed better than either alone. AI leaders in life sciences should benchmark advanced models against simpler baselines and test hybrid approaches before increasing model complexity.