Research demonstrates improved bioavailability and toxicity prediction in drug design using a DeBERTa-based SMILES encoder, highlighting its multi-modal capabilities.
Key Points
The DeBERTa-based SMILES encoder improved predictive capacity for 22 ADMET endpoints.
Achieved a 14–30% improvement in critical drug properties like bioavailability compared to previous models.
Trained on a 300K dataset, utilizing multi-label regression to stabilize learning across diverse properties.
Results indicate that DeBERTa’s disentangled representations enhance structural fluency in AI-driven drug design.