Vikram Sundar

FLIGHTED improves protein fitness model performance by accounting for experimental noise in high-throughput data, showing that data size and architecture matter more than model scale.

This paper introduces DHARMA, an ultra-high-throughput method for measuring biomolecular activities, and benchmarks protein models using large variant datasets.

The paper uses data augmentation with single protein-ligand models to improve drug-target interaction predictions for unseen proteins, achieving AUC > 0.9 on test sets with no experimental data.