Flock: A Negative-Enriched Protein-Protein Interaction Dataset

Protein-protein interactions (PPIs) play a crucial role in biological processes and are crucial for drug discovery. Despite this, predicting PPIs remains difficult due to the scarcity of reliable negative data (non-interacting PPIs). Traditionally, computational models are trained on positive interactions or random pairs designated as negatives, while the available negative datasets are small and contain errors.

To overcome this, researchers have developed Flock, a new dataset with an unprecedented number of PPIs. Flock includes 26,934 positive pairs and 377,643 negative pairs, making it the largest dataset of its kind.

Flock combines data from the Protein Data Bank (PDB) with highly accurate negatives derived from scientific literature. This is achieved through an advanced workflow that uses a state-of-the-art language model. To ensure fair evaluation, Flock also presents a Leakage-Free Set, which strictly limits sequence and interface similarities, as well as historically based cutoffs used in training co-folding models.

When evaluated, the co-folding model ESMFold2 showed a decline in performance on the Leakage-Free set, while the protein language model ESMC demonstrated improved generalization capabilities. With its extensive and challenging dataset, Flock sets a new benchmark for assessing the generalizability of protein structure and language models in predicting PPIs.