AstraBIND: Rapid AI Predicts Ligand Binding Sites for Accelerated Drug Discovery

Predicting where small molecules, or ligands, bind to proteins is central to computational biology and drug discovery. Existing machine learning approaches for this task often face a trade-off: while structure-based deep learning models typically outperform sequence-based methods, they frequently demand high computational cost or ligand-specific data, thereby limiting scalability.

Addressing this challenge, researchers have developed AstraBIND, a lightweight graph neural network designed to bridge this gap. AstraBIND combines protein sequence, structural information (whether experimentally determined or predicted), and homology data to predict both ligand classes and specific binding residues within minutes. Built on a GATv2 architecture with 0.9 million parameters, the model was trained on approximately 250,000 curated protein-ligand complexes spanning 16 different ligand categories.

The core of AstraBIND's success lies in its ability to encode residue-level features and spatial geometry through graph attention mechanisms. This allows it to identify binding residues and classify ligand types while ensuring structural consistency. In benchmarking tests, AstraBIND demonstrated robust performance, achieving a weighted macro-F1 score of 0.47 across all ligand classes, with top results for nucleotides (F1 = 0.79), porphyrins (0.74), and cofactors (0.73). Case studies involving proteins like p53 and CRFR1 further validated its capacity for robust pocket localization for diverse proteins.

AstraBIND's capabilities are highlighted by lead author Goteti, who notes in the paper, "Combined with its minimal inference time and broad ligand coverage, AstraBIND enables rapid in-silico screening and integration into laboratory workflows." This minimal inference time, predicting binding sites within minutes, positions it as a powerful tool for accelerating the drug discovery pipeline. Together with other Astra ML models, AstraBIND represents a step toward real-time protein design and validation pipelines. The Astra models are available online at https://www.orbion.life.