AI agentic platforms predict human and virus protein interactions
A recent preprint on arXiv (2604.23924) presents two innovative AI-driven platforms. The first platform autonomously trains machine learning models to forecast protein-protein interactions (PPIs) between humans and viruses, while the second generates explicit, human-readable rules for these interactions. The initial platform utilizes five AI agents for tasks such as data collection, verification, feature embedding, model design, and training-validation across three-way protein-disjoint cross-fold datasets. It achieves an accuracy of 87.3% for human-human PPIs and 86.5% for human-virus PPIs. The second platform enhances interpretability by substituting ML predictions with rules based on protein embeddings, physicochemical autocovariance descriptors, compartment annotations, and pathway-domain information, showcasing autonomous AI-driven discovery in molecular biology.
Key facts
- arXiv:2604.23924v1
- Two agentic AI platforms built: one for autonomous ML training, one for rule induction
- Five AI agents handle data collection, verification, feature embedding, model design, training/validation
- Three-way protein-disjoint cross-fold datasets used
- Human-human PPI ensemble accuracy: 87.3%
- Human-virus PPI ensemble accuracy: 86.5%
- Second platform uses protein embeddings, physicochemical autocovariance, compartment annotations, pathway-domain data
- Focus on interpretability via human-readable rules
Entities
Institutions
- arXiv