PeptiVerse: A Unified Multi-Agent Platform for Therapeutic Peptide Property Prediction

ABSTRACT

Designing therapeutic peptides requires balancing binding affinity with complex ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties, a task that overwhelms traditional single-model predictors. We introduce PeptiVerse, a unified platform utilizing a swarm of specialized 'Critic Agents,' each fine-tuned to predict specific biochemical traits like membrane permeability, proteolytic stability, and immunogenicity. An 'Aggregator Agent' synthesizes these evaluations to rank de novo peptide candidates. This multi-agent architecture outperforms traditional monolithic sequence-based predictors by dynamically querying chemical language models and structural embeddings as needed, creating a modality-flexible workflow for peptide therapeutics.