Multi-agent reasoning enables predictive design of living materials
Journal:
bioRxiv
Published Date:
Feb 16, 2026
Abstract
Living materials derive function from tightly coupled cellular and material processes to deliver adaptive and therapeutic capabilities, yet their predictive design remains constrained by fragmented, cross-disciplinary knowledge and experience-driven iteration. Here we introduce LiveMat, a multi-agent reasoning framework that reconstructs living materials as a computable design space from unstructured literature. LiveMat curates and standardizes 34,738 living-material records, integrating 16,086 microorganism entries and 18,682 polymer entries into a domain-scale knowledge graph comprising tens of thousands of entities and relationships. Through constraint-driven multi-agent reasoning and expert-anchored evaluation, the system converts implicit design heuristics into explicit, auditable rules. Comparative benchmarking across leading large language models shows that limitations in living materials reasoning arise primarily from cross-domain feature integration rather than coarse-grained classification. In a prospective acute wound-healing task, LiveMat systematically evaluates combinatorial four-component living material configurations under a formalized multi-constraint scoring framework and identifies a top-ranked design whose in vivo performance matches state-of-the-art systems. LiveMat establishes a scalable reasoning infrastructure for cumulative, data-grounded living materials discovery.
