Autonomous Multi-Agent Reconstruction of Prebiotic Reaction Networks Reveals Organizational Features of Amino Acid Synthesis

[chemrxiv.org] Decades of origin-of-life research have uncovered hundreds of prebiotically plausible reactions across hydrothermal, surface, and biochemical environments, yet this knowledge remains fragmented across chemistry, geochemistry, and astrobiology.

Here we present ASTRA, a multi-agent AI system that reconstructs prebiotic reaction networks for biomolecular targets by integrating autonomous literature synthesis, reaction network construction, and iterative mechanistic evaluation.

Applied to the twenty canonical amino acids, ASTRA reconstructed 286 reaction networks encompassing 2,850 plausible pathways and 4,513 reactions, more than 92% of which are not represented in existing curated databases of prebiotic reaction pathways. Across these networks, ASTRA identifies recurrent organizational features of prebiotic amino acid synthesis, including shared network hubs, shared intermediate building blocks, and environment-dependent molecular properties.

For literature-sparse targets such as histidine and tryptophan, ASTRA reconstructs pathways grounded in published experiments, identifies missing mechanistic links, and generates explicit, testable bridging hypotheses.

These results reveal system-level organization in prebiotic amino acid synthesis and demonstrate how autonomous scientific agents can accelerate the reconstruction and exploration of chemical evolution.

ASTRA generates literature-grounded prebiotic reaction networks through three coordinated LLM agents. a Given a target molecule (here, L-alanine), a Deep Research agent performs an automated literature review and compiles a structured report from the prebiotic chemistry literature (top right; displayed snippet cites [9]; Methods). In this study, reports were generated using commercial Deep Research services (ChatGPT, Claude, or Gemini). A SynthNet agent then proposes the molecular inventory, reaction set, and candidate pathways, producing an environment-partitioned bipartite reaction network represented as structured JSON (bottom left). Each reaction node contains reactants, products, catalysts, environmental conditions, mechanistic rationale, and supporting citations. A Critic agent evaluates pathway feasibility, pathway coherence, environmental consistency, and target completion, as well as network-level properties including hub quality and pathway convergence. Networks with confidence below 0.9 are iteratively refined until the confidence threshold is reached or the maximum number of cycles is exhausted. b Representative generated reaction network for L-alanine. Molecule nodes (circles) are colored by role (black, starting material; blue, intermediate; red, target), whereas reaction nodes (rectangles) are colored by inferred prebiotic environment (purple, hydrothermal; orange, surface; teal, biochemical). Solid edges connect reactants to reactions, and dashed edges connect reactions to products. Framed reaction nodes indicate reactions with bestmatch ChemOrigins similarity greater than 0.75. The inset (right) expands a representative reaction node (rxn 6, nitrile hydrolysis of 2-aminopropionitrile to alanine), showing the reaction-level metadata produced by SynthNet: physical conditions, catalysts, proposed mechanism, scientific rationale with supporting citation, and the closest matching reaction in the ChemOrigins database [6]. — chemrxiv.org

Autonomous Multi-Agent Reconstruction of Prebiotic Reaction Networks Reveals Organizational Features of Amino Acid Synthesis, chemrxiv.org

Astrobiology, AI, LLM,