ASTRA

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

Daniel Saeedi
Daniel Saeedi1
Amirali Aghazadeh
Amirali Aghazadeh1*
Amanda Stockton
Amanda Stockton4
Nihit Pokhrel
Nihit Pokhrel2
Leijia Gao
Leijia Gao2
Charley Wen
Charley Wen2
Elizabeth Bruce
Elizabeth Bruce2
José C. Aponte
José C. Aponte3
Georgia Institute of Technology
Microsoft Microsoft Discovery
NASA
User Task · Input
Instructions
⟶
Target Molecule
L-Alanine structure
L-Alanine
2-aminopropanoic acid
Deep Research
Deep Research
Autonomous multi-agent literature review synthesis
researching…
0sources
Subagents · parallel
Live search · 0 results
Sources · 611
Key findings
✦ Research report ready
Literature Report
Literature Report
Deep Research Agent · Report
The Deep Research Agent's structured report — synthesis routes, key intermediates, reaction mechanisms and conditions, mineral catalysts, experimental evidence, and unresolved mechanistic gaps — passed verbatim to the SynthNet Agent as literature context.
literature_report.md generating
SynthNet
SynthNet Agent
Backbone · Generator
Backbone: GPT-4.1 · GPT-5.4 · Claude Opus 4.6
Molecule Nodes Reaction Nodes Named Pathways Hubs & Convergence
Given the literature report, generates a complete bipartite reaction network in a single LLM call — across Hydrothermal, Surface, and Biochemical environments — with molecule nodes (network role + environment), annotated reaction nodes (reactants, products, catalysts, conditions, rationale, citations), named pathways, and network-level hubs and convergence points.
synthnet_output.json single LLM call
Critic
Critic Agent
Backbone · Evaluator
Backbone: GPT-4.1 · GPT-5.4 · Claude Opus 4.6
Chemical Feasibility Pathway Coherence Environmental Consistency Target Achievement Hub Quality Pathway Convergence
Accept if overall confidence ≥ 0.90 · otherwise return feedback to SynthNet and refine (≤ 3 iterations).
critic_output.json scoring
Reaction Network
Converged synthesis graph → L-Alanine
Reactant (solid) Product (dashed) Starting Mol. Intermediate Hub Intermediate Target Molecule High recall (>75%) Hydrothermal rxn Surface rxn Biochemical rxn
rxn 6Reaction Details
Temperature
Pressure
Catalyst / Agents
Aqueous
Mechanism
Scientific Reasoning
🗄️100%
Results
Figure 2a · Reaction-network quality across backbone models
GPT-4.1GPT-4.1
GPT-5.4GPT-5.4
Claude Opus 4.6Claude Opus 4.6
Model Comparison
Averaged across all configurations
Emergent Properties
Figure 2b–d · physicochemical structure across the 8-amino-acid benchmark
H-bond donors / acceptors by role
Hydrogen-bond donor and acceptor counts by molecule role
Crippen logP by environment
Crippen logP distribution by formation environment
20 most-generated molecules
Per-Target Profiles
Figure 3 · reconstruction quality per amino acid · Serine → Histidine
Reconstructed PathwayFigure 4a · L-Alanine · two convergent routes
Hydrothermal Surface Biochemical
Reconstructed PathwayFigure 4b · Glycine · four convergent routes
Hydrothermal Surface Biochemical ASTRA-Inferred
Hypothesis NetworkFigure 5a · Histidine · literature-sparse target
Hydrothermal Surface Biochemical
Hypothesis NetworkFigure 5b · Tryptophan · mechanistic gaps flagged
Hydrothermal Surface Biochemical Mechanistic gap