ASTRA

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

---

Authors

• Daniel Saeedi¹ · Amirali Aghazadeh¹ (corresponding, amiralia@gatech.edu)
• Nihit Pokhrel² · Leijia Gao² · Charley Wen² · Elizabeth Bruce²
• José C. Aponte³
• Amanda Stockton⁴

^{1 — School of Electrical and Computer Engineering, Georgia Institute of Technology · 2 — Microsoft Discovery · 3 — Astrochemistry Laboratory, NASA Goddard Space Flight Center · 4 — School of Chemistry and Biochemistry, Georgia Institute of Technology}

ASTRA is a multi-agent system that reconstructs chemical synthesis networks connecting simple prebiotic feedstocks (H₂O, CO₂, NH₃, HCN, H₂CO, H₂S, CH₄, H₂) to complex biomolecules such as amino acids. A SynthesisNetworkAgent proposes a complete, multi-pathway reaction network for a target molecule; a CriticAgent scores it for chemical feasibility, pathway coherence, environmental consistency, and literature alignment; and an optional deep-research / RAG stage grounds both agents in the origins-of-life literature. Generated networks are then validated against the curated ChemOrigins reference dataset.

The repository doubles as an ablation study: it measures how different research-context sources (Deep Research from Claude / ChatGPT / Gemini, MS Discovery, Perfect RAG, or none) and different LLM backends (Claude Opus, GPT-4.1, GPT-5.4) affect the quality of the reconstructed networks.

ASTRA pipeline: a target molecule and a literature report feed the SynthNet and Critic agents, which iterate until confidence ≥ threshold, producing an environment-tagged reaction network with full per-reaction chemical detail.

Highlights

• Single-shot network generation — one agent call emits a complete SynthesisNetwork (molecules, reactions, and pathways), rather than a step-by-step search.
• Critic-in-the-loop — networks are re-generated with feedback until the critic's confidence crosses a threshold.
• Environment-grounded chemistry — every reaction is tagged with a prebiotic environment (Hydrothermal, Surface, or Biochemical) and annotated with temperature, pressure, catalyst, mechanism, and scientific reasoning.
• Pluggable LLM backends — Claude Opus, GPT-4.1, and GPT-5.4 selectable via a single env var.
• Reproducible validation — networks are matched reaction-by-reaction against ChemOrigins reference data with RDKit/InChI similarity, yielding precision and recall.

Installation

git clone <this-repo-url>
cd astra
python -m venv .venv && source .venv/bin/activate   # optional but recommended
pip install -r requirements.txt

Dependencies include autogen-agentchat/autogen-ext (0.4.x), rdkit, drfp, anthropic, openai, and google-generativeai. Python 3.10+ is recommended.

Configure credentials

All credentials live in config.env. You only need the keys for the backend you select — the others can stay as placeholders. Set MODEL_PROVIDER to the model that should build the networks:

# config.env
MODEL_PROVIDER=claude          # claude | gpt41 | gpt5.4_openai | gpt5.4_agent_azure

ANTHROPIC_API_KEY=sk-ant-...   # required when MODEL_PROVIDER=claude
CLAUDE_MODEL=claude-opus-4-6

MODEL_PROVIDER	Backend	Required keys
claude	Anthropic Claude Opus	ANTHROPIC_API_KEY
gpt41	Azure OpenAI GPT-4.1	AZURE_GPT41_*
gpt5.4_openai	OpenAI GPT-5.4 (Responses API)	OPENAI_DIRECT_*
gpt5.4_agent_azure	Azure AI Projects agent (GPT-5.4)	AZURE_AGENT_*

Quick Start (notebook)

The fastest way to generate your first network is QuickStart.ipynb. It walks through the full generation flow interactively — you supply the deep-research context yourself from a chat assistant, so no research API keys are required (only the key for your chosen MODEL_PROVIDER):

jupyter lab QuickStart.ipynb     # or: jupyter notebook

The notebook takes you through five steps:

1. Pick a target molecule — any common or IUPAC name (Glycine, L-Alanine, Serine, …).
2. Collect deep research — the notebook prints a ready-to-copy prompt; run it in the Deep Research / Research mode of Claude.ai, Gemini, or ChatGPT.
3. Paste the markdown answer back into the notebook.
4. Run the ASTRA pipeline — SynthNet proposes a network, the Critic scores it (up to MAX_RETRIES cycles).
5. Inspect the results — the reconstructed network JSON and its confidence score.

Generated networks are written to outputs/quickstart/.

Full pipeline (CLI)

For batch / ablation experiments, use pipeline.py. It has three stages that chain end-to-end, with output routed per provider (outputs/{provider}/ → outputs_inchi_corrected/{provider}/ → validation_report/{provider}/). Every stage auto-resumes, skipping already-completed (molecule, config, run) combinations.

# 1) Generate synthesis networks
python pipeline.py generate --molecules L-Alanine,Glycine --runs 3 --provider claude

# 2) Backfill canonical InChI from PubChem
python pipeline.py inchi --provider claude

# 3) Validate networks against ChemOrigins reference data
python pipeline.py validate --provider claude

# ...or run all three stages in sequence for one provider:
python pipeline.py all --molecules L-Alanine --provider claude

# per-stage flags:
python pipeline.py generate --help

Useful generate flags: --configs (which research-context ablation to run), --runs, --num-pathways, --confidence-threshold, --max-retries. See CLAUDE.md for the full architecture and output-layout reference.

How it works

ASTRA grounds every reaction in one of three prebiotic environments from the ChemOrigins framework:

• Hydrothermal (deep-sea alkaline vents) — 350–600 K, high pressure, iron-sulfur catalysis; CO₂ reduction, reductive amination, pyruvate synthesis.
• Surface (evaporitic pools & geothermal fields) — 300–370 K, wet-dry cycles, UV; Strecker synthesis, HCN oligomerization, cyanosulfidic pathways.
• Biochemical (prebiotic assembly) — peptide/nucleotide assembly and proto-metabolic cycles that stitch hydrothermal and surface products into biomolecules.

Results

ASTRA reconstructs literature-supported synthesis routes across the standard amino acids and, in several cases, proposes chemically plausible ASTRA-inferred links that fill gaps between known steps.

(a) Expert-metric radar profiles across three LLM backends and four research-context conditions. (b–c) Physicochemical structure of the reconstructed networks — hydrogen-bond donors/acceptors and Crippen logP across starting, hub, intermediate, and per-environment molecules. (d) Feedstock usage across the reconstructed amino-acid networks.

Per-amino-acid evaluation profiles comparing single-prompt baselines against report-grounded runs (Exact, Claude, ChatGPT, Gemini, and No Report).

Reconstructed multi-environment routes to alanine (a) and glycine (b), with each step tagged by environment and annotated with supporting citations; purple = hydrothermal, brown = surface, green = biochemical.

Reconstructed routes to histidine (a) and tryptophan (b). Report snippets show how ASTRA identifies and reasons about the key synthetic bottlenecks (e.g., prebiotic formation of the indole ring).

Citation

If you use ASTRA in your research, please cite:

@article{saeedi2026astra,
  title   = {Autonomous Multi-Agent Reconstruction of Prebiotic Reaction Networks
             Reveals Organizational Features of Amino Acid Synthesis},
  author  = {Saeedi, Daniel and Pokhrel, Nihit and Gao, Leijia and Wen, Charley and
             Bruce, Elizabeth and Aponte, Jos{\'e} C. and Stockton, Amanda and
             Aghazadeh, Amirali},
  journal = {ChemRxiv},
  year    = {2026},
  doi     = {10.26434/chemrxiv.15005927/v1},
  url     = {https://chemrxiv.org/doi/abs/10.26434/chemrxiv.15005927/v1}
}