Plato-Scientific-Research-Autonomous-Agent

Plato

Plato is a multi-agent research workflow that turns a data specification into research ideas, methods, executable analyses, and manuscript drafts. Its verification gates are designed to make evidence and limitations inspectable; human authors remain responsible for scientific validity and publication.

Plato-Bio preprint

Plato-Bio: verification-first biological novelty screening with temporal rediscovery and structural benchmarks is a fork-specific methods and validation study of Plato’s computational-biology workflow. It asks a narrow question: can an AI-assisted research system make its evidence, intermediate claims, calculations, and limitations inspectable enough for a human scientist to verify?

Stefan G. Creadore is the sole author, writing through Praxa Labs, an independent open-source research initiative, United States. The preprint was submitted to arXiv on July 27, 2026, with cs.AI as the primary category and q-bio.QM as a cross-list. It has not been peer reviewed, and this repository does not claim that arXiv submission or moderation validates its conclusions.

Important links

• Public arXiv record
• Current preprint PDF
• Supplementary PDF
• Canonical manuscript
• LaTeX source and BibTeX references
• Reproduction and artifact guide
• arXiv source package and submission metadata
• Experiment scripts, machine-readable results, and figures
• Author ORCID
• GitHub repository
• Media kit and verified plain-language summary

Abstract

Large language model research agents can connect literature retrieval, analysis code, and manuscript preparation, but coherent output does not establish scientific validity. Plato-Bio extends the open Plato/Denario architecture with explicit workflow states, provenance records, citation checks, claim-to-evidence links, scoped file writes, and publication gates. A source audit found and repaired three measurement defects: loss of task domain in the default evaluation factory, omission of declared method signals from scoring, and evidence sidecars that lacked the drafted-claim denominator.

The study then evaluated two deliberately limited use cases. In one frozen historical task, independent pre-1986 literature bridges ranked the later-studied fish-oil/Raynaud relation first. In a separate comparison of AlphaFold models with experimental structures for 15 human proteins, 11 targets had high-confidence-core Cα RMSD below 1 Å (median 0.501 Å), while 27 traceable discrepancy regions were retained as unvalidated hypotheses. These results support reproducible software contracts and auditable screening baselines—not autonomous discovery, general agent efficacy, or established biological novelty.

Overview

Plato represents research as explicit state machines rather than one monolithic prompt. The biology profile routes retrieval and analysis toward biomedical sources and tools; downstream controls emit citation reports, claim/evidence matrices, consistency checks, run manifests, analysis outputs, and manuscript artifacts. Human authors remain responsible for source review, scientific interpretation, authorship, and publication.

The study separates three evidence lanes:

1. Software-contract validation tests whether implemented controls behave as declared.
2. Temporal rediscovery tests whether pre-cutoff literature can recover a relation studied later.
3. Structural hypothesis triage compares declared AlphaFold and experimental structures while preserving confidence and experimental context.

The lanes are complementary but not interchangeable. Passing software tests does not measure biological accuracy, and a retrospective or descriptive benchmark does not establish prospective novelty.

Scientific study and experiments

Experiment	Design	Primary outputs
Source audit and measurement repair	Inspected the evaluation factory, method-signal scoring, and claim/evidence persistence; added regression coverage for each repair	Correct biology routing, method_signal_recall, and Claim plus EvidenceLink JSONL records
Deterministic software validation	Ran biology, genomics-adapter, evidence/citation, adversarial-safety, and full Python suites on a clean recorded revision	931 passes and 6 skips in the full suite; no failures or errors
Temporal rediscovery pilot	Ranked candidates from a frozen set of pre-1986 PubMed records, with later validation held out	Frequency, TF–IDF, A–B/B–C bridge, and evidence-aware rankings with explicit controls and abstention
Original globin comparison	Sequence-aware Cα alignment and Kabsch superposition for hemoglobin α, hemoglobin β, and myoglobin	Sub-ångström RMSD, residue-level error, pLDDT, source URLs, and hashes
Declared 15-protein structural screen	Compared AlphaFold predictions with experimental PDB structures using whole-chain and pLDDT≥70 core fits	Target/residue CSVs, bootstrap summaries, 27 traceable discrepancy regions, and not_established novelty labels
External benchmark alignment	Used ScienceAgentBench, BioDSA-1K, BixBench, and the pinned CompBioBench catalog as design constraints	Executable artifacts, baselines, provenance, abstention, and explicit failure boundaries; no claimed CompBioBench score

Results

• The complete recorded Python suite contained 937 tests: 931 passed, 6 skipped, 0 failed, and 0 errored. Targeted suites overlap with the full suite and therefore are not added to that total.
• In the single historical rediscovery task, corpus frequency ranked the held-out fish-oil/Raynaud relation third (MRR 0.333), TF–IDF ranked it second (MRR 0.500), and both bridge-only and evidence-aware conditions ranked it first (MRR and Recall@1 1.000). This is one retrospectively curated case, not a population estimate.
• The 15-target structure screen contained 2,688 matched residues. Median whole-chain RMSD was 0.520 Å; median high-confidence-core RMSD was 0.501 Å, with 11 of 15 targets below 1 Å.
• Four targets retained core RMSD above 2 Å: KRAS, SUMO1, TP53, and estrogen receptor α. Confidence masking reduced SUMO1 from 16.610 Å over the whole chain to 2.576 Å over 74 high-confidence residues.
• The declared rule emitted 27 high-confidence discrepancy regions, nine spanning at least two adjacent residues. Every region remains an unvalidated hypothesis requiring construct, ligand, oligomeric-state, alternate-structure, and experimental review.

Conclusions

Plato-Bio demonstrates an inspectable route from retrieved literature or predicted structures to candidate hypotheses. Its strongest contribution is methodological: the workflow preserves known controls, abstentions, provenance, and explicit not_established labels rather than treating difference or fluent prose as discovery. General efficacy claims require a larger preregistered benchmark, repeated agent runs, blinded expert adjudication, independent replication, prospective analysis, and experimental validation.

Selected scientific citations

1. King RD, et al. “The automation of science.” Science (2009). doi:10.1126/science.1165620
2. Lu C, et al. “The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery” (2024). arXiv:2408.06292
3. Schmidgall S, et al. “Agent Laboratory: Using LLM Agents as Research Assistants” (2025). arXiv:2501.04227
4. Villaescusa-Navarro F, et al. “The Denario project: deep knowledge AI agents for scientific discovery” (2025). arXiv:2510.26887
5. Varadi M, et al. “AlphaFold Protein Structure Database.” Nucleic Acids Research (2022). doi:10.1093/nar/gkab1061
6. Jumper J, et al. “Highly accurate protein structure prediction with AlphaFold.” Nature (2021). doi:10.1038/s41586-021-03819-2
7. Chen Z, et al. “ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery.” ICLR (2025). Conference paper
8. Wang Z, et al. “BioDSA-1K: Benchmarking Data Science Agents for Biomedical Research” (2025). arXiv:2505.16100
9. Mitchener L, et al. “BixBench: A Comprehensive Benchmark for LLM-Based Agents in Computational Biology” (2025). arXiv:2503.00096

The canonical manuscript contains the complete reference list, including the PDB, Kabsch, lDDT, FAIR, and historical Raynaud literature used by the experiments.

Reproduce the study

The validation package records exact literature cutoffs, PMIDs, candidate rankings, abstentions, input URLs, versions, SHA-256 hashes, target- and residue-level structural results, exact test counts, and figure-generation code:

.venv/bin/python preprint/experiments/run_globin_structure_benchmark.py
.venv/bin/python preprint/experiments/run_globin_structure_benchmark.py --panel-file preprint/experiments/diverse_structure_panel.json --output-dir preprint/results/diverse_structure_benchmark --figures-dir preprint/figures --benchmark-name diverse_structure_panel
.venv/bin/python preprint/experiments/run_temporal_novelty_benchmark.py --fixtures evals/biological_novelty/fixtures/historical_pilot.json --output-dir preprint/results/temporal_novelty_historical_pilot
.venv/bin/python -m evals.biomedical_benchmarks
.venv/bin/python preprint/experiments/run_software_validation.py
.venv/bin/python preprint/experiments/build_summary_figures.py

The historical benchmark contains one manually curated retrospective task, and the 27 structural discrepancy regions are hypothesis candidates with novelty explicitly marked not_established. The evidence does not establish autonomous discovery, independent peer review, prospective biological novelty, or live end-to-end LLM efficacy. The imported 100-task CompBioBench catalog is coverage metadata, not a performance claim.

What’s new in 1.0.1

Phase 5 hardening landed alongside the dashboard’s 13-stream feature push:

• Multi-source retrieval — scholarly-source adapters behind a domain-aware orchestrator with rate-limit backoff, ETag caching, and per-host circuit breakers.
• Citation validation — every reference is resolved against Crossref + Retraction Watch + arXiv before the paper finalizes. The run dir gets a validation_report.json with per-reference pass/fail.
• Claim → Evidence Matrix — the literature pass extracts atomic claims with quote spans and links them to source records. Persisted as evidence_matrix.jsonl per run.
• Reviewer-role revision loop — methodology / statistics / novelty / writing axes feed an aggregator that drives a bounded redraft loop. These roles currently use the drafting client and are self-critique, not independent peer review.
• Research-loop scaffold — plato loop --hours 8 --max-cost-usd 50 provides wall-clock/cost budgeting and git keep/discard checkpoints. The default adapters score existing artifacts; they do not yet execute a complete research cycle.
• Reproducibility manifest primitives — the manifest schema and recorder can capture git/project hashes, models, prompts, seeds, sources, tokens, and cost when supplied by the calling workflow; public-path coverage is not yet complete.
• Observability — opt in by setting LANGFUSE_* env vars; LangFuse callbacks are wired into every LangGraph invocation.
• Pluggable domains — DomainProfile registry exposes retrieval, keyword extractor, journal preset, executor, and novelty corpus as swap points. Astro is the default; biology ships out-of-the-box.
• Multi-tenant dashboard — set PLATO_DASHBOARD_AUTH_REQUIRED=1 and the dashboard reads X-Plato-User from the upstream proxy to scope every project, key store, and run artifact per tenant.

See docs/adr/ for the design decisions behind these changes and dashboard/CHANGELOG.md for the full list.

Resources

• Live Plato demo

• GitHub repository

• Prior Denario systems paper

Installation

To install plato create a virtual environment and pip install it. We recommend using Python 3.12:

python -m venv Plato_env
source Plato_env/bin/activate
pip install "plato[dashboard]"

Or alternatively install it with uv, initializing a project and installing it:

uv init
uv add plato[dashboard]

Then, run the Plato dashboard with:

plato dashboard

Get started

Initialize a Plato instance and describe the data and tools to be employed.

from plato import Plato

p = Plato(project_dir="project_dir")

prompt = """
Analyze the experimental data stored in data.csv using sklearn and pandas.
This data includes time-series measurements from a particle detector.
"""

p.set_data_description(prompt)

Generate a research idea from that data specification.

p.get_idea()

Generate the methodology required for working on that idea.

p.get_method()

With the methodology setup, perform the required computations and get the plots and results.

p.get_results()

Finally, generate a latex article with the results. You can specify the journal style, in this example we choose the APS (Physical Review Journals) style.

from plato import Journal

p.get_paper(journal=Journal.APS)

You can also manually provide any info as a string or markdown file in an intermediate step, using the set_idea, set_method or set_results methods. For instance, for providing a file with the methodology developed by the user:

p.set_method(path_to_the_method_file.md)

Plato Dashboard (new, recommended)

A Linear-themed real-time web dashboard with full pipeline visualization, cost tracking, and live agent log streaming. See dashboard/README.md for setup.

pip install "plato[dashboard]"
plato dashboard

The dashboard supersedes the legacy plato run Streamlit app for new workflows.

For hosted Railway SaaS/Lab deployments with Clerk auth or Clerk Billing, run the local production gates and redacted strict preflight before deploying:

bash dashboard/scripts/check-local-production-gates.sh
bash dashboard/scripts/check-hosted-saas-preflight.sh --railway --service plato --environment production --hosted-required --strict

It verifies the Clerk user/Lab auth contract, PLATO_BACKEND_PROXY_SECRET, public origin, Clerk proxy, and hosted billing flags without printing secret values. Strict mode treats preflight warnings as release blockers. After deploying, use the read-only production readiness check to probe the live app and scan Railway logs:

bash dashboard/scripts/check-production-readiness.sh --service plato --environment production --origin https://discovering.app

If railway variables --json/--kv is unavailable but you have a local Railway variables snapshot, pass it with --variables-file; the checker still prints only redacted key status and lengths:

bash dashboard/scripts/check-production-readiness.sh --service plato --environment production --origin https://discovering.app --variables-file /path/to/railway-variables.json

See dashboard/RAILWAY.md for the full production variable matrix.

Build from source

pip

You will need python 3.12 or higher installed. Clone Plato:

git clone https://github.com/Eldergenix/Plato-Scientific-Research-Autonomous-Agent.git
cd Plato-Scientific-Research-Autonomous-Agent

Create and activate a virtual environment

python3 -m venv Plato_env
source Plato_env/bin/activate

And install the project

pip install -e .

uv

You can also install the project using uv, just running:

uv sync

which will create the virtual environment and install the dependencies and project. Activate the virtual environment if needed with

source .venv/bin/activate

Docker

You can run Plato with Docker using the dashboard compose file:

docker compose -f dashboard/compose.yaml up --build

The local dashboard runs on http://localhost:7878 by default.

You can also build an image locally with

docker build -f docker/Dockerfile.dev -t plato_src .

Contributing

Pull requests are welcome! Feel free to open an issue for bugs, comments, questions and suggestions.

Citation

If you make use of Plato, please cite the following references:

@article{villaescusanavarro2025platoprojectdeepknowledge,
         title={The Plato project: Deep knowledge AI agents for scientific discovery}, 
         author={Francisco Villaescusa-Navarro and Boris Bolliet and Pablo Villanueva-Domingo and Adrian E. Bayer and Aidan Acquah and Chetana Amancharla and Almog Barzilay-Siegal and Pablo Bermejo and Camille Bilodeau and Pablo Cárdenas Ramírez and Miles Cranmer and Urbano L. França and ChangHoon Hahn and Yan-Fei Jiang and Raul Jimenez and Jun-Young Lee and Antonio Lerario and Osman Mamun and Thomas Meier and Anupam A. Ojha and Pavlos Protopapas and Shimanto Roy and David N. Spergel and Pedro Tarancón-Álvarez and Ujjwal Tiwari and Matteo Viel and Digvijay Wadekar and Chi Wang and Bonny Y. Wang and Licong Xu and Yossi Yovel and Shuwen Yue and Wen-Han Zhou and Qiyao Zhu and Jiajun Zou and Íñigo Zubeldia},
         year={2025},
         eprint={2510.26887},
         archivePrefix={arXiv},
         primaryClass={cs.AI},
         url={https://arxiv.org/abs/2510.26887},
}

@software{Plato_2025,
          author = {Pablo Villanueva-Domingo, Francisco Villaescusa-Navarro, Boris Bolliet},
          title = {Plato: Modular Multi-Agent System for Scientific Research Assistance},
          year = {2025},
          url = {https://github.com/Eldergenix/Plato-Scientific-Research-Autonomous-Agent},
          note = {Available at https://github.com/Eldergenix/Plato-Scientific-Research-Autonomous-Agent},
          version = {latest}
          }

@software{CMBAGENT_2025,
          author = {Boris Bolliet},
          title = {CMBAGENT: Open-Source Multi-Agent System for Science},
          year = {2025},
          url = {https://github.com/CMBAgents/cmbagent},
          note = {Available at https://github.com/CMBAgents/cmbagent},
          version = {latest}
          }

License

GNU GENERAL PUBLIC LICENSE (GPLv3)

Plato - Copyright © 2026 Pablo Villanueva-Domingo, Francisco Villaescusa-Navarro, Boris Bolliet