🧬 biology

Plato-Bio: verification-first biological novelty screening with temporal rediscovery and structural benchmarks

Plato-Bio is a verification-first biological research agent that integrates explicit workflow states and provenance tracking to ensure reproducible, auditable screening, as demonstrated by successful software validation and specific benchmarks in historical literature rediscovery and protein structure analysis.

Original authors: Stefan G. Creadore

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stefan G. Creadore

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Detective's Toolkit: Why "Smart" isn't Always "Right"

Imagine you are trying to solve a mystery, but instead of a magnifying glass, you have a super-smart robot assistant. This robot can read millions of books, write code, and even draft a detective story about the solution. But here's the tricky part: just because the robot writes a story that sounds perfect and flows beautifully, doesn't mean the mystery is actually solved. In the world of science, especially biology, getting the story to sound good is easy, getting the facts to match reality is incredibly hard.

Scientists have been building these "AI detectives" to help them find new cures or understand how our bodies work. The big idea is that these robots can connect dots between old research papers to find new answers. However, there's a danger: the robot might get so good at writing that it tricks us into thinking it found a new discovery when it actually just made things up or missed a crucial clue. This paper is about building a special "truth-checker" toolkit for these AI detectives. It's not about making the robot smarter at writing, it's about making sure the robot keeps a perfect, unbreakable diary of where it got every single fact, so we can check its homework.

The Paper's Mission: Building a "Truth-First" Biology Lab

The author of this paper, working on a project called Plato-Bio at Praxa Labs, an independent open-source research initiative, United States, decided to stop pretending that a fancy AI story equals scientific truth. Instead, he built a system that treats every claim like a suspect in a courtroom. In this courtroom, the AI can't just say, "I think this is true." It has to show its ID, its source, and its evidence. If it can't, the claim gets thrown out.

The author started by auditing his own AI system and found three sneaky bugs that were messing up the results. It was like finding out his detective was accidentally using an astronomy textbook to solve a biology case, or that it was forgetting to write down the "why" behind its answers. He fixed these bugs and created a strict set of rules, a "software contract," that the AI must follow. He tested this new, stricter system with two very specific challenges to see if it could actually do the job without hallucinating.

Challenge 1: The Time-Travel Test

First, he tried a "temporal rediscovery" task. Imagine you are a detective in 1989, and you want to know if fish oil helps with a condition called Raynaud's phenomenon. You can only look at books published before 1986.

The AI had to figure out the connection using only old clues. A simple search for "fish oil" and "Raynd's" wouldn't work because no one had written those two words together yet. In this single, manually curated retrospective task, the AI used an "evidence bridge" method to connect the dots. The results showed that the bridge-only and evidence-aware ranking placed the held-out relation, fish oil to Raynaud's, first, followed by TF-IDF second, and corpus frequency third. This showed how the system could rank connections by building logical paths rather than just relying on popular words.

Challenge 2: The 3D Shape Match

Next, the author tested the system's ability to compare existing AlphaFold predictions against real, experimentally measured shapes for 15 different human proteins.

The results were a mix of high agreement and areas needing more work, which is exactly what the author wanted to show:

Higher-agreement targets: For 11 out of 15 proteins, the high-confidence-core C-alpha RMSD was below 1 Ångström. The median across all 15 targets was 0.501 Ångström. For some proteins like hemoglobin, the difference was as small as 0.270 Å.
Targets retaining larger discrepancies: For 4 proteins, the differences were larger. One protein, called SUMO1, had a large difference in the full prediction of 16.61 Å, but when the author focused only on the parts the system was confident about, the error dropped to 2.58 Å.

The key takeaway here is that the system didn't just say, "It's different, so it's a new discovery!" Instead, it flagged the differences as "unvalidated hypotheses." It pointed out exactly where the shapes didn't match and said, "Check this out, but don't celebrate yet." It found 27 specific regions where the prediction and reality disagreed, but it labeled them as things to investigate, not as proven new facts.

What This Means (and What It Doesn't)

The author is very careful not to overhype his results. He isn't saying, "I built a robot that discovered a new cure!" He is saying, "I built a robot that knows how to keep a perfect, auditable diary of its work."

The paper explicitly rules out the idea that a polished, well-written report from an AI means the science is correct. The author argues that without these strict "proof-of-work" steps, like checking citations, linking claims to evidence, and marking uncertain areas as "not established," we can't trust the AI's discoveries.

In the end, Plato-Bio is a tool for reproducibility. It ensures that if you run the same test again, you get the same results, and you can see exactly how the AI got there. It found that for historical puzzles, connecting evidence bridges works better than simple word searches. It found that for protein shapes, the system is great at identifying the core parts of existing predictions but highlights where they deviate from experimental reality.

The author concludes that while this system is a solid foundation, it's not a magic wand. To truly claim a new biological discovery, you still need real-world experiments, human experts to review the work, and a lot more testing. But with Plato-Bio, we finally have a way to make sure the AI is doing its homework correctly before we even start grading the test.

Author: Stefan G. Creadore, Praxa Labs (an independent open-source research initiative, United States).
Paper: https://arxiv.org/abs/2607.23975
Code and data: https://github.com/Eldergenix/Plato-Scientific-Research-Autonomous-Agent
ORCID: https://orcid.org/0000-0003-2268-053X

Technical Summary: Plato-Bio

Problem Statement
The paper addresses a critical gap in the emerging field of scientific research agents: the conflation of high-quality, coherent prose with scientific validity. While large language model (LLM) agents can automate literature retrieval, code generation, and manuscript drafting, polished output does not guarantee evidentiary accuracy. In computational biology, where analyses depend on versioned data, coordinate conventions, and strict data-use conditions, agents must preserve the relationship between claims, sources, and executable artifacts. The author argues that current systems risk mistaking rhetorical quality for scientific truth, necessitating a "verification-first" approach that prioritizes auditable workflows over autonomous discovery claims.

Methodology
The author developed Plato-Bio, a biology-routed extension of the open-source Plato/Denario architecture, developed at Praxa Labs, an independent open-source research initiative, United States. The system is designed not as a monolithic prompt but as explicit LangGraph state machines that separate the research process into distinct, inspectable stages: idea generation, method development, execution, and manuscript drafting.

Key methodological components include:

Explicit Domain Profiling: A biology-specific profile routes tasks to appropriate sources (e.g., PubMed, Europe PMC, OpenAlex) and enforces MeSH-based keyword extraction, distinct from the default astronomy profile used in prior versions.
Provenance and Control Mechanisms: The system implements "publication gates" requiring references, citation-validation reports, and claim-to-evidence matrices. It distinguishes between drafted claims and source-derived claims, classifying relationships as supporting, refuting, neutral, or unclear.
Measurement-Validity Repairs: A source-level audit identified and fixed three critical defects that previously distorted evaluation:
1. The default factory incorrectly applied an astronomy profile to biology tasks.
2. The evaluator failed to score declared method signals.
3. Evidence sidecars lacked drafted claims, causing the "unsupported-claim rate" to default falsely to 0.0.
Deterministic Validation: The system was tested via a full Python suite and targeted suites (biology domain, genomics adapters, evidence/citation, adversarial safety) without relying on live LLM calls or optional cloud services for the core validation.

Key Contributions
The paper makes five specific contributions, framed as software contracts and reproducible baselines rather than biological discoveries:

Biology-Routed Workflow: Implementation of an explicit domain profile for literature sources, keyword extraction, and execution defaults.
Inspectable Artifacts: Documentation of safety controls including citation reports, claim/evidence matrices, and run manifests that record hashes, timestamps, and model parameters.
Regression-Tested Repairs: Correction of the three measurement defects listed above, verified by failing regression tests prior to implementation.
Temporal Rediscovery Benchmark: A frozen historical task (pre-1986 literature) designed to rank candidate relations (e.g., fish oil and Raynaud phenomenon) based on evidence bridges (A–B and B–C paths) rather than direct prior art.
Structural-Biology Case Study: A deterministic screen comparing AlphaFold models against experimental structures for 15 human proteins, utilizing confidence masking (pLDDT ≥ 70) to isolate high-confidence core discrepancies.

Results

Software Validation: The full Python suite completed with 931 passes, 6 skips, and 0 failures. The targeted biology, genomics, evidence, and safety suites also passed without failure. Skips were attributed to missing optional credentials or SDKs, not logic errors.
Temporal Rediscovery: In a single curated historical task, the evidence-aware and bridge-only conditions correctly ranked the fish oil/Raynaud phenomenon relation first (MRR 1.000). In contrast, corpus frequency ranked it third, and TF-IDF ranked it second. This demonstrates the utility of evidence bridging over simple lexical relevance for retrospective ranking.
Structural Agreement (Globins): For three globin targets (Hemoglobin $\alpha$ , Hemoglobin $\beta$ , Myoglobin), AlphaFold models achieved sub-Ångström C $\alpha$ RMSD (0.270–0.520 ˚A) against experimental structures, with 99.3% of residues within 2 ˚A.
Diverse Structural Screen: Across 15 human proteins, 11 targets had high-confidence-core RMSD below 1 ˚A, and the median high-confidence-core C-alpha RMSD was 0.501 Angstrom. Four targets (KRAS, SUMO1, TP53, Estrogen Receptor $\alpha$ ) exceeded 2 ˚A. The system identified 27 traceable discrepancy regions but explicitly labeled them as "unvalidated hypotheses" rather than discoveries. For example, the high discrepancy in SUMO1 was reduced from 16.61 ˚A to 2.58 ˚A when restricted to high-confidence residues, illustrating the importance of confidence masking.

Significance and Claims
The paper maintains a modest scope, explicitly stating that Plato-Bio provides reproducible software contracts and auditable screening baselines, but does not infer autonomous scientific validity from its results.

Reproducibility: The system successfully implements a workflow where claims, sources, and calculations are traceable and versioned.
Retrospective Utility: The temporal pilot suggests that evidence-bridge ranking is superior to frequency or TF-IDF for reconstructing historical relations, though this is a single-task pilot.
Structural Triage: The structural screen demonstrates that confidence masking and context stratification are necessary to distinguish genuine structural discrepancies from artifacts caused by flexible termini or partial constructs.
Limitations: The author emphasizes that broader claims of agent efficacy or biological novelty require preregistered evaluation, independent review, and prospective validation. The current work does not claim to have discovered new biology; rather, it establishes a framework for how such discoveries could be audited. The "reviewer roles" in the system are self-critique mechanisms using the same model, not independent human or external AI review.

In summary, Plato-Bio is presented as a rigorous infrastructure for verifying the scientific workflow, ensuring that the output of research agents is grounded in explicit provenance and that any "novelty" is treated as a hypothesis requiring independent validation.