Biodius Scientific Portfolio: Preprint, Case Pages and Editorial Analysis

TL;DR: The public Biodius evidence in one place: a technical preprint (not peer reviewed), two case pages, three editorials, the offer, the FAQ and the limits.

This page is one path through the public scientific material behind the Biodius Target Challenge, and it says what each item is. There is one technical preprint, two case pages drawn from it, three editorial articles, the Biodius Target Challenge offer and its FAQ. The limits come before the offer.

Why this matters: a biotech team or investor deciding whether to trust this work needs to separate our own preprint from commentary and from a commercial offer.

How to read the labels

• Technical preprint. Our own report, deposited on Zenodo. It is not peer reviewed and it is not a journal paper.
• Technical report, web edition. The same work on this site. Where it and the preprint differ, the preprint is the reference.
• Case page. An account of one experiment from the preprint, taken from our own prospectively frozen analysis of public data.
• Editorial analysis. Articles by First AI Movers that explain a method or a question. They are not peer-reviewed publications. Where they use an example, they say it is hypothetical.
• Offer and FAQ. A description of a service. It is not evidence for the method.

The technical preprint (version 1.1)

Falsification-First AI for Biological Target Assessment: Lessons from Four Prospectively Frozen Experiments. Costa, H. (2026). Preprint v1.1, 1 October 2026, on Zenodo. Open access under CC BY 4.0. The preprint's own status line reads "Technical Report / Preprint. Not peer reviewed."

• Cite all versions with the concept DOI: doi.org/10.5281/zenodo.23057911. This is the identifier our other pages use.
• This version has its own DOI: doi.org/10.5281/zenodo.23100159. The record page is zenodo.org/records/23100159.

What it reports

Four experiments, each bounded in advance, using public data and no new laboratory work:

1. A strong agentic research baseline recovered known biology better than a heavier architecture and matched it on four frozen discipline dimensions. One evaluator judged one domain, and a third arm was not run.
2. A later, smaller safeguard layer preserved its specified safeguards in one fresh-domain transfer. That is one transfer.
3. An ulcerative colitis causal-chain study ended as not evaluable on available public data.
4. A TTPAL challenge: one gut-resident T-cell discovery resource supported TTPAL in 3 of 66 designated groups, with PP4 up to about 0.992. The primary independent blood cohort showed no usable signal, with PP4 at most 0.031 across the six measured blood T-cell classes.

Downloadable evidence on Zenodo

• Preprint v1.1 as PDF (102 KB).
• Preprint v1.1 as Markdown (40 KB).
• Reproducibility supplement, ZIP (406 KB). The preprint describes it as containing the manuscript source, figure sources, claim ledger, public-source register, source-digest lock, typesetting build and validation programs.

What is not in the package: the frozen result records and the programme code that produced them are not distributed with the preprint. Non-public working material and third-party summary statistics are not redistributed either.

The same work is also on this site as a technical report: Falsification-First AI for Biological Target Assessment (technical report; not peer reviewed).

Two case pages from the preprint

Ulcerative colitis causal chain (H3): not evaluable

A biologically coherent ulcerative colitis hypothesis survived an initial evidence bar and two adversarial AI reviewers. A frozen human-genetics adjudication then found that the decisive causal link could not be evaluated from lawful public data. The hypothesis was not promoted, and it was not disproven. Read Not Evaluable Is a Result.

Limit: the result describes what available public data can decide. It does not say the causal chain is false.

TTPAL: a striking signal that did not replicate

An unexpected T-cell signal reached PP4 up to about 0.992 in three gut-resident groups in a discovery dataset. The three groups overlap within one donor lineage, so they are not three independent replications. It did not reproduce in independent public blood T-cell cohorts. Read 0.992 Was Not Replication.

Limit: the discovery cells were gut-resident and the independent cohorts were blood. The result does not prove a tissue-resident effect false. It shows the signal was single-dataset-only. PP4 is a model score, not the probability that TTPAL causes disease.

Editorial analysis: three articles

These are editorial analysis written for decision-makers. They are not peer-reviewed publications and they present no new experimental results.

• How Much Evidence Is Enough Before a Biotech Decision? How to choose a review scope around the decision, with a map of search coverage, study independence and remaining uncertainty. It states that a focused challenge is not a systematic review.
• Before You Trust an AI Scientific Review, Read Its Error Report Six error measures, an independent reference standard and an account of what remains untested. It proposes an evaluation design and reports no new benchmark, independent certification or clinical validation of Biodius.
• Four Papers, One Trial: How to Count Independent Evidence How to map reports to underlying studies before judging biological evidence, with a hypothetical trial example and a source checklist.

For why we built the service, see Why We Built a Service That Tries to Kill Your Biological Target.

The offer and the FAQ

The Biodius Target Challenge is a fixed-scope adversarial review of one target, pathway or mechanism in one indication. It ends in a decision memo: ADVANCE | REWORK | STOP | NOT_EVALUABLE. Scope, deliverables and founding-pilot terms are on that page, and the Target Challenge FAQ answers the questions teams ask before sending a target.

Limitations

These come from the preprint and from the pages above.

• The preprint is not peer reviewed. It is our own report and has not been independently replicated. The method has not been externally validated.
• Small scope. Four experiments in two disease areas, using public data only. The results support a narrow claim about workflow design, not a general claim about AI in biology.
• No laboratory work. Nothing was tested experimentally.
• Tissue mismatch. In the TTPAL challenge, the independent cohorts were peripheral blood and the discovery signal came from gut-resident cells.
• Narrow architecture comparison. One evaluator, one domain, and a third arm that was not run. It does not rank all research architectures.
• AI reviewers are not experts. They generate challenges that were then adjudicated against primary sources. They do not replace qualified scientific judgment.
• Interest declared. The preprint states that its author is affiliated with Hernani Holding B.V., which owns the Biodius research software compared as the heavier architecture in the first experiment.
• Editorial pieces are not research. The three articles are editorial analysis. They use hypothetical examples where stated.
• No clinical or therapeutic claim. Nothing here establishes a therapeutic target, a clinical mechanism, a diagnostic, or the safety or efficacy of any intervention. Biodius does not provide clinical advice, validate safety or efficacy, perform wet-lab work, or replace legal/patent/regulatory review.
• No discovery or prediction claim. Biodius does not discover drug targets or predict clinical success. It is experimental scientific diligence tooling, and our current commercial hypothesis is that it is useful before a costly downstream decision.

Frequently asked questions

Has the Biodius preprint been through peer review?

No. It is a technical preprint on Zenodo, and its own status line says it is not peer reviewed. The editorial articles are not peer-reviewed publications either.

Which identifier should I cite?

Cite the concept DOI, 10.5281/zenodo.23057911, to refer to the work across versions. Cite 10.5281/zenodo.23100159 to refer to version 1.1 specifically.

Where can I download the evidence?

On the Zenodo record: the PDF, the Markdown source and a reproducibility supplement. The frozen result records and programme code are not part of that package.

Does this show that Biodius works?

No. It shows what four of our own bounded experiments on public data did and did not establish. It is not external validation.

Start with one target

Send us one target + indication + rationale. We will tell you whether it deserves the next expensive step.

Send us one target + indication + rationale: tell us the target, pathway or mechanism; the indication; your current rationale; and the next decision you are considering.

Prefer to talk first? Ask for a 20-minute fit call at info@firstaimovers.com. Learn more about First AI Movers.

Our source for the scientific claims is Costa (2026), Falsification-First AI for Biological Target Assessment: Lessons from Four Prospectively Frozen Experiments, DOI 10.5281/zenodo.23057911. This preprint is not peer reviewed and has not been externally validated; it establishes no diagnostic, therapeutic, safety or predictive claim.

AI assistance was used for drafting, source checks and editing. Human editorial responsibility remains with First AI Movers.