This policy applies to every Vemryx Life Sciences project. It exists so that any claim we make can be checked by someone outside the company, and so that nobody, ourselves included, mistakes a hypothesis for a discovery.
1. Three things that are not the same
- An AI-generated hypothesis. A possible disease mechanism, therapeutic target or research question. It is a starting point.
- Computational evidence. The result of a data analysis, with quantified uncertainty, that another person can redo.
- Experimental evidence. Verification of the hypothesis in biological systems and, where appropriate and authorized, in humans.
A molecule suggested by a model is not a treatment. No document, presentation or page of ours uses the words “cure”, “treatment” or “effective” to describe something that lacks the corresponding experimental evidence.
2. The evidence ladder
Every project states which step it is on, in its record and in any public communication.
| Step | Name | What it takes to get here | Who carries it out |
|---|---|---|---|
| 1 | Registered hypothesis | Question, hypothesis and analysis plan written and dated before looking at outcome data | Vemryx |
| 2 | Reproducible analysis | Code, data and environment under version control; a second person redoes the result from the repository | Vemryx |
| 3 | Independent replication | The finding holds in a dataset that was not used to produce it | Vemryx |
| 4 | Expert review | A researcher in the field, who took no part in the analysis, assesses methods and conclusions | Vemryx, with an external reviewer |
| 5 | Experimental validation | Laboratory testing, in appropriate facilities and under appropriate supervision | Partner institutions |
| 6 | Clinical trial | A study in humans with a sponsor and ethics and regulatory approval | Partner institutions |
A project does not skip steps, and it can move down: if replication fails, the project returns to step 2 and the negative result goes on record.
3. What every project keeps
- The scientific question and hypothesis, registered before the analysis.
- The origin, version and license of each dataset.
- The code and environment needed to reproduce the result.
- Statistical tests, metrics and negative results.
- Limitations and a specialist’s assessment.
The record template is at docs/project-record-template.md in the division’s repository.
4. Analysis rules
- Plan before data. Outcomes, covariates and tests are set in the record. Later changes are allowed, as long as they are dated and justified.
- Association is not causation. Every report classifies each claim as an association, a causal inference or a speculative hypothesis.
- Multiple comparisons. Analyses with many tests (for example, thousands of genes) apply a correction and state the method used.
- Confounders. The adjustment variables are listed, along with those that could not be controlled.
- Effect size and uncertainty. We report confidence intervals, not p-values alone.
- Reproduce first. In a new research line, the first piece of work is reproducing a result that has already been published.
5. Use of AI models
Where we use them
- Searching and synthesizing the literature, and generating testable hypotheses.
- Writing and reviewing analysis code.
- Cross-checking statistics, biases and contrary evidence.
- Supporting the drafting of reports.
Rules
- References are checked one by one. No citation enters a document until a person has opened the source by its DOI or PMID. Models invent references.
- AI review is not independent review. Two models can make the same mistake and agree with each other. A second model’s review is useful, but it is never the last one.
- Mandatory human decisions. A named person chooses the hypothesis, interprets the findings and signs off on the conclusions. The project record keeps who decided and when.
- Audit trail. Prompts and responses that influenced an analysis decision are stored with the project, along with the model’s name and version.
- No clinical use. No model or result of ours is offered for diagnosis, prognosis or treatment decisions without the required validation and authorization.
6. Data, privacy and licenses
- We use public data and respect the license and restrictions of each database. Controlled-access data is used only with the database’s authorization.
- Identifiable medical data is not sent to AI tools. Any project involving sensitive personal data requires, before it starts, a legal basis, a contract and security measures that meet the applicable data protection law (in Brazil, the LGPD).
- We do not attempt to re-identify participants in any dataset.
- Questions and requests about personal data go to privacy@vemryx.com.
7. What we do not do alone
| Type of research | Vemryx’s position |
|---|---|
| Literature review and analysis | Done in-house |
| Computational studies with licensed public data | Done in-house |
| Development and evaluation of AI models | Done in-house, with no unvalidated clinical use |
| Experiments on cells and tissues | Only with appropriate facilities and supervision, at partner institutions |
| Studies with human participants, including psychological studies | Only with prior ethics approval under the law of the country where the study takes place (in Brazil, Law 14,874/2024 and its regulations) |
| Clinical trials of drugs and vaccines | Only with a sponsor, a qualified team and the approvals required by the relevant regulator (in Brazil, Anvisa) |
In psychology and neuroscience studies with people, consent, privacy and risk assessment carry the same weight as the statistical model.
8. Publication and authorship
- We publish methods, code, limitations and results, including negative ones, subject to the rights and restrictions of the data.
- AI tools are not authors. The use of AI in analysis and writing is disclosed in every manuscript, in line with the ICMJE recommendations.
- Authorship requires intellectual contribution and responsibility for the content. Whoever signs answers for the work.
- Conflicts of interest are declared, including Vemryx’s commercial interests.
9. Errors and corrections
When we find an error in something we have already released, we correct it in public, in the same place the error appeared, with the date and a description of what changed.
If you find an error in our work, write to research@vemryx.com. We answer every report, including the ones we end up disagreeing with.
10. Review of this policy
This is version 0.1. It should be assessed by a researcher with experience in biomedicine or bioinformatics before the first technical report, and revisited whenever the division takes on a new type of study.