Model card

VesTRA z0.1 (Beta)

A foundation AI model built to understand extracellular vesicle biology.

z0.1 is the first public version. It is based on our initial BioTranslate framework, and we are currently building the VesTRA AI model.

Version: z0.1 (Zen 0.1)
Status: Beta
Data release: 0.1.0
Released: 10 October 2026

01At a glance

Built on: The BioTranslate framework (University of Kansas)
What it is today: A curated base of published EV evidence, scored by fixed, published rules. z0.1 contains no trained neural network yet.
Input: A gene, protein, UniProt accession or miRNA
Output: One page per molecule: evidence records, tissue and blood findings, seven-stage evidence reach, an assessment, reference biology and PDF export
Coverage: 151,973 molecules and 1,517,212 evidence records in data release 0.1.0
Use: Research only

02What z0.1 does

Finds a molecule from any common name or identifier and gathers all its EV evidence on one page.
Shows where it was measured: tissue, tissue EVs, and EVs from blood and other fluids.
Places the evidence on seven stages, from normal baseline to therapeutic convergence.
Gives a class and a next step for proteins with matched tissue and plasma data.
Links every record to the source file, table and cell.

03What it does not do

Make predictions. Every answer comes from published evidence and fixed rules.
Say anything about a molecule no included study has measured.
Act as a diagnostic test, or guide decisions about patients.
Judge study quality beyond the checks in the methodology.

04How the map is drawn

Each point on the homepage map is one molecule. In z0.1, molecules sit close together when they were measured in the same kinds of samples. Colour shows where a molecule has been found: tissue, blood, both, or other body fluids. Brighter points have more evidence.

The z0.1 map is computed from the evidence. It is not drawn by a trained model. From z1, the trained VesTRA AI model will place the molecules itself.

05The seven stages

A stage is reached when at least one evidence record meets its rule. Stages describe how far the evidence reaches, not how good a biomarker is.

01Normal baselineMeasured in a healthy or normal sample, or against a normal comparator
02Precancer perturbationMeasured in a precancerous context or against one
03Tissue discoveryDetected or quantified in tumour tissue or tissue-derived EVs
04Plasma translationDetected or quantified in EVs from blood (plasma or serum)
05Cross-platform validationSupported on more than one measurement platform
06Clinical thresholdReported with a classification performance or a clinical cut-off
07Therapeutic convergenceGiven a registered therapeutic verdict in the framework

06Classes and next steps

Given to 977 proteins with a matched tissue comparison and plasma comparison. Other molecules read “Not classified”.

Class	Meaning	Next step
Concordant	Tissue and plasma directions agree	Prioritize
Attenuated	Same direction, but weaker in blood	Recover or enrich
Lost signal	Detected in plasma, but no clear difference there	Recover or enrich
Lost detection	Seen in tissue, tested in plasma, not detected	Recover or enrich
Plasma emergent	Signal in blood EVs without a matching tissue call	Investigate first
Discordant	Opposite directions in tissue and plasma	Investigate first
Both not significant	No clear signal in either	Deprioritize

07Data and known limits

Source: published, peer-reviewed EV studies and their supplementary data. Data release 0.1.0 holds 151,973 molecules and 1,517,212 evidence records.
Each release is versioned, and the number appears on every molecule page and PDF. New data releases come out as new studies are added.
Coverage follows the studies included so far. Some cancers and body fluids are covered far better than others.
A stage not reached means no included study meets its rule. It does not mean the biology is absent.
Records keep each study's own units and statistics, so values across studies are not directly comparable.
Names are mapped to molecules automatically. Rare or ambiguous names can be mapped wrongly; please report them.

08Versions

z0.1NOWPublic search on the BioTranslate framework, with the evidence map, stages and assessment for every molecule.
z0.2NEXTAdditional features and a large-scale molecular catalogue.
z1PLANNEDThe first trained VesTRA AI model. It draws the map itself and is released with its own model card and evaluation.

09Licence and citation

How to cite: Bhanu P, Mahoney DE, Godwin AK. VesTRA z0.1 (Beta): a foundation AI model for extracellular vesicle biology. Data release 0.1.0. Godwin Lab, University of Kansas Medical Center; 2026. Available at: https://vestra.vumens.bio
Contact: vumens.bio@gmail.com