The Virtual Brain

From a person's variants to the state of their brain. Two models run on the same genome — sPACE-AD for biological pathways, Decima for cell-type expression — and their outputs are combined to give cell states and a brain state, one person at a time.

Request a pilot See the demo

9,796

participants in the reference cohort

184

pathways, the same set for every person

brain regions in the expression reference

1,119,570

variant weights behind each profile

Demo

An individual virtual brain.

One person's genome, read through to cell states, brain regions and cortical depth. sPACE-AD and Decima are computed independently from the same variants and combined at the cell-state stage — the cell states are not derived from the pathways. Select a pathway line, or a table row, to follow it forward. The people shown here are synthetic and exist to demonstrate the readout. Synthetic example

Participant

Order within category

Pathways

	Pathway	Score	z

Cortical layer means

Translucent left hemisphere showing cortical parcellation and subcortical structures. — Left hemisphere, lateral view, on the conte69 surface at 500 parcels per hemisphere (Schaefer atlas). Each parcel is named by the Desikan-Killiany region its vertices fall in. Colour is an identifier only — hues are assigned by shuffled index, not position, so the map carries no spatial gradient and implies nothing about where the three pathway categories act.

Regional pattern for this participant

This participant's profile projected onto cortical parcels — strongest 40 of 396 shown.

Illustrative values. The surface is displayed at parcel resolution; the imaging arm's volumetric analysis is run at Desikan-Killiany region level. In the reference cohort it returned a well-powered null across three cohorts — see Status. Nothing here is an established regional association.

Pathway category against rate of decline, by cognitive domain

Cohort-level enrichment across contributing studies. Does not vary by participant.

Counts are pathways whose association with the rate of decline passes an uncorrected threshold, against 9.2 expected by chance out of 184. Enrichment is the design: no multiplicity correction is applied, and the count — not the signed mean — is the readout.

Cohorts contributing

Domain scores are harmonised composites. The memory result is the one with signal in the reference cohort and is currently being rerun against the corrected cohort definition.

Approved mechanisms and the pathways they act on

This is a mechanism map, not a treatment recommendation. It shows which pathways in the dictionary overlap the target biology of each approved drug class, and this participant's average score on those pathways. It has not been validated as a predictor of response, and must not be used to select therapy for a patient. A response analysis in an independent cohort is listed in Status as exploratory.

Why this matters

Now here are two of them.

Both have the same diagnosis. Their inherited risk runs through different biology, and acts on different cells. A trial that enrols on diagnosis treats them as the same patient.

A risk score cannot separate them

A polygenic score gives each patient one number. It ranks them from lower to higher risk — and says nothing about which biology is driving it.

Treatments act on mechanisms

A drug aimed at a glial mechanism is aimed at the first patient, not the second. Nothing in the enrolment tells the two apart, so both are recruited.

The effect averages away

A real benefit in the patients a drug fits is diluted by everyone it does not. The trial reads as negative, and the mechanism is abandoned.

Applications

Where knowing the mechanism changes the decision.

Trial enrichment

Group a screening cohort by the pathways a programme targets, not by diagnosis and APOE alone. The enrichment rule is written down before the trial opens.

Target prioritisation

Find which pathways separate cases from controls, and which cell types they act through, before committing to a mechanism.

Progression and response

Pair profiles with longitudinal cognitive, imaging, or treatment data to test whether a subgroup declines or responds differently.

Platform

What goes in, and what comes back.

Input	Genotype array or whole-genome sequence, GRCh38
Variant weights	PRS-CS-auto posterior weights, 1,119,570 variants
Cell-type model	Decima, a foundation model that predicts expression per cell type. Run independently of the pathway model on the same variants
Combination	Cell states are defined by combining Decima cell-type output with sPACE-AD pathway scores. Neither is derived from the other
Expression reference	GTEx v8 brain tissue eQTL panels, thirteen regions
Pathway set	184 Reactome Level-2 pathways, fixed across all participants — nothing refitted per cohort
Category split	44 disease-differentiating, 66 orthogonal, 74 partial; checked at load, and a mismatch stops the run
Excluded region	chr19:44.4–46.4 Mb (APOE), removed before scoring and confirmed empty afterwards
Scoring	Alignment between a participant's centred dosage vector and each pathway's weights, independent of overall magnitude
Regions scored today	Cortex. The rest sit in the reference and are reported unscored, never estimated
Output	184 pathway scores, cell-state loadings per scored region, and in cortex a profile across depth
Reference cohort	ADSP Release 5, ancestry-restricted and principal-component-cleaned, 9,796 AD and control participants

Mid-sagittal T1 slice of the MNI152 template showing cortex, corpus callosum, brainstem and cerebellum. — Mid-sagittal MNI152 template. The expression references are sampled across thirteen regions including cortex, hippocampus, amygdala, striatum, hypothalamus, substantia nigra and cerebellum.

Status

Including the analyses that came back negative.

Every number in the demo names its source. Anything without one is shown as unavailable rather than estimated, and work being rerun against a corrected cohort definition is marked rather than quoted as final.

Component	Detail	State
Cohort construction	ADSP Release 5, 15,730 participants, restricted and trimmed to a fixed analysis set of 9,796.	Established
APOE independence	Emptiness test across chr19:44.4–46.4 Mb confirms no variants remain after exclusion.	Established
Region set	Thirteen GTEx v8 brain tissues. Cortex scored; the rest reported unscored.	Established
Memory decline	Whether Category I pathways are over-represented among those associated with memory decline. Enrichment seen, low heterogeneity between cohorts.	Being rerun
Structural imaging	Pathway scores against regional cortical volume change, three cohorts. Negative — well powered, reported as a null.	Being rerun
Pathway score matrix	Per-participant scores regenerating against the corrected cohort definition.	Being rerun
Treatment response	Pathway profiles against cholinesterase-inhibitor and memantine response in an independent cohort. Exploratory; no validated predictor exists, and the mechanism map in the demo is not one.	Exploratory
Cortical region atlas	Volumetric analysis at Desikan-Killiany region level. Surface displayed at 500 parcels per hemisphere (Schaefer atlas) on conte69, each parcel named by the Desikan-Killiany region containing it.	Established
Cell-state definition	How Decima output and sPACE-AD pathway scores are combined. Decima track selection and the aggregation operator are not yet settled, so the demo's cell-state column is illustrative only.	In development

Working together

Three ways to begin.

Scoped pilot

We score a cohort you supply and return the profiles, the analysis, and the provenance record. Fixed scope, fixed deliverable.

Programme licence

The scoring pipeline and the fixed pathway set for a named indication, with version-pinned references so results stay reproducible.

Research collaboration

Joint work on a new tissue, indication, or outcome, with authorship and intellectual property agreed at the start.

Tell us about the cohort and the question.

We will come back with what the platform can and cannot answer for it.

Get in touch