The platform

From a disease indication to a ranked list of cell-type-grounded target hypotheses

Input a disease indication and tissue context. Receive a ranked list of cell populations driving the phenotype and, for each, a set of target gene candidates scored by expression specificity. Built on a foundation model trained on millions of single-cell profiles across human tissues and disease states.

Architecture

How the platform is built

The platform is a four-layer system. Each layer addresses a distinct problem in single-cell-based target identification, from raw data harmonization to structured hypothesis output.

Data ingestion layer
Ingests and harmonizes scRNA-seq data from public repositories (Human Cell Atlas, CELLxGENE, GEO). Handles batch correction, gene vocabulary normalization, and cell-type annotation transfer at scale.
Foundation model encoder
Transformer-based cell encoder trained on the curated corpus. Each cell becomes a vector embedding capturing its gene expression signature in biological context, enabling cross-dataset generalization.
Cell-population retrieval
For a given disease query, the retrieval layer identifies cell populations in the embedding space that are associated with the disease phenotype, drawing on cell state variation across donors and disease conditions.
Target ranking engine
Scores gene candidates by expression specificity in the identified cell populations across tissues, producing a ranked list structured for bioinformatics and discovery teams.
Platform architecture: single-cell cell landscape
Data pipeline

From raw scRNA-seq to harmonized embeddings

The data pipeline addresses three practical obstacles in building robust models from multi-source single-cell data: preprocessing variability, batch effects, and gene vocabulary differences across datasets.

Data pipeline flow diagram
scRNA-seq preprocessing

Quality filtering, normalization (log1p counts), and gene selection per dataset. Standardized across all ingested sources regardless of the original protocol or sequencing depth.

Batch harmonization

Integration of data from different labs, donors, and platforms using harmonization methods that remove technical variation while preserving biological signal. A critical step given the diversity of the training corpus.

Embedding generation

Each cell is encoded to a dense vector by the foundation model encoder. Embeddings are indexed for retrieval and periodically updated as the training corpus expands with new public data.

Output format

What discovery teams receive

The platform produces structured output designed for direct use by bioinformatics and discovery teams. Three components per query.

Cell-population rankings

A ranked list of cell populations associated with the disease phenotype in the query tissue. Each population includes the cell-type annotation, a confidence score based on expression pattern overlap with disease signal, and cross-tissue prevalence data.

Target gene candidates per population

For each ranked cell population, a list of gene candidates with expression specificity scores. Specificity is computed across cell types and tissues, surfacing genes that are preferentially expressed in the disease-associated population and not broadly active across unrelated cell types.

Supporting evidence layer

For each target candidate, the output includes links to the source datasets where the expression evidence was observed, enabling discovery teams to trace each hypothesis to the underlying single-cell observations. No black box: every rank has an interpretable data anchor.

Access

API access and research portal

The platform is available via API for integration into bioinformatics pipelines, and through a research portal UI for discovery teams who prefer an interactive query interface. Both access modes return the same structured output.

We are currently accepting early access requests from drug discovery teams and research collaborators. Access is reviewed by the team and granted on a rolling basis.

Request Access