Research

We build models of human biology. Our long-term goal is a Digital Twin that predicts individual responses.

01 · Bulk RNA → single cells

PBISC

Generate single-cell RNA profiles from one bulk blood sample.

Question

Can widely collected blood bulk RNA provide useful single-cell RNA profiles?

Approach

A diffusion model generates cell-level counts from bulk PBMC RNA-seq, without matched single-cell data at inference.

Status

Manuscript under review at BIBM.

02 · RNA → one token

Let a language model read a whole transcriptome as one token.

Question

Can a language model reason over RNA without reading thousands of gene values as text?

Approach

A learned bridge compresses a single-cell, pseudobulk or bulk profile into one token the model reads.

Status

Manuscript under review at BIBM.

03 · Case reports → diagnostic environments

Turn case reports into step-by-step diagnostic environments.

Workflow

Case PDF → Environment config → Sequential diagnosis → Reward and review.

Actions

Ask · Examine · Order tests · Diagnose.

Status

Early prototype: PDF converter, diagnostic environment, pilot rollouts and local review UI.

Next: multi-case review, reward calibration and an RL trainer adapter.

Models are evaluated on donors and studies excluded from training.

We strictly state each model's limitations.

Data sources, methods and dated corrections stay on record. Research practice

None of this is a clinical tool. VHL does not predict individual treatment outcomes or give medical advice.