Measured biology
Molecular, cellular, clinical, behavioral, and environmental observations collected with time and context.
Research / 2026
From biological state to modeled response.
VHL’s long-term research program is to condition computational models on measurements from an individual and estimate how that biological state may change under a defined perturbation. Any such estimate would need quantified uncertainty and validation against observed outcomes.
Current work begins with RNA: recovering cell-state structure, representing transcriptomes, and building leakage-controlled benchmarks. These projects study components relevant to the long-term aim. They do not yet constitute a virtual-human model or a clinical decision tool.
Research model
A defined biological state. A defined perturbation. A testable response.
A virtual human is not a complete digital replica. It is a task-specific, updateable computational representation of selected aspects of human biology, conditioned on measurements from an individual and evaluated for a defined question.
Individual state
A computational representation that separates what was measured from what the model inferred or compressed.
Defined change
A medicine, nutrient, behavior, exposure, or disease process specified with dose, timing, and assumptions where relevant.
Response + uncertainty
A conditional estimate compared with new measurements, real outcomes, and explicit failure criteria.
Long-term research questions
What would become testable if an individual model were reliable?
These are research directions, not capabilities or services VHL offers today. Each would require longitudinal data, causal evidence, external validation, and prospective testing.
Which change is worth testing?
Compare modeled responses to candidate medicines, nutrition, supplements, sleep, exercise, or other defined exposures without turning an estimate into a personal recommendation.
How might response vary before recruitment?
Explore response subgroups and prioritize trial hypotheses before prospective human testing. Simulation cannot establish safety or efficacy or replace clinical trials.
Which biological paths remain plausible?
Use repeated measurements to estimate possible progression, recovery, or relapse paths while exposing assumptions, time horizons, and uncertainty.
Current work / RNA
Current RNA work asks how to read biological state.
RNA is a dynamic but incomplete view of cellular activity. PBISC, RNA-LLM, and leakage-controlled benchmarks address resolution, representation, and testing questions relevant to broader response models.
Program 01
PBISC
Cross-resolution transcriptomics
PBMC bulk → pseudo-single-cell
Can a bulk immune profile infer useful cell-state structure?
PBISC asks how much cell-level structure an RNA profile can support. It generates an analysis-ready pseudo-single-cell count matrix from PBMC bulk RNA-seq. It is designed to work without a matched target-cohort single-cell reference at inference time.
Can large legacy bulk cohorts support cell-type and cell-state analyses that normally require single-cell measurements?
A generated count matrix for downstream immune-state analysis—not a claim that measured single cells have been recovered.
Cell composition, expression structure, pathways, and downstream biological conclusions must be tested separately.
Program 02
RNA-LLM
Multimodal representation
Transcriptome → learned token
Can a language model reason over RNA without serializing every gene?
RNA-LLM tests an early bridge from a measured biological profile to a language-model interface. It compresses a single-cell, pseudobulk, or bulk transcriptome into one learned RNA token that conditions a language model on the measured transcriptomic profile.
Can a compact RNA representation preserve enough signal for biological reasoning across assay resolutions?
A learned bridge maps transcriptome features into the language model rather than turning thousands of gene values into a text prompt.
Biological plausibility, cross-study transfer, and evidence use matter alongside language quality.
Program 03
Evaluation
Generalization and leakage control
Study-held-out / donor-held-out
Does the model transfer—or recognize the dataset?
A virtual-human model would need to transfer to new people and settings rather than recognize a familiar dataset. VHL develops evaluation protocols for autoimmune, infectious-disease, and vaccine-response transcriptomes that separate donors and studies before performance is measured.
Samples from the same cohort can share collection, processing, site, and batch signals. Random sample splits may reward those shortcuts.
Repeated measurements from one person must stay in one split so the test does not reward subject recognition.
Results should state the split unit, cohort provenance, endpoint, uncertainty, subgroup behavior, and known sources of leakage.
Evidence threshold
Every output needs uncertainty, failure conditions, and a boundary for use.
A convincing output is not enough. The full chain from observation to response must show provenance, uncertainty, transfer, and the conditions under which the model should not be used.
- 01 / ObserveDefined biological state
Record assay, context, cohort, donor, time, and access conditions.
- 02 / EncodeTraceable representation
State what is measured, inferred, compressed, generated, or missing.
- 03 / PerturbConditional response
Define the intervention, assumptions, time horizon, and uncertainty.
- 04 / ValidateNew people + outcomes
Test transfer, calibration, failure modes, and prospective performance.
Research inquiries
Bring a biological question that can be measured and tested.
Include your affiliation, research context, available evidence, and the response or perturbation you want to study. Do not send patient information, controlled-access data, or confidential material.
Contact VHL