Specialist models.
One combined interpretation.
A mixture of experts combines disease-specific models with a healthy-baseline specialist. A gating model weighs their outputs alongside clinical information.
Interactive architecture- InputsProtein panelClinical contextAge, sex, smoking, recent blood count
- Specialists
- Lung
- Breast
- Liver
- Colorectal
- Healthy baseline
- GateWeighs every reading with the person’s context
- One resultFor example: signal, likely lungor indeterminate, with a next step
One blood draw, many measurements.
A panel of proteins is measured from a single sample, alongside routine clinical context: age, sex, smoking and a recent blood count.
A specialist for each cancer.
Each specialist model is trained on the proteins relevant to one organ, and a healthy-baseline specialist reads what normal looks like. Each returns its own reading.
A gate weighs them, in context.
A gating model weighs every specialist’s reading together with the person’s clinical context, so no single signal decides on its own.
One result, or a next step.
The output is one result with its likely tissue of origin. When the readings disagree, it is indeterminate, with a directed next step instead of a forced call.
A technical reading of the diagram
A mixture of experts combines specialist outputs through a gating or meta-model layer. Each organ specialist returns a cancer probability and a tissue-of-origin confidence; the gate weighs them with the clinical context into one score. The same structure can run a full panel, a tiered sequence that adds measurements only when a result is indeterminate, or a single-cancer panel that escalates when needed.
Why specialists?
Different biological questions can call for different model scopes. In a mixture-of-experts design, specialist models contribute outputs that a gating or meta-model layer considers in context.
The specialist scope can follow the biology. The combining layer then learns how to bring those readings together for the research question.
Three ways to explore the idea.
The full-panel concept considers a broader input scope. The tiered/reflex concept allows additional measurements after an indeterminate step. The disease-specific concept focuses on the relevant panel and specialist.
A model is only as useful
as the question it answers.
We begin with a disease and a defined research question. The panel, the specialist models, and the evaluation are built around it.
Sensitivity and specificity matter together. So do cancer stage, population, and how well the approach works beyond its development data.
The questions behind performance