- A gene mutation model scores each sample for risk and ranks the most influential genes or pathways out of thousands, not dozens.
- Several trained models can run on the same sample, so one profile can be checked against more than one condition or treatment response.
- What-if analysis shows how the predicted risk changes when an input changes, for research and planning, not as a treatment recommendation.
- Clinicians and scientists make every decision. No clinical efficacy claim is made for this use case.
Thousands of genes, one sample at a time
Precision medicine puts a patient's genomic and phenotype profile at the center of diagnosis and research. The profile is large, and deciding which of thousands of genes matter for one sample is slow work when it is done by hand.
Your team also needs context: how a sample's risk compares with patients you have already evaluated, and which cellular pathways explain the difference. Any model you add has to show its reasoning. Your scientists and clinicians will not act on a score that cannot name the genes behind it.
How gene mutation risk scoring works
Models learn from samples with known outcomes, then score new samples and explain which genes and pathways drove each score.
Step 1: assemble the data
Combine genomic measurements, pathway annotations and phenotype data for samples whose outcomes you know.
Step 2: build and compare models
H2O Driverless AI runs automatic feature engineering, including features built from gene ontology groupings, and ranks candidate models. GPU acceleration shortens training on wide genomic data.
Step 3: rank the drivers
Shapley values and interpretability methods rank the genes and pathways that most influenced each sample's score.
Step 4: compare with past patients
Similarity search shows how a sample's predicted risk compares with previously evaluated samples.
Step 5: ask what-if questions
Individual conditional expectation shows how the prediction for one sample moves as an input changes, for research and planning.
Step 6: load more models
Load models trained on your own data for other conditions or treatment responses, and score the same sample against each.
What the model does not decide
What-if analysis is for research and planning, not a treatment recommendation. Clinicians and scientists make every decision, and no clinical efficacy claim is made for this use case. Protected health information never has to leave your boundary, in any of four deployment options: on-premises, air-gapped, private cloud or cloud.