Every number has a name
DNAI predictions are not opaque scores. They are auditable chains of biologically-named computations — from raw gene expression through named pathways, physics-constrained parameters, to time-resolved trajectories with calibrated uncertainty.
The trust problem in oncology AI
Most AI platforms produce a single score — response probability: 0.72 — with no way to inspect what drove that number, which biological assumptions it encodes, or where it might be wrong.
When clinicians and researchers cannot trace a prediction back to its source, adoption fails. We have seen this repeatedly: technically capable systems that never leave the pilot stage because no one trusts them enough to act.
Five inspectable stages
A DNAI prediction flows through five stages. At each stage, the computation is decomposable — you can pause, inspect intermediate values, and verify they make biological sense.
Input Data
Named genes, named variants
Up to 6 data modalities per patient. Every gene is HGNC-standardized. The gene list is version-locked with SHA-256 verification — no silent changes between model versions.
Structured Latent Space
328 named dimensions
The VAE does not produce an opaque embedding. Its 328 dimensions are structurally partitioned into biologically-named groups. Each is inspectable.
Proliferation rate (correlates with Ki67, r=0.96)
50 MSigDB Hallmark pathways × 4 dims — named biological processes
Biological context for driver gene identification
Residual biology not assigned to named pathways
Methylation patterns — epigenetic regulatory state
Chromosomal spatial structure
Domain Separation
What's tumor, what's artifact
For preclinical PDX data (Path B), the DSN decomposes the signal into two explicit branches. For direct human predictions (Path A), this stage is bypassed.
Species-invariant tumor biology — 201 dimensions. Cancer subtype accuracy >90%.
Species-specific stroma artifacts — 64–128 dims. Mouse contamination removed.
Physics Parameters
Six numbers with physiological units
The Hypernet transforms the latent into six named ODE parameters. Each has a physical unit, a valid range enforced by architecture, and a plain-English meaning.
Trajectory + Uncertainty
Time-resolved predictions with confidence
The ODE parameters produce time-resolved trajectories — not a single score, but a curve showing how the tumor evolves over time, with calibrated uncertainty.
Trajectory outputs
Uncertainty layers
The platform checks its own work
Before any prediction reaches a clinician, six automatic checks run in the background — providing context, catching problems, and explaining exactly what the model does and does not know.
"Patients Like Yours"
Every prediction comes with context from similar cases
The platform finds the most biologically similar patients in its training data — not just the same cancer type, but the same pattern of active biological pathways. This grounds every prediction in real patient outcomes.
What a clinician sees: "This prediction is based on 47 similar breast cancers with the same pathway profile. The most similar patients had a median survival of 14.2 months."
When the model says "I don't know"
It tells you exactly why — and what would help
Most AI systems either give a confident answer or a vague warning. DNAI gives specific, actionable reasons for uncertainty — so clinicians know exactly what is missing and how to improve the prediction.
What drove this prediction?
For every prediction, the platform shows which biological pathways had the most influence on each output — traced through the actual computation, not added as an afterthought.
Example: DNA Repair drives 85% of the drug sensitivity prediction — consistent with BRCA1 loss making cells vulnerable to DNA-damaging drugs.
Have we seen this before?
Before making any prediction, the platform checks whether this patient's biology falls within its experience. If a patient is unlike anything it has been trained on, it flags this upfront — before the prediction even runs.
Patients whose biology falls far outside the training population are automatically flagged with wider uncertainty bounds — the platform will not overstate its confidence on unfamiliar cases.
Tumor Board Report
A one-page downloadable summary of the complete prediction — designed for printing and discussion at tumor board meetings. Everything a clinician needs to evaluate and discuss the recommendation.
Stability Check
The platform automatically tests whether small, natural variations in the input data would change the prediction. If the answer flips because of measurement noise, something is wrong — and the platform catches it.
If a prediction is sensitive to small input changes, it is flagged as unstable — the platform will not present a fragile result as confident.
End-to-end trace example
A BRCA1-mutant breast cancer patient treated with carboplatin — traced through every stage.
At every stage, a clinician can ask "why?" and get a biologically grounded answer — not "because the neural network said so."
Glass box vs. black box
| Aspect | Typical Oncology AI | DNAI |
|---|---|---|
| Output | Single score (0.72) | Time-resolved trajectory with clonal dynamics |
| Intermediate values | Opaque 512d embedding | 328 named dimensions (pathways, proliferation, methylation) |
| Parameters | None visible | 6 named ODE parameters with physical units |
| Why this prediction? | Post-hoc feature importance | Biologically-named pathway activations at source |
| Missing data | Silent degradation | Explicit reporting + uncertainty adjustment |
| Impossible predictions | Can output negative growth | Architecturally impossible — physics in activation functions |
| Uncertainty | Rarely provided | MC Dropout CI on every output + 'Data Insufficient' flag |
| Audit trail | Minimal | SHA-256 hashed I/O, versioned models, 21 CFR Part 11 |
Three traceability guarantees
Structural
Every latent dimension maps to a named biological concept. There are no hidden layers that learn uninterpretable features. The architecture forces biological structure.
Physics
Every ODE parameter has a physical unit and a valid range enforced by architecture. The model cannot hallucinate biologically impossible dynamics.
Uncertainty
Every prediction has a confidence interval. Missing data, rare cancers, and model disagreement increase reported uncertainty — transparently.
Built for regulatory review
The audit infrastructure is designed to support FDA Class II SaMD requirements and 21 CFR Part 11 electronic records compliance.
Input Hashing
SHA-256 hash of every input tensor ensures exact reproducibility
Model Versioning
Every checkpoint tracked (VAE v5.10, Hypernet v3.2, DSN v1.0)
Deterministic Mode
Reproducible inference — sampling replaced with mean outputs for audit
Request Logging
Timestamp, model version, input hash, output hash, duration for every call
See the trace-through live
Request a demo to see how a real patient prediction decomposes at every stage — from raw omics to trajectory.