Validated on held-out data
Every metric is computed on data the model never saw during training. We report both successes and limitations.
Request DemoExternal Validation by Cohort
Performance depends on which modalities are available and what cancer type is being predicted. We present all results — including where the model correctly identifies it doesn't have enough data.
The biggest lesson: more modalities = better predictions. RNA + Methylation fusion (C=0.839) outperforms DNA-only (C=0.551) by 52%. Brain tumors with methylation data score highest. Breast cancer without histopathology correctly returns chance-level — the model knows what it needs.
CPTAC — Adult Solid Tumors
1,031 patients · 10 cancer types · 10 hospital sites · RNA-seq + WSI
Three DRO checkpoints: TSS-based (C=0.718, CPTAC truly external), E5 DeepHit (C=0.739, 4-env), and E8 (C=0.741 TCGA-val, 5-env with CGGA). GBM and PDAC benefit most from multi-omics.
OpenPedCan — Pediatric Brain Tumors
943 patients · 43 Heidelberg subclasses · RNA-seq + Methylation
Strongest external result. Methylation dominates classification (+11.6pp), both modalities contribute to prognosis, proper fusion wins. Age matters: infants favor RNA, adolescents favor methylation.
CGGA — Glioma
485 held-out pts · Chinese cohort · RNA-seq only · Truly external
E8 5-env DRO (CGGA split: 485 train, 485 held-out). Within-cancer C improved from 0.548 (A1) to 0.630 (+0.082). GBM C=0.570, LGG C=0.690. Cross-ethnic generalization confirmed.
CT/MRI Radiology
132 pts · CT scans · Standalone module
CT imaging adds significant signal beyond molecular data. Highest super-modality Shapley value (+0.293).
TCGA Sarcoma — Solid Tumor Validation
255 patients · Internal held-out · RNA-seq + DNA + CNV + Meth
Sarcoma patients whose treatment matched the model's recommendation survived 1,137 days longer than discordant patients. 85.5% qualify for high-confidence predictions. Full multi-omics profiling enables strong performance on rare solid tumors.
TARGET — Pediatric
2,929 pts · 6 types · RNA-seq
DRO recovers +10.2pp. Adult→pediatric transfer works partially. In E5 training pool.
MMRF — Multiple Myeloma
787 pts · Liquid tumor · RNA-seq
Liquid tumor — fundamentally different biology. E5 improves +12.2pp over baseline. In E5 training pool.
GENIE — Panel Sequencing
227,696 pts · 20 centers · DNA only
DNA mutations alone provide modest signal. Expression adds +0.019. Driver validation: ρ=0.452 vs IntOGen.
SCAN-B — Breast Cancer
3,069 pts · RNA-seq only · No histopathology
Model correctly fails — breast cancer requires histopathology. This validates modality-awareness.
What We Learn: Modalities Matter More Than Algorithms
The single biggest driver of prediction quality is which molecular data types are available — not which model you use.
+ Fusion
(Green-tier)
+ DRO
(wrong type)
Survival prediction
Within-indication ranking accuracy on held-out TCGA data. Proves we rank patients within cancer types, not just between them.
On Green-tier external CPTAC cohort (N=229) where ISS exceeds threshold. DRO-trained model on validated subset.
Trained and validated across all major TCGA cohorts for pan-cancer applicability.
Benchmark vs. Standard of Care
C-index comparison on held-out TCGA pan-cancer cohort
+27% improvement over Cox Proportional Hazards on high-confidence predictions. DNAI's epistemic uncertainty calibration identifies when predictions are reliable.
Representation quality
Proliferation and context subspaces are statistically independent, enabling clean biological interpretation.
Proliferation latent correlates strongly with MKI67 expression, validating biological meaning.
High-fidelity reconstruction across all input modalities.
Trajectory simulation
Model vs. Biology: Physics parameters learned from PDX (patient-derived xenograft) growth curves accurately predict real tumor dynamics.
Math vs. Math: Learned trajectory emulator matches numerical ODE solver, enabling <5ms inference.
400-1000x faster than numerical solver, enabling real-time treatment optimization.
Note: We do not validate trajectories on TCGA (snapshot data) to avoid temporal paradoxes. PDX data provides true longitudinal measurements.
Known limitations
Phase 0 trained on TCGA (9,393 patients, 33 cancer types). Production DRO trained on 6 cohorts: TCGA+MMRF+TARGET+CPTAC+CGGA+OpenPedCan (~17,500 patients). Validated on 9 external cohorts including 2,697 pediatric patients. Truly external cohorts for E8: CGGA held-out (485 glioma, C=0.729) and SCAN-B (3,069 breast, C=0.504). Performance on rare cancers or non-standard sample preparation may vary.
Not approved for clinical decision-making. Intended for research and pilot deployments.
TSS-DRO: CPTAC C=0.718 (truly external, 1,031 patients, 10 cohorts). E8 DRO: CGGA C=0.729 (held-out, 485 glioma, within-cancer macro C=0.630). SCAN-B C=0.504 (breast, needs histopathology). MMRF C=0.609, TARGET C=0.621 (in training pool).