Our Research
Peer-reviewed publications, patent portfolio, and technical documentation underlying the DNAI physics-constrained cancer digital twin platform.
Research Publications
Original research from the DNAI project. All preprints are freely available for download. Journal submissions in progress.
Domain Shift is a Feature, Not a Bug: Distributionally Robust Optimization Outperforms Harmonization in Clinical Foundation Models
Feb 2026
We challenge the harmonization paradigm, demonstrating that site-specific variance encodes critical prognostic information. Group DRO with Pooled Cox achieves C=0.718 on CPTAC, outperforming all baselines including ComBat and CORAL.
The Plural Twin: Quantifying Treatment Policy Stability via Set-Valued Cancer Digital Twins
Feb 2026
We introduce Plural Twins, a set-valued framework where each patient is represented as a distribution of outcomes. 82.9% of patients show policy instability; for 1 in 6, the optimal treatment depends on the algorithm's risk tolerance.
The Sim-to-Real Scaling Paradox: Biological Heterogeneity Reverses Transfer Learning Gains in Cancer Digital Twins
Feb 2026
We report a counterintuitive data scaling paradox: expanding PDX training data from 128 to 573 samples degrades clinical prediction. Multi-cancer alignment erases biology-preserving variance through shortcut domain adaptation.
Runtime Reliability Labeling for Safe Deployment of Oncology AI Under Distribution Shift
Feb 2026
A unified framework for certifying when a clinical AI prediction is reliable enough for decision-making. Per-patient transportability certificates, structured abstention, and evidence-completion recommendations.
The Glass Cannon Phenotype: High Predicted Benefit but Low Robustness to Biological and Dosage Perturbations
Feb 2026
We identify a clinically actionable phenotype defined by the intersection of high predicted treatment benefit and low robustness to perturbation. These patients (7.0%) show median OS of 478d with 71.1% event rate — the worst outcomes despite high expected benefit.
Practical Identifiability Failure in Physics-Constrained Cancer Models: The Therapeutic Controllability Index
Feb 2026
We report a fundamental identifiability failure: drug sensitivity (beta) occupies 0.14% of its allowed range under survival-only supervision. Rather than treating this as a defect, we introduce the Therapeutic Controllability Index to quantify treatment authority per patient.
Clonal Architecture-Aware Digital Twins: Bridging VAF-Based Deconvolution with Physics-Constrained Tumor Simulation
Mar 2026
A complete pipeline bridging VAF-based clonal deconvolution with physics-constrained tumor simulation. Introduces the Resistance Sentinel strategy that preserves clinically critical minor subclones in fixed-compartment ODE models, and knowledge-grounded per-clone drug sensitivity from OncoKB/CIViC replacing under-identified learned parameters.
Latent-Space Clonal Decomposition: Attention-Weighted Variational Inference for Per-Subpopulation Biological State Estimation
Mar 2026
A novel architecture extending variational autoencoders to jointly decompose bulk tumor omics into per-subpopulation biological states in structured latent space. Unlike mutation-only deconvolution, LSCD infers per-clone pathway activities, growth rates, and epigenetic states — resolving the under-identification of drug sensitivity parameters.
Resistance Forecasting via Clonal Digital Twins: A Retrospective Validation Protocol
Mar 2026
A pre-registered validation protocol for testing clone-aware digital twin predictions against longitudinal patient data. Compares predicted resistance emergence timing and dominant resistant clone identity against observed outcomes in patients with serial biopsies from GLASS, TRACERx, and GENIE BPC cohorts.
Complementary Foundation Model Distillation for Multi-Omics Cancer Digital Twins
Mar 2026
We distill domain-specific foundation models (Geneformer, ESM-2) into the structured latent space of a multi-omics VAE, achieving R^2=0.774 on full latent reconstruction. The complementary distillation strategy preserves biological interpretability while incorporating protein structure and gene regulatory knowledge.
Zero-Shot Drug Combination Discovery via Orthogonal Pathway Sensitivity Profiles
Mar 2026
We predict drug combination efficacy from monotherapy data alone using pathway sensitivity orthogonality. Validated on 1,209 GDSC drug pairs, the method achieves Spearman rho=0.800 between predicted and actual combination patterns, enabling zero-shot discovery of synergistic combinations without combination screens.
Identity Memorization in Drug Synergy Prediction: A Cautionary Analysis of Cell-Line Overlap
Mar 2026
We demonstrate that standard drug synergy benchmarks suffer from identity memorization: models learn cell-line fingerprints rather than genuine drug interaction biology. Cell-line-holdout evaluation reveals dramatic performance drops compared to random splits, calling for stricter evaluation protocols in combination prediction.
Treatment-Gated Gradient Isolation Reduces Prognostic Leakage in Neural ODE Survival Modeling
Apr 2026
We demonstrate that mechanistic Neural ODEs trained on observational survival data suffer from a fundamental identifiability failure: drug response parameters are either excluded from the computation graph or hijacked as prognostic proxies. Our V4.1 architecture enforces untreated gradient isolation via treatment-gated graph surgery, hard biological parameterization, and two-stage PDX system identification, achieving causally identified drug sensitivity (Spearman rho=0.416 vs RECIST, p<0.0001) without sacrificing survival prediction (C-index 0.742).
Data Diversity Outperforms Feature Injection for Cross-Institutional Cancer Survival Prediction
Apr 2026
We demonstrate that the within-cancer survival prediction ceiling (C-index 0.57) is caused by institutional domain shift, not feature poverty. Discrete-time hazard modeling (DeepHit) with multi-institutional Group DRO across 6 environments (17,526 patients) improved the held-out CGGA within-cancer C-index from 0.548 to 0.630. Foundation Model injection (-0.043) and proteomics LUPI (-0.041) both degraded external generalization. Adding diverse training cohorts (+0.045) was 10× more effective than adding features.
Patent Portfolio
13 U.S. Provisional Applications filed. Covering physics-constrained simulation, domain adaptation, uncertainty quantification, runtime safety, risk-averse optimization, knowledge-grounded annotation, and drug combination prediction.
#1: Sim-to-Real Transfer
Physics-Constrained Sim-to-Real Transfer Learning
#2: Collapse Prevention
Preventing Metabolic Scaling-Induced Collapse
#3: PESD & Missing Modalities
Uncertainty-Calibrated Missing Modality Imputation
#4: Pathway-Aware GRL
Ontology-Guided Autogradient Modulation
#5: ExecutionToken & Solver Interlock
Adjoint Sensitivity & Physics-Constrained Gradient Topologies
#6: Group DRO for Digital Twins
Distributionally Robust Training (DRO)
#7: Risk-Averse Stochastic Optimization
Stabilized Stochastic Inference and Risk-Averse Optimization in Physics-Constrained Oncology Digital Twins
#8: Runtime Gating & Solver Interlock
Cryptographically Enforced Runtime Resource Gating in Differential Equation Solvers via Memory Allocation Interlock
#9: Fail-Closed Integrity Gating
System and Method for Stabilized Stochastic Inference and Fail-Closed Integrity Gating in Neural Networks
#10: Knowledge-Grounded Parameter Annotation
Systems and Methods for Knowledge-Grounded Parameter Annotation and Execution Control of Computational Dynamics Simulators
#11: Safety Assurance & Execution Control
System and Method for Cryptographically Enforced Execution Mode and Automated Safety Assurance of Calibrated Machine Learning Pipelines
#12: Resistance Sentinel Architecture
System and Method for GPU Execution Control via Privilege-Separated Resistance Sentinel Architecture in Fixed-Compartment Tumor History Modeling
#13: Drug Combination Efficacy Prediction
System and Method for Predicting Drug Combination Efficacy Using Symmetric Bilinear Interaction of Monotherapy-Derived Pathway Sensitivity Profiles
#15: Cross-Domain Parameter Anchoring
System and Method for Cross-Domain Interventional Parameter Anchoring for Mechanistic Digital Twin Calibration
#16: Runtime Execution Controller
Runtime Execution Controller for Differentiable Simulators with Autodiff-Verified Parameter Exclusion and Fail-Closed Enforcement
Whitepapers
Deep-dive technical documentation on the architecture, algorithms, and validated performance of the platform.
Foundation White Paper
Neuro-Symbolic architecture overview. H-BDVAE, dual-path system, and physics-constrained simulation.
The Dark Matter of Oncology
Why genomics alone cannot predict resistance. Non-Mutational Resistance and the failure of standard AI.
Sim-to-Real Architecture
Split-Source Transfer Learning for predictive oncology. From organoids to patients via domain separation.
Presentations & Media
Visual walkthroughs of the platform architecture and patent portfolio.
Physics-Constrained Cancer Dynamics
15-slide visual walkthrough of the sim-to-real architecture, separated-state ODEs, and patent portfolio.
Using Physics to Translate Mouse Cancer
AI-generated podcast covering patent innovations, cross-species transfer learning, and clinical implications.
Validated Performance
Key metrics across internal and external cohorts
Research Use Only
All publications and methods described here are for research purposes only and have not been cleared or approved by any regulatory authority for clinical use. The DNAI platform is not a medical device. Patent applications are U.S. Provisional Applications; no patents have been granted.
Interested in collaborating?
We welcome academic collaborations, validation partnerships, and licensing discussions.