DNAI Biotech

Patent Pending

U.S. Provisional Application No. 64/036,627

System and Method for Cross-Domain Interventional Parameter Anchoring for Mechanistic Digital Twin Calibration

A system and method for calibrating mechanistic digital twin simulators using cross-domain interventional parameter anchoring. The system resolves parameter confounding by utilizing paired preclinical time-series data to identify intrinsic dynamics parameters from untreated control curves and interventional response parameters from treated curves of the same biological model identifier. A Domain Separation Network aligns preclinical and clinical latent representations, and a Conditional Prior neural network imputes missing modalities. A hypernetwork conditioned on the domain-invariant shared representation generates predicted mechanistic parameters for execution in a differentiable simulator.

18 Claims
← Back to Whitepapers

CROSS-REFERENCES

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims the benefit of and priority to U.S. Provisional Patent Applications No. 63/967,576, No. 63/974,083, No. 63/974,099, No. 63/988,460, No. 63/988,475, No. 63/988,480, No. 63/991,254, No. 63/991,263, No. 64/029,329, No. 64/029,334, No. 64/029,335, No. 64/029,336, and No. 64/029,337. All prior provisional applications are incorporated herein by reference in their entireties.

I.FIELD OF THE INVENTION

[0001]The present disclosure relates generally to computational biology, machine learning, and mechanistic modeling. More specifically, it relates to systems and methods for executing mechanistic neural network simulators utilizing cross-domain parameter anchoring and fail-closed execution control via structural computational graph disconnection, conditional argument binding, and runtime gradient verification.

II.BACKGROUND OF THE INVENTION

1.Parameter Confounding in Observational Data

[0002]Mechanistic digital twin models of human biology, particularly in oncology, require the accurate identification of intrinsic biological parameters (e.g., tumor growth rate, ρ) and treatment-response parameters (e.g., drug sensitivity, β). In standard machine learning architectures trained on observational clinical data, these parameters are fundamentally confounded because treatment assignment is biased by disease severity.

2.Cross-Domain Transfer Challenge

[0003]While preclinical data such as Patient-Derived Xenografts (PDXs) can identify β interventionally through controlled experiments, preclinical models operate in a fundamentally different biological domain. Existing tumor growth inhibition models fit parameters to xenograft curves but fail to mechanistically transfer these parameters into a human clinical digital twin. Standard domain adaptation techniques align predictive representations but do not anchor mechanistic ODE parameters, nor do they resolve missing modality disparities between preclinical and clinical datasets.

III.SUMMARY OF THE INVENTION

[0004]The present invention provides a specific technical improvement to the functioning of computer-based mechanistic simulators by introducing a cross-domain interventional parameter anchoring system. The training system isolates the intrinsic growth rate ρ by fitting vehicle control curves with the immune kill parameter ω hard-zeroed for athymic preclinical models. The system freezes ρ via a stop-gradient operation and fits drug sensitivity β from treated curves of the same tumor model identifier using a constrained slope-subtraction estimator.

[0005]A Domain Separation Network (DSN) with shared and private encoders, trained with an adversarial discriminator and gradient reversal layer, aligns a 201-dimensional RNA latent slice across species. A Conditional Prior neural network imputes missing molecular modalities (methylation and CNV) to construct a 281-dimensional aligned vector for hypernetwork conditioning. The hypernetwork generates predicted mechanistic parameters for execution in a differentiable simulator.

[0006]Empirical validation demonstrates: vehicle ρ fitting across 140 PDX models (mean=0.060), treated β fitting across 2,594 curves and 62 drugs (mean=0.032), β vs RECIST response Spearman ρ=0.416 (p<0.0001), β/ρ ratio improvement from 0.0017 to 0.83, and human C-index of 0.742.

IV.CLAIMS (18)

18 claims covering method (Claims 1-9), system (Claims 10-14), and computer-readable medium (Claims 15-18) embodiments. Independent claims center on the cross-domain interventional calibration pipeline. Runtime verification of parameter exclusion invariants is covered in dependent claims.

Independent Claim 1 (Method)

[0007]A method for calibrating a mechanistic simulator using cross-domain interventional parameter anchoring, comprising: identifying an intrinsic dynamics parameter from untreated control curves while constraining a domain-specific parameter to zero; freezing said parameter and identifying an interventional response parameter from treated curves via a constrained estimator; aligning preclinical and clinical latent representations via a domain separation network; generating predicted mechanistic parameters via a conditioned hypernetwork; and executing a differentiable simulator to generate a candidate output.

Key Dependent Claims

[0008]Claim 2: Parameter exclusion invariant verification with fail-closed enforcement. Claim 3: DSN with shared/private encoders and adversarial training. Claim 4: Conditional prior for missing modality imputation. Claim 5: Slope-subtraction estimator for interventional response parameter. Claim 6: Immune kill parameter zeroed for athymic preclinical models.