System and Method for Physics-Constrained Sim-to-Real Transfer Learning in Computational Oncology

The pharmaceutical industry currently faces a translational challenge in oncology drug development. Despite the availability of therapeutic candidates and preclinical models, the attrition rate for oncology drugs entering human clinical trials remains high. A significant factor in this attrition is the difficulty in translating efficacy signals from preclinical models -- specifically the Patient-Derived Xenograft (PDX) -- to human clinical responses.

A contributing factor to predictive failure is the biological phenomenon of “Stroma Replacement”. Upon engraftment of human tumor tissue into a murine host, human stromal cells are frequently replaced by murine host cells. Consequently, machine learning models trained on genomic data from these models may inadvertently learn correlations between tumor growth and murine stromal signals (e.g., murine angiogenesis drivers) that are distinct from human biology.

When transferred to a human context, such models may suffer from “Negative Transfer”due to the distributional shift between the source domain (Murine Stroma) and target domain (Human Stroma).

Furthermore, a data asymmetry exists between preclinical and clinical environments. PDX models generate high-density, longitudinal time-series data, whereas human clinical trial data is typically sparse, with tumor assessments occurring infrequently (e.g., every 6 to 12 weeks). Additionally, metabolic rates differ between species, requiring allometric correction during translation. Finally, preclinical datasets often lack full multi-modal coverage (e.g., missing methylation or copy number data) compared to comprehensive human datasets like TCGA, creating a feature mismatch that hinders robust model training.

There is a need for a computational system capable of: (1) disentangling conserved intrinsic tumor drivers from species-specific artifacts; (2) imputing missing modalities to enable multi-modal simulation from single-modal inputs; (3) utilizing high-density murine data to learn growth physics; and (4) estimating parameters from sparse human data to generate valid tumor trajectory simulations.

[0001]The invention provides a computer-implemented system (100) and method for generating validated, physics-constrained tumor volume trajectories. The system accepts as input a molecular feature vector x from a target patient and outputs a time-series trajectory $V(t)$ representing tumor burden over time. In some embodiments, the state variable $V$ represents volumetric tumor size (e.g., mm³); in other embodiments, $V$ represents a sum-of-longest-diameters (SLD) burden derived from RECIST measurements. The system is configured to bridge the translational gap between preclinical models and clinical outcomes by disentangling conserved biological drivers from host-specific microenvironmental artifacts.

[0002]The system processes two distinct datasets: a Source Dataset (Murine) and a Target Dataset (Human). In preferred embodiments, the system applies domain-specific standardization (e.g., Z-scoring) to independently align the statistical distributions prior to encoding. In embodiments utilizing gene expression data, preprocessing includes ortholog mapping to align murine genes with human equivalents, retaining only one-to-one orthologs.

Source Dataset (D_S)

Comprises molecular feature vectors (e.g., RNA-seq) and longitudinal tumor volume measurements sampled at a high-frequency cadence (e.g., every 2–3 days) from xenograft subjects. Preprocessing includes ortholog mapping (e.g., via Ensembl or HomoloGene), TPM/FPKM normalization, and optional batch correction.

Target Dataset (D_T)

Comprises molecular feature vectors and sparse clinical endpoints sampled at a low-frequency cadence (e.g., every 6–12 weeks) from human subjects.

[0003]To address host microenvironment shifts (e.g., Stroma Replacement), the system employs a Domain Separation Network (DSN). The DSN comprises:

Shared Encoder (112)

In one embodiment, an MLP with architecture 201 (input) → 128 (hidden) → 201 (output), utilizing LeakyReLU(0.2) activation and LayerNorm.

Private Encoder (114)

In one embodiment, an MLP with architecture 201 (input) → 128 (hidden) → 64 (bottleneck), with projection to 201 dimensions, utilizing LeakyReLU(0.2) activation and LayerNorm.

Domain Discriminator

In one embodiment, an MLP with architecture 201 → 128 → 64 → 1, utilizing LeakyReLU(0.2), Spectral Normalization, and Dropout(0.3).

Training Objectives: The network is trained to minimize a composite loss function:

In specific embodiments: $\lambda_{\text{adv}} = 1.0$; $\lambda_{\text{ortho}} = 0.5$; $\lambda_{\text{recon}} = 0.1$. Gradient Reversal Layer (GRL) schedule: Reverse sigmoid function.

• Reconstruction Loss ($L_{\text{recon}}$): A decoder reconstructs the original input from the concatenation of $z_{\text{shared}}$ and $z_{\text{private}}$.
• Adversarial Loss ($L_{\text{adv}}$): A Domain Discriminator classifies domain based solely on $z_{\text{shared}}$. A Gradient Reversal Layer (GRL) reverses the gradient to remove species-specific information.
• Orthogonality Loss ($L_{\text{ortho}}$): Enforced via mean absolute cosine similarity or squared Frobenius norm of the covariance between shared and private latent representations.

[0004]To address data sparsity where modalities (e.g., DNA Methylation, CNV) are absent, the system employs Probabilistic Encoder Self-Distillation (PESD).

Architecture

A teacher encoder (full modalities, frozen) trains a student encoder (masked modalities) via KL divergence matching.

Loss Function

$L_{\text{total}} = \lambda_{\text{kl}} \cdot D_{\text{KL}} + \lambda_{\text{recon}} \cdot L_{\text{recon}} + \lambda_{\text{calib}} \cdot L_{\text{calib}}$

Missingness Embeddings

The system utilizes learned missingness embeddings [MISSING_METH] and [MISSING_CNV] to represent unobserved modalities, distinct from zero-imputation.

Calibration Constraint

The student log-variance is constrained to be $\geq$ teacher log-variance when modalities are missing, preventing overconfident hallucination.

[0005]The latent vectors parameterize a differential equation governing intrinsic tumor growth:

where $f$ = Logistic | Gompertz | von Bertalanffy

ODE Input Vector: In one embodiment, the input vector comprises disentangled latent variables: z_prolif, z_pathway, z_ctx, z_res, z_meth, and z_cnv. In a specific embodiment, the total input dimension is 328, comprising:

The z_prolif variable exhibits a variance of approximately 4.95, indicating well-separated proliferative signal in the latent space.

Parameter Generation Network (Hypernetwork):

head_rho (growth rate)

Sigmoid activation scaled to $[0,\, 0.3] \; \text{day}^{-1}$

head_beta (drug sensitivity)

Sigmoid activation scaled to $[0,\, 1]$

head_omega (immune clearance)

Softplus activation, range $(0, +\infty)$

head_sigma (observation noise)

Sigmoid activation scaled to $[0.01,\, 0.2]$

Constraint Enforcement Layer: Rather than detecting constraint violations post-hoc, the system enforces physical constraints a priori via bounded activation functions. Non-negativity is ensured for $\rho$ and $\omega$ via Softplus/Sigmoid activations. Carrying capacity is fixed at physiological upper bounds (e.g., $10^{11}$ cells). This ensures the coupled ODE remains numerically stable during integration, preventing stiffness and ensuring solver convergence.

[0006]To correct for metabolic time dilation between species, the system initializes a learnable time-scaling parameter $\beta$ (denoted as $\tau$ in implementation) based on the quarter-power law:

The learnable parameter is initialized at 1.0 (stored as log_tau = 0.0) and fine-tuned during training on the target dataset to adapt to tumor-specific dynamics beyond mass scaling.

[0007]The system accounts for immune-mediated clearance via:

In a preferred embodiment, $\omega$ is predicted via head_omega (neural network head with Softplus activation) to ensure non-negative clearance rates. Alternatively, $\omega$ may be estimated via bounded 1-D optimization on endpoint residuals.

[0008]During inference, the system generates trajectories using an adaptive-step ODE solver (e.g., dopri5) with relative tolerance 1e-4 and absolute tolerance 1e-6, utilizing the adjoint method for O(1) memory backpropagation.

[0009]Experimental Setup: A cohort of Patient-Derived Xenograft (PDX) models and matched human clinical data was utilized to validate the Split-Source architecture. All primary metrics reported below are from version v5.10 (production). Older versions are cited only where they provide negative-transfer evidence.

Proliferation Disentanglement:

All 16 RNA private dimensions active.

ODE Parameter Prediction (5-fold CV):

Imputation Quality (PESD):

Domain Separation Validation:

Allometric Scaling:

Ablation Study Results:

These results demonstrate that the Conditional Imputation Module provides substantial predictive value, with all ablation drops statistically significant.

[0010]In one specific embodiment, training hyperparameters comprise: