System and Method for Preventing Metabolic Scaling-Induced Representational Collapse in Cross-Species Oncology Models

Preclinical oncology models, including patient-derived xenografts (PDX), provide dense longitudinal tumor measurements, while human clinical datasets are comparatively sparse and irregular. Standard domain adaptation techniques apply adversarial training uniformly across all representation channels to enforce domain invariance between a source domain and a target domain.

The inventors identified a failure mode termed representational collapse that is particularly harmful in coupled machine learning and dynamical simulation pipelines. When proliferation-associated signals, which correlate with interspecies metabolic scaling differences, are forced into adversarial domain alignment, the induced gradient conflicts can trigger: (i) variance collapse in modality-specific latent channels, (ii) degradation in downstream hypernetwork parameter predictions, and (iii) numerical instability in ODE integration due to inadmissible or stiff parameterizations. Such collapse can degrade computational functionality of the simulator and is not readily remedied without substantial retraining or architectural modification.

Prior art in partial domain adaptation and feature disentanglement teaches various methods for excluding or down-weighting statistically dissimilar subsets during alignment. Such methods generally rely on distributional metrics or task-driven heuristics and do not account for biophysical constraints that arise in cross-species oncology modeling.

In cross-species tumor modeling, proliferation markers can appear statistically similar across domains while remaining mechanistically incompatible due to allometric scaling laws. As a result, purely statistical selection criteria can fail to identify which signals must be protected from alignment to preserve stability of coupled dynamical solvers.

There is a need for systems that procedurally identify which latent representation channels should remain domain-private based on biophysically grounded criteria, and that structurally exclude those channels from adversarial gradients so that downstream dynamical parameter generation and numerical integration remain stable.

[0001] A computer-implemented system 100 comprises:

• an encoder $f_{\phi}$: $X \to Z$ mapping molecular features $\mathbf{x}$ to a latent representation $\mathbf{z}$;
• a partitioning mechanism that produces $z_{\text{aligned}}$ and $z_{\text{private}}$;
• a domain discriminator $D_{\psi}$: $Z^* \to [0,1]$;
• a parameter generator (hypernetwork) $h_{\eta}$: $Z \to \Theta$;
• a constraint mapping $c: \Theta \to \Theta_{\text{valid}}$;
• an ODE solver $S$ integrating $dy/dt = F(y, t; \theta)$ with $\theta = c(h_{\eta}(z))$;
• and a solver instrumentation module 140 that records integration diagnostics to non-transitory memory.

[0002]The following terms are used throughout this application:

Source domain

preclinical tumor models including PDX.

Target domain

human clinical or population-scale datasets.

Latent representation channels

latent dimensions, tensor slices, encoder heads, or indexed feature groups. In one embodiment, $z \in \mathbb{R}^k$ and channels correspond to indices $i \in \{1,\ldots,k\}$.

Representational collapse

loss of effective information capacity in a subset of channels during training, evidenced by one or more of: (i) variance collapse in one or more channels, (ii) degraded downstream parameter prediction, or (iii) numerical instability during ODE integration.

Non-aligned subset / private subset

channels designated to be excluded from adversarial alignment. In index form, the private subset corresponds to $I_p \subset \{1,\ldots,k\}$, and the aligned subset corresponds to $I_a = \{1,\ldots,k\} \setminus I_p$. In gating form, a gate vector $g \in \{0,1\}^k$ or $g \in [0,1]^k$ indicates membership.

Private pathway

a structural mechanism ensuring that gradients of a discriminator loss do not propagate to parameters that generate the private subset. Implementations include gradient stopping, architectural separation, and input masking, and can also include restricting discriminator input to aligned channels.

Parameters that generate the private subset

parameters of the encoder component(s) whose outputs define $z_{\text{private}}$. In embodiments with separate encoder heads, these parameters comprise weights exclusive to the private head and do not include weights exclusive to the aligned head.

Predetermined range

a bound stored in non-transitory memory (for example in configuration memory, model metadata, or a database) used by differentiable constraints to enforce admissibility and improve solver stability.

[0003] Interspecies tumor growth kinetics can be influenced by allometric scaling. In one embodiment, metabolic rate scales with body mass according to Kleiber's law $B \propto M^{3/4}$, and rate-like biological processes scale approximately with $r \propto M^{-1/4}$. As a result, proliferation-associated signals can exhibit mechanistically different scaling regimes across species even when pathway activation appears superficially similar.

The system uses $s$ to modulate the selection rule. In one embodiment, the size of the private subset, the exclusion threshold, or a rank cutoff changes monotonically with $|\log s|$, which provides consistent behavior under swapping of source and target domains.

Mass sourcing: In embodiments, $M_{\text{source}}$ and $M_{\text{target}}$ are obtained from stored species metadata in non-transitory memory, user-provided configuration, or a database of typical body masses.

Mechanistic integration into method: in one embodiment, $s$ is used to adjust an exclusion threshold $\tau(s)$ or a selected proportion $q(s)$ of channels designated as private, such that larger interspecies scaling disparities yield stronger exclusion of proliferation-linked channels from adversarial gradients.

[0004] The system computes a proliferation metric $p(x)$ indicating cell-cycle activity. Embodiments include:

• gene set scoring using cell-cycle genes including MKI67, PCNA, TOP2A, MCM6, CDK1, CCNB1, CCNA2, CCNE1, PLK1, AURKA, and BUB1;
• a learned latent component supervised to correlate with Ki67 expression.

A selection rule $R$ maps the proliferation metric to a private-channel designation. Example embodiments include:

• Thresholding: $I_p = \{ i : s_i(p) \geq \tau \}$ where $s_i$ is an attribution or dependence score and $\tau$ is a threshold;
• Ranking: selecting top-$q$ channels by dependence with $p$;
• Gating: learning $g(x)$ and designating channels as private based on gate values.

Fully specified example algorithm (enablement anchor):

1. Compute proliferation score $p(x)$ from a cell-cycle gene set score.
2. Compute channel dependence scores $a_i$ where $a_i$ is mutual information or a gradient-based attribution magnitude relating channel $z_i$ to $p$.
3. Define $I_p$ as the indices of the top-$q$ channels by $a_i$, or all channels where $a_i \geq \tau$.
4. Optionally set $q$ or $\tau$ as a function of $|\log s|$, such that larger scaling disparities produce a larger private subset or stricter exclusion.

[0005] The system trains a discriminator $D_{\psi}$ to predict domain labels from discriminator input $u$. The discriminator loss is denoted $L_D(D_{\psi}(u))$. An adversarial objective is implemented via gradient reversal or equivalent sign inversion applied to encoder gradients for aligned channels.

Structural exclusion requirement:

The private subset is structurally excluded from adversarial alignment such that gradients of the discriminator loss do not propagate to parameters that generate the private subset.

Embodiments include:

Clarification for discriminator inputs that include private channels: In embodiments where discriminator input includes the private subset (for example via concatenation), the private subset is provided through a gradient-blocking operation such that discriminator-loss gradients do not update parameters that generate the private subset.

[0006] In one embodiment, training uses losses that update disjoint parameter subsets:

• task loss $L_{\text{task}}$ updates parameters of the encoder generating $z_{\text{aligned}}$ and $z_{\text{private}}$, and updates the parameter generator $h_{\eta}$;
• discriminator loss $L_D$ updates discriminator parameters $\psi$;
• adversarial gradients derived from $L_D$ update only parameters that generate $z_{\text{aligned}}$, and are structurally blocked from updating parameters that generate $z_{\text{private}}$.

This explicit routing is implemented by at least one of: discriminator input restriction, stop-gradient, masking, or separate encoder heads, as described herein.

[0007] To reduce leakage between aligned and private subsets, the system may include independence regularization, including:

• HSIC regularization between $z_{\text{aligned}}$ and $p$;
• orthogonality penalties or covariance penalties between $z_{\text{aligned}}$ and $z_{\text{private}}$;
• probe-based minimization where a probe attempts to predict proliferation from aligned channels and is discouraged.

[0008] A parameter generator $h_{\eta}$ outputs dynamical parameters which are constrained by differentiable mappings to predetermined ranges stored in memory. In one embodiment, $h_{\eta}$ comprises parameter heads, including $h_{\eta,\rho}$, $h_{\eta,\beta}$, and $h_{\eta,\omega}$, and outputs parameters that are constrained as follows:

In embodiments where inclusive endpoints are desired, the system additionally applies a clamp, projection, or stored-range saturation rule to enforce membership in a closed interval $[L, U]$ stored in memory.

[0009] The ODE integrated by solver $S$ may be selected from a family including:

Logistic growth:

Gompertz:

von Bertalanffy:

Immune-mediated tumor dynamics:

where $I$ is an immune effector state variable, $\epsilon$ is a saturation parameter, and $\delta$ is a decay or loss-rate parameter.

[0010] In coupled representation-dynamical systems, instability can manifest as a compound failure mode: adversarial alignment modifies latent geometry, which perturbs constrained parameter generation, which increases stiffness or inadmissibility of ODE parameterizations, which then triggers numerical instability (for example step rejections, excessive function evaluations, or integration failures). The invention addresses this coupled failure mode by mechanistically selecting proliferation-linked channels for structural exclusion from adversarial gradients and by constraining ODE parameters to predetermined ranges stored in memory, thereby improving numerical solver stability.

[0011] The solver instrumentation module 140 logs to non-transitory memory at least two of:

• number of function evaluations (NFE);
• wall-clock integration time;
• step rejection events;
• constraint violation events.

Concrete feedback policy example (technical loop):

In one embodiment, logged instrumentation is used to automatically update solver control parameters for a subsequent trajectory generation cycle, within bounds or thresholds stored in memory, for example:

• if step rejections exceed a stored threshold $R_{\max}$, decrease a maximum step size parameter selected from a stored set, or adjust an error-control heuristic within a stored envelope;
• if NFE exceeds a stored threshold $N_{\max}$, switch solver selection among a predefined set stored in memory, or adjust tolerances within a bounded envelope;
• if constraint violations are detected, adjust constraint application to occur earlier in the parameter-generation pipeline or clamp offending parameters before integration.

[0012] The following metrics are provided from evaluated implementations and are included as concrete indicators of technical effect. Where comparisons are described, the invention is not limited to a particular baseline implementation.

Collapse evidence (cross-version):

• DNA methylation latent variance collapse: $0.955 \to 0.0002$ (selective alignment version vs forced full alignment version).
• ODE growth-rate prediction degradation: $\rho$ $R^2$ $0.899 \to 0.537$ (selective alignment version vs forced full alignment version).
• Model-internal implied PDX doubling-time divergence: median 4.49 days vs 939 days (raw selective vs standardized/aligned). This doubling time value is a model-internal implied timescale under a standardized alignment transform and is not a biological measurement.

Solver efficiency and stability (evaluated):

No other numerical comparator metrics are asserted in this section.