TumorPred: Integrated Computational Oncology

• TumorPred is a family of computational oncology frameworks that integrate mechanistic models with machine, deep, and graph learning to predict tumor dynamics, heterogeneity, and molecular profiles.
• It employs diverse data modalities, including time-series, imaging, and molecular data, to model longitudinal growth, spatial patterns, and clinical prognoses across various cancer types.
• The framework is exemplified by an R/Shiny application for CNS pharmacokinetics, demonstrating parameter estimation and sensitivity analysis in a permeability-limited four-compartment brain model.

TumorPred denotes a set of computational tumor-prediction frameworks that combine mechanistic modeling, machine learning, deep learning, graph learning, and interactive software for tasks including longitudinal tumor dynamics prediction, tumor heterogeneity estimation, spatial localization, prognosis, molecular profiling, and central nervous system pharmacokinetics. In its most explicit naming, "TumorPred: A Computational Framework Implemented via an R/Shiny Web Application for Parameter Estimation and Sensitivity Analysis in Compartmental Brain Modeling" defines TumorPred as an R/Shiny-based system for simulation, sensitivity analysis, and pharmacokinetic parameter calculation in a permeability-limited four-compartment brain model (Wickramasinghe et al., 5 Sep 2025). In the surrounding literature, the same label is also used as a unifying integration point for pipelines that predict tumor volume trajectories from treatment histories, classify patch-level heterogeneity from spatial cell graphs, generate tumor heatmaps from whole-slide images, and infer molecular or structural tumor properties from microscopy and protein structures (Chattopadhyay et al., 27 May 2025, Njifon et al., 8 Feb 2025, Isett et al., 9 Mar 2026, Binder et al., 2018).

1. Conceptual scope and data modalities

TumorPred spans multiple data regimes rather than a single modality. Longitudinal systems operate on tumor volume time series, imaging-derived features, treatment labels, organ context, and multimodal pharmacology; examples include XGBoost plus exponential/logistic growth models for nanoparticle therapy comparison, heterogeneous graph encoders combined with Neural-ODEs for patient-derived xenograft data, mixed-effects models for lung cancer trajectories, voxel-wise pancreatic tumor progression from CT/PET, and time-conditioned neural fields for vestibular schwannoma growth (Chattopadhyay et al., 27 May 2025, Bazgir et al., 2023, Nasiri et al., 2018, Zhang et al., 2017, Chen et al., 2024).

A second group of TumorPred systems is spatial and cell-centric. These frameworks encode tumors as KNN graphs of cells, tile grids from hematoxylin and eosin whole-slide images, or lattice processes over organ geometries. Representative instances are Block Graph Neural Networks for patch-level heterogeneity classification, DenseNet169-based multi-cancer tile classification with heatmap reconstruction, and an Ising energy model with jump, duplicate, and die events for probabilistic prediction of tumor islets in prostate cancer (Njifon et al., 8 Feb 2025, Isett et al., 9 Mar 2026, Amoudruz et al., 28 Aug 2025).

A third group addresses downstream clinical interpretation. Radiomics-based survival models classify glioma overall survival from MRI; cascaded segmentation plus transfer learning regresses survival in BraTS; bag-of-words plus kernel SVM systems infer protein expression, gene expression, copy number variation, methylation, and somatic mutation states from histology; and a DCNN plus dynamic programming pipeline predicts whether a cancer gene behaves as an oncogene, tumor suppressor gene, or fusion from PDB structures (Yousaf et al., 2020, Cabezas et al., 2018, Binder et al., 2018, Anandanadarajah et al., 2021). This breadth indicates that TumorPred is better understood as a computational oncology framework family than as a single algorithm.

2. Longitudinal dynamics and treatment-response prediction

One mechanistic branch of TumorPred couples supervised learning with explicit tumor-growth equations. In "Hybrid Machine Learning and Mathematical Modeling for Tumor Dynamics Prediction: Comparing SPIONs against mNP-FDG," the model combines Extreme Gradient Boosting with the exponential law

$\frac{dV}{dt}=rV,\qquad V(t)=V_0 e^{rt},$

and the logistic law

$$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$

using initial tumor volume $V_0$, treatment type, and longitudinal time points to forecast tumor volume under SPIONs and mNP-FDG. The paper reports cross-validation with $R^2 > 0.95$ for both SPION and mNP-FDG datasets, while also noting that the parameter estimation procedure is not explicitly described and that zone-specific ray therapy is referenced conceptually rather than operationalized spatially. The comparative results are internally inconsistent but directionally stable: the abstract reports mNP-FDG control within 2 days and SPIONs in 18 days, whereas Figures 3–4 report containment within approximately 4 days and approximately 23 days; complete termination is reported as SPIONs at approximately 20–23 days and mNP-FDG at more than 40 days. On that basis, the study proposes mNP-FDG first for rapid containment, followed by SPIONs for eradication, and packages the workflow in a Tkinter/Python GUI with inputs for time, species type, and treatment type (Chattopadhyay et al., 27 May 2025).

A more multimodal formulation appears in "Integration of Graph Neural Network and Neural-ODEs for Tumor Dynamic Prediction." That framework builds a heterogeneous graph over drugs, diseases, and genes, encodes drug-gene edges from DGIdb, disease-gene edges from DisGeNET, and tissue-specific gene-gene networks from TissueNexus, and conditions a Neural-ODE

$\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$

on both graph-derived embeddings and an RNN encoding of early tumor volumes. Applied to a PDX subset with RNA-seq for 191 unique tumors and 59 treatments, yielding 3,470 PDX experiments across 5 tumor types, the method improves on an empirical Tumor Growth Inhibition baseline: reconstruction reaches $R^2=0.96$ versus $0.71$ and Spearman $0.96$ versus $0.86$; under a 7-day observation window, future dynamics prediction improves from $23.3\%\pm 5.2$ without the graph encoder to $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$0 with it (Bazgir et al., 2023).

Personalized image-based progression prediction has also been framed as a TumorPred workflow. "Personalized Pancreatic Tumor Growth Prediction via Group Learning" trains AlexNet on three-channel patches composed of SUV, ICVF, and tumor mask, combines the learned outputs with time interval, tumor-level geometry, and patient-level clinical factors, and then personalizes a linear-kernel SVM on the target patient. On a pancreatic neuroendocrine tumor cohort of seven patients, it reports Dice $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$1 and RVD $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$2, outperforming a previous model-based method with Dice $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$3 and RVD $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$4 (Zhang et al., 2017).

For irregular follow-up intervals, "Vestibular schwannoma growth prediction from longitudinal MRI by time conditioned neural fields" represents each tumor by a signed distance function conditioned on a latent grid and propagates the latent state with a time-conditioned 3D ConvLSTM. The temporal encoding is

$$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$5

with empirical best performance at $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$6. Across 131 vestibular schwannoma patients in 5-fold cross-validation, DeepGrowth reaches Dice $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$7 and $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$8 HD $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$9 mm, versus Dice $V_0$0 and $V_0$1 HD $V_0$2 mm for the best baseline; in the top $V_0$3 most changing tumors, the improvement is at least $V_0$4 Dice score and at least $V_0$5 mm $V_0$6 Hausdorff distance (Chen et al., 2024).

Longitudinal prediction can also be cast statistically rather than with deep latent dynamics. "Mixed-Effect Modeling for Longitudinal Prediction of Cancer Tumor" models lung cancer trajectories with patient-specific random effects:

$V_0$7

and explores both linear and quadratic time effects on volume, mean Jacobian, and variance of the Jacobian. The study uses 19 survived and 19 deceased patients, with leave-one-out cross validation, and concludes that mixed-effect modeling has superior performance on some extracted features and similar or worse performance on the others (Nasiri et al., 2018).

3. Spatial heterogeneity, localization, and microscopic spread

TumorPred’s spatial heterogeneity branch is exemplified by "Block Graph Neural Networks for tumor heterogeneity prediction." That framework begins with an agent-based, continuous-space, continuous-time stochastic birth–death process, defines subclonal heterogeneity by normalized entropy

$V_0$8

and converts thin biopsy-like cuts into cell graphs with 10-nearest-neighbor connectivity. Classification uses a cut-off $V_0$9, with patches in $R^2 > 0.95$0 discarded to enhance class separability. The dataset contains 200 tumors total, split into 160 train, 20 validation, and 20 test, with 6 cuts per tumor and 100 candidate patch centers per cut. Node features include local intensity, local density, birth and death indicators, Voronoi cell volume, and local birth and death intensities; the BGNN itself comprises a dense embedding block with GraphNorm, 3 stacked Graph Attention layers with GELU, and global average pooling. The reported best test accuracy is $R^2 > 0.95$1 on the small-dataset, 4-head variant, while the full-feature configuration reaches $R^2 > 0.95$2 on the main feature-ablation setup (Njifon et al., 8 Feb 2025).

Whole-slide localization extends the same spatial logic to digital pathology. "A Lightweight Multi-Cancer Tumor Localization Framework for Deployable Digital Pathology" trains DenseNet169 on 79,984 non-overlapping 224×224 tiles from melanoma, hepatocellular carcinoma, colorectal cancer, and non-small cell lung cancer, using PathML for tiling, Macenko normalization, OpenCV-based artifact filtering, and a >70% tissue-content threshold. The model achieves tile-level ROC-AUC $R^2 > 0.95$3, F1 $R^2 > 0.95$4, sensitivity $R^2 > 0.95$5, and specificity $R^2 > 0.95$6 on validation data from the four training cancers, and ROC-AUC $R^2 > 0.95$7, F1 $R^2 > 0.95$8, sensitivity $R^2 > 0.95$9, and specificity $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$0 on an independent pancreatic ductal adenocarcinoma cohort of 7,346 tiles. A slide-level probability grid is then smoothed by Gaussian filtering, thresholded at $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$1, and exported as heatmaps and contours compatible with QuPath (Isett et al., 9 Mar 2026).

Microscopic spread beyond radiologically visible tumor has been modeled explicitly with statistical physics. "Ising energy model for the stochastic prediction of tumor islets" places tumor and healthy voxels on a 3D lattice with the Hamiltonian

$\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$2

and defines three elementary stochastic events—jump, duplicate, and die—with propensities

$\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$3

Parameters are fit with CMA-ES to histologic data from 23 prostate cancer patients, yielding $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$4, $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$5, $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$6, and $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$7. The model reproduces islet median diameters in the range 1–3.5 mm versus clinical 1.5–4 mm, predicts that approximately $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$8 of patients have no islets versus approximately $\frac{dy(t)}{dt}=f_\theta(y(t),\beta)$9 clinically, and estimates sub-volume involvement by Monte Carlo probabilities

$R^2=0.96$0

Its principal mechanistic claim is that the Ising interaction term acts as a surface tension, producing regular, approximately spherical islets (Amoudruz et al., 28 Aug 2025).

4. Prognosis, response classification, and clinical stratification

A prognostic TumorPred formulation based on handcrafted MRI descriptors is given by "Brain Tumor Survival Prediction using Radiomics Features." Using BraTS 2019 data and 90 radiomic features per tumor-containing 2D slice across four modalities, the study frames overall survival as a three-class classification problem and evaluates five standard classifiers. On 166 high-grade glioma subjects, the Random Forest classifier achieves accuracy $R^2=0.96$1, precision $R^2=0.96$2, and recall $R^2=0.96$3, outperforming SVM, Decision Tree, Discriminant Analysis, and k-NN. Haralick texture features are identified as the most important predictors, with GLCM contrast, inertia, entropy, correlation, and homogeneity dominating first-order and shape descriptors (Yousaf et al., 2020).

A less successful but conceptually important route is the cascaded segmentation-plus-transfer-learning strategy of "Survival prediction using ensemble tumor segmentation and transfer learning." This pipeline first segments whole tumor, tumor core, and enhancing tumor with a cascaded ensemble of 3D CNNs, then combines subregion volumes, clinical age, and VGG-16-derived image features from 20 tumor-centered slices to regress overall survival in days. The segmentation stage attains validation Dice $R^2=0.96$4 for whole tumor, $R^2=0.96$5 for tumor core, and $R^2=0.96$6 for enhancing tumor, but the survival regressor remains weak: training-set accuracy is $R^2=0.96$7 with Spearman $R^2=0.96$8, and validation accuracy is $R^2=0.96$9 with Spearman $0.71$0 (Cabezas et al., 2018).

Response classification can also be derived from continuous dynamics. In the heterogeneous GCN plus Neural-ODE framework, predicted trajectories are thresholded by mRECIST, while the graph encoder embeddings can be passed to a three-layer MLP for categorical response prediction. In 5-fold cross-validation, the pretrained GCN attains balanced accuracy $0.71$1, AUROC $0.71$2, and F1 $0.71$3, outperforming the non-pretrained GCN, MLP, and Random Forest baselines. The paper further states that mRECIST category prediction improves when heterogeneous graph embeddings are included across all observation windows (Bazgir et al., 2023).

Taken together, these studies show that TumorPred-style prognostic systems range from interpretable radiomics and segmentation-derived covariates to multimodal continuous-time latent models. Their clinical value lies less in any single metric than in how they couple prediction targets—survival, response category, or growth trajectory—to image- and biology-derived intermediates that can be inspected or recalibrated.

5. Molecular and structural tumor profiling

TumorPred has also been extended from geometry and prognosis to latent molecular state. "Towards computational fluorescence microscopy: Machine learning-based integrated prediction of morphological and molecular tumor profiles" uses bag-of-words histology features, kernel SVMs, and Layer-wise Relevance Propagation to infer protein expression/phosphorylation, gene expression, copy number variation, DNA methylation, and somatic mutation states from breast cancer microscopy. The TCGA cohort contains approximately 565 patients with matched H&E images and molecular profiles, while the in-house B-CIB morphology resource contains more than 200,000 expert-annotated cells across more than 1,000 slides and TMAs. The study reports 19/190 predictable proteins with balanced accuracy up to $0.71$4, 5,274/24,775 significant CNV endpoints with balanced accuracy from $0.71$5 to $0.71$6, 7,076/20,530 significant RNASEQ endpoints with balanced accuracy from $0.71$7 to $0.71$8, and 5,946/19,953 significant methylation endpoints with balanced accuracy from $0.71$9 to $0.96$0. Among somatic mutations, CDH1/E-Cadherin reaches balanced accuracy $0.96$1, TP53 reaches $0.96$2, GATA3 reaches $0.96$3, and PIK3CA reaches $0.96$4 (Binder et al., 2018).

A structurally distinct TumorPred formulation appears in "An Integrated Deep Learning and Dynamic Programming Method for Predicting Tumor Suppressor Genes, Oncogenes, and Fusion from PDB Structures." Here, experimentally determined X-ray PDB structures are converted into 24 orthogonal projections with 21 biochemical channels per projection, processed by a "Brain inception residual" DCNN with approximately 2,925,751 trainable parameters, and then aggregated to the gene level by dynamic programming over sequence-coverage intervals. In Approach 1, after cleaning, the pooled dataset contains 1,031 ONGO, 706 TSG, and 371 Fusion PDBs; the DCNN stage reaches AUROC $0.96$5 for OG vs TSG, while the final gene-level classification reaches AUROC $0.96$6. In Approach 2, where all PDBs for a gene are kept together in train or test, the OG vs TSG AUROC drops to $0.96$7 at the DCNN stage and $0.96$8 at the final gene level (Anandanadarajah et al., 2021).

These systems do not predict tumor burden directly. Instead, they convert morphology or structure into probabilistic statements about tumor-associated biology. A plausible implication is that TumorPred can function not only as a trajectory predictor but also as a bridge between observable tissue architecture and otherwise latent molecular programs.

6. The R/Shiny TumorPred application for CNS pharmacokinetics

The most explicit software realization of TumorPred is the R/Shiny web application introduced in "TumorPred: A Computational Framework Implemented via an R/Shiny Web Application for Parameter Estimation and Sensitivity Analysis in Compartmental Brain Modeling." It implements a permeability-limited four-compartment brain model with compartments for brain blood/vascular space, brain mass, cranial CSF, and spinal CSF, driven by a user-supplied plasma concentration–time profile. TumorPred is built in R with Shiny, uses the LSODA ODE solver via deSolve and Differential Evolution via DEoptim, allows users to visualize and download simulated plots and data tables, and computes compartment-specific $0.96$9, $0.86$0, and AUC (Wickramasinghe et al., 5 Sep 2025).

The state equations are mass-balance ODEs over $0.86$1, $0.86$2, $0.86$3, and $0.86$4 with flows $0.86$5 and $0.86$6, permeability–surface area terms $0.86$7 and $0.86$8, partition coefficients $0.86$9 and $23.3\%\pm 5.2$0, and elimination or reabsorption terms such as $23.3\%\pm 5.2$1. Parameter estimation minimizes the variance-weighted least-squares criterion

$23.3\%\pm 5.2$2

while local sensitivity is represented by

$23.3\%\pm 5.2$3

The app parameterizes 27 quantities in the 4-compartment model and supports fixing a subset of parameters while estimating others to mitigate identifiability issues (Wickramasinghe et al., 5 Sep 2025).

Validation is performed against the Simcyp Simulator using a sample oral abemaciclib 10 mg dataset. When six parameters were estimated—$23.3\%\pm 5.2$4, $23.3\%\pm 5.2$5, $23.3\%\pm 5.2$6, $23.3\%\pm 5.2$7, $23.3\%\pm 5.2$8, and $23.3\%\pm 5.2$9—the reported absolute errors were $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$00, $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$01, $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$02, $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$03, $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$04, and $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$05, respectively. The study further reports that brain mass concentration $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$06 was highly sensitive to $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$07, denoted PSB in the app figure (Wickramasinghe et al., 5 Sep 2025).

Within the TumorPred literature, this application is notable because it replaces image- or graph-based tumor-state inference with mechanistic CNS exposure modeling. Its central prediction target is not tumor size or class label but drug concentration in brain-relevant compartments when direct measurement is difficult or infeasible.

7. Reporting gaps, assumptions, and future directions

Across the TumorPred literature, predictive performance is often stronger than methodological reporting depth. The hybrid XGBoost plus growth-model study reports $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$08 but does not specify exact sample sizes, train/validation/test splits, hyperparameters, RMSE, MAE, confidence intervals, or the precise parameter estimation procedure linking XGBoost to $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$09 and $$\frac{dV}{dt}=rV\left(1-\frac{V}{K}\right),\qquad V(t)=\frac{K}{1+\left(\frac{K-V_0}{V_0}\right)e^{-rt}},$$10; it also treats zone-specific ray therapy conceptually rather than through spatial PDEs, segmentation-based dosing strategies, or dose-volume rules (Chattopadhyay et al., 27 May 2025). The heterogeneous GCN plus Neural-ODE framework does not provide the exact ODE solver, adjoint details, or hyperparameters, and reports no explicit uncertainty quantification or calibration analyses (Bazgir et al., 2023). The BGNN heterogeneity framework is trained exclusively on synthetic data and does not report precision, recall, F1, ROC-AUC, or calibration, while explicitly identifying synthetic-to-real domain shift as a central deployment problem (Njifon et al., 8 Feb 2025).

Several TumorPred pipelines are also constrained by small or specialized datasets. The pancreatic group-learning model is built on seven patients; the BraTS 2018 survival regressor is evaluated on 59 training and 28 validation cases; the radiomics-based glioma survival classifier treats some subjects with missing survival information as low survival; and DeepGrowth is trained on exactly three consecutive scans per vestibular schwannoma patient (Zhang et al., 2017, Cabezas et al., 2018, Yousaf et al., 2020, Chen et al., 2024). These constraints do not invalidate the methods, but they restrict claims about generalization across institutions, cancer types, and acquisition protocols.

Reported future work converges on several themes. For tumor-dynamics prediction, the proposed extensions include validation across diverse cancer types, incorporation of additional biological parameters such as immune response and vascular density, spatial PDEs or reaction-diffusion models for true zone-specific ray therapy planning, multi-compartment models, Gompertz or reduced Gompertz fits, Bayesian inference for parameter estimation and uncertainty quantification, and rigorous external validation with standardized datasets (Chattopadhyay et al., 27 May 2025). For spatial heterogeneity and digital pathology, the reported directions include domain adaptation, self-supervised pretraining, multi-modal alignment with spatial transcriptomics and Ki-67 or apoptosis markers, threshold calibration to clinical endpoints, stain-invariant representations, uncertainty estimation, and broader multi-institution validation (Njifon et al., 8 Feb 2025, Isett et al., 9 Mar 2026). For the pharmacokinetic TumorPred application, planned enhancements include tumor-specific compartments and spatial heterogeneity, active transport and nonlinear binding, population variability and covariate models, global sensitivity methods, uncertainty quantification, and additional PK metrics (Wickramasinghe et al., 5 Sep 2025). This suggests that TumorPred is increasingly oriented toward calibrated, multimodal, and deployable computational oncology systems rather than isolated predictive models.