Cancer remains a leading cause of death globally, despite significant advancements in treatment. One key challenge in oncology clinical trials is the precise timing of on-treatment biopsies, which are crucial for verifying a drug's mechanism of action and informing treatment strategies. Currently, biopsy timing relies on expert guesswork, lacking a quantitative, scientific basis. This study addresses this limitation by developing a quantitative method integrating digital pathology and mathematical disease modeling. The approach uses spatial biomarker data from immunohistochemistry (IHC) images, which have shown prognostic value for cancer recurrence and treatment response. While statistical and machine-learning methods have been applied to these data, this study employs a mechanistic, agent-based model (ABM) to leverage the spatiotemporal resolution of paired biopsy samples. The ABM, implemented in Matlab and based on previous work by Kather et al., simulates tumour-immune cell interactions on a grid. Previous ABMs focused on qualitative validation, while this study uses paired pretreatment and on-treatment biopsy data for quantitative validation, addressing the dynamic component. The data originates from phase 1 clinical trials of Simlukafusp, an immunomodulatory molecule. A novel spatial agreement measure (SAM) is introduced to compare simulated and observed biopsy samples quantitatively. The model uses the spatial distribution of immune cells from baseline biopsies to predict the on-treatment distribution, aiming to identify the optimal time for biopsy collection and to assist in the design of combination therapies and personalized treatment strategies.

The introduction cites several studies demonstrating the prognostic value of spatial biomarkers derived from IHC images in predicting cancer outcomes and treatment responses. These studies utilize statistical analysis and machine learning techniques. The authors acknowledge the existence of agent-based models representing tumour-immune interactions, noting that their validation has been largely qualitative. They highlight the novelty of their approach in using paired clinical samples for quantitative validation of the model's dynamic components. The study also mentions the lack of a widely accepted comparative measure for spatial cell distributions in clinical images, leading to the development of the SAM.

The study uses data from two clinical trials of Simlukafusp (NCT03063762 and NCT02627274), focusing on 44 patients with paired pre- and on-treatment biopsies suitable for analysis. An in-house machine learning algorithm, trained by a pathologist, analyzes digitized IHC slides to identify and locate CD8+ T-cells and tumour cells. This spatial information is mapped onto a grid for input into the ABM. The ABM simulates cell proliferation, migration, killing, and death based on probabilistic rules and local neighbourhood interactions. The model is optimized and validated using a novel spatial agreement measure (SAM) based on the radial distribution function (RDF). The SAM quantifies the overlap between RDFs derived from observed and simulated biopsies. A secondary metric, VarSAM, weights the agreement in the immediate vicinity of CD8 cells. A sensitivity analysis identifies key model parameters influencing CD8 and tumour cell dynamics. The model optimization process involves scanning a parameter space, selecting parameter sets meeting predefined SAM and VarSAM thresholds, and validating the selected parameter sets on a held-out test dataset. The last observation carried forward (LOCF) method provides a null benchmark against which the ABM's predictive accuracy is compared. The model’s application in clinical protocol design and combination therapy assessment is demonstrated.

The model optimization identified 12 parameter sets that reliably reproduced the spatial features of on-treatment biopsies, achieving a mean accuracy of 77% in validating on a held-out test dataset compared to 35% with randomly selected parameters and 41% with the LOCF method. This highlights the predictive power of the model, demonstrating that the spatial distribution of CD8 cells at baseline is predictive of on-treatment distribution. The sensitivity analysis revealed that parameters related to immune cell dynamics (proliferation, death, killing, migration, and influx) exhibited the highest sensitivity. The model predicted that biopsies taken around 30 days post-treatment would capture maximum CD8 infiltration. Simulations of hypothetical combination therapies showed differing effects on CD8 cell numbers and spatial distribution, depending on whether the combination partner increased CD8 proliferation or influx. The study demonstrates the ability of the model to provide quantitative information to support clinical decision-making regarding biopsy timing, the design of combination therapies, and personalization of treatments. The study shows how spatial features vary widely based on initial conditions, proposing model application in a personalised fashion to predict the optimal time point for biopsy scheduling for individual patients.

The study demonstrates a successful integration of digital pathology and mathematical modelling to predict spatial biomarker dynamics in cancer immunotherapy. The high accuracy (77%) of the model in predicting on-treatment CD8+ T-cell distributions from baseline data is noteworthy. This suggests that the spatial arrangement of immune cells at baseline is a powerful predictor of treatment response. The ability to predict the optimal timing of biopsies can significantly improve the efficiency and cost-effectiveness of clinical trials. The model’s application to hypothetical combination therapies provides insights into how different therapeutic strategies might impact immune cell dynamics. This highlights the potential use of this approach for preclinical investigation and optimization of novel combination treatments. The development of the SAM provides a useful tool for comparing spatial biomarker data across different sources. This methodology can be applied broadly in research involving spatial analysis of digital pathology images. The limitations of the study are discussed below, alongside considerations for future improvements and expansion of the model’s capabilities.

This study presents a novel and robust method for predicting spatial biomarker dynamics in cancer immunotherapy using an integrated digital pathology and mathematical modeling approach. The high predictive accuracy of the model demonstrates its potential to optimize clinical trial design, particularly concerning biopsy scheduling. Furthermore, the model can facilitate the exploration of combination therapies and inform personalized treatment strategies. Future work should focus on expanding the model to incorporate additional factors, such as drug concentration and patient-specific characteristics, to further enhance its predictive power and clinical utility.

The study acknowledges limitations concerning the relatively small sample size due to the constraints of early-phase clinical trials and the challenges associated with obtaining sufficient tissue from needle biopsies. The sensitivity analysis was not exhaustive and might not fully capture parameter co-dependencies. The model’s computational complexity limited the parameter space exploration during optimization. The current model does not account for drug concentration in the tumour, which would enhance its biological realism. Finally, while the SAM is a novel approach, further validation across larger and more diverse datasets is warranted.