Figure S3 describes bar plots for area normalized cell counts across SD and PD patients, pre and during treatment.

Figure S4 describes the statistical 'difference network analysis'

Supplementary Table 1 describes the classification statistics using LOOCV for SVM and NN SDM when using cells and quadrats.

Supplementary Table 2 shows different metastatic sites used for biopsy collection in this study.

Figure S8 describes bar plots for area normalized cell counts across SD and PD patients, pre and during treatment, for each cellular/quadrat neighborhood

Figure S7 displays example marker panels for a FoV from the 9-patient NSCLC cohort (FoVdata).

Figure S2 describes the image t-SNE generated using Mistic

Figure S5 shows the RoGM analysis for the non-spatial setting.

Multiplexed imaging of tissues is an approach that holds promise for improving early detection, diagnosis, and treatment of cancer. Here, we investigated multiplexed histological images of paired pre- and on-treatment samples from nine patients with immunotherapy-refractory non-small cell lung cancer (NSCLC) treated with an oral HDAC inhibitor (vorinostat) combined with a PD-1 inhibitor (pembrolizumab). Patient responses comprised of either stable disease (SD) or progressive disease (PD). An extensive multiplexed-image analysis pipeline involving both cell segmentation and quadrats, coupled with spatial statistics, machine learning, and deep learning was built to analyze the spatial and temporal features that predict disease progression and identify potential clinical biomarkers. Distinct spatial immune ecologies existed between SD and PD patients, and tumors from PD patients were already characterized by an immune-suppressive environment prior to treatment. Finally, the learned spatial ecologies predicted disease progression better than PD-L1 status alone, suggesting these ecologies could be used as potential companion biomarkers with PD-L1 in NSCLC. These findings will be investigated in a larger-cohort study generated from an ongoing clinical trial (NCT02638090) that includes a wider range of responses including complete and partial responders. Together, this study developed a computational infrastructure for analyzing multiplex imaging to predict immunotherapy response in NSCLC, which can potentially be generalized to any type of cancer.

Targeted therapies induce strong clinical responses but fail to eliminate advanced cancers, as a subset of tumor cells survives within residual disease and eventually develops resistance. While numerous cell-intrinsic and microenvironmental mechanisms have been implicated in this survival, their relative contributions remain poorly defined. Using spatial histological inferences from ALK + NSCLC models, we show that peristromal niches protect tumor cells from elimination, enabling in vivo persistence. This spatially restricted sheltering provides an ecological rescue mechanism that sustains residual populations, enabling their eventual evolutionary escape. Mechanistically, this protective effect reflects an integrated action of multiple juxtacrine and paracrine signals. This complexity limits the utility of targeting individual mechanisms of protection, favoring a shift towards exploiting orthogonal collateral sensitivities of residual disease. We find that adaptive HER2 upregulation, associated with both cell-intrinsic and stroma-mediated persistence, can be exploited by the antibody-drug conjugate T-DXd to dramatically enhance therapeutic responses and suppress tumor relapse.

Figure S6 depicts the NN SDM architecture

Figure S1 describes the image processing pipeline and convex hull analysis

Despite inducing strong and durable clinical responses, targeted therapies do not eliminate advanced cancers, as a subset of tumor cells survives within residual tumors, eventually developing resistance. The ability of tumor cells to avoid therapeutic elimination can be mediated both by cell-intrinsic and microenvironmental mechanisms. Whilst the specific molecular mediators of cell-intrinsic and microenvironmental resistance are well understood, their relative contribution to in vivo therapeutic responses remains poorly defined. Using spatial histological inferences from experimental models of ALK+ NSCLC, we found that peristromal niches protected tumor cells from therapeutic elimination in vivo, enabling in vivo persistence. Whereas the development of bona fide resistance is associated with the development of the development of cell-intrinsic resistance, relapse of tumor growth reflects a combined effect of both cell-intrinsic and microenvironmental mechanisms. Mechanistically, the protective effect of the peristromal niche is not reducible to a single mechanism, instead reflecting a combined effect of multiple juxtacrine and paracrine mediators. The lack of reducibility to a single molecular mediator presents an obvious challenge to the therapeutic paradigms of targeting individual resistance mechanisms. We found that this challenge could be mitigated by shifting the therapeutic focus to orthogonal collateral sensitivities of residual tumors. Exploiting adaptive upregulation of HER2, associated with both cell-intrinsic and microenvironmental persistence, using the antibody-drug conjugate T-DXd strongly enhanced the effect of targeted therapies and suppressed the development of resistance.

Our study examines tumor growth dynamics and immunosuppression under the presence of immune attack to identify optimal immune pulsing treatment strategies. While immunotherapy has shown remarkable success in treating many cancers like non-small cell lung cancer (NSCLC), not all patients respond well to immunotherapy. This mixed response can be attributed to the complexity of the tumor-immune system dynamics. TIL therapy is an emerging immunotherapy where activated T cells are injected into the patient. This therapy can fail due to tumor-induced immunosuppression, for example via the PDL1/PD1 axis. PDL1 expression has been studied and can increase under IFN-γ, released by activated T cells, but PD-L1 relaxation is not understood. We hypothesize that a) the regimen of TIL therapy administration should impact the dynamics of PDL1, and that b) there may be better strategies of therapy delivery that maximize tumor kill while minimizing immune suppression. We investigate these complex PD-L1 driven dynamics through a unique combination of in vitro studies and mathematical modeling. We have collected in vitro and temporal confocal microscopy images of PD-L1 expression on NSCLC cells, which were either untreated or treated for 48 hours with high dose IFN-γ followed by chronic low dose IFN-γ for 21 days. To reinforce our in vitro experiments and findings, we developed a hybrid agent-based model (ABM) to study PD-L1 dynamics. The agents in the ABM are tumor cells having variable PD-L1 expression, and immune cells that secrete IFN-γ and can kill tumor cells. Tumor cells can be randomly- or cluster-seeded. IFN-γ is modeled as either a binary well-mixed pulse (to simulate the in vitro experimental setup), or as a diffusible via a reaction-diffusion process (to simulate the tumor-immune cell interactions in an in vivo spatial manner). We model the TIL therapy as a pulse of immune cells, given at regular or irregular intervals and with different quantities of cells. Our ABM is calibrated to capture in vitro data dynamics. Under certain combinations of spatial configurations of the tumor and intermittent immune dosing schedules, we observe the presence of ‘sweet spots’ where tumor extinction is possible and immune exhaustion is avoided. This indicates that an appropriate pulsing of TIL therapy may lead to better overall immune efficacy than a bolus injection or continuous immunotherapy, by preventing sustained suppression that increases immune cell exhaustion, despite the potential for tumor regrowth between pulses. We have built an ABM to study tumor-immune evolution under various pulsing strategies of TIL therapy. Our novel findings can potentially benefit clinical cancer research by giving multiple insights related to the tumor extinction, equilibrium and escape phenomena, and lead to improved schedules in patients receiving immunotherapy. Sandhya Prabhakaran,Mark Robertson-Tessi,Kimberly Luddy,Rafael R. Bravo,Taylor M. Bursell,ulian Pineiro,Jeffrey West,Megan Johnson,Jhanelle A. Gray,Amer A. Beg,Scott J. Antonia,Robert A. Gatenby,Alexander R. Anderson. Evolutionary immunotherapy in NSCLC: identifying optimal dosing strategies in TIL therapy using agent-based modeling [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3696.

Several methods for cell cycle inference from sequencing data exist and are widely adopted. In contrast, methods for classification of cell cycle state from imaging data are scarce. We have for the first time integrated sequencing and imaging derived cell cycle pseudo-times for assigning 449 imaged cells to 693 sequenced cells at an average resolution of 3.4 and 2.4 cells for sequencing and imaging data respectively. Data integration revealed thousands of pathways and organelle features that are correlated with each other, including several previously known interactions and novel associations. The ability to assign the transcriptome state of a profiled cell to its closest living relative, which is still actively growing and expanding opens the door for genotype-phenotype mapping at single cell resolution forward in time.

Glioblastoma (GBM) is the most aggressive form of primary brain tumor. The infiltrative nature of GBM makes complete surgical resection impossible. The selective forces that govern gliomagenesis are strong, shaping the composition of tumor cells during the initial progression to malignancy with late consequences for invasiveness and therapy response. Here, we developed a mathematical model that incorporates ploidy level and the nature of the brain tissue microenvironment to simulate the growth and invasion of GBM and used the model to make inferences about GBM initiation and response to the standard-of-care treatment. The spatial distribution of resource access in the brain was approximated through integration of in silico modeling, multiomics data, and image analysis of primary and recurrent GBM. The in silico results suggested that high-ploidy cells transition faster from oxidative phosphorylation to glycolysis than low-ploidy cells because they are more sensitive to hypoxia. Between surgeries, simulated tumors with different ploidy compositions progressed at different rates; however, whether higher ploidy predicted fast recurrence was a function of the brain microenvironment. Historical data supported the dependence on available resources in the brain, as shown by a significant correlation between the median oxygen levels in human tissues and the median ploidy of cancers that arise in the respective tissues. Taken together, these findings suggest that the availability of metabolic substrates in the brain drives different cell fate decisions for cells with different ploidy, thereby modulating both gliomagenesis and GBM recurrence.Significance: Ploidy viewed in the context of the resources in the microenvironment has the potential to inform whether modulation of energetic availability can delay tumor progression and could help guide clinical decision making.

The National Cancer Institute (NCI) supports numerous research consortia that rely on imaging technologies to study cancerous tissues. To foster collaboration and innovation in this field, the Image Analysis Working Group (IAWG) was created in 2019. As multiplexed imaging techniques grow in scale and complexity, more advanced computational methods are required beyond traditional approaches like segmentation and pixel intensity quantification. In 2022, the IAWG held a virtual hackathon focused on addressing challenges in analyzing complex, high‐dimensional datasets from fixed cancer tissues. The hackathon addressed key challenges in three areas: (1) cell type classification and assessment, (2) spatial data visualization and translation, and (3) scaling image analysis for large, multi‐terabyte datasets. Participants explored the limitations of current automated analysis tools, developed potential solutions, and made significant progress during the hackathon. Here we provide a summary of the efforts and resultant resources and highlight remaining challenges facing the research community as emerging technologies are integrated into diverse imaging modalities and data analysis platforms.

Mathematical models have played a significant role in the development of current chemo- and radiotherapy treatment protocols. The widespread use of cytotoxic drugs has shaped the paradigm of uniformly administering a "maximum tolerated dose" to patients; however, this approach fails to account for the dynamic and heterogeneous nature of challenging cancers, including metastatic disease. Recent clinical trials and regulatory decisions have aimed to address these issues by integrating mathematical models to drive preclinical experiments and personalize treatment schedules. By capturing mechanisms of dose-response dynamics, ecological dynamics such as tumor-immune interactions or competition dynamics, and evolutionary dynamics across different therapeutic regimens, mathematical models hold the potential to advance current therapeutic strategies. As more preclinical and clinical data become available, the integration of mathematical models with "virtual patient" frameworks, including "digital twins," and artificial intelligence methods could further advance the mechanistic complexity and decision support capabilities of such models. Nonetheless, translating mechanistic models to routine clinical workflows will require overcoming current translational barriers, notably access to clinical data in standardized formats and regulatory constraints. Overall, recent trials demonstrate the promise of the field of mathematical oncology in translating predictive dynamics into treatment decision-making beyond the "maximum tolerated dose" approach. This article is part of a special series: Driving Cancer Discoveries with Computational Research, Data Science, and Machine Learning/AI .

Abstract Purpose: Our study examines tumor growth dynamics and immunosuppression under the presence of immune attack to identify optimal immune pulsing treatment strategies. Motivation: Checkpoint inhibitors disrupting the PD-1/PD-L1 axis are used in multiple cancer subtypes. However, not all patients respond, and predictive biomarkers are needed to better stratify patients. Therapy for PD-L1, the most studied biomarker, has given mixed results in non-small cell lung cancer (NSCLC). The complexity of the immune response and plastic nature of PD-L1/PD-1 expression on tumor and immune cells contributes to the heterogeneity of success. IFN-γ, a cytokine released by activated immune cells, induces PD-L1 expression in some tumor cells. Chronic IFN-γ results in long-term PD-L1 expression. However, relaxation dynamics following removal of IFN-γ remains unexplored. In this study, we investigate these complex PD-L1 dynamics through a unique combination of in vitro studies and mathematical modeling. METHODS: In vitro experimental data: We have collected in vitro and temporal confocal microscopy images of PD-L1 expression on NSCLC cells, which were either untreated or treated for 48 hours with high dose IFN-γ followed by chronic low dose IFN-γ for 21 days.Agent-based model: To reinforce our in vitro experiments and findings, we developed a hybrid agent-based model (ABM) to study PD-L1 dynamics. The agents in the ABM are tumor cells having variable PD-L1 expression, and immune cells that secrete IFN-γ. Tumor cells can be randomly seeded or seeded as a cluster of cells. IFN-γ is modeled as either a binary well-mixed pulse (to simulate the in vitro experimental setup), or as a diffusible via a reaction-diffusion process (to simulate the tumor-immune cell interactions in an in-vivo spatial manner). This allows us to investigate the effects of various pulsing strategies of immunotherapy on tumor evolution where each pulse can be a variable dose of immune cells, given at regular or irregular intervals. RESULTS: Our in vitro experimental data and microscopy images confirm that PD-L1 expression continues to increase within the cells after a 48-hour IFN-γ exposure followed by low levels of IFN-γ. Our ABM has been calibrated to capture in vitro data dynamics for different conditions. Our initial results indicate that pulsing immunotherapy may lead to better overall immune efficacy than continuous immunotherapy, by preventing sustained suppression that increases immune cell exhaustion, despite the potential for tumor regrowth between pulses. We also hypothesize that pulsing immunotherapy will have different immunosuppression effects when tumors are clustered or randomly distributed. CONCLUSION: We have quantified the temporal dynamics of PD-L1 expression in NSCLC cells under different conditions of IFN-γ, and built an ABM to study tumor-immune evolution under various pulsing strategies. The results of this work could lead to improved therapy schedules in patients responsive to checkpoint inhibitors that exploit the complex dynamics of immune response and suppression. Citation Format: Sandhya Prabhakaran, Taylor M. Bursell, Kimberly Luddy, Julian Pineiro, Rafael Bravo, Jeffrey West, Megan Johnson, Mark Robertson-Tessi, Amer A. Beg, Jhanelle E. Gray, Scott Antonia, Robert A. Gatenby, Alexander R. A. Anderson. Evolutionary immunotherapy in NSCLC: An integrated approach [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Translating Cancer Evolution and Data Science: The Next Frontier; 2023 Dec 3-6; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_2):Abstract nr A021.