Renal cell carcinoma (RCC) is more common in men than women, and emerging evidence shows that biological sex influences not just who gets kidney cancer, but how the tumor behaves, how the immune system responds, and how well various treatments work.
Sex hormones play a central but often overlooked role in shaping the immune system's anti-cancer activity. Estrogen tends to boost immune responses by activating T cells, NK cells, and dendritic cells, while testosterone generally suppresses immune activity, potentially creating an environment more favorable to tumor growth.
Existing treatments for advanced RCC include immune checkpoint inhibitors (ICIs) that help the immune system recognize and attack cancer cells, and tyrosine kinase inhibitors (TKIs) that block signals driving tumor blood vessel growth and proliferation. Real-world data suggest that males and females may respond differently to these therapies.
Despite growing awareness of sex differences in cancer, most computational models used to study tumor-immune dynamics do not incorporate sex hormones, immune cell composition, or tumor genetic evolution simultaneously. This study set out to fill that gap by building a model that reflects biological sex as a core feature.
The research team developed an Agent-Based Learning Model (ALM), a type of computational simulation in which individual cells, called agents, behave according to biological rules and interact with each other and their environment. This approach can capture complex, emergent behaviors that simpler mathematical models cannot.
The model includes multiple types of immune agents: CD8+ cytotoxic T cells, NK cells, regulatory T cells (Tregs), and dendritic cells (DCs). It also models tumor cells that can mutate over time to evade immune attack, mimicking the real-world process of tumor evolution under immune pressure.
Sex-specific effects are encoded through hormone-sensitive mechanisms that adjust cell activation rates, apoptosis probabilities, antigen recognition thresholds, and killing efficiency in response to local estrogen or testosterone levels. These are modeled as three-level thresholds (low, normal, high) applied to all sex hormones.
The model also incorporates effects of ICIs (which block PD-L1/PD-1 interactions and reawaken exhausted T cells) and TKIs (which reduce tumor proliferation and angiogenesis and prevent T cells from being converted into immunosuppressive Tregs), mirroring how these drugs work in real patients.
Parameters were optimized against real clinical data from the ARON dataset using Optuna, an automated hyperparameter tuning framework. This adaptive learning component allows the model to iteratively refine itself to better match actual patient survival outcomes.
Each virtual tumor cell in the model carries an internal genetic representation organized into gene segments corresponding to real biological genes known to drive cancer. Each gene has a promoter sequence (governing expression level) and a gene sequence (governing protein function), mirroring actual molecular biology.
Tumor cells can accumulate mutations during simulation, affecting key cancer-relevant processes including antigen presentation (controlled by MHC genes, JAK1, JAK2, B2M), checkpoint signaling (CD274/PD-L1), DNA repair (BRCA1/2), cell cycle control (TP53, RB1, CDKN2A), and proliferation drivers (MYC, RAS, EGFR, HER2).
The fitness of each tumor cell is determined by how well its mutations help it survive. Mutations that improve immune evasion or prevent cell death increase a tumor cell's chance of reproducing, while mutations that impair proliferative signaling (such as EGFR mutations in this context) reduce fitness and are gradually eliminated.
This genetic algorithm within the model is described as real-time learning, because the tumor population adapts to immune pressure during a single simulation run, just as real tumors evolve under treatment in patients over time. This makes the model considerably more realistic than static tumor representations.
Simulations consistently showed distinct treatment response patterns between female and male models. Female models initially responded more slowly to ICI-TKI combination therapy, showing weaker immediate immune engagement following treatment initiation.
However, over the longer term, female models mounted a stronger immune response with sustained therapy support. When immune stimulation was maintained, particularly with TKI co-treatment, female simulations showed more rapid tumor regression, less immune cell exhaustion, and a higher likelihood of immune system restoration.
Male models showed more stable early responses to treatment, with less volatile tumor progression and relatively consistent efficacy especially with ICI monotherapy. However, tumor cells in male simulations demonstrated greater long-term resilience, driven by more frequent and advantageous genetic adaptations over time.
The ICI component was particularly critical in male simulations: removing ICIs from male treatment regimens often led to significantly diminished treatment effectiveness, suggesting that the male immune system, while initially more engaged, requires sustained checkpoint blockade to overcome androgen-driven T cell exhaustion mechanisms.
These sex-specific response patterns are described as exploratory and hypothesis-generating. They align with known biology around estrogen's delayed but strong immune boosting effects and testosterone's immunosuppressive effects on CD8+ T cells, but require further empirical validation.
Using data from seven RCC patients in the ARON dataset (3 males, 4 females), the researchers created virtual patient models by mapping clinical characteristics such as immune cell counts and body mass index to simulation parameters.
The optimization procedure ran 26 trials with each fitness function (concordance index and mean squared error), iteratively adjusting learnable parameters to better match real patient survival outcomes. This adaptive learning process showed measurable improvement over manually tuned default parameters.
For several patients, such as p4 and p5, the trained model showed better agreement between simulated and observed survival outcomes compared to the untrained baseline. This confirms that even with a very small clinical dataset, data-driven tuning can meaningfully improve model behavior.
The researchers observed increased variability in predictions for some patients following training, which they attribute to biological heterogeneity among patients rather than a model failure. This highlights the importance of domain expertise in selecting which parameters should be allowed to vary during optimization.
Real-world clinical data confirm that men have a higher incidence of RCC and more aggressive disease at presentation. Immune profiling studies suggest that male patients have higher CD8+ T cell infiltration but these cells are often deeply exhausted, driven by androgen receptor (AR) signaling that promotes T cell dysfunction.
Research in mouse models and human cohorts has shown that AR signaling inhibits cytotoxic T cell function and promotes exhaustion-like states, suggesting that combining androgen-blocking therapies with PD-1/PD-L1 checkpoint inhibitors could synergize to improve outcomes in male RCC patients specifically.
Female patients show greater baseline immune competence but also a higher tendency toward immune-related adverse events from immunotherapy, suggesting that sex-specific dosing or combination strategies may be needed to balance efficacy and tolerability.
The model also incorporates BMI as a proxy for metabolic state, acknowledging that adipose tissue and its hormone-modulating effects represent another dimension of sexual dimorphism influencing immunotherapy outcomes. Obesity-related inflammation and adipokine signaling differ by sex and may further shape treatment responses.
This study represents one of the first agent-based models specifically designed for RCC that simultaneously incorporates sex hormone signaling, tumor genetic evolution, drug pharmacodynamics, and immune cell interactions. It positions computational simulation as a tool for exploring hypotheses that would be impossible to test with conventional clinical trials alone.
A key strength is the model's adaptive learning pipeline, which treats clinical data as a driver of iterative model refinement rather than a simple validation benchmark. This approach aligns with broader trends in the field toward personalized computational patient twins.
The study's primary limitation is the small size of the ARON dataset subset used for training: only 7 patients. This constrains the statistical power of adaptive learning and limits confidence in the generalizability of optimized parameters. Larger, more detailed datasets with longitudinal immune profiling and hormone measurements are needed.
Future research directions include expanding the model with prospective clinical data that include sex-specific immune profiles and hormone levels, validating findings in larger multicenter cohorts, and eventually using the ALM as a virtual clinical trial platform to design sex-stratified treatment studies before conducting expensive real-world trials.
This study demonstrates that biological sex is not just a demographic variable in kidney cancer but a fundamental biological factor that shapes tumor genetics, immune behavior, and treatment response in ways that computational models must explicitly capture to be useful for personalized medicine.
The agent-based learning model offers a flexible platform for generating testable hypotheses about how estrogen, testosterone, immune cell composition, and tumor genetic evolution interact under different treatment combinations. This kind of systems-level thinking is essential for moving beyond one-size-fits-all oncology.
For patients and families, the key insight is that your sex biology may matter as much as tumor staging or biomarker expression when predicting how well a particular cancer therapy will work. This underscores the importance of asking oncologists about sex-specific considerations in treatment planning.
While this model is not yet ready for clinical use, it represents a step toward a future where computational simulations run alongside clinical decision-making, helping oncologists identify which combination of therapies is most likely to work for a specific patient based on their unique biological profile.