Immunotherapy (checkpoint inhibitors, CAR‑T cells, cytokines like IL‑15 agonists, cancer vaccines) has shown that the immune system can eliminate even advanced cancers in some patients. But responses are highly variable:

• Only a subset of patients respond.
• Many develop resistance.
• Toxic immune‑related side effects can be severe.
• Choosing the right combination and dose is complex.

Advanced artificial intelligence (AI) is now being used to:

1. Design better immune‑based therapies (vaccines, CAR‑T/NK, antibody constructs).
2. Personalize them to each patient’s tumor and immune system.
3. Predict and monitor responses and toxicities.
4. Optimize combinations with chemo, radiation, targeted drugs.

In short, AI is shifting cancer immunotherapy from empirical trial‑and‑error toward a data‑driven, model‑guided engineering discipline.

2.1 AI‑Designed Personalized Cancer Vaccines

Goal: Train a patient’s T cells to recognize and attack their own tumor using vaccines that encode neoantigens—mutated peptides present on cancer cells but not on normal tissues.

What AI does:

1. Tumor sequencing and mutation calling
   • Next‑generation sequencing (NGS) identifies somatic mutations in a patient’s tumor and matched normal tissue.
2. Neoantigen prediction
   AI models analyze:
   • Which mutations create altered peptides.
   • Which of these will bind strongly to the patient’s specific HLA (MHC) molecules.
   • Which are likely to be processed and actually presented on tumor cell surfaces.
   • Which peptides are immunogenic (can trigger strong CD8+ and CD4+ T‑cell responses).
3. Vaccine design
   • Selected top neoantigens are encoded into an mRNA, DNA, peptide mix, or biomaterial‑based vaccine platform.
   • AI helps rank and combine up to ~20–34 neoantigens per vaccine to maximize breadth and potency.

Evidence and recent advances:

• A 2025–2026 wave of clinical studies shows personalized mRNA cancer vaccines combined with checkpoint inhibitors (e.g., KEYNOTE‑942 in melanoma) substantially reduce recurrence risk and boost neoantigen‑specific T‑cell responses compared to checkpoint inhibitors alone [1][2].
• Reviews on RNA‑based cancer vaccines highlight that individualized neoantigen mRNA vaccines can encode dozens of patient‑specific targets, with early trials reporting notable reductions in recurrence when combined with PD‑1 blockade [1][2].
• A 2026 Nature paper describes individualized neoantigen RNA vaccines using NGS plus AI prediction to select T‑cell targets, inducing durable T‑cell immunity in treated patients [3].
• A first‑in‑human trial of a personalized biomaterial‑based cancer vaccine (Harvard Wyss Institute) showed feasibility, safety, and immune activation, and is now moving toward combinations with checkpoint blockade [4].

How this “boosts” the immune system against cancer:

• Increases tumor‑specific T‑cell clones that can recognize tiny remnants of disease.
• Converts “immune‑cold” tumors into “hot” ones more responsive to checkpoint inhibitors.
• Provides immune memory, helping prevent relapse.

Actionable implications:

• For high‑risk or relapsed solid tumors (melanoma, certain lung, pancreatic, and colorectal cancers), AI‑designed neoantigen vaccines combined with checkpoint blockade are rapidly becoming a central research and early clinical strategy.

---

2.2 AI‑Guided Engineering of CAR‑T and CAR‑NK Cells

Goal: Genetically engineer a patient’s immune cells (T or NK cells) to express synthetic receptors (CARs) that target cancer cells with high specificity and potency.

Where AI helps:

1. Optimizing CAR design
   AI models analyze large datasets of CAR configurations and outcomes to:
   • Predict which antigen‑binding domains (scFvs) bind tumor antigens strongly without cross‑reacting with healthy tissue.
   • Tune signaling domains to balance activation, persistence, and safety (reducing exhaustion, CRS, neurotoxicity).
   • Identify sequence motifs linked to superior proliferation and survival.
2. Selecting targets and combinations
   • AI integrates tumor genomics and proteomics to find ideal target antigens (e.g., CD19, BCMA, EGFRvIII) and combinations (dual CARs) to reduce antigen escape.
   • Recent work describes AI‑guided CAR designs and pathway modulation strategies to enhance long‑term efficacy of CD19 CAR‑T cells [5].
3. In vivo and ex vivo engineering strategies
   • Emerging in vivo CAR‑T approaches use viral or non‑viral vectors to reprogram T cells directly in the patient’s body. AI helps select integration sites, vector doses, and control circuits to maximize safety and efficiency [6][7].
   • AI‑assisted CRISPR editing identifies optimal genomic loci for CAR insertion to avoid insertional mutagenesis and preserve native T‑cell function.

Recent developments:

• Reviews in 2025–2026 describe how AI is intersecting with CAR‑T development—guiding CAR construction, improving manufacturing, and predicting efficacy and safety [5][8].
• 2026 reports on in vivo site‑specific CAR engineering in patients highlight CRISPR‑based insertion of CAR constructs into specific T‑cell loci, reducing manufacturing time and potentially broadening access [7].
• Preclinical and early clinical data on next‑generation CAR‑T for solid tumors show designs that secrete IL‑12 or block PD‑L1 locally within the tumor microenvironment; these designs are informed by AI analysis of TME features [9].

How this boosts anti‑cancer immunity:

• Dramatically expands high‑affinity tumor‑specific effector cells.
• Overcomes some natural tolerance and exhaustion mechanisms that blunt endogenous responses.
• Potentially extends CAR‑based therapies from blood cancers into solid tumors by optimizing trafficking, persistence, and on‑target effects in the TME.

Actionable implications:

• For refractory hematologic malignancies and experimental solid tumor applications, AI‑assisted CAR‑T/NK design is a fast‑moving area that can improve both response rates and safety compared to earlier CAR generations.

---

2.3 AI‑Driven Optimization of Chemo‑Immunotherapy and Radio‑Immunotherapy

Goal: Use AI to find synergistic schedules and combinations of chemotherapy, radiotherapy, and immunotherapy that maximize immune activation and tumor death while minimizing systemic toxicity.

Key functions:

1. Modeling immunogenic cell death (ICD)
   • Certain chemotherapies and radiation exposures cause tumor cells to die in a way that releases danger signals and antigens, effectively acting as an in situ vaccine.
   • AI models integrate multi‑omics and clinical data to predict which regimens produce the strongest ICD and how they reshape “cold” tumors into “hot,” T‑cell–inflamed tumors [10].
2. Personalizing combination regimens
   • AI systems evaluate patient‑specific tumor genomics, TME signatures, and lab parameters to:
     • Predict benefit from chemo‑immunotherapy vs. immunotherapy alone.
     • Choose optimal drugs, doses, and timing (e.g., when to give checkpoint inhibitors relative to radiation fractions).
3. Adaptive adjustment
   • Longitudinal data (labs, imaging, symptoms) feed ML models that dynamically adjust regimens—escalating or de‑escalating immune stimulation based on early response or toxicity.

Evidence:

• A 2026 review on AI‑enhanced synergistic chemo‑immunotherapy emphasizes:
  • AI can identify optimal chemo‑immunotherapy combinations by integrating multi‑omics and TME features.
  • Chemo can remodel tumors into more immunogenic states, enhancing checkpoint inhibitor efficacy when optimally scheduled [10].
• A parallel review on AI‑driven immunotherapy and radiotherapy combinations describes AI‑based predictive biomarkers and treatment planners that align radiation dose and field design with expected immune activation [11].

How this boosts immunity:

• CD8+ and NK cells gain more and better antigens from ICD.
• AI‑optimized sequencing avoids deep lymphopenia that would otherwise cripple immune‑based strategies.
• More patients convert from non‑responders to responders by reconditioning the tumor microenvironment.

---

2.4 AI‑Enhanced Diagnosis, Prognosis, and Immune Monitoring

Goal: Use AI to read signals of immune activity or suppression from complex data (pathology, imaging, blood tests) and guide immunotherapy decisions.

Applications:

1. Pathomics and digital pathology
   • Deep learning models analyze routine H&E slides to quantify:
     • Density and spatial distribution of T cells, B cells, macrophages.
     • Immune‑excluded vs. immune‑inflamed patterns.
   • Recent work presented at AACR 2026 showed a “pathomics” AI platform predicting response to immunotherapy in lung cancer using standard pathology images [12].
2. Radiomics
   • AI extracts features from CT/MRI/PET imaging that correlate with immune infiltration and response to checkpoint blockade [13].
3. Response prediction for checkpoint inhibitors
   • Models trained on large cohorts predict probabilities of benefit from immunotherapy using:
     • PD‑L1 expression.
     • Tumor mutational burden.
     • Gene expression signatures.
     • Clinical and lab data.
   • Reports in 2026 discuss deep learning models that can predict response to immune checkpoint inhibitors in advanced NSCLC, helping select first‑line immunotherapy vs. chemo‑immunotherapy [13].
4. Toxicity and safety monitoring
   • AI can detect patterns in labs, vital signs, and clinical notes that precede immune‑related adverse events (e.g., colitis, myocarditis), enabling earlier intervention.
   • A 2025 review of AI in cancer immunotherapy notes “promising and multifaceted” future trends for AI in safety monitoring [14].

How this boosts effective immunity:

• By putting the right patients on the right immunotherapies, AI ensures that the patients whose immune systems can be effectively mobilized actually receive those treatments.
• Early detection of toxicity allows clinicians to maintain immunotherapy where beneficial, instead of stopping too early or causing severe harm.

---

2.5 AI‑Assisted Discovery of New Immune Targets and Mechanisms

Goal: Use AI as a “co‑scientist” to discover previously unrecognized immune checkpoints, antigens, and regulatory circuits that can be targeted to boost anti‑tumor immunity.

Examples from recent work:

• A 2026 article in Cancer Discovery describes AI “co‑scientists” generating drug candidates, prioritizing immunotherapy targets, and shaping discovery pipelines in pharma and academia [15].
• City of Hope’s 2026 AACR presentations highlight AI‑driven discovery efforts mapping:
  • Why cancers develop earlier in some populations.
  • How tumors develop immune resistance.
  • How AI can reveal new targetable resistance pathways [16].

Specific mechanistic advances:

• Identifying novel glycans and glycan‑binding interactions that tumors use to evade immunity, and exploiting lectins that, when used therapeutically, can “dramatically boost the immune system’s response to cancer cells,” as shown in a 2025 MIT immunotherapy study [17].
• Designing next‑generation antibodies that recruit immune cells more efficiently to tumors; early 2026 research reported “supercharged” antibodies that rally the immune system to hit cancer harder and more effectively [18].

How this boosts immunity:

• Broadens the range of drug targets beyond PD‑1/PD‑L1/CTLA‑4.
• Enables combination strategies that modulate multiple immune pathways simultaneously, tuned to each tumor’s escape mechanisms.

Across all these technologies, AI boosts anti‑cancer immunity by:

1. Increasing the quality of immune targets
   • Better neoantigen selection → more specific and potent T‑cell responses.
   • Safer CAR/TCR targets → strong tumor killing, less off‑tumor toxicity.
2. Enhancing effector cell quantity and fitness
   • CAR‑T/NK, TILs, and vaccines elevate numbers of functional CD8+ T cells and NK cells.
   • AI‑optimized designs reduce exhaustion and improve memory formation.
3. Reprogramming the tumor microenvironment
   • AI guides use of chemo/radiation/cytokines to make tumors less immunosuppressive and more accessible to immune cells.
4. Maintaining balance to avoid harmful autoimmunity
   • Improved prediction and monitoring of immune‑related toxicities helps maintain beneficial anti‑tumor immunity while limiting damage to normal tissues.

Despite the promise, several clinical and systemic challenges remain:

• Data bias and generalizability
  AI models may be trained on limited demographic or disease subsets and may underperform in under‑represented populations.
• Standardization and validation
  Many AI predictors of response/toxicity need robust prospective validation before routine clinical use.
• Cost and infrastructure
  Personalized vaccines and engineered cell therapies require expensive sequencing, manufacturing, and computational resources.
• Regulation and interpretability
  Regulatory agencies are still evolving frameworks for AI‑driven decision support and for therapies whose design pipeline is heavily AI‑mediated.
• Access and equity
  High‑resource academic centers are often first adopters; translating these methods to community oncology practices and lower‑income regions is an open challenge.

For researchers, clinicians, and strategists thinking about boosting the human immune system against cancer with AI:

1. Prioritize AI‑enabled personalization where possible
   • Consider enrolling eligible patients in trials of AI‑designed neoantigen vaccines plus checkpoint blockade, especially in melanoma and potentially lung and pancreatic cancers [1–4].
2. Use AI‑based prediction tools to guide immunotherapy
   • Incorporate validated AI models for predicting checkpoint inhibitor response and immune‑related toxicity in NSCLC and other cancers as they become clinically available [12–14].
3. Engage in multidisciplinary teams
   • Combine oncologists, immunologists, data scientists, and computational biologists to co‑develop and interpret AI systems, avoiding blind reliance on models.
4. Focus on interpretable, clinically integrated AI
   • Favor models that provide insight into why a given therapy is recommended (e.g., specific immune and tumor features) to improve trust and refine biological understanding.
5. Invest in longitudinal immune monitoring
   • Use AI to analyze serial pathology, imaging, and blood data to detect early resistance or toxicity and to adapt treatment plans appropriately.
6. Promote equitable data collection
   • Ensure that clinical data used to train AI models is diverse across ethnicity, geography, and cancer types to avoid widening disparities in immunotherapy outcomes.

Research over the last few years has made it increasingly clear that AI is becoming a central enabler of next‑generation cancer immunotherapy. Its key contributions are:

• Designing more precise, personalized immune attacks (vaccines, CAR‑T/NK, TILs).
• Identifying optimal drug combinations and schedules that transform tumors into immune‑responsive states.
• Predicting who will benefit and who is at risk of serious toxicity.
• Discovering new immune targets and mechanisms that can be therapeutically exploited.

These advances do not replace the human immune system; rather, they re‑engineer and guide it, making it far more effective at recognizing and destroying cancer while preserving healthy tissues. Over the coming decade, widespread integration of AI into immunotherapy research and clinical practice is likely to be one of the most important drivers in moving many cancers from lethal diseases to controllable or even curable conditions.

What "AI‑native" means in this context

　

In oncology, AI‑native technology refers to platforms where artificial intelligence is the central organizing principle of the system rather than an add‑on analytics layer. Key characteristics:

• End‑to‑end learning from raw data (histology images, spatial transcriptomics, cfDNA methylation, time‑lapse imaging) rather than hand‑crafted features.
• Foundation or large multimodal models pre‑trained on massive heterogeneous datasets and then adapted to tasks such as detecting growth patterns, not just classifying static tumor types.
• Intrinsic modeling of context — the tumor microenvironment (TME) and experimental culture conditions are directly encoded as part of the learned representation, so that growth patterns are learned as functions of environment.

The central scientific goal is: to detect and distinguish how cancer cell growth patterns (proliferation, invasion, dormancy, trajectory, spatial organization) change across different cellular environments (e.g., core vs margin, immune‑hot vs immune‑cold, 2D vs 3D culture, hypoxic vs normoxic conditions).

Broadly, recent work (2024–2026) falls into three interacting categories:

1. Multimodal foundation models for spatial growth pattern detection from routine pathology and spatial omics.
2. Dynamic and spatiotemporal AI systems that infer cell trajectories and evolving growth patterns.
3. Virtual cell and tissue simulators (digital twins) that use AI to simulate growth under different microenvironments.

Each both detects and contextualizes tumor growth patterns in distinct cellular environments.

3.1 GigaTIME: Translating H&E into virtual multiplex proteomics

A flagship AI‑native system is GigaTIME, a multimodal AI framework that learns to translate routine H&E pathology slides into virtual multiplex immunofluorescence (mIF) maps across 21 tumor‑immune microenvironment (TIME) protein markers at near‑cellular resolution [1].

3.1.1 Architecture and data scale

• Task: Cross‑modal translation: H&E image patches → 21‑channel virtual mIF patches (per protein channel: per‑pixel activation).
• Model: A compact NestedUNet encoder–decoder with densely nested skip connections (~9.16M parameters) trained end‑to‑end on paired H&E/mIF data.
• Training scale:
  • ~40 million cells with matched H&E and mIF channels.
  • Real‑world Providence dataset: 14,256 patients, 24 cancer types, 306 subtypes, 51 hospitals, >1,000 clinics; ~299,000 virtual mIF slides generated.
  • Independent validation on ~10,200 TCGA tumors with 214,000+ virtual mIF slides [1].

This scale is characteristic of AI‑native foundation modeling: the model is trained to represent general tumor and microenvironment biology, not a single task.

3.1.2 How GigaTIME detects growth‑related patterns

GigaTIME's outputs are per‑pixel activation maps per protein channel. From them, several pattern metrics are derived:

• Activation density: fraction of activated pixels per channel → surrogate for marker‑positive cell density (e.g., Ki‑67, PHH3 for proliferation).
• Spatial organization metrics:
  • Entropy: heterogeneity of activation; higher entropy often reflects more disordered/invasive growth.
  • Signal‑to‑noise ratio (SNR) and sharpness: discriminate sharply bounded proliferative niches vs diffuse, infiltrative patterns.
• Combinatorial activations: Boolean OR of channels (e.g., PD‑L1 + caspase‑3, CD138 + CD68) revealing synergistic or antagonistic patterns in proliferation, immune evasion, and apoptosis.

These metrics are computed per patch and aggregated across slides or cohorts, producing growth‑pattern spectra per tumor or subtype.

3.1.3 Distinguishing patterns across different environments

Because GigaTIME was trained on multi‑institution, multi‑tumor data and validated in external cohorts, it can compare growth patterns across multiple microenvironmental contexts:

• Tumor stage & burden:
  • Primary tumor size (T stage) correlates positively with immune checkpoint markers (PD‑L1, PD‑1) and infiltration markers (CD68, CD138), indicating co‑evolving growth and immunosuppressive niches as tumors enlarge [1].
  • Nodal and metastatic status show more context‑specific pattern shifts; in some advanced cancers, PD‑L1 associations invert, suggesting alternative immune‑evasion mechanisms in late stages.
• Genomic context:
  • TMB‑high and MSI‑high tumors exhibit distinctive patterns: increased CD138/CD20/CD68/CD4 activation, consistent with immunogenic, inflamed microenvironments that support specific spatial growth patterns [1].
  • KRAS mutations associate with reduced CD3/CD8 infiltration and altered PD‑L1 engagement; KMT2D mutations show the opposite trend.
• Histological context:
  • Brain vs lung vs other cancers show distinct TIME–growth coupling; e.g., brain tumors display strong CD3/CD8/caspase‑3 associations with TP53 and KMT2A mutations, reflecting unique growth‑immune dynamics in CNS microenvironments.

In practical terms, GigaTIME does not merely “detect tumors”; it maps how growth‑relevant protein programs spatially reconfigure as the environment changes (site, stage, genomic background).

3.1.4 Clinical impact on growth pattern stratification

By integrating all 21 channels into a “GigaTIME signature”, the system stratifies patients into subgroups with distinct survival trajectories and progression risks:

• Combined 21‑channel signatures outperform any single marker in predicting survival across multiple cancer types.
• Unsupervised clustering on the signature reveals ecologically distinct growth patterns (e.g., immune‑inflamed, immune‑excluded, hypoxic‑proliferative), directly linked to outcomes [1].

This is AI‑native pattern detection: data‑driven discovery of survival‑linked growth patterns conditioned on microenvironmental context.

---

3.2 Spatial EcoTyper: AI‑defined spatial ecotypes across cancers

A 2026 Nature study introduces Spatial EcoTyper, an AI framework to map conserved spatial ecotypes (SEs) across tumors using spatial transcriptomics (ST) and single‑cell RNA‑seq [2].

3.2.1 Methodological overview

• Discovery data: ~844,000 tumor microenvironment cells profiled with high‑plex ST (MERSCOPE, Xenium, Visium, etc.) across multiple cancer types.
• Key steps:
  1. For each cell type, aggregate gene expression within 50 µm spatial neighborhoods.
  2. Compute neighborhood‑to‑neighborhood similarity matrices per cell type, then fuse them via Similarity Network Fusion (SNF).
  3. Cluster fused networks (Louvain) to identify sample‑level SEs.
  4. Integrate sample‑level clusters across cohorts using NMF to define nine conserved SEs (SE1–SE9).

Each SE is associated with specific cell states (e.g., hypoxic malignant, proliferative malignant, suppressive myeloid, activated T‑cell) and characteristic gene programs.

3.2.2 Environmental trajectories and growth patterns

Spatial EcoTyper uncovers spatial trajectories of SEs:

• SEs align along gradients from tumor core to adjacent stroma:
  • Core‑enriched SEs often show proliferative, hypoxic, and immune‑excluded programs.
  • Margin/stromal SEs show immune aggregates, fibroblast activation, or angiogenic states.
• Distance‑to‑margin analyses show SEs cluster within ~250 µm of specific margins, forming ecological strata:
  • For instance, SE9 may represent aggressively proliferative core communities, whereas SE1 marks stromal/immune interface states.

Thus, AI reveals that growth patterns are stratified into conserved ecologies that recur across cancer types, each tied to specific microenvironments.

3.2.3 Bulk and liquid inference of growth ecotypes

EcoTyper is extended in two ways:

1. Bulk RNA‑seq deconvolution: Using pseudo‑bulk mixtures, NMF is trained to estimate SE abundances from ordinary bulk transcriptomes.
2. Liquid EcoTyper: A CpG Set Binary Network (CSBN) infers SE levels from cfDNA methylation in plasma [2].

These tools allow non‑invasive monitoring of growth ecotypes, enabling detection of shifts in growth patterns (e.g., increased SE5—often associated with poor prognosis—during therapy).

3.2.4 Clinical associations of SE‑defined growth patterns

SE abundances correlate with:

• Overall survival across multiple cancer types (e.g., SE5 often associated with shorter survival; SE7/SE8 with better outcomes in some contexts).
• Immune checkpoint inhibitor response:
  • Certain SEs outperform classic biomarkers like PD‑L1 or TMB in predicting benefit.
  • Plasma‑derived SE levels (via cfDNA) mirror tissue SEs and predict durable benefit, enabling early detection of growth pattern transitions under immunotherapy [2].

Practically, EcoTyper is an AI‑native map linking where and how cancer grows (SE pattern) to how the environment supports or opposes that growth and how it responds to treatment.

Static snapshots reveal environment‑specific spatial patterns, but growth is inherently temporal. Several AI‑native efforts focus on cell trajectories and dynamic patterning.

　

4.1 AI‑based inference of tumor cell trajectories from images

Recent work on image‑based inference of tumor cell trajectories uses deep learning to predict how cell populations move and evolve over time from fixed pathology slides [3]. Key ideas:

• Self‑supervised or foundation models (e.g., Phikon‑style models) encode morphology and microenvironment context.
• These embeddings are used to infer likely cell differentiation status and future spatial trajectories, enabling:
  • Identification of regions likely to become invasive fronts.
  • Prediction of which subclones will dominate under a given microenvironment.

Though based on static images, such models capture implicit growth directions, effectively turning “photographs” into approximations of “videos” of tumor evolution.

　

4.2 Live cell imaging and AI phenotyping in different culture systems

AI‑enabled live‑cell imaging frameworks (e.g., for T‑cell mediated killing, organoid growth, and spheroid dynamics) provide:

• Trajectory and event detection:
  • Automatic detection of mitosis, apoptosis, migration tracks in large time‑lapse datasets.
  • Quantification of generation times, cell cycle transitions, and spatial expansion rates.
• Comparative analysis across environments:
  • 2D monolayers vs 3D spheroids vs organoids cultured in different matrices (collagen, Matrigel, synthetic hydrogels).
  • Hypoxic vs normoxic conditions; with or without stromal or immune co‑culture.

AI models combining morphology (CNNs) and motion (optical flow or recurrent units) can detect that:

• 3D spheroids exhibit different mitotic geometries, spindle orientations, and division symmetry than 2D cultures [4].
• Environments with high fibroblast or macrophage content induce distinct motility patterns and growth front morphologies compared with cancer‑cell‑only cultures.

These systems qualify as AI‑native because:

• They are trained directly on raw time‑lapse data.
• They learn environment‑conditioned growth phenotypes (e.g., growth arrest vs invasive sprouting) rather than simple counts.

Grammar‑based virtual cell models

A 2025 Cell study describes a plain‑language “hypothesis grammar” and AI framework for building digital twins of multicellular tissues that simulate cancer growth, immune response, and therapy effects [5].

5.1 Design and inputs

• Inputs:
  • Patient‑specific genomics.
  • Spatial genomics (e.g., spatial transcriptomics).
  • Microenvironment composition (immune cells, fibroblasts, ECM properties).
• Hypothesis grammar:
  • Uses human‑readable statements (e.g., “If TGF‑β is high and oxygen is low, then malignant cells increase EMT rate”) to encode rules.
  • AI uses these rules plus data to run thousands of simulations, exploring growth scenarios.

5.2 Detection of environment‑dependent growth patterns

By varying environmental parameters, the system reveals specific growth patterns that only emerge under certain cellular environments, for example:

• In breast cancer, simulations show immune systems that fail to restrain growth can instead promote invasive spreading patterns when specific myeloid populations dominate [5].
• In pancreatic cancer, simulations replicating a real immunotherapy trial show distinct “virtual patients” whose tumors:
  • Continue expanding despite high immune infiltration in dense, fibrotic microenvironments.
  • Remain quiescent or shrink only in microenvironments where CAF–tumor signaling is disrupted.

Because these virtual models are data‑driven yet environment‑parameterized, they effectively test causal hypotheses: changing the cellular environment (e.g., fibroblast density, oxygen, immune composition) and observing AI‑predicted shifts from:

• Clustered, nodular growth → diffuse, infiltrating growth.
• Uniform proliferative patterns → mixed proliferative/dormant mosaics.

This is fundamentally different from classical static modeling: the system uses AI to constantly recalibrate growth behavior as microenvironmental states evolve.

6.1 Representing the environment as part of the model

Across these projects, the cellular environment is explicitly encoded in one or more of the following ways:

• Spatial coordinates and distances (to tumor margin, vessels, stroma) as features for SEs and spatial patterns [2].
• Cell‑type composition and cell–cell interaction graphs (immune cells, CAFs, endothelial cells, etc.).
• Modality‑specific proxies for microenvironmental states:
  • Protein markers (e.g., hypoxia markers, immune checkpoints).
  • Gene expression programs (hypoxia, EMT, metabolism).
  • cfDNA methylation patterns indicative of specific ecotypes.

AI‑native models learn joint latent spaces where both tumor cell states and environmental features co‑embed. Distinct growth patterns emerge as trajectories or clusters in this space.

　

6.2 Distinguishing environment‑specific growth patterns

Examples of environment‑conditioned differences that AI can detect:

• Core vs margin:
  • Core SEs (e.g., SE9) exhibit high proliferation and hypoxia; margins (SE1–SE3) show more immune infiltration and stromal remodeling [2].
  • GigaTIME detects different PD‑L1/proliferation relationships in core vs margin: in some cancers, PD‑L1 activation in cores associates with reduced apoptosis; at margins, it may co‑localize with caspase‑3, indicating ongoing immune attack.
• Immune‑hot vs immune‑cold:
  • EcoTyper and GigaTIME capture that immune‑hot tumors can show dispersed, fragmented growth with pockets of regression, while immune‑cold tumors maintain compact, expanding fronts.
• Fibrotic vs non‑fibrotic stroma:
  • Virtual simulators and spatial models indicate that dense CAF‑rich stroma often correlates with sheathed or “sandwich” patterns, where proliferative malignant zones are encased by fibroblast layers, limiting immune access but enabling invasive outgrowth at specific weak points [2,5].
• 2D vs 3D vs organoid cultures:
  • AI‑driven image analysis reveals that in 3D culture, cancer cells adopt new division geometries, migration paths, and resistance phenotypes mirroring in vivo behavior more closely than 2D [4].
  • Growth patterns in 3D (e.g., spheroid compaction, necrotic core formation, invasive sprouting) are distinct and can be classified and quantified by ML models.

In all cases, AI is used to compare and classify growth patterns as a function of environment, rather than treating growth as an environment‑agnostic feature.

7.1 For researchers

1. Design experiments with explicit environment variation
   Use AI‑native models (e.g., GigaTIME, EcoTyper) to compare growth patterns across:
   • Multiple 3D culture conditions (matrix stiffness, composition).
   • Co‑cultures (CAF, immune cells, endothelial cells).
   • Controlled gradients (oxygen, pH, nutrients).
2. Leverage multimodal integration
   Combine histology, spatial omics, time‑lapse imaging, and cfDNA methylation to obtain consistent growth pattern signatures across compartments (tissue and blood).
3. Adopt foundation models and transfer learning
   Start from released pretrained models (e.g., GigaTIME weights, spatial foundation models) and fine‑tune for:
   • Specific tumor types.
   • Niche environments (e.g., bone, brain, liver metastasis).
4. Use virtual twins to test environmental interventions
   Simulate the effect of modulating microenvironmental features (CAF depletion, angiogenesis inhibition, immune cell recruitment) on predicted growth patterns before designing in vivo experiments.

7.2 For translational and clinical applications

1. Risk stratification by growth pattern, not just mutation
   Integrate SE‑level or GigaTIME‑like signatures into prognostic models:
   • Identify patients with high‑risk growth ecotypes (e.g., SE5‑dominant) even with “favorable” genomic markers.
   • Tailor surgical margins and follow‑up imaging based on predicted invasion patterns.
2. Therapy selection and sequencing
   • Use environment‑specific growth signatures to decide when to prime the microenvironment (e.g., anti‑fibrotic, vascular normalization) before immunotherapy or chemotherapy.
   • Detect early pattern shifts (via cfDNA SEs or AI on serial biopsies) that herald resistance, enabling treatment adaptation.
3. Non‑invasive monitoring
   Apply liquid eco‑typing and plasma‑based AI models to track real‑time changes in growth patterns and distinguish pseudo‑progression from true progression.

7.3 Key future research directions

• True 4D modeling: Fully integrating space, time, and multimodal biology into unified AI models to move from inferred to directly observed growth trajectories.
• Causal AI in the TME: Going beyond pattern correlation toward explicit modeling of why a given environment produces a specific growth pattern, enabling rational microenvironmental therapies.
• Standard benchmarks and ontologies: Creating shared definitions of growth patterns (e.g., “diffuse infiltrative front”, “mosaic dormancy”, “sandwich proliferation”) so that AI‑derived findings are comparable across studies.

AI‑native technologies have shifted the study of cancer cell growth from static, environment‑agnostic snapshots to rich, context‑aware pattern discovery. Multimodal foundation models like GigaTIME and Spatial EcoTyper integrate histology, spatial omics, and cfDNA methylation to map how growth patterns vary across microenvironments and cancer types. Dynamic and virtual cell modeling frameworks extend this further, simulating how tumors will grow and respond under altered environmental conditions.

Together, these approaches allow researchers and clinicians to:

• Detect distinct growth patterns (proliferative, invasive, dormant, regressive) with high spatial and temporal precision.
• Link those patterns directly to specific cellular environments (immune composition, stromal architecture, hypoxia, culture geometry).
• Predict and manipulate the evolution of tumors by targeting not just the cancer cells, but the ecologies in which they grow.

This emerging AI‑native landscape is moving oncology toward a future where cancer growth is modeled, monitored, and modulated as an ecosystemal, environment‑dependent process, rather than a static property of malignant cells alone.