Immunotherapy has fundamentally changed oncology, offering the possibility of long-term, durable survival. However, these therapies present a profound clinical paradox: they are highly effective in only a minority of patients. For the non-responding majority, administering these treatments results in severe immune-related toxicities, high financial costs, and critical delays in seeking alternative care.
Predicting who will respond to treatments like immune checkpoint inhibitors (ICIs) traditionally relies on tissue biopsies to evaluate markers such as Programmed Death-Ligand 1 (PD-L1) expression or Tumor Mutational Burden (TMB). Yet, physical biopsies are invasive, risky, and capture only a single snapshot of a tumor that is highly heterogeneous across space and time.
To overcome these limitations, a major research direction has emerged focused on computational patient stratification. By training machine learning models on non-invasive imaging, digital pathology, and circulating blood biomarkers, researchers are uncovering sub-visual, systemic signatures that accurately map a patient's immune landscape and predict therapeutic outcomes.
Radiomics as a Macro-Scope for the Tumor Microenvironment
Why can a standard medical scan predict whether an immune cell will attack a tumor? The answer lies in the relationship between a tumor's macroscopic structure and its microscopic biology. As a tumor interacts with the immune system, changes in cell density, vascularization, and tissue architecture alter its appearance on medical imaging. While these subtle patterns are invisible to the human eye, computer vision algorithms can extract thousands of quantitative features—a field known as radiomics.
[Standard Medical Scan (CT/PET/MRI)]
│
▼
[Radiomics Feature Extraction] (Texture, Shape, Heterogeneity)
│
▼
[Deep Learning / Machine Learning Model]
│
▼
[Microscopic Inferences] (PD-L1 Expression, TMB, Immune Infiltration)
Using this approach, Wei Mu et al. developed a "Deeply Learned Score" (DLS) from standard PET/CT scans in their paper, Non-invasive measurement of PD-L1 status and prediction of immunotherapy response using deep learning of PET/CT images. The model successfully predicted PD-L1 status and survival outcomes as effectively as traditional tissue biopsies, acting as a digital surrogate for invasive procedures.
This capability spans multiple imaging modalities and cancer types:
- CT Imaging: In non-small cell lung cancer (NSCLC), Jiexiao Wang et al. demonstrated in CT radiomics-based model for predicting TMB and immunotherapy response in non-small cell lung cancer that CT textures can predict genomic features like TMB. For oral squamous cell carcinoma, Qifan Ma et al. used CT radiomics to achieve high accuracy in forecasting immunotherapy responses in Predicting response to immunotherapy in oral squamous cell carcinoma via a CT-based radiomics model.
- MRI Imaging: In liver cancer, Rui Zhang et al. leveraged MRI scans in MRI radiomics model for predicting tumor immune microenvironment types and efficacy of anti-PD-1/PD-L1 therapy in hepatocellular carcinoma to map the tumor immune microenvironment and distinguish responders from non-responders before treatment begins.
- Cellular Therapies: Radiomics also extends to advanced cellular treatments. H Peters et al. showed in [18F]FDG-PET/CT in DLBCL-patients treated with CAR-T cell therapy: potential for defining patient prognosis that pre-treatment metabolic parameters from PET/CT scans can predict early failure in CAR-T cell therapies.
Pathomics and Weakly Supervised Digital Pathology
When tissue samples are available, computational models can optimize how they are interpreted. Traditional pathology relies on manual annotation and visual scoring, which are prone to inter-observer variability. "Pathomics" uses deep learning to analyze whole slide images (WSIs) of standard patient tissue, converting spatial cellular arrangements into predictive statistical signatures.
To streamline this process, Marta Ligero et al. introduced a weakly supervised deep learning framework in Weakly Supervised Deep Learning Predicts Immunotherapy Response in Solid Tumors Based on PD-L1 Expression. Instead of requiring pathologists to manually outline thousands of individual cells, their model learns directly from raw, slide-level immunohistochemistry images. It automatically identifies the features that correlate with PD-L1 expression and clinical response, making the workflow highly reproducible.
Similarly, Abdou Khadir Dia et al. bypassed specialized staining altogether in Computational analysis of whole slide images predicts PD-L1 expression and progression-free survival in immunotherapy-treated non-small cell lung cancer patients. By extracting texture features from cell density maps on standard, inexpensive H&E-stained slides, their interpretable pipeline successfully correlated digital spatial patterns with actual biological immune activity and progression-free survival.
Circulating Liquid Biomarkers and Systemic Signals
The tumor microenvironment does not exist in a vacuum; it interacts continuously with the host's systemic immune and metabolic systems. Tumors and immune cells shed proteins and metabolites into the bloodstream, making liquid biopsies an attractive avenue for dynamic, non-invasive monitoring.
[Tumor-Immune Interaction] ──► [Systemic Shedding] ──► [Serum Biomarkers]
├── suPAR (Infiltration/Prognosis)
├── CRP (Inflammatory Baseline)
└── N-(3-Indolylacetyl)-L-alanine (Metabolism)
Several key circulating factors have been identified as indicators of immunotherapy efficacy:
- Soluble Receptors: Sven H. Loosen et al. demonstrated in Serum Levels of Soluble Urokinase Plasminogen Activator Receptor Predict Tumor Response and Outcome to Immune Checkpoint Inhibitor Therapy that baseline serum levels of soluble urokinase plasminogen activator receptor (suPAR) serve as a potent indicator of systemic inflammation. High suPAR levels strongly correlate with poor treatment response and shortened overall survival.
- Metabolomics: In Association of metabolomics with PD-1 inhibitor plus chemotherapy outcomes in patients with advanced non-small-cell lung cancer, the authors discovered that a specific serum metabolite, N-(3-Indolylacetyl)-L-alanine, acts as an independent predictor of response to combination chemo-immunotherapy, providing a convenient blood-based alternative to genomic sequencing.
- Inflammatory Markers: Systemic inflammatory baselines also dictate response durability. For instance, Mingxiu Ju et al. identified C-reactive protein (CRP) as a critical predictor of outcomes in recurrent or metastatic cervical cancer within The efficacy of first and second immunotherapy exposure in patients with recurrent or metastatic cervical cancer.
Multi-Omic Integration and Advanced Computational Modeling
Because cancer resistance mechanisms are multifaceted, single-modality predictions can miss critical biological context. The frontier of patient stratification lies in integrating these disparate data streams—radiomics, genomics, and clinical data—into unified computational frameworks.
In A multi-omics-based serial deep learning approach to predict clinical outcomes of single-agent anti-PD-1/PD-L1 immunotherapy in advanced stage non-small-cell lung cancer, Yi Yang et al. introduced SimTA, a deep learning model that combines serial CT scans with laboratory and clinical data. This multi-modal approach proved capable of stratifying patients with stable disease—a subgroup that is traditionally difficult to classify using standard response criteria.
[Serial CT Scans] ──┐
[Laboratory Data] ──┼─► [SimTA Deep Learning Framework] ─► Precise Stratification (Even for Stable Disease)
[Clinical Records] ──┘
To manage these diverse, non-tabular clinical data types, researchers are also adapting natural language architectures. Zekai Chen et al. demonstrated in Boosting Transformers and Language Models for Clinical Prediction in Immunotherapy that transformer-based large language models (LLMs), when fine-tuned with domain-specific medical knowledge, can accurately predict immunotherapy responses even in few-shot learning scenarios where patient data is scarce.
Concurrently, genomic and transcriptomic modeling continues to refine our understanding of the molecular baseline. Shenglei Song et al. constructed a prognostic model using pairs of immune-related long noncoding RNAs in Identification of an Immune-Related Long Noncoding RNA Pairs Model to Predict Survival and Immune Features in Gastric Cancer, while Ruo Qi Li et al. used transcriptomic clustering in Identification of tumor antigens and immune subtypes in breast cancer for mRNA vaccine development to map "immune-hot" tumor subtypes.
By anchoring these advanced computational models in observable biological mechanisms—whether through the lens of radiomics, pathomics, or circulating biomarkers—this research direction moves oncology closer to a non-invasive, highly personalized paradigm for immunotherapy selection.
Key Foundational Work
Autophagy: novel applications of nonsteroidal anti‐inflammatory drugs for primary cancer (2017) — contributed 1 concept elements to this direction.
MiR-770-5p inhibits cisplatin chemoresistance in human ovarian cancer by targeting ERCC2 (2016) — contributed 1 concept elements to this direction.