Cancer immunotherapy has been built around a surprisingly small number of extraordinarily successful biological discoveries. CTLA-4, PD-1 and PD-L1 demonstrated that removing inhibitory signals from T cells could produce durable tumor control, while subsequent work expanded the field toward additional checkpoints, engineered immune cells and increasingly sophisticated approaches to the tumor microenvironment.
Yet the human immune system contains thousands of potentially druggable interactions across tumor cells, T cells, B cells, macrophages, dendritic cells, natural killer cells and stromal populations. These interactions are further shaped by somatic mutations, antigen presentation, transcriptional states, spatial organization and treatment-induced evolutionary pressure. Finding the next clinically meaningful immunotherapy target within this complexity has become increasingly difficult using one dataset or experimental system at a time.
Artificial intelligence offers a different strategy. Rather than beginning with a single candidate molecule and asking whether it matters, computational models can integrate genomic, transcriptomic, single-cell, spatial, proteomic, immunopeptidomic and functional perturbation data to search for biological relationships that would be difficult to identify manually.
In 2026, that possibility moved closer to experimental reality. A study in Nature Machine Intelligence used a multimodal machine-learning framework to identify potential immunotherapy targets and then tested one of its predictions directly in patient-derived tumor tissue. The study provides an unusually clear example of what AI could contribute to immuno-oncology: not simply predicting which patient will respond to an existing therapy, but helping discover what we should target next.
From Predicting Response to Discovering Biology
Much of the early enthusiasm around AI in immunotherapy has focused on prediction. Machine-learning models have been developed to estimate checkpoint inhibitor response from pathology images, radiological scans, gene expression, tumor mutational burden, immune signatures and combinations of clinical and molecular variables.
That is valuable, but target discovery is a fundamentally different problem. Predicting that a patient is unlikely to respond to anti-PD-1 therapy does not explain which molecular interaction should be altered to make that tumor sensitive. A target-discovery system must move beyond recognizing patterns associated with response and identify biological nodes whose manipulation could actually change the tumor-immune interaction.
This distinction becomes especially important because immuno-oncology data are fragmented across biological scales. A potentially important gene might appear in tumor sequencing, be selectively expressed in a suppressive macrophage population, influence T-cell dysfunction in single-cell data, participate in an antigen-presentation network and produce an immune phenotype when perturbed by CRISPR. Each observation alone may be insufficient to justify drug development. Together, however, they may reveal a therapeutic vulnerability.
The challenge is connecting those observations.
MIDAS: Asking AI to Find New Immunotherapy Targets
One of the most important recent examples comes from Marcellus Augustine, Nuno Rocha Nene, Kevin Litchfield and colleagues, who reported “Immunotherapy Drug Target Identification Using Machine Learning and Patient-Derived Tumour Explant Validation” in Nature Machine Intelligence in May 2026.
The investigators developed MIDAS, Mining Immunotherapy Drug tArgetS, a multimodal graph neural network specifically designed for immuno-oncology target discovery. Rather than relying on a single molecular dataset, MIDAS combined multiple layers of evidence relevant to antitumor immunity.
These included:
- exomic and transcriptomic profiles from patients treated with checkpoint inhibitors;
- single-cell RNA sequencing of tumor-infiltrating immune populations;
- HLA peptidomics describing peptides presented through antigen-presentation machinery;
- immuno-oncology gene-phenotype relationships;
- gene-interaction networks;
- and causal genetic perturbation data from CRISPR tumor-T-cell coculture experiments.
This is where AI becomes particularly interesting for immunotherapy. A human researcher can examine each of these datasets individually, but identifying relationships across all of them becomes increasingly difficult as dimensionality grows. MIDAS represented these different sources of information within a biological network and used graph machine learning to prioritize genes according to their potential relevance as immunotherapy targets.
Importantly, the investigators did not evaluate the system only by asking whether it could rediscover known targets in the same data on which it had been trained. In time-sliced analyses designed to approximate prospective discovery, MIDAS outperformed comparator approaches including Open Targets and preferentially ranked targets that had ultimately reached approval above those remaining in clinical development. It also recovered genes associated with immunotherapy response in previously unseen patient data.
But ranking targets computationally is still not target discovery.
The more important question is what happens when the prediction enters biology.
Artificial Intelligence in Cancer Immunotherapy
From an Algorithmic Prediction to a Human Tumor
Among the pathways prioritized by MIDAS was signaling involving oncostatin M and its receptor OSMR.
Oncostatin M is a member of the IL-6 cytokine family and can participate in inflammatory signaling within the tumor microenvironment. MIDAS implicated the OSM-OSMR axis as a potential immunotherapeutic vulnerability, but the investigators then performed the step that many AI studies never reach: experimental validation.
They perturbed OSM-OSMR signaling in patient-derived melanoma explants from the TRACERx study, preserving important elements of the native tumor microenvironment rather than relying exclusively on computational associations or conventional cell lines. Disruption of this pathway reduced dysfunctional CD8-positive T-cell populations associated with immunotherapy response and decreased CCL4 levels. Analysis of human tumor transcriptomes also connected OSM and OSMR expression with altered T-cell and macrophage states.
The study does not establish OSM-OSMR as a clinically validated immunotherapy target. That would require substantially more functional work, pharmacological development and ultimately prospective clinical testing. What it demonstrates is arguably more important for the AI field: a computationally prioritized target can be taken back into human tumor biology and experimentally challenged.
That closes part of the loop between data and mechanism.
AI Needs Experiments, and Experiments Need AI
The MIDAS study reflects a broader evolution in immuno-oncology discovery. High-throughput functional screening has already shown how systematically perturbing genes can reveal immune vulnerabilities that would be difficult to anticipate from expression data alone.
A landmark example came from Juan Dubrot and colleagues in Nature Immunology, who performed genome-scale in vivo CRISPR screens across cancer models exposed to immune checkpoint blockade. Their experiments identified multiple pathways involved in immune escape and highlighted the non-classical MHC class I molecule HLA-E and its murine counterpart Qa-1b as an inhibitory mechanism operating through NKG2A/CD94. The study also revealed the context-dependent complexity of interferon signaling, showing that tumor IFNγ sensing can, under some conditions, generate inhibitory programs rather than simply strengthening antitumor immunity.
More recently, CRISPR screening has moved directly into engineered cellular therapy. In 2025, Anne-Christine Orts and colleagues reported “Systematic Discovery of CRISPR-Boosted CAR T Cell Immunotherapies” in Nature. Their CELLFIE platform systematically screened genetic perturbations intended to improve CAR T-cell function across multiple therapeutic objectives, illustrating how functional genomics can search the enormous design space of immune-cell engineering rather than modifying one candidate gene at a time.
These studies are not simply examples of AI, and that distinction matters. CRISPR generates experimental evidence about what happens when genes are altered. AI can then help integrate those perturbational data with patient genomics, immune-cell states, molecular networks and clinical outcomes.
The most powerful discovery system may therefore be neither computational nor experimental alone. It may be an iterative cycle in which AI prioritizes a biological hypothesis, perturbation experiments test it, the resulting data refine the model and increasingly informative experiments follow.
Single-Cell Biology Has Changed What a “Target” Means
Traditional drug discovery often asks whether a molecule is overexpressed in cancer and whether inhibiting it reduces tumor growth. Immunotherapy introduces a more complicated problem because the same molecule can perform different functions depending on which cell expresses it and where that cell is located.
A receptor expressed on exhausted CD8 T cells may have a different therapeutic meaning when expressed on regulatory T cells, macrophages or tumor cells. A cytokine pathway may promote immunity during early T-cell priming but contribute to immune suppression during chronic inflammation. Even cells assigned the same broad label, such as “macrophage” or “CD8 T cell,” can contain multiple functional states with opposing effects on tumor control.
Single-cell RNA sequencing has exposed this complexity by resolving tumor ecosystems into individual cellular states. Spatial technologies add another layer by showing where those states exist relative to malignant cells, blood vessels, stromal barriers and one another.
For AI-based target discovery, this changes the question from “Which gene is associated with cancer?” to “Which molecular interaction in which cell state, in which spatial context, is causally maintaining immune escape?”
That is a much harder question, but it is also closer to the biology that a successful immunotherapy needs to manipulate.
Spatial Biology Could Reveal Targets Hidden by Bulk Sequencing
Bulk tumor sequencing averages molecular signals across millions of cells. A potentially critical immune interaction occurring within a small cellular neighborhood may therefore disappear when the entire biopsy is analyzed together.
Spatial transcriptomics and multiplex imaging can instead identify localized immune ecosystems: T-cell-rich regions adjacent to antigen-presenting cells, macrophage-dominated suppressive niches, immune-excluded tumor borders, tertiary lymphoid structures and vascular regions that regulate immune-cell trafficking.
For target discovery, spatial information is not merely descriptive. If two cell populations repeatedly occupy the same microenvironment and communicate through a specific ligand-receptor pair, that interaction becomes a candidate therapeutic mechanism. AI can analyze thousands of such cellular relationships simultaneously and identify patterns associated with immune exclusion, dysfunction or treatment resistance.
The future target may therefore not be the most highly expressed protein in a tumor. It could be a context-dependent interaction between two rare cellular populations occupying a specific spatial niche.
The Immunopeptidome Adds Another Layer
Immunotherapy ultimately depends on immune recognition, making antigen presentation another major target-discovery space.
Genomic sequencing can identify mutations capable of generating candidate neoantigens, but a mutation does not automatically become an immune target. The mutated protein must be expressed, processed into peptides, loaded onto an HLA molecule, displayed on the cell surface and recognized by an available T-cell receptor.
Mass-spectrometry-based immunopeptidomics can directly characterize peptides presented by HLA molecules, while computational models can integrate sequence information, HLA-binding properties, gene expression and antigen-processing features. MIDAS incorporated HLA peptidomic information into its multimodal target-discovery framework, illustrating how antigen presentation can be considered alongside immune-cell biology, genetics and functional perturbation rather than treated as an isolated dataset.
This integration could eventually expand target discovery beyond conventional cell-surface checkpoints. The relevant vulnerability might instead involve antigen processing, peptide presentation, HLA regulation or another mechanism controlling whether immune cells can see malignant cells at all.
AI Can Also Rediscover Biology We Already Know
An important test of any discovery algorithm is whether it can recover established biology without being explicitly instructed to find it.
MIDAS preferentially ranked clinically successful immunotherapy targets and recovered genes associated with checkpoint inhibitor response in patient datasets that were not used to build the relevant predictions. Its interpretability analyses suggested that the model was drawing information from biologically plausible domains including autoimmunity, regulatory networks and known immuno-oncology pathways.
This type of retrospective validation is necessary, but it creates an interesting tension. If an AI system only rediscovers PD-1, CTLA-4 and other established biology, it has limited value for target discovery. If it proposes completely unfamiliar targets without mechanistic evidence, those predictions may simply represent sophisticated correlations.
The real test lies between these extremes: can the model learn enough established immunology to be credible while still identifying unexpected biology that survives experimental perturbation?
OSM-OSMR is interesting precisely because the investigators attempted that next step.
Prediction Is Not Causation
This remains the central limitation of AI-driven target discovery.
Cancer datasets contain enormous numbers of correlations. A gene can be associated with poor immunotherapy response because it drives resistance, because resistant tumors induce its expression, because it marks another causal cellular state or because it is correlated with an entirely different biological process.
A machine-learning model can identify the association without distinguishing among these possibilities.
That is why functional perturbation remains indispensable. CRISPR knockout or activation, organoids, tumor-immune cocultures, patient-derived explants, genetically engineered models and eventually clinical intervention are required to determine whether manipulating the proposed target actually changes antitumor immunity.
The distinction is particularly important in immunology because many pathways are bidirectional and context dependent. Interferon signaling provides a good example. IFNγ is central to antitumor immune activity, yet the Nature ImmunologyCRISPR study showed that tumor IFN sensing can also induce immune-evasion programs involving classical and non-classical MHC class I molecules.
A model that simply labels interferon signaling as “good” or “bad” would therefore miss the biology that matters most.
AI Is Beginning to Learn Immune Concepts, Not Just Genes
Another major development appeared in Nature Medicine in July 2026. Wanxiang Shen, Marinka Zitnik and colleagues introduced COMPASS, a pan-cancer foundation model designed to predict immunotherapy response from tumor transcriptomes.
COMPASS was pretrained on 10,184 tumors representing 33 cancer types and evaluated using 1,133 patients from 16 immunotherapy cohorts spanning seven cancers and six checkpoint inhibitor regimens. Instead of treating the transcriptome as an uninterpretable collection of thousands of gene measurements, the model organized expression data into biologically grounded representations of immune cells, signaling pathways and tumor-microenvironment interactions.
Across the evaluated cohorts, COMPASS outperformed 22 comparator methods on average, improving accuracy by 8.5% and area under the precision-recall curve by 15.7%. More interesting for target discovery, however, were its personalized response maps, which linked individual gene-expression patterns to interpretable immune concepts. Among immune-inflamed tumors that nevertheless failed to respond, the model highlighted different candidate mechanisms, including TGFβ signaling, endothelial exclusion, CD4 T-cell dysfunction and B-cell deficiency.
COMPASS is primarily a response-prediction and hypothesis-generation system rather than a validated target-discovery engine, and its mechanistic concepts have not yet been experimentally established as causal. The authors explicitly caution that its explanations should therefore be regarded as hypothesis generating.
But the conceptual direction is important. AI is beginning to move from asking which genes predict response toward asking which immune state explains why this particular tumor does or does not respond.
That transition could ultimately make target discovery much more biologically meaningful.
COMPASS: Can AI Predict Immunotherapy Response Across Multiple Cancers?
Could AI Find Targets That Humans Would Never Think to Test?
This is where the question becomes genuinely provocative.
Traditional target discovery is inevitably influenced by existing knowledge. Researchers study pathways already implicated in cancer, immune checkpoints resembling known checkpoints and molecules for which experimental tools or drugs already exist. That approach has produced enormous advances, but it also creates a form of biological selection bias.
AI can potentially search a much larger hypothesis space. A graph model does not need a molecule to be famous before considering it important. It can connect weak signals distributed across genetics, immune-cell states, protein interactions, antigen presentation and functional screens that individually might never justify an experimental program.
This does not mean AI possesses biological intuition. In many cases, the opposite is true: the model has no intrinsic understanding of cancer unless that structure is provided through data, architecture or biological constraints.
Its advantage is different. AI can evaluate combinations of evidence at a scale that exceeds human manual reasoning.
The scientist’s role then changes from generating every candidate hypothesis to determining which computational hypotheses are biologically plausible, experimentally testable and clinically meaningful.
The Data Problem May Be Harder Than the Algorithm Problem
Even highly sophisticated models cannot recover biological information that is absent or systematically biased in their training data.
Most immunotherapy datasets remain relatively small compared with datasets used to train foundation models in other domains. Different studies use different sequencing platforms, biopsy sites, treatment regimens, response definitions and sample-processing methods. Patients enrolled in clinical trials may not represent the populations ultimately treated in routine practice.
Tumors also evolve over time. A pretreatment biopsy provides only one spatial and temporal sample of a dynamic immune ecosystem, while metastatic lesions within the same patient can contain different immune states and different resistance mechanisms.
These limitations create a danger that an apparently powerful model learns technical artifacts or cohort-specific patterns rather than transferable immunology.
COMPASS illustrates both the progress and the remaining problem. Its pan-cancer architecture generalized across multiple cohorts, cancers and checkpoint therapies, but the investigators noted that the absence of non-ICI comparator arms prevented complete separation of predictive from prognostic signals. They also emphasized that prospective validation remains necessary before such models could influence treatment decisions.
For target discovery, the standard should be even higher. A computational ranking is only the beginning.
Explainability Matters More When AI Is Discovering Drugs
A black-box model may be acceptable for some classification tasks if its performance is independently validated. Target discovery is different because researchers need to understand why a candidate was prioritized before investing years in biological validation and drug development.
Graph-based and concept-based AI systems offer one potential solution because their predictions can be connected to biological networks, cell states or pathways. MIDAS used interpretability analyses to identify the types of biological evidence contributing to its predictions, while COMPASS routes transcriptomic information through interpretable tumor-immune concepts.
Even then, explainability should not be confused with mechanism. A model can provide a convincing explanation for its prediction without proving that the underlying pathway causes immune resistance.
Ultimately, the strongest form of interpretability in drug discovery remains experimental: perturb the predicted target and determine whether the biology changes in the predicted direction.
From Target Discovery to Target Validation
The emerging workflow for AI-driven immunotherapy discovery may therefore look very different from traditional computational biomarker research.
Patient tumors provide genomic, transcriptomic, single-cell, spatial, proteomic and antigen-presentation data. Functional screens add information about what happens when individual genes or pathways are disrupted. AI integrates these layers and prioritizes candidate immune vulnerabilities. The strongest candidates then return to experimental systems, ideally including patient-derived material, where their causal effects can be tested.
Promising targets would subsequently move through increasingly stringent validation: independent patient cohorts, mechanistic experiments, pharmacological perturbation, toxicity assessment, preclinical models and ultimately early-phase clinical trials.
Each stage can generate new data that return to the computational system.
The result is not AI replacing experimental immunology. It is a closed discovery loop between computation, human tumor biology and functional validation.
The Hardest Question: Can AI Discover a Target That Becomes a Drug?
As of 2026, this remains largely unanswered.
AI systems can prioritize immunotherapy targets, identify response-associated pathways and generate experimentally testable hypotheses. MIDAS represents an important advance because a machine-learning-derived target was taken into patient-derived tumor explants for functional validation. COMPASS demonstrates that foundation models can organize enormous transcriptomic datasets into interpretable immune concepts that generate mechanistic hypotheses. CRISPR platforms such as CELLFIE show how functional genomics can systematically explore modifications capable of improving immune-cell therapies.
But identifying an interesting pathway is very different from developing a successful medicine.
The history of oncology is filled with biologically compelling targets that failed because they were not sufficiently causal, could not be safely inhibited, were redundant with other pathways or worked only in narrow biological contexts. AI does not remove those constraints.
A 2026 Nature Reviews Drug Discovery analysis of AI in drug discovery makes precisely this broader point: despite intense interest in AI-based approaches, evidence of clinically relevant impact remains limited, and translation must become a central consideration during model development rather than an afterthought.
For immuno-oncology, the decisive milestone will therefore not be another algorithm with better retrospective performance. It will be a therapy against a previously unrecognized immune target whose discovery was materially enabled by AI and whose clinical benefit can be demonstrated prospectively.
We are not there yet.
But the route toward that milestone is becoming visible.
