Tumor changes may anticipate reaction to immunotherapy
Predicting Immunotherapy Response: From MSI to AI and Personalized Vaccines (2026 Update)
As a medical oncologist, I have witnessed the revolutionary impact of immunotherapy firsthand. I've seen patients with advanced, treatment‑resistant cancers achieve remarkable and durable remissions after receiving immune checkpoint inhibitors. Yet, I have also seen many patients who, despite our best hopes, derive no benefit from these powerful drugs. This fundamental challenge—predicting who will respond and who will not—has been a central focus of cancer research for the past decade.
This post, originally published in 2019, highlighted a landmark study from Johns Hopkins that identified microsatellite instability (MSI) intensity as a potential "crystal ball" for immunotherapy response. That research laid a crucial foundation. But the field has advanced at a breathtaking pace. In 2026, we have an entire arsenal of new tools—from more sophisticated biomarkers to artificial intelligence and personalized vaccines—that are transforming how we predict and enhance immunotherapy outcomes.
🔬 The 2019 Foundation: MSI, Indel Mutations, and Neoantigens
The 2019 study, published in the journal Science by researchers from Johns Hopkins Kimmel Cancer Center and Memorial Sloan Kettering Cancer Center, addressed a critical question: why do some patients with mismatch repair‑deficient (dMMR) tumors respond dramatically to immunotherapy, while others do not?[reference:0]
Their key findings were:
- MSI intensity matters: Tumors with a higher degree or "intensity" of microsatellite instability (MSI‑H) were more likely to respond to anti‑PD‑1 checkpoint inhibitors than tumors with lower MSI.[reference:1]
- Indel mutations are key: These MSI‑high tumors harbored a greater number of insertion/deletion (indel) mutations—changes in the DNA where small segments are either added or removed.[reference:2]
- Neoantigens drive the response: These indel mutations generate "neoantigens"—abnormal proteins that the immune system can recognize as foreign and target for destruction.[reference:3]
💡 Clinical Perspective: The "Crystal Ball" Analogy
The study's lead author, Dr. Rajarsi Mandal, described this genetic signature as a potential "crystal ball" to predict which patients would benefit from immunotherapy.[reference:4] This was a powerful concept because it suggested that a simple biopsy and DNA sequencing test could guide treatment decisions, sparing non‑responders from the potential side effects and costs of ineffective therapy.
The study demonstrated this in both mouse models and human patient data. Mice implanted with MSI‑high tumors showed a significant reduction in tumor volume and a pronounced increase in tumor‑infiltrating lymphocytes after anti‑PD‑1 treatment.[reference:5] In human cohorts, patients with higher MSI levels had better responses and improved survival compared to those with lower levels.[reference:6]
This research was a major step forward, but as Dr. Mandal noted at the time, it required validation in larger studies. Since then, the science has not only validated these findings but has expanded them in remarkable ways.
📈 The 2025‑2026 Landscape: A New Era of Precision Prediction
In the years since 2019, our understanding of immunotherapy response prediction has evolved from a single biomarker (MSI) to a multi‑faceted approach that integrates genomics, computational biology, and even artificial intelligence.
1. Tumor Mutational Burden (TMB): A Broader Biomarker
While MSI remains a powerful predictor in specific cancer types like colorectal and endometrial cancers, its utility is limited in the vast majority of cancers that are microsatellite stable (MSS). Enter Tumor Mutational Burden (TMB)—a measurement of the total number of mutations present in a tumor's DNA, calculated per million base pairs.[reference:7]
TMB is now an FDA‑approved biomarker for immune checkpoint inhibitor therapy.[reference:8] A 2026 study published in BMC Cancer demonstrated that TMB can predict neoantigen profiles and immunotherapy response in MSS tumors across various cancer types, effectively identifying a biologically distinct subset of patients who can benefit from immunotherapy despite being MSS.[reference:9]
However, TMB is not perfect. Its predictive value varies among tumor types, and researchers are now exploring "Very High" TMB (TMB‑VH) as a distinct category that may offer even greater precision.[reference:10]
2. Advanced Neoantigen Prediction: The NeoPrecis Framework
The 2019 study highlighted the role of neoantigens. Today, we have sophisticated computational tools to identify and prioritize these neoantigens. One of the most exciting developments is NeoPrecis, a framework that integrates clonality (whether a mutation is present in all cancer cells or just a subset) and predicted immunogenicity to refine response prediction.[reference:11]
In a study published in Nature Communications, NeoPrecis improved response prediction in melanoma by 11% and in non‑small cell lung cancer (NSCLC) by 20% compared to TMB alone.[reference:12] This is a clinically meaningful improvement that could help oncologists make more informed treatment decisions.
💡 Clinical Perspective: Clonality Matters
A key insight from NeoPrecis is that clonal neoantigens—those present in all cancer cells—are more likely to trigger a robust and durable immune response than subclonal neoantigens found only in a subset of cells. This explains why some tumors with high TMB still fail to respond: their mutations may be too diverse and present in too few cells to mount a coordinated immune attack.[reference:13]
Another breakthrough, the neoIM model, focuses exclusively on predicting CD8 T‑cell response rather than just MHC binding, offering a more direct measure of immunogenicity.[reference:14] Additionally, researchers have discovered that reduced activity of the nonsense‑mediated mRNA decay (NMD) pathway is a predictor of improved checkpoint inhibitor response, opening yet another avenue for patient stratification.[reference:15]
3. Liquid Biopsy: Monitoring Response in Real Time
Predicting response before treatment is only half the battle. Monitoring response during treatment is equally critical. Liquid biopsy—the analysis of circulating tumor DNA (ctDNA) from a simple blood draw—has emerged as a powerful tool for this purpose.[reference:16]
In December 2025, the IMMUNOMICS‑VHIO platform published results demonstrating that ultrasensitive ctDNA analysis can predict and monitor immunotherapy response with high accuracy.[reference:17] Analyzing ctDNA before treatment can help select patients most likely to benefit, while tracking ctDNA dynamics during therapy provides an early indication of whether the treatment is working—often months before changes are visible on CT scans.
In a major step forward, the Cancer Research Institute launched a Phase 2/3 clinical trial in 2025 that uses the Labcorp Plasma Focus™ liquid biopsy test to measure ctDNA response after just two cycles of pembrolizumab. This approach could allow clinicians to switch non‑responders to alternative therapies much earlier, sparing them from prolonged ineffective treatment.[reference:18]
4. Artificial Intelligence: The Next Frontier
Perhaps the most transformative development since 2019 has been the integration of artificial intelligence (AI) and machine learning into immunotherapy response prediction. These computational approaches can analyze vast amounts of genomic and clinical data to identify patterns that are invisible to the human eye.
Several AI models have shown remarkable promise:
- HAPIR (Hallmark gene set‑based Approach for Predicting Immunotherapy Response) uses a refined set of cancer hallmark genes to predict checkpoint inhibitor response and guide treatment strategies.[reference:19]
- PathHDNN, a pathway‑informed deep neural network, has demonstrated higher prediction accuracy than other state‑of‑the‑art algorithms in multiple immunotherapy cohorts.[reference:20]
- TxGemma, a generative language model, can predict anti‑PD‑1 response using only a small set of immune genes and the right prompting strategy, without requiring extensive retraining.[reference:21]
💡 Clinical Perspective: AI as a Decision‑Support Tool
While these AI models are not yet ready to replace clinical judgment, they are rapidly becoming invaluable decision‑support tools. In the near future, I envision a scenario where a patient's tumor biopsy is sequenced, the data is fed into an AI model like PathHDNN, and within hours, the oncologist receives a personalized prediction of immunotherapy response probability—along with insights into which specific pathways are driving resistance and might be targeted with combination therapies.
💊 Clinical Advances: New Treatments and Approvals
The improved ability to predict response has been accompanied by an expansion in the therapeutic options available for biomarker‑defined patient populations.
Expanded Approvals for MSI‑H/dMMR Cancers
In April 2025, the FDA approved the combination of nivolumab (Opdivo) and ipilimumab (Yervoy) for adult and pediatric patients aged 12 and older with unresectable or metastatic MSI‑H/dMMR colorectal cancer.[reference:22] This dual checkpoint blockade (targeting both PD‑1 and CTLA‑4) has shown superior efficacy compared to single‑agent therapy in this highly immunogenic subset of cancers.
The CheckMate‑8HW trial, which led to this approval, demonstrated significant improvements in progression‑free and overall survival, establishing immunotherapy as the first‑line standard of care for dMMR/MSI‑H colorectal cancer.[reference:23][reference:24]
Furthermore, a Phase III study (AZUR‑2) is now evaluating perioperative dostarlimab—a chemotherapy‑free option—for patients with resectable dMMR/MSI‑H colon cancer, potentially redefining the role of surgery in this disease.[reference:25][reference:26]
Targeting TMB‑High Cancers
In early 2025, the FDA granted Fast Track designation to invikafusp alfa (STAR0602), a first‑in‑class selective dual T‑cell agonist, for the treatment of advanced colorectal cancer with high tumor mutational burden (TMB‑H).[reference:27][reference:28] This designation reflects the growing recognition of TMB as a valid biomarker for guiding therapy, even in cancers that are not MSI‑H.
💉 The Future: Personalized Cancer Vaccines
If predicting response is the first step, the ultimate goal is to create a response where none exists. This is the promise of personalized neoantigen vaccines—therapies designed to train a patient's own immune system to recognize and attack their unique tumor mutations.
Recent breakthroughs include:
- Neoantigen vaccines for kidney cancer: A study published in Nature in 2025 demonstrated that a neoantigen vaccine generated antitumor immunity in patients with renal cell carcinoma.[reference:29]
- Nous‑209 for Lynch syndrome: This vaccine targets 209 shared neoantigens found across MSI‑high neoplasms and is being tested in a Phase 1b/2 trial for cancer prevention in Lynch syndrome carriers.[reference:30]
- Personalized mRNA vaccines for triple‑negative breast cancer: A 2026 trial showed that an individualized neoantigen mRNA vaccine induced durable, functional anti‑tumor T‑cell responses in patients with this aggressive breast cancer subtype.[reference:31]
- Oncolytic viruses encoding neoantigens: Researchers are now engineering viruses that selectively infect and kill cancer cells while simultaneously delivering neoantigens to stimulate a systemic immune response.[reference:32]
📊 Evolution of Immunotherapy Response Prediction: 2019 vs. 2026
| Biomarker/Tool | 2019 Status | 2026 Status |
|---|---|---|
| MSI (Microsatellite Instability) | Emerging evidence for MSI intensity as a predictor | Established biomarker; guides first‑line therapy in colorectal and other dMMR cancers |
| TMB (Tumor Mutational Burden) | Concept recognized but not widely used clinically | FDA‑approved biomarker; used to identify responders in MSS tumors; "Very High" TMB under investigation |
| Neoantigen Prediction | Basic understanding that indels create neoantigens | Sophisticated tools (NeoPrecis, neoIM) integrate clonality and immunogenicity; improve prediction by up to 20% |
| Liquid Biopsy (ctDNA) | Primarily a research tool | Clinically available assays (e.g., Labcorp Plasma Focus™) used to predict and monitor response in real time |
| AI/Machine Learning | Not applied | Models like HAPIR, PathHDNN, and TxGemma demonstrate high accuracy in predicting response |
| Personalized Cancer Vaccines | Experimental | Early‑phase clinical trials show durable immune responses in kidney, breast, and colorectal cancers |
📋 The Bottom Line: Key Takeaways for 2026
🔬 MSI and TMB are now standard biomarkers: MSI‑H/dMMR status guides first‑line immunotherapy in colorectal and other cancers. TMB is FDA‑approved and helps identify responders in MSS tumors.
🧬 Neoantigen prediction has matured: Tools like NeoPrecis and neoIM integrate clonality and immunogenicity to improve response prediction by up to 20% over TMB alone.
🩸 Liquid biopsy is changing monitoring: ctDNA analysis can predict response before treatment and detect progression months before imaging, enabling earlier treatment adjustments.
🤖 AI is enhancing prediction: Machine learning models like HAPIR, PathHDNN, and TxGemma are demonstrating high accuracy in predicting immunotherapy response across cancer types.
💊 New therapies are approved: Dual checkpoint blockade (nivolumab + ipilimumab) is now approved for MSI‑H/dMMR colorectal cancer. Fast‑tracked agents like invikafusp alfa target TMB‑H cancers.
💉 Personalized vaccines are on the horizon: Neoantigen vaccines are showing durable immune responses in early‑phase trials for kidney, breast, and colorectal cancers, with potential for combination with checkpoint inhibitors.
⚕️ The future is integrated: The optimal approach will combine genomic biomarkers (MSI, TMB), computational neoantigen prediction, AI‑driven risk stratification, liquid biopsy monitoring, and personalized vaccines—all tailored to the individual patient.
🔔 Enjoy evidence‑based oncology insights? Subscribe for free weekly updates — no spam, just science.