Developed by researchers at City of Hope and Memorial Sloan Kettering, the tool uses routine blood tests to flag patients at high risk of relapse after CAR T therapy.

Chimeric antigen receptor T-cell therapy (CAR T) has dramatically altered the treatment landscape for relapsed or refractory non-Hodgkin lymphoma (NHL), offering a beacon of hope where few options remain. Yet, a persistent challenge remains: more than half of patients treated with CAR T experience disease progression or relapse within six months.

In a landmark study published in Nature Medicine (2025), a multidisciplinary team led by Roni Shouval, Marcel R.M. van den Brink, Miguel-Angel Perales, and Hyung C. Suh, with contributions from over 50 co-authors across City of Hope, Memorial Sloan Kettering Cancer Center (MSK), and other institutions, introduced InflaMix—a novel machine learning model trained to assess systemic inflammation and predict CAR T outcomes in NHL.

A Quantitative Tool to Measure a Qualitative Risk

InflaMix, short for INFLAmmation MIXture Model, is an unsupervised Gaussian mixture model that evaluates 14 pre-treatment laboratory and cytokine markers, including CRP, ferritin, IL-6, and LDH, among others. By recognising hidden patterns in these markers, the tool stratifies patients into inflammatory and non-inflammatory profiles prior to CAR T-cell infusion.

According to lead author Sandeep S. Raj and colleagues, the inflammatory signature defined by InflaMix was strongly associated with treatment failure. Patients in the inflammatory group had a hazard ratio of 2.98 (95% CI, 1.60–4.91; P < 0.001) for death or disease relapse compared to the non-inflammatory cluster.

Trained on One Cohort, Validated Across Many

InflaMix was first trained on a cohort of 149 patients with large B-cell lymphoma treated at MSK. The model was then rigorously validated across three independent cohorts comprising 688 additional patients with diverse clinical profiles and CAR T products, from institutions including Sheba Medical Center (Israel), Hackensack Meridian Health (New Jersey), and City of Hope (California).

Notably, even when using only six standard lab tests—albumin, AST, ALP, CRP, haemoglobin, and LDH—the model maintained high predictive performance, demonstrating its value as a real-world, point-of-care tool.

Beyond Prognosis: A Predictive and Actionable Tool

Unlike many risk scores trained on clinical outcomes, InflaMix is unsupervised—built without any outcome labels. Despite this, it proved highly predictive of treatment failure, even after adjusting for known risk factors such as tumour burden and primary refractory disease. The tool also outperformed conventional inflammation markers like CRP alone and more complex regularised models.

The authors showed that integrating InflaMix into clinical prediction models significantly improved area under the ROC curve (AUROC) for progression-free survival at 6 months and enhanced decision curve analyses for evaluating consolidation therapies after CAR T.

The Signature of Inflammation and Its Dynamics

The model does more than assign risk—it also reveals how inflammation evolves across treatment stages. Patients who transitioned from the inflammatory to the non-inflammatory cluster between early time points (e.g. apheresis or lymphodepletion) and infusion showed improved outcomes, suggesting that systemic inflammation is not fixed and may be modifiable through pre-treatment interventions.

This insight opens the door to new treatment strategies. If inflammation can be resolved before CAR T infusion—potentially through bridging therapy, corticosteroids, or targeted anti-inflammatory agents—it may improve the therapy’s success rate.

Accessible and Open-Source

To facilitate clinical adoption, the researchers have released the InflaMix calculator as a free, open-source tool on GitHub: https://github.com/vdblab/InflaMix. Clinicians can input available blood test data and receive a patient’s inflammatory cluster assignment along with the associated risk estimate—even if some values are missing.

A New Era for Personalised CAR T Decision-Making

The study, titled “An inflammatory biomarker signature of response to CAR-T cell therapy in non-Hodgkin lymphoma”, represents one of the most thorough validations to date of a predictive model for CAR T response. It reflects a growing shift in oncology: using artificial intelligence not just to describe outcomes, but to anticipate and act on them.

As co-senior author Dr Marcel van den Brink, President of City of Hope Los Angeles, noted:

“By using machine learning and blood tests, we could develop a highly reliable tool that helps predict who will respond well to CAR T therapy. This is one of the most validated tests we have for lymphoma patients and could support oncologists in risk assessment worldwide.”

With CAR T therapy expanding to other malignancies and new T-cell-based treatments entering the clinic, InflaMix may soon become a cornerstone of personalised oncology—ensuring that advanced therapies are matched with the right patients at the right time.

Study reference:
Sandeep S. Raj, Teng Fei, Shalev Fried, et al. An inflammatory biomarker signature of response to CAR-T cell therapy in non-Hodgkin lymphoma. Nature Medicine (2025). https://www.nature.com/articles/s41591-025-03532-x