The Pre-Infusion Bet
If you have ever managed a CAR T-cell infusion clinic, you know the bottleneck: waiting for cytokine release syndrome (CRS) to manifest before deciding to intervene. By then, the damage is done, ICU beds are scarce, and the patient is unstable. A preprint from MD Anderson and Moffitt Cancer Center argues that the answer lies in the blood drawn before the cells enter the vein. They report a 5-protein panel that predicts severe CRS with an AUC of 0.85 in discovery and 0.76 in external validation. An expanded 8-protein panel predicts neurotoxicity (ICANS) with similar metrics. The claim is bold: that a static snapshot of the host’s proteome, taken days before therapy, captures a latent "endothelial–immune priming" state that dictates toxicity risk. If this holds, it shifts the entire operational model of CAR T-cell therapy from reactive crisis management to proactive risk stratification.
The Problem
Current predictive tools for CAR T-cell toxicity are blunt and late-acting. The EASIX score relies on lactate dehydrogenase and platelet counts, which are nonspecific stress markers. Post-infusion cytokine panels (IL-6, IL-8, GM-CSF) are informative but arrive too late to prevent the initial cascade. By the time IL-6 spikes, the endothelial glycocalyx is already shedding, and capillary leak is underway. This forces universal hospitalization for high-risk patients, driving up costs and exposing patients to nosocomial risks. There is no reliable, externally validated tool that uses pre-infusion data to distinguish a patient who will tolerate therapy from one who will require intensive care. The gap is not just clinical; it is a fundamental limitation of relying on downstream inflammatory markers rather than upstream susceptibility.
How It Works
The authors’ approach is a classic translational proteomics pipeline: discover, lock, validate. They did not train a black-box neural net on raw spectra; they built a logistic regression model on curated protein abundance.
Step 1: Unbiased Profiling. They collected plasma (MD Anderson, n=39) and serum (Moffitt, n=59) from patients with relapsed/refractory lymphoma immediately before CAR T-cell infusion. They used data-independent acquisition (DIA) mass spectrometry on a timsTOF HT system. Crucially, they depleted high-abundance proteins (albumin, IgG, etc.) to access the lower-abundance proteome where the signal lives. This yielded ~1,000 quantified proteins per sample.
Step 2: Discovery and Consensus. On the MD Anderson cohort, they identified 58 proteins predictive of Grade ≥2 CRS and 70 for ICANS. They then intersected these with the Moffitt cohort to find consistency. The goal was not just statistical significance, but reproducibility across institutions and biofluid types (plasma vs. serum). This cross-cohort integration yielded a 17-protein CRS signature and a 21-protein ICANS signature. See for the initial volcano plots and for the consensus intersection.
Step 3: Model Locking. From the 17 CRS-associated proteins, they applied logistic regression with backward feature selection to derive a minimal 5-marker panel: SCRIB, MYL6, MTHFD1L, HSP90B1, and MMP2. The model coefficients were fixed on the MD Anderson discovery set. They then applied this exact model to the Moffitt validation set without any refitting or hyperparameter tuning. This is the gold standard for external validation in biomarker studies.
Step 4: Biological Interpretation. The authors used Ingenuity Pathway Analysis (IPA) to map these proteins to upstream regulators. The consensus points to AKT-driven inflammation and endothelial activation. They propose a "dual-anatomy" framework: severe CRS is driven by peripheral macrophage priming and endothelial barrier failure, while ICANS is driven by cerebrovascular junction stripping and hepatic synthetic suppression. See for the schematic of this dual-anatomy model.
Numbers
The performance metrics are strong for a clinical biomarker study, but the drop-off in validation requires scrutiny.
- CRS Panel (5 markers): AUC 0.85 (95% CI 0.72–0.98) on MD Anderson discovery; AUC 0.76 (95% CI 0.63–0.89) on Moffitt validation. The authors report that high-risk tertile patients were 13.84-fold more likely to develop Grade ≥2 CRS in the combined dataset.
- ICANS Panel (8 markers): Adds SPOCK2, SLC3A2, and CD84 to the CRS panel. AUC 0.91 (95% CI 0.81–1.00) on discovery; AUC 0.67 (95% CI 0.51–0.84) on validation. High-risk tertile patients were 8.59-fold more likely to develop Grade ≥2 ICANS.
See for the ROC curves comparing discovery and validation performance.
The drop from 0.91 to 0.67 for ICANS is notable. While 0.67 is better than chance, it lacks the discrimination power seen in the discovery set. The authors attribute this partly to the smaller validation cohort and the difference between plasma and serum matrices. The CRS panel, however, maintained a tighter confidence interval, suggesting greater robustness. The odds ratios are compelling: a high-risk score multiplies the probability of severe toxicity by nearly 14x for CRS. In a clinical setting, this magnitude of shift is sufficient to justify prophylactic tocilizumab or enhanced monitoring.
What's Missing
No study of this scale is complete without acknowledging its blind spots.
- Sample Size and Power: The total cohort is 98 patients. For a high-dimensional proteomics study, this is modest. The confidence intervals on the validation AUCs are wide, particularly for ICANS. A larger, multi-center trial is needed to narrow these bounds and ensure the signal isn’t driven by outliers.
- Matrix Effects: The discovery set used plasma; the validation set used serum. These are not interchangeable. Serum contains clotting factors and platelet-derived proteins that plasma lacks. The fact that the CRS panel transferred well is encouraging, but the ICANS panel’s degradation warrants investigation. Is it a biological difference, or a technical artifact of the collection tube?
- Generalizability to Other CAR T Products: Most patients received axicabtagene ciloleucel (axi-cel). The panel’s performance on other CD19-targeted therapies (e.g., lisocabtagene maraleucel, ciltacabtagene autoleucel) or BCMA-targeted therapies is not reported. Different CAR constructs may trigger distinct cytokine profiles, potentially altering the baseline proteomic landscape.
- Cost and Turnaround Time: The paper does not detail the cost per sample or the turnaround time for the DIA-MS workflow. In a clinical setting, if the test takes 48 hours and costs $5,000, it is unusable. Rapid immunoassays (ELISA/Luminex) for these 5–8 proteins are implied as the next step, but the feasibility of such an assay is not discussed.
Should You Prototype This
Depends on X.
If you are a clinician or researcher involved in CAR T-cell therapy, this paper is worth prototyping, but not with mass spectrometry. The value proposition is the biomarkers, not the platform. The 5-protein CRS panel (SCRIB, MYL6, MTHFD1L, HSP90B1, MMP2) is small enough to implement as a multiplex ELISA or Luminex assay.
The paper states that source data underlying the panel scoring and risk-stratification analyses are provided in the Supplementary Information. Code used to fit and apply the predictive panels is available from the corresponding author upon reasonable request. The mass-spectrometry proteomic data generated during this study are available from the corresponding author upon reasonable request and will be deposited in the ProteomeXchange Consortium (PRIDE partner repository) at the time of publication. However, the immediate next step is not to reproduce the MS workflow, but to validate the protein abundances in your own cohort using a cheaper, faster platform. If you can replicate the 0.76 AUC with an off-the-shelf immunoassay, this becomes a viable clinical tool. If not, the signal may be fragile.
Given the modest sample size and the matrix mismatch, I would not bet a clinical trial on this yet. But I would bet a pilot study on validating these specific proteins in-house. The biological narrative—endothelial priming as a precursor to toxicity—is plausible and mechanistically grounded. The question is whether the signal survives the noise of real-world clinical variability. Start there.