The study, led by Ahmad Al Khleifat and Ammar Al-Chalabi of King’s College London together with collaborators across the United Kingdom and France, examined data from MIROCALS, a phase 2b randomised, double-blind, placebo-controlled trial of low-dose interleukin-2 in ALS. That trial, which enrolled 220 participants, had previously produced a striking finding: the concentration of phosphorylated neurofilament heavy chain, or pNFH, in the cerebrospinal fluid was not only a powerful predictor of survival in its own right, but also modified the treatment effect of the immunomodulatory drug. In other words, the amount of this protein, which leaks into the fluid surrounding the brain and spinal cord when axons are damaged, told researchers which patients were likely to respond to the therapy. The question the new analysis posed was whether a purely clinical tool, the TRICALS risk score, could do the same job without the need for an invasive lumbar puncture.

The TRICALS survival prediction model is an impressive piece of clinical epidemiology in its own right. Derived from more than 10,000 individuals across international cohorts, it combines age, the delay between symptom onset and diagnosis, scores on the revised ALS Functional Rating Scale, the site of disease onset, and respiratory status into a single validated index that robustly predicts how long a patient is likely to live. Because it requires nothing more than routine clinical assessment, it is cheap, quick, and easy to apply in any trial centre in the world. If it could replace a molecular biomarker, trial designers could streamline their statistical models considerably, avoiding the complexity and cost of collecting and assaying cerebrospinal fluid samples. That possibility is what made the new analysis worth doing.

The researchers designed two prespecified comparisons. In the first, the substitution analysis, they took the published MIROCALS biomarker model, which included the functional rating score, age, regulatory T-cell frequency, plasma CCL2, cerebrospinal fluid pNFH, treatment arm, and a pNFH-by-treatment interaction, and simply swapped the neurofilament measurement for the TRICALS risk score. In the second, the augmentation analysis, they kept the full biomarker model intact and added the TRICALS score on top, testing whether the clinical phenotype contributed anything beyond what the molecular marker already captured. Model performance was judged using likelihood ratio statistics and the negative two log likelihood, standard measures of how well a statistical model describes the observed survival data.

The results of the substitution analysis were unambiguous. Replacing pNFH with the TRICALS score caused model performance to deteriorate sharply, with the negative two log likelihood rising from 793.36 in the published biomarker model to 928.25 under the substituted specification. The TRICALS score did interact significantly with treatment, yielding a hazard ratio of 0.72 with a 95 percent confidence interval of 0.54 to 0.96 and a p value of 0.025, which means a treatment effect would still have been detected. But the contribution of that interaction to overall model fit was modest, adding a change in the chi-squared statistic of only 4.76, compared with 14.24 for the pNFH-by-treatment interaction in the original model. The clinical score captured some of the treatment-response signal, but only a fraction of it.

The correlation analysis explained why. The TRICALS risk score and cerebrospinal fluid pNFH were only moderately correlated, with a Pearson coefficient of 0.357 across the 220 participants. The coefficient of determination, the square of that value, comes to just 0.127, meaning the clinical score accounts for roughly 12.7 percent of the variance in the neurofilament concentration. The remaining 87 percent reflects biology that the clinical assessment simply cannot see. Neurofilament levels track the ongoing destruction of motor axons at a molecular level, a process that unfolds at different rates in different patients and is only loosely reflected in how quickly their symptoms have progressed so far. Two patients with identical age, onset site, functional scores, and diagnostic delay can carry very different burdens of active neurodegeneration, and it is that burden, rather than the clinical phenotype, that appears to govern response to the immune-targeting therapy.

The augmentation analysis told the complementary story. When cerebrospinal fluid pNFH and its interaction with treatment were added to the TRICALS-based model, the fit improved substantially, with a change in chi-squared of 14.24 and a p value below 0.001. The interaction term itself was highly significant, and once it entered the model, the treatment effect became robust, with a hazard ratio of 0.28 and a 95 percent confidence interval of 0.12 to 0.63, a p value of 0.002. Notably, both the TRICALS score and pNFH remained independently associated with survival in the combined model, indicating that each variable carried information the other did not. The final framework simultaneously captured baseline clinical severity, biological disease activity, and the heterogeneity of treatment response, three dimensions that no single measure could encompass alone.

The authors are careful to acknowledge the limitations of their work. This was a post-hoc analysis of an existing trial dataset rather than a formal reanalysis, designed specifically to test whether a phenotypic score could substitute for a fluid biomarker of neurodegeneration. The complete-case approach, the relatively modest sample size of 220 participants, and the focus on a single immunomodulatory intervention all constrain how far the conclusions can be generalised. Yet the biological logic of the finding is compelling, and it aligns with a growing body of evidence linking neurofilament concentrations to the rate and extent of neuroaxonal degeneration in ALS. The study was conducted in compliance with Good Clinical Practice and the Declaration of Helsinki, with all analyses performed on fully anonymised data under the trial’s data-sharing agreements.

The implications for trial design are immediate and practical. Composite clinical risk scores remain valuable for adjusting for baseline differences between treatment groups, and the new results confirm that they do carry some information about treatment-response heterogeneity. But they should not be treated as surrogates for molecular stratification. As ALS therapies increasingly target specific biological pathways, from immune modulation and protein aggregation to RNA metabolism and mitochondrial dysfunction, no single biomarker is likely to serve every trial. Instead, the authors argue, optimal design will require identifying molecular markers that directly index the biological process each intervention is meant to hit. For an immune-targeting drug like low-dose interleukin-2, the burden of neuroaxonal injury, read out through pNFH, proved to be that marker.

The broader message is a pragmatic blueprint for precision medicine in neurodegenerative disease. Clinical phenotype should anchor baseline risk adjustment; disease-activity biomarkers such as neurofilaments should identify the patients most likely to respond; and integrating both within a unified survival modelling framework offers mechanistic insight, statistical efficiency, and a clearer path to interpretable, therapeutically informative trials. For a disease that has defeated dozens of experimental therapies over three decades, the lesson is sobering but constructive: the shortcuts that seem most convenient, the clinical scores that require no needle and no laboratory, are precisely the ones that cannot carry the full weight of prediction. The biology, inconvenient as it may be, must be measured directly.

Subject of Research: Comparison of the TRICALS clinical risk score and cerebrospinal fluid neurofilament biomarkers for prognosis and treatment stratification in amyotrophic lateral sclerosis trials

Article Title: TRICALS risk score augments but cannot replace neurofilament as a prognostic biomarker in ALS

Article References: Al Khleifat, A., Malaspina, A., Kirby, J., Tree, T., Shaw, P. J., Leigh, P. N., & Al-Chalabi, A. (2026). TRICALS risk score augments but cannot replace neurofilament as a prognostic biomarker in ALS. Journal of Neurology, 273(10), Article 643. https://doi.org/10.1007/s00415-026-14040-4

Image Credits: AI Generated

DOI: 10.1007/s00415-026-14040-4

Keywords: amyotrophic lateral sclerosis, ALS, TRICALS risk score, neurofilament, pNFH, biomarkers, clinical trials, MIROCALS, interleukin-2, prognosis, precision medicine, neurodegeneration