Fig. 4: Random forest approach predicts natalizumab treatment outcome in MS patients.

A Schematic representation of the machine learning approach built to predict response to natalizumab treatment based on the analysis of pretreatment PBMCs with the HCI pipeline. A random forest (RF) model was trained on the HCI dataset from a discovery cohort composed of 19 responders and 21 non-responder patients. The elaborated model, which runs with 1,048,575 combinations of a restricted set of 20 features, was then challenged with the HCI dataset from a validation cohort composed of 14 responders and 13 non-responder patients. Created in BioRender. Chaves, B. (2025) https:// BioRender.com/3068pko. B Schematic drawing of the F1 score parameter, which maximizes the performance of correct response classification and limits incorrect classification of patients. Created in BioRender. Chaves, B. (2025) https://BioRender.com/ mx3ulna. C, E Parameters of RF performance evaluation of discovery (C) and validation (E) cohorts. Prediction was performed with the 20 most important features pointed out by the RF. All possible combinations of these features were tested, yielding 1,048,575 combinations. Parameters are shown for combinations that achieved an F1 score of ≥85% in both cohorts, to ensure robustness. Data were obtained from 8 to 6285 CD8⁺ T cells and calculated on the normalized delta values. The dashed line depicts a threshold of 85% above which the parameters are associated with high performance. D, F Feature importances (black) and frequencies on RF combinations with F1 score ≥85 % (pink) for the discovery (D) and validation (F) cohorts. Arrows point to common features on RF combinations with an F1 score ≥85% across the two cohorts. G–K Relative delta values are shown for actin profile 2/5 SER-saddle and SER-edge (G, H), actin symmetry 03 (I), nucleus area (J), and cell area (K). The relative delta values represent the in vitro effect of natalizumab on CD8⁺ T cells from responder (R, red dots) and non-responder patients (NR, blue dots). Data were obtained from CD8⁺ T cells from 67 MS patients, corresponding to 8 to 6285 CD8⁺ T cells per condition. Data stems from the mean of three independent experiments per patient. Bars represent the mean of the considered patient subgroup. p values were calculated by a two-sided unpaired t-test with Bonferroni correction (significance = p ≤ 0.0025). Source data for all panels are provided as a Source Data file.