Fig. 1: Schematic for NeuroPM-box software workflow and practical guidelines.

a The primary software data inputs include molecular (e.g., RNA and proteins concentration arrays), multimodal imaging (e.g., tau, amyloid-β and glucose metabolism PET, vascular, functional, and structural MRI), whole-brain connectomics (e.g., structural and vascular networks), cognitive/clinical evaluations, and therapeutic interventions (e.g. medication). There is no restriction on the number of modalities that can be analyzed. b The NeuroPM-box interface allows users to select from four analytical methods (tools), apply auxiliary applications, and access the visualization tool (NeuroPM-viewer) and software tutorial. c Main software modules supporting the data-driven analysis of the multimodal data. d NeuroPM-viewer enables detailed exploration of the human cortex of both real and modeled spatiotemporal brain dynamics (see also Fig. S2). e Practical guidelines for methods users (available methods are further described in Table 1 and subsections below). Essentially, the analytical methods belong to two main categories, empirical and mechanistic. The former is purely data driven and focus on identifying and interpreting intrinsic patterns in the data without making strong a priori biological assumptions. Specifically, the included algorithm (see contrasted Trajectory Inference subsection and summary on Table 1) provide individualized quantitative scores reflective of disease progression and assign each subject to distinctive subpopulations (tentatively reflecting different disease subtrajectories). Any type of quantitative data can be used as input (e.g. transcriptomic, proteomic, histopathological, metabolomics, multimodal imaging, clinical), while each data-feature’s contribution to the subjects’ final stratification is quantified, revealing the most informative features (e.g. specific genes, brain regions, clinical evaluations) and associated data modalities (e.g. RNA, imaging, clinical). However, the user should avoid performing causal interpretations based on empirical modeling, because the intrinsic limitation to distinguish between direct and indirect biological effects. Mechanistic models, by the contrary, aims to decode cause-effects in terms of biological factors alterations spreading through physical brain connections and/or synergistic factor–factor interactions contributing to spatiotemporal brain reorganization. The two implemented generative models focus on uni-modal or multimodal imaging data, i.e. ESM considers the intra-brain spreading of a unique biological factor measured with an specific imaging modality (e.g. tau-PET, or amyloid-β PET), while MCM considers the direct (causal) interactions and concurrent intra-brain spreading of multiple biological factors’ alterations quantified with different imaging modalities (tau, amyloid-β, and glucose metabolism PET; cerebrovascular flow, functional activity indicators, and structural atrophy measured with MRI). Notably, mechanistic approaches (ESM, MCM, pTIF) can be informed by the empirical data-driven outputs (cTI stratification), allowing the incorporation of a wide range of possible multiscale biological information (e.g. molecular and clinical stages and subtypes) on the imaging-based generative brain models (see “cTI → ESM, MCM, pTIF” method on Table 1). Finally, personalized causal brain models identified by the multimodal mechanistic approach (see MCM) can be interrogated to identify individual therapeutic needs in terms of biological deformations required to stop/revert factor(s)-specific (imaging modalities) alterations or clinical deterioration (see pTIF on Table 1 and subsequent subsection).