Abstract Background Early Psychosis patients (EP, within 3 years after psychosis onset) show significant variability, making outcome predictions challenging. Currently, little evidence exists for stable relationships between neural microstructural properties and symptom profiles across EP diagnoses, limiting the development of early interventions. Methods A data-driven approach, Partial Least Squares (PLS) correlation, was used across two independent datasets to examine multivariate relationships between white matter (WM) properties and symptomatology, to identify stable and generalizable signatures in EP. The primary cohort included EP patients from the Human Connectome Project-Early Psychosis (n=124). The replication cohort included EP patients from the Feinstein Institute for Medical Research (n=78). Both samples included individuals with schizophrenia, schizoaffective disorder, and psychotic mood disorders. Results In both cohorts, a significant latent component (LC) corresponded to a symptom profile combining negative symptoms, primarily diminished expression, with specific somatic symptoms. Both LCs captured comprehensive features of WM disruption, primarily a combination of subcortical and frontal association fibers. Strikingly, the PLS model trained on the primary cohort accurately predicted microstructural features and symptoms in the replication cohort. Findings were not driven by diagnosis, medication, or substance use. Conclusions This data-driven transdiagnostic approach revealed a stable and replicable neurobiological signature of microstructural WM alterations in EP, across diagnoses and datasets, showing a strong covariance of these alterations with a unique profile of negative and somatic symptoms. This finding suggests the clinical utility of applying data-driven approaches to reveal symptom domains that share neurobiological underpinnings.