Background: In 2009, the National Institute of Mental Health launched the Research Domain Criteria (RDoC), a paradigm shift in psychiatry highlighting the need to move beyond the currently used diagnostic categories, and ultimately promoting precision psychiatry. RDoC is a research framework integrating different levels of measures (e.g. brain imaging and behaviour) with the aim of understanding the basic dimensions of functioning from normal to abnormal. Methods: Here, we propose an innovative machine learning framework combined with sparse partial least squares (SPLS) to identify hidden dimensions of brain-behaviour associations, therefore a potential analytic tool to subserve the RDoC ideal. To illustrate the approach, we investigate structural brain-behaviour associations in an extensively phenotyped developmental sample of 345 participants (312 healthy, 33 clinically depressed). The brain data consisted of whole brain grey matter volumes, the behavioural data included item-level self-report questionnaires, IQ and demographic measures. Results: SPLS captured two hidden dimensions of brain-behaviour relationships: one related to age and drinking and the other one to depression. The applied machine learning framework indicates that these results are stable and generalize well to new data. Indeed, the identified brain-behaviour associations are in agreement with previous findings in the literature concerning age, alcohol use and depression-related changes in brain volume. Conclusion: SPLS embedded in our novel framework is a promising tool to link behaviour/symptoms to neurobiology, thus it has a great potential to contribute to a biologically grounded definition of psychiatric diseases.