Prediction of antidepressant treatment response with thalamo-somatomotor functional connectivity revealed by generalizable stratification of depressed patients

Abstract

Major depressive disorder (MDD) is diagnosed based on symptoms and signs without relying on physical, biological, or cognitive tests. MDD patients exhibit a wide range of complex symptoms, and it is assumed that there are diverse underlying neurobiological backgrounds, possibly composed of several subtypes with relatively homogeneous biological features. Initiatives, including the Research Domain Criteria, emphasize the importance of biologically stratifying MDD patients into homogeneous subtypes using a data-driven approach while utilizing genetic, neuroscience, and cognitive information. If biomarkers can stratify MDD patients into biologically homogeneous subtypes at the first episode of depression, personalized precision medicine may be within our scope. Some pioneering studies have used resting-state functional brain connectivity (rs-FC) for stratification and predicted differential responses to various treatments for different subtypes. However, to our knowledge, little research has demonstrated reproducibility (i.e., generalizability) of stratification markers in independent validation cohorts. This issue may be due to inherent measurement and sampling biases in multi-site fMRI data, or overfitting of machine learning algorithms to discovery cohorts with small sample sizes, i.e., a lack of appropriate machine learning algorithms for generalizable stratification. To address this problem, we have constructed a multi-site, multi-disorder fMRI database with prospectively and retrospectively harmonized data from thousands of samples and proposed a hierarchical supervised/unsupervised learning strategy. In line with this strategy, our previous research first developed generalizable MDD diagnostic biomarkers using this fMRI database of MDD patients via supervised learning. The MDD diagnostic biomarker determines the importance of thousands to tens of thousands of rs-FCs across the whole brain for MDD diagnosis. In this study, we constructed stratification markers for MDD patients using unsupervised learning (Multiple co-clustering) with a subset of top-ranked rs-FCs in the MDD diagnostic biomarker. We developed a method to evaluate the clustering stability between two independent datasets as a generalization metric of stratification biomarkers. To discover stratification biomarkers with high stability across datasets, we utilized two multi-site datasets with substantial differences in data acquisition facilities and fMRI measurement protocols (Dataset-1: a dataset of 138 depressed patients obtained with a unified measurement protocol across three facilities; Dataset-2: a dataset of 181 depressed patients obtained with non-unified measurement protocols across four facilities, distinct from Dataset-1). Starting from several diagnostic biomarkers, we constructed some stratification markers and identified the stratification biomarker with the highest clustering stability between the two datasets. This stratification biomarker was based on several rs-FCs between the thalamus and the postcentral gyrus, and the MDD subgroups stratified by this biomarker showed significantly different treatment responsiveness to a selective serotonin reuptake inhibitor (SSRI). By narrowing down whole-brain rs-FCs using MDD diagnostic biomarkers and further dividing the rs-FCs using multiple co-clustering, the feature dimension was significantly reduced, thereby avoiding overfitting to the training data and successfully constructing stratification biomarkers that are highly stable between independent datasets, i.e., have generalizability. Furthermore, the correlation between MDD subgroups and antidepressant treatment response was demonstrated, suggesting the potential for achieving personalized precision medicine for MDD.