Deep Learning-Based Multimodal Clustering Model for Endotyping and Post-Arthroplasty Response Classification using Knee Osteoarthritis Subject-Matched Multi-Omic Data

Background Primary knee osteoarthritis (KOA) is a heterogeneous disease with clinical and molecular contributors. Biofluids contain microRNAs and metabolites that can be measured by omic technologies. Deep learning captures complex non-linear associations within multimodal data but, to date, has not been used for multi-omic-based endotyping of KOA patients. We developed a novel multimodal deep learning framework for clustering of multi-omic data from three subject-matched biofluids to identify distinct KOA endotypes and classify one-year post-total knee arthroplasty (TKA) pain/function responses.

Materials and Methods In 414 KOA patients, subject-matched plasma, synovial fluid and urine were analyzed by microRNA sequencing or metabolomics. Integrating 4 high-dimensional datasets comprising metabolites from plasma (n=151 features), along with microRNAs from plasma (n=421), synovial fluid (n=930), or urine (n=1225), a multimodal deep learning variational autoencoder architecture with K-means clustering was employed. Features influencing cluster assignment were identified and pathway analyses conducted. An integrative machine learning framework combining 4 molecular domains and a clinical domain was then used to classify WOMAC pain/function responses post-TKA within each cluster.

Findings Multimodal deep learning-based clustering of subjects across 4 domains yielded 3 distinct patient clusters. Feature signatures comprising microRNAs and metabolites across biofluids included 30, 16, and 24 features associated with Clusters 1-3, respectively. Pathway analyses revealed distinct pathways associated with each cluster. Integration of 4 multi-omic domains along with clinical data improved response classification performance, with Cluster 3 achieving AUC=0·879 for subject pain response classification and Cluster 2 reaching AUC=0·808 for subject function response, surpassing individual domain classifications by 12% and 15% respectively.

Interpretation We have developed a deep learning-based multimodal clustering model capable of integrating complex multi-fluid, multi-omic data to assist in KOA patient endotyping and test outcome response to TKA surgery.

Funding Canada Research Chairs Program, Tony and Shari Fell Chair, Campaign to Cure Arthritis, University Health Network Foundation.