Reproduce “POEMS: Product of Experts for Interpretable Multi-omic Integration using Sparse Decoding” on OpenResearch

About this paper

Mihriban Kocak Balik, Pekka Marttinen, Negar Safinianaini

The POEMS framework, developed by Aalto University, integrates diverse molecular data using a Product of Experts approach and sparse decoding to achieve robust cancer subtyping while providing comprehensive biological interpretability. It delivers strong predictive performance, with a K-means accuracy of 0.63 on the TCGA BRCA dataset, alongside interpretable insights into biomarker associations and omic contributions.

Key insights

• The Product-of-Experts approach effectively integrates diverse omic modalities into a coherent shared latent space, emphasizing agreement and robustness.
• Sparsity in decoder mappings enables direct and interpretable identification of specific omics features (biomarkers) associated with underlying biological processes.
• An adaptive gating mechanism provides novel insights into the per-sample contribution of each omic, revealing their dynamic importance.
• Interpretability can be achieved in deep generative models without sacrificing nonlinear expressive power or predictive performance through optimized sparse decoding.

Summary by alphaXiv.