DeepMF

DeepMF performs deep neural network-based matrix factorization to denoise, impute, and embed high-throughput omics data (mRNA, miRNA, and protein) into low-dimensional spaces to reveal latent biological patterns.

Key Features:

• Deep Neural Network Architecture: Uses a deep neural network-based matrix factorization approach to model omics matrices and generate low-dimensional embeddings.
• Tolerance to Noisy and Incomplete Data: Handles noisy and incomplete mRNA, miRNA, and protein datasets to maintain robust performance.
• Cancer Subtype Discovery: Achieved clustering accuracies of 76% (medulloblastoma), 100% (leukemia), 92% (breast cancer, TCGA BRCA), and 100% (small-blue-round-cell, SRBCT) for cancer subtype discovery.
• High Recovery Capacity: Outperformed other matrix factorization tools with up to 70% missing entries, yielding silhouette values of 0.47 (medulloblastoma), 0.6 (leukemia), 0.28 (TCGA BRCA), and 0.44 (SRBCT).
• Enhanced Embedding Strength: Produced embedding-driven clustering accuracies of 88% (medulloblastoma), 100% (leukemia), 84% (TCGA BRCA), and 96% (SRBCT), improving over previous best results of 76%, 100%, 78%, and 87%, respectively.
• Denoising, Imputation, and Embedding: Performs denoising, imputation, and embedding of quantitative omics profiling matrices (mRNA, miRNA, protein).

Scientific Applications:

• Cancer Subtype Discovery: Supports identification and clustering of cancer subtypes across datasets such as medulloblastoma, leukemia, TCGA BRCA, and SRBCT using low-dimensional embeddings.
• Latent Biological Pattern Discovery: Reveals latent biological processes and associations in molecular profiles (mRNA, miRNA, protein) through low-dimensional representation.

Methodology:

Applies deep neural network-based matrix factorization that disentangles molecular feature-associated and sample-associated latent matrices and maps quantitative omics profiling matrices into low-dimensional embedding spaces.