DeepMF: deciphering the latent patterns in omics profiles with a deep learning method.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-throughput technologies generate large omics datasets.
• Matrix factorization (MF) methods map omics data to low-dimensional spaces for biological insights.
• Current MF tools struggle with noisy and missing data, limiting their real-world application.

Purpose of the Study:

• To develop a robust matrix factorization model capable of handling noisy and missing omics data.
• To improve the accuracy of cancer subtype discovery and data imputation using omics profiles.
• To provide a deep learning-based approach for uncovering underlying biological processes.

Main Methods:

• Proposed DeepMF, a deep neural network-based matrix factorization model.
• DeepMF disentangles molecular feature-associated and sample-associated latent matrices.
• The model is designed to be tolerant to noisy and missing values in omics data.

Main Results:

• DeepMF achieved high cancer subtype discovery accuracy (up to 100%) across various cancer types (medulloblastoma, leukemia, breast, small-blue-round-cell).
• Demonstrated superior data recovery capacity (silhouette values up to 0.6) with 70% missing entries compared to state-of-the-art MF tools.
• Showcased improved embedding strength (clustering accuracy up to 100%) on multiple cancer datasets.

Conclusions:

• DeepMF exhibits robust denoising, imputation, and embedding capabilities for omics data.
• The model provides valuable insights for biological process discovery, particularly in cancer subtype identification.
• An implementation of DeepMF is publicly available for research use.