Dimensionality reduction methods for extracting functional networks from large-scale CRISPR screens

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• CRISPR-Cas9 screens are vital for discovering gene functions and cancer dependencies.
• The Cancer Dependency Map (DepMap) is a large dataset of these screens but contains a mitochondria-associated bias.
• This bias can obscure important biological signals, necessitating improved normalization techniques.

Purpose of the Study:

• To evaluate unsupervised dimensionality reduction methods for normalizing the DepMap dataset.
• To improve the accuracy of co-essentiality networks derived from CRISPR screen data.
• To develop and validate a novel normalization strategy for large-scale cancer dependency datasets.

Main Methods:

• Exploration of autoencoders, robust principal component analysis (PCA), and classical PCA for data normalization.
• Development of a novel 'onion' normalization technique to integrate multiple normalized data layers.
• Benchmarking of normalization methods against existing approaches using DepMap data.

Main Results:

• Robust PCA combined with the proposed onion normalization significantly outperformed other methods.
• Dimensionality reduction effectively removed confounding low-dimensional signals, such as mitochondrial bias.
• Normalized DepMap data yielded improved functional gene networks.

Conclusions:

• Unsupervised dimensionality reduction is valuable for normalizing large CRISPR screen datasets like the DepMap.
• The robust PCA and onion normalization method offers a superior approach for enhancing cancer dependency network analysis.
• These tools provide a generalizable strategy for improving the interpretation of functional genomics data.