FAVA: High-quality functional association networks inferred from scRNA-seq and proteomics data

Abstract

Protein networks are commonly used for understanding the interplay between proteins in the cell as well as for visualizing omics data. Unfortunately, most existing high-quality networks are heavily biased by data availability, in the sense that well-studied proteins have many more interactions than understudied proteins. To create networks that can help elucidate functions for the latter, we must start from data that are not affected by this literature bias, in other words, from omics data such as single cell RNA-seq (scRNA-seq) and proteomics. While networks can be inferred from such data through simple co-expression analysis, this approach does not work well due to high sparseness (many transcripts/proteins are not consistently observed in each cell/sample) and redundancy (many similar cells/samples are analyzed) of such data. We have therefore developed FAVA, Functional Associations using Variational Autoencoders, which deals with both issues by compressing these high-dimensional data into a dense, low-dimensional latent space. We demonstrate that calculating correlations in this latent space results in much improved networks compared to the original representation for large-scale scRNA-seq and proteomics data from the Human Protein Atlas, and from PRIDE, respectively. We show that these networks, which given the nature of the input data should be free of literature bias, indeed have much better coverage of understudied proteins than existing networks.

The STRING database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest

“…To address that, in the latest version we have utilized a novel method called FAVA (Functional Associations using Variational Autoencoders) ( 68 ) to build STRING’s co-expression network. This deep-learning model reduces the dimensionality of the data into lower-dimensional latent spaces using variational autoencoders (VAE).…”

Toward an Integrated Machine Learning Model of a Proteomics Experiment

“…Both approaches rely on an autoencoder architecture to reduce data noise and ambiguity of high-dimensional data. Notable uses of these approaches includes exploring relevant clusters in single cell proteomics data as well as inferring protein–protein interaction networks from high-dimensional combined RNA-Seq and proteomics data . In addition to dimensionality reduction techniques, multimodel representation learning is well suited for proteomics due to its ability to integrate features from different modalities and identify attributes that are shared and different.…”

“…A complex remaining task is the prediction of tissue- or organism level changes in protein abundance or post-translational modification, which are intrinsically linked to protein function. Machine learning may complement databases and ontologies to improve functional annotation of proteins, predicting function based on similarity to proteins with known functions in protein families, inferring function from coregulation of proteins found in large-scale proteomics studies or integrating protein and RNA-Seq data . Machine learning has already been used to predict functional relevance of phosphorylation by combining multiple databases and repositories .…”

“…Notable uses of these approaches includes exploring relevant clusters in single cell proteomics data as well as inferring protein–protein interaction networks from high-dimensional combined RNA-Seq and proteomics data. 11 In addition to dimensionality reduction techniques, multimodel representation learning 12 is well suited for proteomics due to its ability to integrate features from different modalities and identify attributes that are shared and different. Successful applications, such as Contrastive Language–Image Pretraining (CLIP 13 ), can integrate image and text data into a shared embedding by using a contrastive loss function that trains a model with pairs of positive and negative examples from the two modalities (image and text).…”

“…Machine learning may complement databases and ontologies to improve functional annotation of proteins, predicting function based on similarity to proteins with known functions in protein families, 54 inferring function from coregulation of proteins found in large-scale proteomics studies 55 or integrating protein and RNA-Seq data. 11 Machine learning has already been used to predict functional relevance of phosphorylation by combining multiple databases and repositories. 56 The proteome scale availability of three-dimensional protein structures from AlphaFold enables systematically investigating PTMs in their structural context to further improve our understanding of their functional relevance, and hence our ability to predict their changes.…”