Evaluation of input data modality choices on functional gene embeddings

Abstract

Functional gene embeddings, numerical vectors capturing gene function, provide a promising way to integrate functional gene information into machine learning models. These embeddings are learnt by applying self-supervised machine-learning algorithms on various data types including quantitative omics measurements, protein–protein interaction networks and literature. However, downstream evaluations comparing alternative data modalities used to construct functional gene embeddings have been lacking. Here we benchmarked functional gene embeddings obtained from various data modalities for predicting disease-gene lists, cancer drivers, phenotype–gene associations and scores from genome-wide association studies. Off-the-shelf predictors trained on precomputed embeddings matched or outperformed dedicated state-of-the-art predictors, demonstrating their high utility. Embeddings based on literature and protein–protein interactions inferred from low-throughput experiments outperformed embeddings derived from genome-wide experimental data (transcriptomics, deletion screens and protein sequence) when predicting curated gene lists. In contrast, they did not perform better when predicting genome-wide association signals and were biased towards highly-studied genes. These results indicate that embeddings derived from literature and low-throughput experiments appear favourable in many existing benchmarks because they are biased towards well-studied genes and should therefore be considered with caution. Altogether, our study and precomputed embeddings will facilitate the development of machine-learning models in genetics and related fields.

Analysis of 3760 hematologic malignancies reveals rare transcriptomic aberrations of driver genes

“…We also explored models that include complementary features from external datasets. Specifically, we incorporated co-essentiality modules from DepMap [ 59 ] and a 256-dimensional functional gene embedding that integrates protein–protein interactions (PPI), genome-wide deletion screen results from the DepMap project, co-expression from bulk RNA-seq and single-cell RNA-seq compendia [ 60 ] (see Methods). This resulted in an enhancement of overall performance (average precision increased from 23 to 36%, Figure S 22 , Table S 11 and S 12 ).…”

“…The gene-level features consisted of 21 features from seven gene-level metrics of seven IntOGen tools, nine features from AbSplice-DNA scores, 22 features from OUTRIDER obtained using combinations of fold-change direction, significance, and effect size cutoffs, and, similarly, 11 features from NB-act and 22 features from FRASER ( Supplementary Materials and Methods ). For those models integrating external gene functional data, co-essential modules from DepMap [ 59 ] and 256-dimensional functional gene embeddings [ 60 ] were further included as features. In total, 377 genes listed among the hematologic panel genes ( Supplementary Materials and Methods ) or the hematologic malignancy driver genes from CGC GRCh37 v97 [ 54 ] were used as the positive class for the classifiers.…”

Functional gene embeddings improve rare variant polygenic risk scores

“…For each sample, the covariates age (UKBB data field 21003), age 2 , sex (UKBB data field 31), age:sex, age²:sex, and the first 20 genetic principal components (PCs) (UKBB data field 22009_0. [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20]) were obtained directly from the UK Biobank. All of these were used as covariates for the burden test and for FuncRVP.…”

“…Moreover, FuncRVP led to more robust gene effect estimates and to increased gene discoveries notably among genes that are more genetically constrained. Altogether, these results demonstrate that the integration of functional information across genes improves rare-variant phenotype prediction and gene discovery.Recently, functional gene embeddings, numerical vectors capturing gene function such that genes with similar functions are close in the embedding space, have been proposed as an alternative to gene sets to represent gene functions [18][19][20] . Since they are vector representations rather than sets, functional gene embeddings can retain some quantitative information about functional similarities and are more straightforwardly integrated into machine learning algorithms.…”

“…Recently, functional gene embeddings, numerical vectors capturing gene function such that genes with similar functions are close in the embedding space, have been proposed as an alternative to gene sets to represent gene functions [18][19][20] . Since they are vector representations rather than sets, functional gene embeddings can retain some quantitative information about functional similarities and are more straightforwardly integrated into machine learning algorithms.…”