Source-linked AI summary
ProDiGe: PRioritization Of Disease Genes with multitask machine learning from positive and unlabeled examples
Fantine Mordelet, Jean-Philippe Vert
TL;DR
Disease-gene discovery commonly yields large candidate lists whose causal genes are costly and time-consuming to identify. ProDiGe integrates heterogeneous gene information with positive-unlabeled learning and shares disease-gene information across diseases. On real OMIM data, it outperformed established prioritization methods, with performance depending on the disease-sharing variant and ranking depth.
Problem
Disease-gene discovery often produces tens to hundreds of candidates, making causal-gene identification expensive and time-consuming.
Method
ProDiGe uses positive-unlabeled learning to integrate heterogeneous gene information and can share known disease-gene information across diseases, including through phenotypic similarity.
Results
ProDiGe outperformed Endeavour and PRINCE on real OMIM data, ranking correct genes in the top 5% for 69% of diseases with another known causal gene and 67% with none known.
Takeaways & Limitations
ProDiGe provides disease-specific and cross-disease variants whose relative utility depends on ranking depth and the number of known genes.
Takeaways & Limitations
The disease kernel strongly influences performance, leaving room to improve the prior used to represent disease relatedness.
Abstract
from arXiv · showhide
Elucidating the genetic basis of human diseases is a central goal of genetics and molecular biology. While traditional linkage analysis and modern high-throughput techniques often provide long lists of tens or hundreds of disease gene candidates, the identification of disease genes among the candidates remains time-consuming and expensive. Efficient computational methods are therefore needed to prioritize genes within the list of candidates, by exploiting the wealth of information available about the genes in various databases. Here we propose ProDiGe, a novel algorithm for Prioritization of Disease Genes. ProDiGe implements a novel machine learning strategy based on learning from positive and unlabeled examples, which allows to integrate various sources of information about the genes, to share information about known disease genes across diseases, and to perform genome-wide searches for new disease genes. Experiments on real data show that ProDiGe outperforms state-of-the-art methods for the prioritization of genes in human diseases.
Introduction
Disease-gene discovery often leaves tens to hundreds of candidates, making causal-gene identification expensive and time-consuming. ProDiGe addresses this prioritization problem by integrating heterogeneous gene information through positive-unlabeled machine learning, and ranks known disease genes effectively in real-data tests.
- Motivation: Chromosomal and high-throughput approaches often produce tens to hundreds of candidate genes, complicating causal-gene identification.Finding the causal gene among these candidates is expensive and time-consuming.
- Existing approach: Gene prioritization exploits biological functions, expression patterns, and gene interactions under a guilt-by-association strategy.Candidates are prioritized according to similarity with the disease or with genes already associated with it.
- Challenge: Integrating heterogeneous information from databases and literature into a ranking of candidate genes is difficult for human experts.Large-scale biological datasets create opportunities for computational prioritization methods.
- ProDiGe: ProDiGe jointly ranks candidate genes using positive-unlabeled learning rather than scoring each candidate independently against known disease genes.The method formulates prioritization as ranking candidates by similarity to positive examples.
- Evidence: On OMIM data, ProDiGe ranked the correct gene in the top 5% for 69% of diseases with another known causal gene and 67% with none known.It outperformed Endeavour and PRINCE in these tests.
Results
ProDiGe consistently prioritizes disease genes effectively, with performance depending on whether information is shared across diseases and how narrowly the ranking is evaluated. Its variants outperform existing methods, while phenotype-aware sharing is especially useful beyond the very top of genome-wide lists.
- Gene prioritization without sharing of information across diseases: ProDiGe1 variants consistently recover held-out disease genes at better ranks than MKL1class in LOOCV, with Wilcoxon P-values of 6.1e−29 and 8.8e−28.Both variants use positive-unlabeled learning, whereas MKL1class uses one-class learning with similar SVM and MKL engines.
- Gene prioritization without sharing of information across diseases: Roughly one third of correct associations appear in ProDiGe1’s top 20 genes among almost 19540, with no significant difference between mean and MKL variants (P-value=0.619).The authors therefore use the mean kernel in subsequent experiments because MKL is more time-consuming.
- Gene prioritization with information sharing across diseases: Average ranks for global approaches were ProDiGe4 (1682), ProDiGe3 (1817), ProDiGe2 (2246), and PRINCE (3065).Phenotypic similarity improves information sharing, while ProDiGe4 additionally keeps phenotypically similar diseases distinct.
- Gene prioritization with information sharing across diseases: Using only PPI data, ProDiGe-PPI outperforms PRINCE at low ranks, indicating that PU learning with SVM contributes beyond label propagation.The comparison uses the same input data for both methods; broader ProDiGe gains also reflect integrating multiple data sources.
- Is sharing information across diseases beneficial?: ProDiGe1 finds 49.2% of genes first in 100-candidate lists, while ProDiGe4 finds 68.9% within the top five; genome-wide, ProDiGe1 finds 27.8% within the top 10.The preferred method depends on the candidate-list setting and the depth of ranking considered.
- Is sharing information across diseases beneficial?: Sharing information is not beneficial within the top 10 of 19540 candidates, but becomes useful beyond the top 2%; ProDiGe4 is the best sharing strategy.Across settings, some ProDiGe variant outperforms existing methods, although only 12.3% of all disease genes are found in the top 10 by the best method.
- Predicting causal genes for orphan diseases: For orphan diseases, ProDiGe3-4 retrieve the true causal gene in the top 10 in 13.1% of cases, compared with 14.6% for diseases with known disease genes.The orphan-disease results are similar to the multitask setting in both method ordering and absolute performance.
- Predicting causal genes for orphan diseases: With only two known genes, precision is much smaller, but sharing information across diseases still retrieves new disease genes.The evaluation covered eight diseases with two known genes in the training set.
Discussion
ProDiGe provides local and cross-disease variants that integrate heterogeneous information through positive-unlabeled multitask learning, and both outperform established prioritization methods. Variant choice depends partly on ranking depth, while the disease kernel remains an important avenue for improvement.
- Discussion: ProDiGe integrates heterogeneous gene information in a unified positive-unlabeled learning strategy and can share information across diseases.The multitask formulation supports optional information transfer between diseases.
- Discussion: ProDiGe1 learns disease-specific rankings, whereas ProDiGe4 transfers information across diseases using phenotypic similarity.ProDiGe1 relies on known causal genes for each disease; ProDiGe4 additionally weights cross-disease sharing by phenotypic similarity.
- Discussion: Both ProDiGe variants outperform Endeavour and PRINCE on real OMIM data.The comparison was performed using real disease-gene data from the OMIM database.
- Discussion: The disease kernel improves phenotypic-similarity-based sharing, but its design still offers substantial room for improvement.Additional phenotype descriptors could provide further priors on task relatedness.
- Discussion: ProDiGe1 has higher recall in the top 1 or 2% of the list, whereas ProDiGe4 performs better after that range.This ranking-level difference is presented as a practical criterion for choosing between the variants.
- Discussion: The positive-unlabeled learning perspective is novel for this gene-prioritization problem apart from prior work cited by the authors.The authors connect this strategy to broader bioinformatics tasks beyond disease-gene identification.
Material and Methods
ProDiGe prioritizes disease genes by learning from positive and unlabeled examples, integrating heterogeneous gene data and sharing information across diseases. Its multitask formulation scores disease-gene pairs, while kernel choices control how data sources and diseases contribute.
- Positive-unlabeled learning: ProDiGe learns candidate-gene scores from both known disease genes and unlabeled genes rather than using positive examples alone.The PU-learning formulation exploits the distribution of unlabeled examples, and bootstrap aggregation discriminates positives from random unlabeled subsamples.
- Multiple data sources: The algorithm integrates heterogeneous gene descriptors by either averaging their kernels or learning a weighted convex combination through multiple kernel learning.The integrated kernel gives each source equal averaging weight, whereas MKL can automatically emphasize predictive sources or reduce the influence of irrelevant ones.
- Multiple diseases: For multiple diseases, ProDiGe jointly learns scores over disease-gene pairs so known associations can inform prioritization for diseases with few or no known causal genes.Disease kernels determine how information is shared, including general sharing through the multitask kernel and phenotype-based sharing through the phenotype kernel.
- Multiple diseases: The phenotype+Dirac kernel preserves disease-specific distinctions when incomplete phenotype descriptions make different diseases appear similar.The added Dirac component gives greater importance to genes associated with the disease of interest than to genes associated with phenotypically similar diseases.
- Evaluation: Evaluation uses leave-one-out cross-validation on OMIM disease-gene associations and compares ProDiGe with MKL1class and PRINCE.The gene dataset contains 19,540 genes described by nine information sources, including expression, functional, interaction, motif, domain, pathway, and literature data.