[Paper Review] SL$^2$MF: Predicting Synthetic Lethality in Human Cancers via Logistic Matrix Factorization

Yong Liu, Min Wu+3 more|arXiv (Cornell University)|Oct 20, 2018

Bioinformatics and Genomic Networks10 citations

TL;DR

This paper proposes SL$^2$MF, a logistic matrix factorization approach that predicts synthetic lethality in human cancers by learning latent gene representations from observed SL data, incorporating importance weighting for known SL pairs and integrating biological knowledge from PPI networks and Gene Ontology. The method achieves AUC scores of 0.7330 and 0.6701 in two validation scenarios, demonstrating strong performance as a complementary tool for identifying novel anti-cancer drug targets.

ABSTRACT

Synthetic lethality (SL) is a promising concept for novel discovery of anti-cancer drug targets. However, wet-lab experiments for detecting SLs are faced with various challenges, such as high cost, low consistency across platforms or cell lines. Therefore, computational prediction methods are needed to address these issues. This paper proposes a novel SL prediction method, named SL2MF, which employs logistic matrix factorization to learn latent representations of genes from the observed SL data. The probability that two genes are likely to form SL is modeled by the linear combination of gene latent vectors. As known SL pairs are more trustworthy than unknown pairs, we design importance weighting schemes to assign higher importance weights for known SL pairs and lower importance weights for unknown pairs in SL2MF. Moreover, we also incorporate biological knowledge about genes from protein-protein interaction (PPI) data and Gene Ontology (GO). In particular, we calculate the similarity between genes based on their GO annotations and topological properties in the PPI network. Extensive experiments on the SL interaction data from SynLethDB database have been conducted to demonstrate the effectiveness of SL2MF.

Motivation & Objective

• To address the high cost and inconsistency of wet-lab screening for synthetic lethal (SL) interactions in human cancers.
• To overcome the scarcity of reliable human SL data by developing a computational method that leverages known SL pairs more effectively.
• To improve prediction accuracy by integrating biological knowledge from protein-protein interaction (PPI) networks and Gene Ontology (GO) annotations.
• To provide a robust, data-driven baseline for future SL prediction methods in human cancer genomics.

Proposed method

• SL$^2$MF employs logistic matrix factorization to model the probability of synthetic lethality between gene pairs as a linear combination of learned gene-specific latent vectors.
• It applies importance weighting to prioritize known SL pairs during training, assigning higher weights to reliable interactions and lower weights to unknown pairs.
• Gene similarity is computed using both GO semantic similarity and topological features from the PPI network to enrich latent vector learning.
• The model is trained on SL data from SynLethDB, with two evaluation scenarios: one excluding DAISY-predicted pairs and one including them to assess generalization.
• The final predictions are ranked by estimated interaction probabilities, and performance is evaluated using AUC and AUPR metrics.
• The method is compared against DAISY, a state-of-the-art SL prediction method, under identical settings to validate its effectiveness.

Experimental results

Research questions

• RQ1 Can logistic matrix factorization effectively model synthetic lethality probabilities using incomplete and noisy SL data from human cancers?
• RQ2 How does importance weighting of known SL pairs improve prediction performance compared to uniform weighting?
• RQ3 To what extent does integrating PPI network topology and GO annotations enhance the accuracy of SL predictions?
• RQ4 How does SL$^2$MF compare to existing methods like DAISY in terms of predictive performance and overlap in top-ranked predictions?
• RQ5 Can SL$^2$MF serve as a reliable baseline for future SL prediction methods in human cancer genomics?

Key findings

• SL$^2$MF achieved an AUC of 0.7330 in scenario 1, where DAISY-predicted SL pairs were excluded from training, indicating strong generalization capability.
• In scenario 2, which included DAISY-predicted pairs in the training set, the AUC dropped to 0.6701, likely due to reduced data diversity and potential overfitting.
• The AUPR scores for both scenarios were below 0.005, indicating that many high-ranked predictions were false positives, though the model still captured some true SL pairs.
• Among the top 5,740 predicted SL pairs, 14 were verified in the SynLethDB database in scenario 1, while DAISY had no overlap, suggesting SL$^2$MF's superior recall in this setting.
• In scenario 2, 27 of SL$^2$MF's top predictions were validated via SynLethDB and text mining, compared to only 3 for DAISY, highlighting SL$^2$MF's improved predictive reliability.
• The overlap between SL$^2$MF and DAISY predictions was minimal—only 5 common pairs in scenario 1 and 3 in scenario 2—indicating distinct and complementary prediction mechanisms.