Improved genome-scale multi-target virtual screening via a novel collaborative filtering approach to cold-start problem

Abstract

Conventional one-drug-one-gene approach has been of limited success in modern drug discovery. Polypharmacology, which focuses on searching for multi-targeted drugs to perturb disease-causing networks instead of designing selective ligands to target individual proteins, has emerged as a new drug discovery paradigm. Although many methods for single-target virtual screening have been developed to improve the efficiency of drug discovery, few of these algorithms are designed for polypharmacology. Here, we present a novel theoretical framework and a corresponding algorithm for genome-scale multi-target virtual screening based on the one-class collaborative filtering technique. Our method overcomes the sparseness of the protein-chemical interaction data by means of interaction matrix weighting and dual regularization from both chemicals and proteins. While the statistical foundation behind our method is general enough to encompass genome-wide drug off-target prediction, the program is specifically tailored to find protein targets for new chemicals with little to no available interaction data. We extensively evaluate our method using a number of the most widely accepted gene-specific and cross-gene family benchmarks and demonstrate that our method outperforms other state-of-the-art algorithms for predicting the interaction of new chemicals with multiple proteins. Thus, the proposed algorithm may provide a powerful tool for multi-target drug design.

Figure 1. The components of the loss function.

INPUT: the sum of the interaction and impute matrices R + Q; the weight matrix W; the protein similarity matrix M; the chemical similarity matrix N. OUTPUT: the matrix of protein latent preferences F; the matrix of chemical latent preferences G; the matrix of predicted interaction probabilities .

Figure 2. ZINC benchmark MCS.

The True Positive Rate (TPR) at top r% predictions with varying number of (maximal) chemical structural similarity (MCS).

Figure 3. ZINC benchmark LT.

The True Positive Rate (TPR) at top r% predictions with varying number of ligands per target (LT).

Figure 4. Accuracy over iterations.

The accuracy of COSINE (TPR at top 1%; y-axis) as a function of the number of iterations (x-axis) in different subsets of the ZINC benchmark (1–5, 6–10, 11–15, 26–20, and >20 ligands per target).

Figure 5. Accuracy as a function of noise.

The accuracy of COSINE in the 10-fold CV Yam08 benchmark as a function of the amount of missing interaction data (left) and as a function of the amount of incorrect interaction data (right).