Machine learning and network medicine approaches for drug repositioning for COVID-19

Abstract

We present two machine learning approaches for drug repurposing. While we have developed them for COVID-19, they are disease-agnostic. The two methodologies are complementary, targeting SARS-CoV-2 and host factors, respectively. Our first approach consists of a matrix factorization algorithm to rank broad-spectrum antivirals. Our second approach, based on network medicine, uses graph kernels to rank drugs according to the perturbation they induce on a subnetwork of the human interactome that is crucial for SARS-CoV-2 infection/replication. Our experiments show that our top predicted broad-spectrum antivirals include drugs indicated for compassionate use in COVID-19 patients; and that the ranking obtained by our kernel-based approach aligns with experimental data. Finally, we present the COVID-19 repositioning explorer (CoREx), an interactive online tool to explore the interplay between drugs and SARS-CoV-2 host proteins in the context of biological networks, protein function, drug clinical use, and Connectivity Map. CoREx is freely available at: https://paccanarolab.org/corex/.

Keywords: COVID-19; SARS-CoV-2; drug repurposing; graph visualization; kernels on graphs; network medicine; non-negative matrix factorization.

Figure 1

Drug-virus dataset statistics We used the dataset manually curated by Andersen et al. (Left) Number of drug-virus associations grouped by their known developmental status. The development of broad-spectrum antivirals (BSA) starts with in vitro experiments (e.g., cell culture), moves to animal models, and then to clinical trials in humans (phases I–IV). It terminates with the approval of the drug for commercial use (in red). (Middle) Number of drugs (BSAs) associated to each virus in the dataset. Inset: the word cloud shows the 14 viruses with most associations. The size of the word is proportional to its number of associations and the five most popular viruses among drugs are colored blue. (Right) Number of viruses associated to each drug in the dataset. Inset: the word cloud shows the 18 drugs with most associations and the five most popular drugs among viruses are colored blue.

Figure 2

Overview of our matrix decomposition model for predicting effective drug-virus associations Totals of 850 associations for $n=126$ different BSAs and $m=80$ distinct viruses were collected from the Andersen et al. database. The observed associations were arranged into an $n\times m$ matrix Y by setting $y_{ij}=1$. Unobserved associations were encoded with zeros. Our algorithm decomposes the matrix Y into the product of two matrices, P (of size $n\times k$) and Q (of size $k\times m$). By multiplying the matrices P and Q, we obtain $\hat{Y}$, which models Y, where all the entries are replaced with real numbers—these correspond to our predicted scores. Rows of P are the BSA feature vectors (or BSA signature); columns of Q are the virus feature vectors (virus signature). The lower illustration depicts how our model discovers a low-dimensional signature vector for the antiviral drug zanamivir, and a low-dimensional signature vector for SARS-CoV-2. The dot product of these two signatures is the predicted efficacy of zanamivir against SARS-CoV-2.

Figure 3

Performance at predicting approved/phase IV BSAs for 28 viruses Percentage of approved or phase IV BSA drugs found for a specific virus in the top K predictions retrieved. The performance of our method is compared with different matrix decomposition algorithms in a leave-one-out fashion. NMF, non-negative matrix factorization; tSVD, truncated singular value decomposition. A baseline based on random scores sampled from a uniform distribution is also included.

Figure 4

Overview of our network medicine approach (A) The human interactome containing both host proteins (red) and drug targets (blue). (B) The totals of 14,941 drug target associations between N = 2,197 FDA-approved drugs and n_V = 18,505 proteins are represented by a binary matrix T (blue matrix). Multiple graph kernels are calculated on the interactome, resulting in $n_V\times n_V$ matrices (green matrices). The host proteins are represented by a vector h of size $n_V$ (red vector) indicating their weights (based on gene expression data). (C) Our kernel score is calculated using a matrix multiplication to obtain a prediction score for each drug. (D) The obtained ranking is evaluated using different types of evidence: in vitro efficacy against SARS-CoV-2, Connectivity Map, and clinical trials.

Figure 5

Analysis of the predictions for COVID-19 We used three different sources of evidence: in vitro (A, C, and F), clinical trials (B, D, and G), and CMAP (E and H). We compared scores for drugs with evidence of efficacy against SARS-CoV2 versus scores for the remaining drugs. Our matrix factorization model (A and B) and kernel-based methods (F, G, and H) provide scores that are significantly different between the two groups of drugs in every case (Wilcoxon-Mann-Whitney p < 0.05). We formulated a binary classification problem to discriminate between drugs with evidence of efficacy against SARS-CoV2 and the remaining drugs. (C, D, and F) Comparison of precision and recall at top 150 for our kernel-based methods (commute time, diffusion, p-step, regularized Laplacian, and inverse cosine kernels, and avgRank), DSD, and Guney’s distance. The highest values are colored.

Figure 6

Screenshot of CoREx displaying a functional module for Sulfasalazine (highlighted in green in the “Drug” list) The module is depicted as a network on the top left where nodes represent proteins, edges represent shared functional characteristics, and the thickness of the edges represents the strength of such functional similarity. Host proteins are depicted as diamonds, drug targets are colored. The list of drugs with at least one target in this functional module is presented in the center, alongside CMAP scores for five cell lines (on the left), and an indicator of whether the drug is currently in clinical trials (on the right). The bar plots on the right part correspond to the functional enrichment scores for each GO domain. The bar plot on the bottom left section of the image summarizes the ATC categories of the drugs targeting this functional module.