Functional microRNA-targeting drug discovery by graph-based deep learning

Abstract

MicroRNAs are recognized as key drivers in many cancers but targeting them with small molecules remains a challenge. We present RiboStrike, a deep-learning framework that identifies small molecules against specific microRNAs. To demonstrate its capabilities, we applied it to microRNA-21 (miR-21), a known driver of breast cancer. To ensure selectivity toward miR-21, we performed counter-screens against miR-122 and DICER. Auxiliary models were used to evaluate toxicity and rank the candidates. Learning from various datasets, we screened a pool of nine million molecules and identified eight, three of which showed anti-miR-21 activity in both reporter assays and RNA sequencing experiments. Target selectivity of these compounds was assessed using microRNA profiling and RNA sequencing analysis. The top candidate was tested in a xenograft mouse model of breast cancer metastasis, demonstrating a significant reduction in lung metastases. These results demonstrate RiboStrike's ability to nominate compounds that target the activity of miRNAs in cancer.

Keywords: RNA-targeting drug discovery; artificial intelligence; deep learning; drug toxicity evaluation; graph convolutional neural network; in silico drug screening; microRNA inhibition; microRNA-21.

Graphical abstract

Figure 1

Overview of the RiboStrike pipeline (A) Stages of discovery from input molecular data and deep-learning techniques to candidate selection and experimental validation. Molecular data are collected from multiple sources for training virtual screening models in multitask learning mode where different datasets are grouped together and share learned representations. A task recommender algorithm helps choose the grouping of the tasks for multitask learning to maximize performance. Nine million candidate molecules are filtered based on the predictions of the models on bioactivity, interactions, and toxicity and the uncertainty in those predictions. After clustering for a diverse selection, eight of the top candidates are experimentally validated. (B) Computational pipeline and the flow of data within the GCNN network. Molecules are represented as graphs with calculated node and bond features. The convolution layers learn distinctive representations, which are pooled into fixed-length vectors and used for classification by the fully connected network. Once a multitask learning model is trained, an in-house algorithm is used to recommend task grouping for another round of multitask learning.

Figure 2

The AP score of different sub-models for the task recommender algorithm The tasks above the threshold line make predictions matching the miR-21 ground truth to a higher degree than the rest of the tasks. These tasks are the recommended tasks and are selected for training a new multitask model with narrower scope.

Figure 3

Confusion matrix for models trained using different learning methods (A) Single task, (B) multitask for all tasks, and (C) multitask for recommended tasks. Balance between true positives (predicted correctly as positive) and false negatives (predicted incorrectly as positive) is needed for virtual screening. Due to imbalanced data, false negatives hurt candidate pool quality drastically.

Figure 4

UMAP of the inner features of the model for the inference and training data for (A) ZINC data predicted to be active as the candidate space to select from, (B) the training data and the positive molecules from the FDA-approved set, showing that the candidate space spans the same area as the training set, and (C) the 10 clusters applied to this space for the inference sets (ZINC and Asinex) and the final selected molecules from these clusters to ensure diversity in selection.

Figure 5

Experimental validation results (A) Dosage response assay in reporter cell lines for Ribo21D-1, Ribo21D-2, and Ribo21D-3. (B) Enrichment and depletion patterns of putative miR-21 targets across gene expression changes for anti-miR21 as well as Ribo21D treatments (left). For this analysis, gene expression changes (log2 fold changes) were first sorted and divided into equally populated bins, from downregulation (left) to upregulation (right). For each analysis, the mutual information value (in bits) and its associated Z score are provided. The heatmaps show the enrichment and depletion patterns from the onePAGE package. Blue marks depletion and gold marks enrichment. Expression bins with statistically significant depletion or enrichment of miR-21 putative targets are marked by a dark blue or red border. As an alternative visualization, we have also shown the logFC values for the miR-21 targets and the background set of genes (right). The p values were calculated using t test. (C) Expression of empirically determined miR-21 targets (microCLIP plus anti-miR transfections) in response to the top three compounds. (D) Expression of miR-21 determined using small RNA-seq across the three treatments. The p values were calculated using DESeq2. (E) In vivo lung colonization assays were used to measure impact of Ribo21D-1 on lung metastasis. Normalized lung photon flux, which measures luciferase activity in labeled cancer cells, as a function of time for each cohort (n = 5) is shown. Two-way ANOVA and Mann-Whitney U tests were used to assess statistical significance. Also shown is a representative mouse and lung image.