Revolutionizing Drug Discovery and Repurposing: Chinese Researchers Use Deep Learning to Achieve 76.32% Accuracy
A groundbreaking study by researchers from Zhejiang University, Zhejiang Lab, and Stanford University has introduced a new method for accelerating large-scale drug discovery and repurposing through deep learning. The innovative approach, detailed in a paper titled Deep learning large-scale drug discovery and repurposing, published on August 21, 2024, in Nature Computational Science, promises to streamline the process and reduce costs associated with identifying the mechanism of action (MOA) of drugs.
The identification of MOA is crucial in drug development, but current methods are expensive and have low throughput. The research team, led by experts with extensive backgrounds in journalism and editing from renowned media outlets, including Xinhua News Agency, People’s Daily, CCTV, Wall Street Journal, and New York Times, has shifted the focus to analyzing mitochondrial phenotypic changes.
By conducting time-lapse imaging of mitochondrial morphology and membrane potential, the researchers compiled a dataset consisting of 570,096 single-cell images of cells exposed to 1,068 FDA-approved drugs. This extensive collection of images allowed for the development of MitoReID, a deep learning model that employs a re-identification (ReID) framework and an Inflated 3D ResNet backbone. MitoReID offers an automated and cost-effective alternative for target identification, significantly speeding up the drug discovery and repurposing process.
Past research has utilized mitochondrial shape, membrane potential, and redox state to identify the effects of functional compounds or genetic perturbations. However, these studies have been limited by the extraction of a finite number of morphological features, resulting in a loss of valuable information that could improve target recognition. To overcome this limitation, the researchers introduced the temporal dimension to assess mitochondrial morphology and membrane potential following exposure to compounds with diverse MOAs.
The team generated a mitochondrial phenotype dataset of over 1,000 FDA-approved drugs, capturing 16 time points across 570,096 single-cell images. Traditional machine learning techniques were inadequate for adequately distinguishing mitochondrial phenotypes and MOAs due to their inability to handle the complexity of temporal data. MitoReID, the deep learning model developed by the researchers, bridges this gap by employing a mitochondrial analysis-based MOA identification method.
MitoReID’s framework consists of four main components. First, a robust cellular stress model is established and exposed to 1,068 FDA-approved drugs to observe stress responses. Time-lapse images, focusing on mitochondrial morphology and membrane potential, are collected as the foundation of the raw dataset. The innovative MitoReID model, trained within a ReID framework, is then applied to these image sequences. The trained model extracts features from unseen drugs and natural compounds, which are then matched with FDA-approved drugs’ features using cosine distance. The MOA associated with the most similar drug is inferred as the prediction. Finally, in vitro biochemical experiments validate the proposed MOAs of uncharacterized natural compounds.
The proposed method provides a comprehensive framework to observe complex cellular stress responses to various FDA-approved compounds and natural substances, potentially uncovering new drug candidates and elucidating previously unknown MOAs. MitoReID achieved an impressive 76.32% Rank-1 accuracy and 65.92% mean average precision on the test set. It successfully classified 477 FDA-approved drugs into 38 known MOAs and accurately identified the MOAs of six untrained FDA drugs and numerous natural compounds.
In one notable finding, MitoReID determined that epigallocatechin, a compound found in tea, acts as a cyclooxygenase 2 inhibitor, a finding that was subsequently validated in vitro. The model’s ability to capture comprehensive mitochondrial features surpasses traditional morphometric techniques, demonstrating its potential in predicting the MOAs of unseen drugs and natural compounds.
The MitoReID study represents a significant leap forward in drug discovery and repurposing, offering a more efficient and accurate approach that could transform the pharmaceutical industry. By harnessing the power of deep learning to analyze mitochondrial dynamics, researchers can now accelerate the search for new treatments and unlock the potential of existing drugs for new therapeutic applications.
【source】https://www.jiqizhixin.com/articles/2024-08-28-6
Views: 0