Deep learning using molecular image of chemical structure

Since the first appearance of artificial intelligence (AI) in the 1950s, it has gone through a fascinating history, changing its appearance until it has been recognized. The use of AI is receiving a lot of attention in the field of elucidation and evaluation of the physiological action mechanisms of toxic chemical compounds. This is because toxicity tests on experimental animals are generally used in the toxicity evaluation of chemical substances, and it is required to develop an in silico toxicity prediction method based on time reduction and consumption and the 3Rs perspectives. On the other hand, deep learning is a promising technique for achieving advanced prediction in quantitative structure-activity relationship (QSAR) toxicity prediction. However, QSAR did not fully exploit the capability of deep learning, which can directly analyze the molecular structure because molecular descriptors have traditionally been used to transfer chemical structure information to AI. Therefore, this study develops a novel structural information input method, “DeepSnap,” to learn the characteristics of the entire molecules as image data. In this study, the physiological activity value associated with each molecule was identified by inputting an image file generated from a three-dimensional (3D) molecular structure into a deep learning system developed in the field of image analysis, demonstrating that excellent predictive performance can be obtained. In this way, AI attempts to think beyond human judgment using deep learning. The advancement of this technology is projected to continue; however, the ambiguity of AI's judgment criteria has proven to be a “black box problem.” To dispel such concerns, the technique “explainable AI (XAI),” which explains the judgment basis of deep learning models, has grown in popularity recently. Although XAI has been studied in various directions in response to its high needs, the operating principle of the AI model itself has not yet been clarified. Therefore, comprehensive analysis of various chemical substances using large-scale reliable and explanatory toxicity information is expected to enable the development and enhancement of toxicity prediction methods using machine learning.