Machine learning application in molecular science research field

In recent years, big data has occupied a dominant position in the discussion, but the importance of small data is becoming more and more obvious. Managing small data has become an important challenge in the field of molecular science. Although small data sets are often used as the basis for scientific and technological research, they are limited in scale due to time constraints, financial constraints, technical barriers, and ethical privacy and security, thus forming a major problem. At the same time, small data sets often have problems such as noise interference, data imbalance, high-dimensional characteristics, and interpolation processing is required if necessary. These challenges have greatly increased the difficulty and complexity of data management. However it is worth noting that in the recent field of big data research, breakthroughs in machine learning (ML), deep learning (DL) and artificial intelligence (AI) have provided potential solutions to solve these problems. The AlphaFold tool stands out with high precision and high efficiency in the application of small data molecular science. Although in several key reviews, these issues have been comprehensively addressed. But there is a lack of data to generate a large amount of information from the source to solve the problem, and the problems caused by Transformer, graph neural network (GNN), generative adversarial network (GAN) and other models in the case of small data sets have not been further solved. Therefore, this article will focus on the application of ML and DL in the molecular science of chemistry and biology, and attach great importance to the recent innovative achievements aimed at meeting the challenges of small data. We have conducted systematic research on various basic ML algorithms. It covers random forest (RF), gradient enhancement tree (GBT), k-nearest neighbor (KNN), support vector machine (SVM) and kernel learning (KL), logical regression (LR) and linear regression. Meanwhile, our research has also expanded to include cutting-edge technologies such as convolutional neural networks (CNN) and U-Net architecture, GNN, GAN and long-term memory (LSTM) networks, as well as self-encoders and transformer models. In addition, we have also explored a variety of strategies to deal with the problem of limited samples of small data sets, including cross-verification, quantification of uncertainty, data enhancement and migration learning. This research not only reveals the future development direction of small data research in the field of molecular science, but also highlights the latest breakthrough progress in this area. However, this study has some shortcomings in improving the techniques for processing noisy and unbalanced small data sets, and in developing universal molecular descriptors suitable for complex structures. The data in this paper are primarily sourced from databases such as Elsevier, ACS, and Springer, as well as websites including SciFinder and Computational Chemistry .