How does GNN find new drugs?Dr. MIT's thesis "Molecular Charter Express the Drug Discovery of Learning and Generated"

Source: Specialty

This article is introduced in the paper. It is recommended to read for 5 minutes

We have explored a Transformer -type architecture for molecules and provides new tools for applying these models to graphics structure objects.

Machine learning methods have been widely used in the field of drug discovery, making more powerful and efficient model possible. Before the appearance of the depth model, the modeling molecules are largely driven by expert knowledge; in order to express the complexity of the molecular structure, these manual design rules are proven to be not enough. Deep learning models are powerful because they can learn important statistical characteristics of problems -but only the correct induction deviation. We solve this important problem in the context of the two molecules: representation and generation. The typical success of deep learning is that it can map the input domain to a meaningful representation space. This is particularly sharp for molecular issues, and the "correct" relationship between molecules is delicate and complicated. The first part of this paper will focus on discussing molecular representation, especially nature and reaction forecast. Here, we have explored a Transformer -type architecture for molecules, which provides new tools for applying these models to graphical structure objects. Aside from the traditional Tu neural network paradigm, we show the effectiveness of the molecular representation of the prototype network, which allows us to reason the molecular learning prototype. Finally, we said in the context of improving reaction forecast. The second part of this thesis will be concentrated in molecular generation. This is a crucial drug discovery as a means, and proposes promising drug candidates. We have developed a new method of multi -nature molecular generation to first learn the distributed vocabulary of molecular fragments. Then, using this word, we investigated the effective exploration method of chemical space.

https://dspace.mit.edu/handle/1721.1/143362

Machine learning has quickly changed the traditional channels of drug discovery, providing new tools for each step of the process. Many traditional and professional knowledge of knowledge have been solved through deep learning tools to make them more efficient and cheaper. Previous chemical information methods use many manual design rules to model small molecules. These technologies are used to solve problems such as properties prediction, and the task is the nature of predictive molecules. However, trying to solve these traditional methods that indicate problems lack good generalization capabilities due to their uncomfortable characteristics. The change of deep learning model is the ability of the model to directly learn and extract important features from the data. However, this can only be realized under the modification assumptions of the correct structural deviation and model. The naive application depth method on the molecular issue will limit the ability or usefulness of the model, hindering their promotion ability and the usefulness in practice. Therefore, the importance of using the correct induction deviation cannot be undervalued.

Before the appearance of deep learning methods, the molecular modeling requires heavy engineering and fixed representation, which is usually called quantitative constructive relationship (QSAR) method. In these methods, fingerprint technology is very popular, and it can be roughly divided into structural [30], topology [1], cycle [8] and drug -effective group fingerprints [91]. Some of these fingerprints (such as structure -based MACCS [30] fingerprints) are highly specific representations, consisting of a set of fixed predetermined structure indicator functions. Other fingerprints, topology and round, including JP Morgan fingerprints are more flexible. These fingerprints capture local topology through enumeration path or ring -shaped neighborhoods. However, the problem still exists in the essence of the generation method: if these predetermined rules do not capture correctly for the task, they will not work well. For example, for many small molecular issues, the Property Cliff is still a challenging problem, which is a phenomenon that similar to molecules shows different properties. This problem is particularly sharp for molecular fingerprints, because the characteristics are fixed. However, the use of depth models cannot solve this problem, because the depth model is easy to overfit with data and provides poor generalization.

Therefore, it is crucial to incorporate our deep learning model into the correct type of structural deviation. The neural network operates through the iterative aggregation scheme, and at each step, the node is aggregated from its neighbor. In order, a node should contain more and more information about neighborhoods. The node represents the final agglomeration as a single vector of the graph. Although this simple paradigm is sometimes effective, it may not always include prejudice of the correct molecular task type. For example, when the characteristics of the molecule are considered, this local neighbor aggregation may not be able to capture very important long -range dependencies. More importantly, the aggregation on the two -dimensional molecule is not suitable for ideal molecules that we should observe the three -dimensional structure. There are many considerations for the development of the molecules, but they need the correct structure to be effective. The fingerprint is simple, but it is not flexible, and often involves many rules of human design. On the other hand, the depth model is easy to fit and cannot capture the correct structure representation.

not

- END -