Knowledge Extraction from Reaction Databases

Clearly, the extraction of knowledge on chemical reactions from the information contained in reaction databases (see Section 5.12) is quite a challenging problem. The huge amount of information stored in reaction databases - the larger ones contain several million reactions - renders them quite attractive for knowledge extraction. However, only a few attempts have been made to extract knowledge automatically. 

The reasons for this lack of work are manifold The problem is quite complex and difficult to tackle. The information in reaction databases is inherently biased only known reactions, no reactions that failed, are stored. However, any learning also needs information on situations where a certain event will not happen or will fad. The quality of information stored in reaction databases often leaves something to be desired reaction equations are incomplete, certain detads on a reaction are often incomplete or missing, the coverage of the reaction space is not homogeneous, etc. Nevertheless, the challenge is there and the merits of success should be great  

One of the first attempts to build a knowledge base for synthetic organic reactions was made by Gelernter s group, through inductive and deductive machine learning [1]. Important work on this topic was also performed by Funatsu and his group [2]. 

The main characteristics of the method, developed in our group for reaction classification arc 1) the representation of a reaction by physicochemical values calculated for the bonds being broken and made during the reaction, and 2 use of the unsupervised learning method of a self-organi2ing neural network for the perception of similarity of chemical reactions [3, 4], 

The method is incorporated into the CORA (Classification of Organic Reactions for Analysis) system [Sf Here, wc want to illustrate the merits of this approach by an example of its application to a specific problem, the prediction of the regioselec-tivity of a ring closure reaction. This is detailed in the following tutorial. 

---

The next question is how to represent the reacting bonds of the reaction center. We wanted to develop a method for reaction classification that can be used for knowledge extraction from reaction databases for the prediction of the products of a reaction. Thus, we could only use physicochemical values of the reactants, because these should tell us what products we obtain. 

A wider variety of reaction types involving reactions at bonds to oxygen atom bearing functional groups was investigated by the same kind of methodology [30]. Reaction classification is an essential step in knowledge extraction from reaction databases. This topic is discussed in Section 10.3.1 of this book. 

Before concluding this section it should be emphasi/cd that knowledge extraction from reaction databases is still a challenging problem having many important applications. There is still room for new approaches to this task, Furthermore, groat efforts should be made to improve the depth of information stored in reaction databases. With the introduction of electronic lab journals, the primary information on a chemical reaction gained in the laboratoiy becomes directly available. 

In spite of the importance of reaction prediction, only a few systems have been developed to tackle this problem, largely due to its complexity it demands a huge amount of work before a system is obtained that can make predictions of sufficient quality to be useful to a chemist. The most difficult task in the development of a system for the simulation of chemical reactions is the prediction of the course of chemical reactions. This can be achieved by using knowledge automatically extracted from reaction databases (see Section 10.3.1.2). Alternatively, explicit models of chemical reactivity will have to be included in a reaction simulation system. The modeling of chemical reactivity is a very complex task because so many factors can influence the course of a reaction (see Section 3.4). 

More elaborate scheme.s can he envisaged. Thus, a. self-organizing neural network as obtained by the classification of a set of chemical reactions as outlined in Section 3,5 can be interfaced with the EROS system to select the reaction that acmaliy occurs from among various reaction alternatives. In this way, knowledge extracted from rcaetion databases can be interfaced with a reaction prediction system,... 

Knowledge of fhe presence and stoichiometry of metabolic reactions in a particular microorganism can be extracted from various information sources (eg, annotated genome information, biochemical textbooks, and the published literature and pathway databases) (Patil etal., 2004). 

---