Knowledge-based Expert Systems in Chemistry: Artificial Intelligence in Decision Making

by Philip Judson

Overview

There have been significant developments in the use of knowledge-based expert systems in chemistry since the first edition of this book was published in 2009. This new edition has been thoroughly revised and updated to reflect the advances.

The underlying theme of the book is still the need for computer systems that work with uncertain or qualitative data to support decision-making based on reasoned judgements. With the continuing evolution of regulations for the assessment of chemical hazards, and changes in thinking about how scientific decisions should be made, that need is ever greater. Knowledge-based expert systems are well established in chemistry, especially in relation to toxicology, and they are used routinely to support regulatory submissions. The effectiveness and continued acceptance of computer prediction depends on our ability to assess the trustworthiness of predictions and the validity of the models on which they are based.

Written by a pioneer in the field, this book provides an essential reference for anyone interested in the uses of artificial intelligence for decision making in chemistry.

Product Details

ISBN-13:	9781788014717
Publisher:	RSC
Publication date:	02/15/2019
Series:	Theoretical and Computational Chemistry Series , #15
Edition description:	2nd ed.
Pages:	304
Product dimensions:	6.15(w) x 9.20(h) x (d)

About the Author

The author studied chemistry at the University of Manchester before working on the synthesis of novel herbicides and fungicides for Fisons Ltd at Chesterford Park Research Station near Saffron Walden. His PhD at the University of Surrey was on chemical synthesis. He took an interest in knowledge-based computer systems and became Head of Chemical Information and Computing for Schering Agrochemicals Ltd. He was one of the founders of Lhasa Limited, a not-for-profit company specialising in knowledge-based expert systems in chemistry including the widely-used Derek, Meteor, and Zeneth systems for predicting chemical toxicity, metabolism, and chemical degradation. Although semi-retired, he continues to contribute to research and development work at Lhasa Limited in his role as Scientific Advisor and is working in a project on synthetic accessibility led by scientists at the US Niational Institutes of Health. He developed and maintains software for chemical hazard classification and chemical safety data sheet management, Harmoneus and Prometheus, which are supplied by Hibiscus plc. He has published over eighty scientfic papers, posters and book chapters. His hobbies include climbing and caving and he has published articles about international caving expeditions that he has taken part in.

Read an Excerpt

CHAPTER 1

Artificial Intelligence – Making Use of Reasoning

The first edition of this book began with the flight of a three-metre long paper aeroplane. It served to illustrate a point. So, let us set it airborne it again. Launched by half a dozen young men at a run, it flies successfully, dare we even say "gracefully", the length of a research station canteen before making an unfortunate landing in the director of research's Christmas lunch. It was just a question of getting the aerodynamics right.

My school mathematics teacher reminded us on most days (several times on some) that all science is mathematics. But was it only the power of numbers he had in mind? Does science come down to the mechanical crunching of numbers, real and imaginary?

Contrary to the perceptions of many people outside science, as well as too many inside it, science is not about proving facts: it is about testing hypotheses and theories; ultimately, it is about people and their opinions. In many fields, human decision making may best be supported by reasoned argument or the use of analogy and not much helped by numerical procedures or answers. The minimum braking distance for a car travelling at 40 miles per hour is 24 metres, according to the UK Highway Code. Assuming you can countenance the required mixing of miles and metres, does this information help you to drive more safely? Have you any more idea than I have how far ahead an imaginary 24 metre boundary-marker precedes you along the road?

And there is a further problem. "Numbers out" implies "numbers in", so what do you do if you have no numbers to put in? A regrettably popular solution is to invent them – or at least to come up with dubious estimates to feed into a model that demands them, which is close to invention. It is the only option if you want to apply numerical methods and to give numbers to the people asking for solutions. That numbers make people feel comfortable is a bigger problem than it may at first appear to be, too. Uncritical recipients of numerical answers tend to believe them, and to act on them, without probing very deeply. More sceptical recipients want to judge for themselves how meaningful the answers are but often find that the supporting evidence associated with a numerical method is not much help. Many are the controversies over whether this or that numerical method is more precise but they are missing the point if the data are far less precise than the method. Perhaps numbers are unnecessary – even unsuitable – for expressing some kinds of scientific knowledge.

There are circumstances in which numerical methods are highly reliable. Aeroplanes stay up in the sky and make it safely to earth where they are supposed to do. Chemical plants run 24 hours a day, year in year out. Numerical methods work routinely in physical chemistry laboratories, and toxicology and pharmacology departments. But it is unlikely that the designers of the three-metre paper dart that took flight at the start of this chapter did any calculations at all. My guess is that they depended on analogy, drawing on years of experience making little ones.

This book is about uses of artificial intelligence (AI) and databases in computational chemistry and related science, in cases where qualitative output may be of more practical use than quantitative output. It touches on quantitative structure–activity relationships (QSAR) and how they can inform qualitative predictions, but it is not about QSAR. Neither is it a book about molecular modelling. Both subjects are well-covered in too many books to list comprehensively. A few examples are given in the references at the end of this chapter. This book focuses on less widely described and yet, probably, more widely-used applications of AI in chemistry.

The term "artificial intelligence" carries with it notions of thinking computers but, as a radio personality in former times would have had it, it all depends on what you mean by intelligence. If you type "Liebig Consender" into the Google™ search box, Google™ responds with "Showing results for Liebig Condenser". That is worryingly like intelligent behaviour whether it is intelligent behaviour or not (it is also very irritating if you really do want to look for consenders). Arguments continue about whether tests for artificial intelligence such as the Turing test are valid and whether a categorical test or set of tests can be devised. Perhaps it is sufficient to require that to be intelligent a system must be able to learn, be able to reason, be creative, and be able to explain itself persuasively. Currently, no AI system can claim to have all of these characteristics. Individual systems typically have two or three.

To count as intelligent, solving problems needs to involve a degree of novel thinking, i.e. creativity. Restating the known, specific answer to a question requires only memory. Compare the following questions and answers. The first answer merely reproduces a single fact. Generating the second answer, simple though it is, requires reasoning and a degree of creativity.

"Where's the sugar?"

"In the sugar bowl".

"Where will the sugar be in this supermarket?"

"A lot of supermarkets put it near the tea and coffee, so it could be along the aisle labelled 'tea and coffee'. Alternatively, it might be in the aisle labelled 'baking'. Let's try 'baking' first – it is nearer".

One of the first computer systems to behave like an expert using a logical sequence of questions and answers to solve a problem was MYCIN, a system to support medical diagnosis.

"Doctor, I keep getting these terrible headaches".

"Sorry to hear that. Is there any pattern to when the headaches occur?"

"Now you ask, they do seem to come mostly on Sunday mornings".

"And what do you do on Saturday evenings?"

The doctor's questions are not arbitrary. You can see how they are directed by the patient's responses. You can probably see where they are leading, too, but the doctor would still want to ask further questions to rule out all the possibilities before jumping to the obvious conclusion about the patient's Saturday nights out on the town. The aim of the MYCIN experiment was to design a computer system capable of choosing appropriate sequences of questions similarly, in order to reach a diagnosis.

This kind of reasoning is common throughout science although it often does not involve a dialogue; the questions may be implicit in a process of thought rather than consciously asked. Suppose you know that:

• many α, β-unsaturated aldehydes cause skin sensitisation;

• for activity to be expressed a compound must penetrate the skin;

• compounds with low fat/water partition coefficients do not penetrate the skin easily:

• many imines can be hydrolysed easily in living systems to generate aldehydes.

Actually, the story for skin sensitisers is better understood and can be more fully and more usefully described than this, but what we have will do for the purposes of illustration. Suppose you are shown the structure of a novel α, β-unsaturated imine and asked for an assessment of its potential to cause skin sensitisation. You will be aware that the imine might be converted into a potentially skin-sensitising aldehyde. If you have access to suitable methods you will get an estimate of the fat/water partition coefficient for the imine in order to make a judgment about whether it will penetrate the skin. Most likely you will use a calculated log P value as a measure of fat/water partition coefficient (also known as log K), but there is more about that later in this book. You will presumably have the gumption to consider the partition coefficient for the aldehyde as well, in case the imine is unstable enough to hydrolyse on the surface of the skin.

Depending on the information, you will come up with conclusions and explanations such as:

• "the query substance is likely to be a skin sensitiser because it has the right partition coefficient to penetrate the skin and the potential to be converted into an α, β-unsaturated aldehyde – a class of compounds including many skin sensitisers";

• "the query substance is not likely to be a skin sensitiser because although it is an imine which could be converted into an α, β-unsaturated aldehyde – a class of compounds including many skin sensitisers – both compounds have such low fat/water partition coefficients that they are unlikely to penetrate the skin";

• "the situation is equivocal because the imine has too high a fat/water partition coefficient to penetrate the skin easily but the related aldehyde has a lower fat/water partition coefficient and I do not know how readily the imine will hydrolyse to the aldehyde on the skin surface".

Systems in which a reasoning engine solves problems by applying rules from a knowledge base compiled by human experts were originally called "expert systems", on the grounds that they behave like experts. In this book they are distinguished by being called "knowledge-based systems". They use reasoning to varying degrees and they are creative in the sense that they solve novel problems and make predictions. The particular strength of the best of them is their ability to explain themselves. For example, there is fairly good understanding of why α, β-unsaturated aldehydes are skin sensitisers. The human compilers of a knowledge base can include that information so that the expert system can present it to a user when it makes a prediction and can explain how it reached its conclusion.

Given access to structures and biological data for lots of compounds, you might discover the rule that α, β-unsaturated aldehydes are often skin sensitisers, assuming you were not overwhelmed by the quantity of data. Knowledge-based systems as defined here make no attempt to discover rules from patterns in data – they simply apply the rules put into them by human experts. In terms of the criteria for intelligence, they are unable to learn for themselves. The more general term, "expert system", was later extended to include systems that generate their own models by statistical methods and apply them without any human interpretation. While these systems are perhaps nearer to all-rounders in the stakes for showing intelligence than knowledgebased systems, they fall down on explaining themselves. They cannot go beyond presenting the statistical evidence for their rules.

A speaker remarked at a meeting I attended that "An expert system is one that gives the answers an expert would give ... including the wrong ones". It might be fairer to compare consulting a knowledge-based system (which is what he was talking about at the time) with consulting a group of human experts rather than one, since knowledge bases are normally compiled from collective knowledge, not just individual knowledge, but his warning stands. Other people have, only half-jokingly, suggested that an expert system is one suitable only for use by an expert. That may be over-cautious but users of expert systems should at least be thinking and well-informed: it is what you would expect of someone taking advice from a team of experts.

CHAPTER 2

Synthesis Planning by Computer

Organic synthesis chemists are used to working with ideas and rules of thumb. They do not normally plan reaction sequences to novel compounds on the basis of kinetic or thermodynamic calculations – they are rarely in the position to do so because data of sufficient reliability are not available for the calculations – but they have a reasonable success rate. This raised the question, how do they do it? Could a computer emulate the thinking of a chemist who works out a practical synthesis route to a complicated organic compound?

The tale is told of a conversation over a few beers one evening between three eminent chemists famed for their work in organic synthesis – Elias J. Corey, Alexander R. Todd and Robert B. Woodward. Corey, it is said, expressed the view that computers would eventually be capable of matching or even outclassing human reasoning; soon there would be machines capable of designing chemical syntheses just as well as chemists do. Todd and Woodward were sceptical, it is said, arguing that chemical synthesis was an art more than a science, calling for imagination and creativity well beyond the capacity of a computer. Corey saw how a computer might reason like a chemist and he proposed to set up a project to demonstrate the feasibility of his ideas. The story may be apocryphal but it does not matter if it is. The exciting thing is that Corey recognised a new challenge well beyond the everyday goals of most researchers and took it on. He was not alone in seeing and taking up the challenge – there were others who will feature in this chapter and the next – but his project proliferated like the mustard tree in the parable so that by now every chemist is familiar with at least one spin-off computer application that roosts in its branches.

Corey's project to develop a synthesis-planning program, OCSS (organic chemical simulation of synthesis), started in the 1960s and was described in a paper in Science in 1969. By 1971, when a paper was submitted to the Journal of the American Chemical Society, the program had been re-implemented as LHASA (logic and heuristics applied to synthetic analysis) and the project was expanding.

Right from the start the plan was to develop a computer system that did not just think like a chemist, but communicated like one, too. Computer graphics was in its infancy. The computer mouse was yet to come to public notice – Douglas Engelbart filed his application for a patent in 1967 – but there were systems that linked a graphics tablet, or "bit pad", to a vector graphics screen (a line is displayed on a vector graphics screen by scanning the electron beam between the coordinates of the ends of the line, whereas in a television or a modern personal computer system the screen is scanned systematically from side to side and top to bottom and the beam is activated at the right moments to illuminate the pixels on the screen that lie on the line). Other researchers interested in using computers for chemistry were developing representations of chemical structures to suit computers, but in this project the computer would be expected to use the representations favoured by organic chemists – structural diagrams. In their paper in 1969, Corey and Wipke wrote, "The following general requirements for the computer system were envisaged at the outset: (i) that it be an 'interactive system' allowing facile graphical communication of both input and output in a form most convenient and natural for the chemist ...".

A structural diagram is full of implicit information for a chemist that would not be perceived by someone not trained in chemistry. It is not a picture of a molecule, in as much as there can be a picture of one; it tells you what is connected to what, and how, but it does not tell you the three dimensional locations of atoms: like the map of the London Underground it is a graph. To make useful inferences, the computer needs to be able to "see" the graph like a chemist sees it, and so a chemical perception module in LHASA fills checklists for the atoms and bonds in a molecule for use in subsequent processing. For example, if a carbon atom is found to be bonded through a double bond to one oxygen atom and through a single bond to another oxygen atom which itself bears a hydrogen atom, the carbon atom can be flagged as the centre of a carboxylic acid group; if an atom is at a fusion point between two rings (which would have implications for its reactivity) it can be flagged as a "fusion atom".

Computer perception of a molecule may put the computer in the position to think about it the way a chemist would, but how does a chemist think of ways to synthesise even a simple molecule? The question embodies a host of others each of which probably has more than one answer. Corey would have been well-placed to look for answers suited to computer-implementation, having formulated his ideas for the retrosynthetic approach to chemical synthesis design for which he was later to receive a Nobel Prize in Chemistry, and his thinking on the subject and his work on a computer system must surely have fed each other.

The essence of the retrosynthetic approach is that the target molecule contains the clues to the ways in which it might be constructed. That might be obvious but stating something explicitly and letting it lead your thinking can completely change the way you tackle a problem. To take a simple example, the product of the aldol condensation is an α, β-unsaturated aldehyde or ketone (see Scheme 2.1). So, if there is an α, β-unsaturated ketone in a compound you want to make, perhaps it could be made via the aldol condensation from the appropriate pair of ketones (or a ketone and aldehyde) as illustrated in Scheme 2.2. There is an obvious problem with this synthesis. The aldol condensation is likely to produce a mixture of products, only one of which will be the target, unless R2, R3, R4, are all the same as R1·CH2–, in which case reactants 2.1 and 2.2 will be molecules of the same chemical and the retrosynthetic reaction can be written as shown in Scheme 2.3. To decide whether the aldol condensation is a good or bad choice for a synthesis, a chemist or a computer using the retrosynthetic approach needs access to a set of rules about the effects of appendages in the target represented by R groups in these schemes.

(Continues…)

---

Excerpted from "Knowledge-based Expert Systems in Chemistry"
by .
Copyright © 2019 Philip Judson.
Excerpted by permission of The Royal Society of Chemistry.
All rights reserved. No part of this excerpt may be reproduced or reprinted without permission in writing from the publisher.
Excerpts are provided by Dial-A-Book Inc. solely for the personal use of visitors to this web site.

---

Table of Contents

Artificial Intelligence – Making Use of Reasoning;
Synthesis Planning by Computer;Other Programs to Support Chemical Synthesis Planning;
International Repercussions of the Harvard LHASA Project;
Current Interest in Synthesis Planning by Computer;
Structure Representation;
Structure, Substructure and Superstructure Searching;
Protons That Come and Go;
Aromaticity and Stereochemistry;
DEREK – Predicting Toxicity;
Other Alert-based Toxicity Prediction Systems;
Rule Discovery;
The 2D–3D Debate;
Making Use of Reasoning: Derek for Windows;
Predicting Metabolism;
Relative Reasoning;
Predicting Biodegradation;
Other Applications and Potential Applications of Knowledge-based Prediction in Chemistry;
Combining Predictions;
The Adverse Outcome Pathways Approach;
Evaluation of Knowledge-based Systems;
Validation of Computer Predictions;
Artificial Intelligence Developments in Other Fields;
A Subjective View of the Future