Chemistry, Chess and Computers

Sometime in the 1970’s the Amer. Chem. Soc. published a review of Computers in Chemistry (cannot remember date or title and I’ve lost my copy) and it has remained an inspiration ever since. In it was summarised the work of the Stanford (DENDRAL, CONGEN) and Harvard (LHASA) groups on the applications of artificial intelligence to chemistry (structure elucidation and organic synthesis). Both have heavy elements of problem-solving, coupled with pattern recognition. The systems effectively contained:
* a knowledge base of chemistry
* a set of heuristics (rules)
* formal deterministic procedures (e.g. tree searches).
The accomplishment was remarkable. The systems worked. They weren’t as good as a professional synthetic chemist, but in small areas they were better than me. It seemed obvious to me that with sufficient work on all components, but especially the knowledge base these systems would be able to do organic chemistry at the level of all except the best in the field. Certainly I expected that with the passage of 30 years the chemist/machine combination would be common. (Admittedly I sometimes believed too much hype about AI – now that I work in one branch (language processing) I know how difficult it is).
At the same time very similar work started to be done on chess. Again, when the first programs came out I could easily beat them (and I am a weak player). But gradually they improved and now they can beat essentially all humans.
It seemed to me that chemistry and chess would be quite similar. They are formal systems, too complex for brute force, and where a knowledgebase is essential. In chess all significant games have been captured in a database, and a large number of endgames have been exhaustively worked out. What is interesting is that the chess grandmasters have formed a symbiosis with computer programmers and machines and are still exploring what aspects machines can and cannot do. (I’m not an expert here and comments would be welcome).
By contrast there has been no significant work on chemistry and AI in, perhaps, 15 years. When I was in the pharma industry my boss used to speak of “another outbreak of Lhasa fever” (sic) – meaning that someone had suggested that machine synthesis should be explored. The Lhasa organisation has effectively stopped supplying synthesis methodology and turned to toxicology prediction (albeit it highly valuable).
So I feel a considerable feeling of sadness. I am sure that if synthetic chemists had embraced computers in the same way as chess players we would be sgnificantly better off. This is, of course, an act of faith but it’s borne out by the knowledge revolution taking place in many disciplines. The bioscientists are eagerly exploring the S/semantic W/web witn formal ontologies and reasonaing – another approach to “AI”.
I’ve just been at the UK eScience meeting (cyberinfrastructure) meeting for 3 days. (I’ll probably hark back in future posts). One keynote was given by Stephen Emmott (Director, Eur. Sci. Programme) Microsoft Research, Cambridge). Stephen talked about 2020 and gave a vision when computing could be based on biology – where molecular computers have already been injected into cells. Microsoft is hiring bioscientists who are also computer-able (i.e. they can make their ideas happen through code, rather than requiring comput/er/ational scientists to write the code for them.) He stressed that he did not want a mixture of computer scietists and biologists, he wanted scientists with a mixture of computing and biology. Since his future involves molecules, maybe he’s also hiring chemist/computerScientists…
But we are actively discouraging the sort of work envisioned By Lederberg and Corey 30 years ago. There are exceptions – I spent 3 hours with my colleague Steve Ley discussing how we can bring modern informatics into synthetic chemistry. I am sure that our biggest problem is the lack of an immediate Open global knowledge base in chemistry. It’s all there on paper, but to get it into a machine is a mighty task. It will need new methods of computing – including social computing and I’ll explore these ideas systematically in this blog. We might even achieve something with your help.
So I am pleased to see the quality of the chemical blogs, even if Tenderbutton is retiring. With lightweight mashup-like approaches we may be able to use the new approaches to informatics that are being developed in social computing. Biology has control of its knowledgebase – it had to fight to keep it in the genome information wars- but it’s vibrant and innovative. Chemistry has surrendered its knowledgebase to commercial and quasi-commercial interests who point in the direction of pharma rather than the information revolution. I will show in a week or two how we might be able to start regaining some of it.

This entry was posted in chemistry, open issues, programming for scientists. Bookmark the permalink.

8 Responses to Chemistry, Chess and Computers

  1. When I was a grad student 15 years ago I assumed that chemical synthetic design in the new millenium would involve massive components of automation. That clearly has not happened. After seeing several talks at the ACS about computer assisted synthesis planning I think I understand why more clearly than ever: too much proprietary entanglement. The databases are closed, the articles are closed and the software is closed. If I can’t go back to my lab and play with a new software system without worrying about violating patent and copyright issues, then it might as well not exist for all practical purposes. When we will have computer assisted synthetic design software that is open source and enough data about chemicals and reactions that are truly open then I think we will see the human-automaton synergy explode and reach its true potential.

  2. pm286 says:

    I have just discovered an article in Chem Soc Rev (RSC) comparing Organic Synthesis and Chess.
    Since I am at home and it is Closed source I can’t read it for several days so I can’t comment on it

  3. Peter – I actually cited that article in the context of games I use to teach organic chemistry.
    But I just realized that Mat Todd, the author of that article, is the same guy from the Synaptic Leap who has been commenting on our experiments:
    He also wrote an article about Open Source Research in Chemistry
    How’s that for a small chemical world?

  4. pm286 says:

    First an apology to Mat that I didn’t cite his article. But I didn’t know about it and I don’t naturally read Closed Access papers.
    I love the idea of games for teaching. I used to do things like this – not organic synthesis.

  5. There is a big difference between organic synthesis design and chess: in chess, the legal moves are clearly specified and the number of legal moves is relatively small. In organic synthesis, there is an unspecified (usually large and growing…) number of possible reactions that consitute a “move”, and it is hard to predict whether they work or not. However, I agree that much progress could be made if the structure and reaction databases were open.

  6. Mat Todd says:

    Ivan, you’re correct. The rule set for organic synthesis grows. The discovery of metathesis, for example, is like inventing a new move where a pawn can move 3 squares – it can change one’s whole strategy. There are other similarities/differences I talk about in the article – e.g. moves in chess don’t have yields.
    The two things I stressed though were that:
    a) You need a large, updated knowledgebase. These things exist in the private sector (e.g. Scifinder), but only as comprehensive collections of reported reactions. An advantage of interrogating Steve Ley rather than a computer is that Steve can supply important value judgements. Hence the knowledgebase for computer-aided organic synthesis really needs to be heuristic, which is where an open source effort is probably required. e.g. a comprehensive matrix of reagents vs. functional groups with which they react to identify conflicts in a suggested sequence.
    b) You need intelligent programming. I spoke with Murray Campbell about the success of the Deep Blue project, who was at pains to point out that Kasparov lost due in part to computational power, but mainly due to good algorithms. e.g. when to stop looking at a certain route. I mention in the article a very nice paper by Gelernter where an existing program is parallelized. This has advantages, but the way it’s programmed is very wasteful (as he fully acknowledges). i.e. power is nothing without control.

  7. pm286 says:

    (5) (6)
    Mat mailed me with the review – very nice and comprehensive. I can’t post it, of course, since it’s Closed. Here’s my thoughts to Mat…
    Thanks, very good summary – though depressing that it all ended 15 years ago
    The key question is what problem are you trying to solve. Deep Blue is programmed to win games against human or machines, not necessarily to emulate human thought. What is CAOS trying to do? There are at least the following:
    (a) emulate the current syntheses in JACS/JOC, etc. It is unclear what the point of these are. The choice of target is not that it is useful or interesting (which most aren’t) but that they are difficult. It’s more like rock-climbing than science. It’s argued that this is what the pharma industry wants. I actually think it is harmful as it teaches students to be narrow-minded, oblivious to the world outside (e.g chemical biology, which is of far more value). Drugs fail not because they can’t be made but because they are bad drugs. So no, computers are unlikely to emulate some of this, but does this matter.
    (b) help synthesize useful compounds. Here we have a much greater and exciting opportunity based on both reactant-based knowledge and reaction-based knowledge. In industrial syntheses we need to know about the reactions, rates, solvents temperatures etc.
    If we concentrated on automating the boring bits of organic chemistry – reading the literature, finding reagents, computing the physical properties of substances and reactions then there would be a great enhancement. This is possible today – we can extract reaction conditions automatically from the literature – but it is held back by the regressive approaches to IPR.
    But these ideas are not popular!

  8. Denis Bucher says:

    I liked the analogy between chess and organic chemistry.
    I think “tree”-type algorithm are not optimal to predict outcomes in chemistry. Not only the chemical space is huge (choice of atoms & their combinations), but also the atoms move in a continous space. Instead, Neural networks and genetic algorithms can handle the complexity of such system.
    While, it’s true that today’s computer can beat chess champion in a match using raw calculations, humans are still stronger if you give them enough time (let’s say a day) to think about a position. At the moment the computer is only faster not better.

Leave a Reply

Your email address will not be published. Required fields are marked *