InkSpot. Science. On Demand

Automating Science

• Author: David Leahy
• September 8, 2008
• Category: Uncategorized, drug discovery

Of course scientists try to be objective and rational, but are only human. So when someone says they can automate what scientists do, and the decisions they make, when those decisions are derived from years of study and practical experience, it doesn’t go down very well. So I am going to get myself in trouble and might have to leave.

I think we can automate much of what many scientists do in software. I think that when we do that right, quality improves.

I’ll get my coat.

There was a flavour of this is a recent provocative post by Chris Anderson of Wired which basically said that when you have enough data you don’t need models, and by implication, the scientists that create models. Not surprisingly this upset some people. I don’t see why, since in chemistry at least we are used to this and always have been. It goes back to the similarity principle, i.e. an hypothesis that says that similar chemical structures tend to have similar properties. This is a model-free approach that has practical benefits, although it is limited by the “lumpy” nature of chemical space. In other words, individual chemicals occupy discrete spaces, and not all space is occupied. But that isn’t what I meant, although it is an interesting topic.

Many scientists, particularly in Universities, are doing exploratory science where the experiments and analysis methods are uncertain, that is why it is research. There is still something intuitive and very human about this investigative process, when new approaches are being applied. However, in most cases the methods have “settled down”, in other words they have reached some kind of consistency. At that point they can be automated, at the point where the scientist is starting to repeat the same process over and over again. In my previous company, Cyprotex, the challenge we faced was to take low throughput, slow and expensive assays for ADMET and make them high throughput, very low cost and high quality. We succeeded because of a super-human effort on the part of our scientists and IT team which was often painful. Our scientists ferociously guarded quality and our IT team stubbonly insisted on automation. My belief was that if a scientist could explain how they made a decision in data analysis then the software team could write software to reproduce those decisions. If they couldn’t explain it, then their decisions were subjective and wouldn’t be reproduced from day to day and between different scientists. It was a difficult thing for us to do, but very successful and as a result Cyprotex today has the best (highest quality, highest throughput, lowest cost) ADMET screening facility in the world, probably …

You can do this of course where the science becomes more mature and we see this in the rise of workflow technologies such as Pipeline Pilot, Knime, Taverna and others. MyExperiment is a new site aimed at sharing workflows for data management and analysis which capture scientists working practice and in so doing automate the parts of the process they do regularly.

We took the workflow concept for automating scientific decision making further in the development of “Competitive Workflow”, and implemented this as the Discovery Bus. This workflow system has been developed for automating QSAR and we are currently working on the “Mother of All QSAR’s” with it. It automates what hundreds of QSAR experts do, but it deserves its own post, so more later.�

• 2 Responses
• Leave a comment...
• Comment Feeds