An intelligent robot that could free genomics researchers from routine lab chores has proven as effective as a human scientist. The robot not only performs genetics experiments, it also decides which ones to do, interprets the results and comes up with new hypotheses.
Fields such as genomics are crying out for better and more intelligent automation because they are generating data much faster than it can be analysed. Stephen Muggleton, a computer scientist at Imperial College London, UK, and a member of the team that developed the system, says that scientists in genomics are becoming overwhelmed. Data is increasing almost exponentially, he adds, making more automation inevitable.
Now, Muggleton and his colleagues report how their machine, which they call Robot Scientist, fared when set a typical genomics task – to determine the function of a set of genes in brewer’s yeast, Saccharomyces cerevisiae.
Having been programmed with the necessary genetic background information and a model of the type of experiment to be carried out, their robot successfully confirmed the function of a set of well-known genes in yeast. The Robot Scientist is the first to both “reason” the best experimental approach, and carry it out, say its architects.
Its “brain” is a PC running two novel software packages developed by the team: an experiment selection system called ASE and a hypothesis generator called Progol. The PC is linked up to standard computer-controlled lab equipment. The only human intervention needed is to carry test plates to and from an incubator.
The robotic system was asked to find the function of yeast genes involved in the biosynthesis of some essential amino acids. It was provided with mutant strains of yeast, each lacking a particular gene, and set to go.
Like a researcher tackling a new problem, the system does not come empty-headed. It is pre-programmed with an incomplete model of the biochemical pathways and metabolic networks of yeast, and information on the genes that encode the proteins involved. It uses this knowledge to generate hypotheses about what the missing genes might do.
From its database, the ASE software devised experiments in which particular mutant and normal strains were mixed with different nutrients or intermediates in the biosynthetic pathway. The growth of the yeast was then measured and compared.
For each mutant yeast strain ASE will predict whether or not it can survive without particular nutrients. Progol then tests the predictions by planning the precise experiment that will compare the survival of the mutant to that of the normal strain.
If Progol is correct – and it was 98 per cent of the time – the software uses this knowledge to help formulate the next set of hypotheses – and the next set of experiments to firm up its ideas.
The team compared the performance of the Robot Scientist with that of a graduate student doing the same experiments. Not only were the results just as accurate, but the system did not need to perform as many experiments because its hypothesis generator found solutions more quickly, so its costs were about two-thirds lower.
Automating this drudge work, the team say, will leave scientists to “make the high-level creative leaps at which they excel”.