Can a technology that has been used for such wide-ranging purposes as setting efficient bus routes and scheduling the operation of industrial washing machines also help researchers track down microRNAs? One company that has been working in the field of computational algorithms thinks so.
Natural Selection, headquartered in La Jolla, Calif., was founded in 1993 by Lawrence Fogel in order to use computational evolution as a way to solve real world problems in areas such as factory scheduling at Levi Strauss, bus routing at Greyhound, and breast cancer detection from mammograms.
Computational evolution uses algorithms to select which of a series of possible solutions is best suited to a particular problem. “The idea for this came from looking at nature, where natural organisms are faced with environments that change with time and are highly complex,” Gary Fogel, vice president of Natural Selection and son of the company’s founder, told RNAi News.
According to Fogel, for any given problem, one starts off with a collection of random solutions, and these are scored based on their fitness for solving the problem. Based on the score, the various solutions are ranked, with the lowest scorers being discarded and the top scorers saved.
“Now, you’re going to try to regenerate the full population [of solutions] by making variations off of the good, useful solutions you found in the first variation,” Fogel said. The original top scorers, termed parents, are then evaluated against new solutions, called offspring, which have been created from a combination of the features of the parents and a small amount of variation.
“You’ve regenerated your population and you’ve completed one generation,” he said. “Now, you rescore them all again, and iterate that process of eliminating a percentage of them, using variation to regenerate the population, and then rescoring … until you have plateaued at some answer that you think is useful enough or you run out of time.”
The idea behind Natural Selection’s miRNA work is to apply the computational evolution approach to pattern recognition models, called neural networks, designed to look for the regulatory RNAs. “Neural networks … take in input descriptors, or statistics, and translate them in a non-linear fashion to an output of decision,” Fogel explained.
“The classical statistical approach is to generate statistics about a process and then linearly to say: ‘Here’s descriptor number one. How correlated is that to the output decision,’” he said. “You go through each of the descriptors one by one on their own and see how well they match the output in terms of predicting.”
The problem with the approach, according to Fogel, is that it doesn’t allow for the evaluation of the different descriptors all together or in different combinations. The classical approach, he said, treats all the descriptors independently, and “very rarely in the real world are things independent.”
But neural networks have their limitations — specifically when it comes to ranking the descriptors by their importance. “For you to write down the weights of the importance for all of those descriptors individually, or their combinations, there’s this huge list of possible weights,” Fogel said. “How do you think about that problem? How do you know how important descriptors one, two, five, six, and 42 are relative to 42 on its own?”
While there are existing ways to tackle this problem, most commonly using a method called backpropagation — which involves adjusting the weights of the neural network until its output more closely matches the desired output — Fogel said that “we’ve noticed that if you allow evolution to [assign] the weights associated with the neural network, it can outperform the weight assignment that backpropagation was doing.”
He stressed that Natural Selection isn’t discounting backpropagation. “For real-world problems we’ve faced, typically we can find a better solution [with computational evolution],” Fogel said. However, backpropagation could be used in conjunction with the company’s approach by acting as a starting point for evolution, so that “you don’t have to start completely at random,” he added.
By applying computational evolution to neural networks, “it’s no longer a population of [solutions], it’s a population of neural networks, each with their own weight assignment,” Fogel said. “I’m going to select the ones, over time, that do the best job of doing the pattern recognition on the training set. Now I’m going to take the best model at the end and say: ‘How well do you do on stuff you’ve never seen before?’”
This pattern recognition optimization, Fogel said, “can be used for gene array expression analysis, or gene discovery … or for microRNA detection because they are all pattern recognition problems.”
As such, Natural Selection’s biomedical focus has recently expanded to include the development of a neural network method for the detection of microRNAs, and the company recently was awarded a National Science Foundation SBIR phase I grant for the project. The grant is worth almost $100,000 and runs from July 1 until the end of 2004.
For miRNAs, Fogel said that Natural Selection is using descriptors including base composition and “sequences that are known to be important in RNA structure,” but noted that the key features that could be used to identify miRNAs are still being worked out in-house using a database of all known miRNAs.
“Once we have those descriptors, we can push the button on the evolution and let it go,” he said. “After some period of time [we will] see how well it’s doing, adjust the descriptors, adjust the model, and off we go again — we’ll keep reiterating that until we get a good model.”
Fogel said that it is Natural Selection’s hope to be able to commercialize its miRNA neural network. “There’s lots of genomic sequence and there’s only a few places that are really important to look at,” he said. “So, we’re taking the best model we can evolve and then applying it to genomic sequences and saying: ‘Where are the microRNAs that we really want to find’ or ‘Where are areas in the genome that are likely to be microRNAs?’
“That’s the key that will come out of this — a tool for pharmaceutical research that can be used to hone in on new, unannotated sequence[s] [to find] regions that we should be looking at,” Fogel said.
Although Fogel declined to comment on whether Natural Selection has had discussions with any parties from the RNAi sector that might be interested in the technology, he noted that Isis Pharmaceuticals — which has worked with the company in the past, including on a project using computational evolution to discover RNA structural elements — is “aware” of the miRNA project and the NSF grant.