Months ago (last fall?) a post-doc and I spent what seemed like a lot of time at the whiteboard in the hall, considering different ways our planned USS model might score the sequences it was considering for their similarity to a USS motif.
We eventually settled on the crude system shown on the left (yellow table). It evaluates how well the DNA sequence in a 10-base window matches the USS core consensus. Each match to the consensus earns a point, with the total score for the sequence being the sum of the points it's earned. At the time, we realized that this way of scoring had two (or three?) big problems, but we needed something simple to get the model working so we settled for this.
The first problem is that the score is not very sensitive to how good the match is. The yellow numbers beside the table show the scores earned by specific sequences. A sequence matching at all 10 positions is only 11% better than a sequence matching at 9 positions, even though we know from real uptake experiments that some single base changes can reduce uptake by more than 95%. The second problem is that this method treats all 10 positions in the motif equally. But again our uptake experiments have shown that some positions in the motif affect uptake much more strongly than others.
The third problem is that random sequences have very high scores, and adding a single perfect-match USS to it increases this baseline score only slightly.
This morning the post-doc and I reconsidered the scoring system. We expected that finding a solution to these problems would be very difficult, but we quickly came up with a much better way, illustrated by the blue table on the right of the figure. The new method is to multiply the scores of the individual positions rather than summing them. This causes the scores of well-matched sequences to be dramatically higher than those of poorer matches. And we expect (though we haven't tested this yet), that the baseline score of a random sequence will be much smaller. For now we've given all but the consensus base scores of 1, but these could be larger or smaller; for example some bases at some positions could be worth only 0.1 of a point.
Now that the program is working, implementing a multiplicative scoring system should be simple. I'm tempted to try it right now, but I have lots of other things I should be doing, and I'd probably just get bogged down in technical problems anyway.
Why are unfalsifiable beliefs so attractive?
3 days ago in Epiphenom