Table of Contents
There are three major kinds of evidence we consider in BayesFold 1.0, and sometimes several kinds of experiment within each type:
Thermodynamics: What is the energy of each of the sequences when folded into each of the structures? Calculations of thermodynamic pair probability are described in Section 2.2.1, “Pair Probability”.
Covariation: How likely would we be to see the pattern of changes across the alignment if each of the structures were true? Covariation includes mutual information (Section 2.2.3, “Mutual Information”) and fraction pairable (Section 2.2.2, “Fraction Pairable”).
Chemical Mapping: How likely would we be to see the pattern of light and dark bands when a sequence is mapped with each chemical if each of the structures were true? Calculation details are described in Section 2.1.1, “Chemical Mapping”.
We explain how to calculate Pr(Di| Hj) for measurements derived from each of these criteria in the following sections. Since the list of structures under consideration is assumed to be exhaustive, the general method is to find Hmax, the structure that has the best statistical support. We then get Pr(Di| Hj) by assuming that some test statistic calculated for Hmax is the true value, and asking how surprising it would be to find the observed value of the test statistic for each of the Hj. This limits the effects of sampling error, because Pr(Di| Hmax) is always 0.5 by definition (since, if the true value were that calculated for Hmax, chance predicts that we would find a higher value half the time and a lower value the other half of the time just by sampling error). Consequently, even a measurement (such as a difference in means) based on a very small sample can confer apparent support of at most twice that for Hmax, and such artifacts are unlikely to be consistent across independent types of data.