Noob question: Is there a standard/straightforward way of devising a score function which takes into account the likelihood/joint probability of its arguments?

I'm a hobbyist with an interest in bioinformatics, and I've been trying to come up with a modification to a classical Dynamic Programming algorithm to predict RNA secondary structure. In the algorithm you're given a string of symbols, and in the DP step you fill out an N² table of symbol pairings (so Aij is the table entry for the evaluation of symbols i and j). The evaluation is strictly concerned with whether the symbols are compatible, so the score is 1 if they are and 0 if they aren't.

This compatibility only takes into account the identities of the individual symbol pairings, which is unrealistic. I'd like it to depend on the neighbors of the symbols as well, and I have access to a corpus of training data (RNA sequences with ground truth structures) where I could learn this information. But I'm not sure how to use this information in a good scoring function.

So my question is, I can estimate the conditional probability of each pairing given the neighboring symbols from data, but is that a sensible measure of compatibility? Any resources with regards to how to use probabilities in these types of problems would be very much appreciated.

Thanks a lot!