A max-margin model for efficient simultaneous alignment and folding of RNA sequences

Abstract

Motivation: The need for accurate and efficient tools for computational RNA structure analysis has become increasingly apparent over the last several years: RNA folding algorithms underlie numerous applications in bioinformatics, ranging from microarray probe selection to de novo non-coding RNA gene prediction.In this work, we present RAF (RNA Alignment and Folding), an efficient algorithm for simultaneous alignment and consensus folding of unaligned RNA sequences. Algorithmically, RAF exploits sparsity in the set of likely pairing and alignment candidates for each nucleotide (as identified by the CONTRAfold or CONTRAlign programs) to achieve an effectively quadratic running time for simultaneous pairwise alignment and folding. RAF's fast sparse dynamic programming, in turn, serves as the inference engine within a discriminative machine learning algorithm for parameter estimation.Results: In cross-validated benchmark tests, RAF achieves accuracies equaling or surpassing the current best approaches for RNA multiple sequence secondary structure prediction. However, RAF requires nearly an order of magnitude less time than other simultaneous folding and alignment methods, thus making it especially appropriate for high-throughput studies.Availability: Source code for RAF is available at:http://contra.stanford.edu/contrafold/Contact: chuongdo@cs.stanford.edu

Computational approaches for RNA energy parameter estimation

“…Following the ''large margin'' approach from machine learning (see, e.g., Do et al 2008), we do not just require that the free energy of the true structure be less than all the others; instead, we require that the size of the margin or difference be proportional to the similarity between the true structure and the predicted structure. The intuition is that it is desirable for structures that are similar to the reference structure to have similar energies, but the more a structure deviates from the reference structure, the more different its energy should be.…”

“…As can be seen from Table 3, the accuracy obtained by CONTRAfold 1.1 on S-STRAND2 is only slightly higher than that achieved using Turner99 parameters, while on S-Full-Test the accuracy is slightly poorer. A subsequent version, CONTRAfold 2.0, was trained (by Do et al 2007) on S-Processed, the set we had developed and used for training CG 1.1. CONTRAfold 2.0 achieves an average F-measure of 0.672 on S-STRAND2, which is 0.026 higher than for CG 1.1, possibly because of the differences in the parameter estimation algorithms or energy models (in terms of RNA structural features considered), the maximum-expected accuracy prediction algorithm used by CONTRAfold (Do et al 2006), or their sophisticated algorithm for multihyperparameter learning (Do et al 2007).…”

“…A subsequent version, CONTRAfold 2.0, was trained (by Do et al 2007) on S-Processed, the set we had developed and used for training CG 1.1. CONTRAfold 2.0 achieves an average F-measure of 0.672 on S-STRAND2, which is 0.026 higher than for CG 1.1, possibly because of the differences in the parameter estimation algorithms or energy models (in terms of RNA structural features considered), the maximum-expected accuracy prediction algorithm used by CONTRAfold (Do et al 2006), or their sophisticated algorithm for multihyperparameter learning (Do et al 2007). However, CONTRAfold 2.0 does not reach the prediction accuracies achieved by the best new parameter sets presented in this work; in particular, the F-measure on the comprehensive S-STRAND2 set is 0.031 lower than that obtained by the BL-FR* parameters.…”

“…A disadvantage of this approach is that parameters for some features were already fixed before other parameters were calculated; this limits the range of possibilities considered for the overall parameter set. Do et al (2006Do et al ( , 2007 considered a model similar to the Turner model, but in which parameters are not free-energy changes, because they are not normalized. To estimate parameters, they maximize the conditional likelihood of a structural data set and use a gradient-based method to inform the estimation procedure.…”

Mean-Field Theory of Meta-learning

“…The current trend of machine learning theory and applications is to develop meta-learning techniques, since no single paradigm is superior to others in all possible situations [13][14][15][16][17][18]. What is the exact definition of Meta-Learning?…”