Anchors: a technique of pre-classification and its effects on Hidden Markov Models

Hidden Markov Models have been prominent in the categorization of biological data in recent years. Genemark and Genscan, which incorporate Hidden Markov Models, are popular tools used to identify regions in DNA sequences.Hidden Markov Models make predictions about an observable sequence based on transition and emission probabilities of its states. A pre-classification of part of the observable sequence is proposed in order to increase the accuracy of prediction of the HMM. To pre-classify the sequence, we analyze overrepresented and under-represented substrings in the observable sequence and make a classification if the substring is significantly unique to one of the classification categories. With the addition of these 'anchors' in the sequence, the predictive accuracy of a HMM has been found to be improved.A previous experiment which studied the predictive accuracy of a first order Hidden Markov Model on a mix of English and Spanish (based upon consonant and vowel transitions) yielded favorable results at low probabilities of a language transition. The experiment is revisited with the addition of the pre-classification, which classifies approximately 1% of the sequence at a high accuracy (over 90%). This yields a considerable improvement in classification by the Hidden Markov Model. This addition to new or current HMM techniques of DNA sequence classification should improve accuracy.