A new algorithm to train hidden Markov models for biological sequences with partial labels.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Hidden Markov models (HMMs) are essential for biological sequence analysis, including protein family profiling and domain identification.
• Standard HMM training methods include maximum likelihood (for labeled data) and expectation maximization (e.g., Baum-Welch for unlabeled data).
• Partially labeled biological sequence data presents a challenge for existing HMM training methodologies.

Purpose of the Study:

• To develop a novel training method for HMMs that effectively utilizes partially labeled biological sequence data.
• To address limitations in current HMM training approaches when dealing with incomplete sequence annotations.
• To enhance the accuracy and applicability of HMMs in biological sequence analysis.

Main Methods:

• A new training method was designed, building upon the Baum-Welch algorithm.
• The method is specifically tailored for HMM training when only partial labeling is available.
• The approach was evaluated using both synthetic and real biological sequence data.

Main Results:

• The developed method significantly improves HMM model training compared to a similar active learning approach.
• Higher accuracy in decoding was achieved when tested with synthetic and real biological data.
• The method demonstrates enhanced performance in model training utilizing partial labels.

Conclusions:

• A novel training method for HMMs using partially labeled data has been successfully developed.
• This method is expected to advance the detection of de novo motifs and signals in biological sequences.
• The method will be applied to active learning for detecting plasmodesmata targeting signals, with validation through wet-lab experiments.