Gene prediction in eukaryotes with a generalized hidden Markov model that uses hints from external sources

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene prediction relies on extrinsic evidence (genome comparisons, EST/protein alignments), which is often incomplete and uncertain.
• Extrinsic evidence alone is insufficient for complete and accurate gene structure recovery.
• Balancing extrinsic evidence with sequence-intrinsic data is crucial for improving gene prediction.

Purpose of the Study:

• To develop a general method for integrating external information into gene prediction.
• To enhance the ab initio gene prediction program AUGUSTUS using extrinsic evidence.
• To create a versatile tool, AUGUSTUS+, for more accurate gene structure identification.

Main Methods:

• Utilized a Generalized Hidden Markov Model (GHMM) to evaluate hints from intrinsic and extrinsic data.
• Integrated extrinsic hints from EST and protein database matches.
• Developed AUGUSTUS+ by extending the AUGUSTUS program.

Main Results:

• The new method, AUGUSTUS+, effectively integrates intrinsic and extrinsic information.
• The approach is robust to the length of database matches and utilizes information from absent matches.
• AUGUSTUS+ can predict genes with user-defined constraints, such as known exon positions.
• Achieved 89% accuracy in predicting exons on human chromosome 22 using EST and protein database hints.

Conclusions:

• Probabilistic modeling of extrinsic evidence, like sequence database matches, significantly enhances gene prediction accuracy.
• Sequence interval matches should be treated as compound information, not just positional data, for improved gene prediction.