Enhancing the prediction of protein coding regions in biological sequence via a deep learning framework with hybrid encoding

Protein coding regions prediction is a very important but overlooked subtask for tasks such as of complete gene structure, coding/noncoding RNA. Many machine learning methods have been proposed this problem, they first encode biological sequence into numerical values and then feed them classifier final prediction. However, encoding schemes directly influence the classifier's capability to capture features how choose proper scheme remains uncertain. Recently, we protein region method in transcript sequences based on bidirectional recurrent neural network with non-overlapping 3-mer feature, achieved considerable improvement over existing methods, there still much room improve performance. First, feature that counts occurrence frequency trinucleotides only reflects local order information between most contiguous nucleotides, which loses almost all global information. Second, kmer length k larger than three (e.g., hexamer) may also contain useful Based two points, here present deep framework hybrid sequences, effectively exploit information, gapped (gkm) statistical dependencies among labels. 3-fold cross-validation tests human mouse demonstrate our significantly outperforms state-of-the-art methods. • Coding labels dependency can extend from genomic sequence. Hybrid single scheme. Global be captured by CNN.