Although remarkable advances have been reported in high-throughput sequencing, the ability to aptly analyze a substantial amount of rapidly generated biological (DNA/RNA/protein) sequencing data remains a critical hurdle. To tackle this issue, the application of natural language processing (NLP) to biological sequence analysis has received increased attention. In this method, biological sequences are regarded as sentences while the single nucleic acids/amino acids or k-mers in these sequences represent the words. Embedding is an essential step in NLP, which performs the conversion of these words into vectors. Specifically, representation learning is an approach used for this transformation process, which can be applied to biological sequences. Vectorized biological sequences can then be applied for function and structure estimation, or as input for other probabilistic models. Considering the importance and growing trend for the application of representation learning to biological research, in the present study, we have reviewed the existing knowledge in representation learning for biological sequence analysis.

MotivationA cancer genome includes many mutations derived from various mutagens and mutational processes, leading to specific mutation patterns. It is known that each mutational process leads to characteristic mutations, and when a mutational process has preferences for mutations, this situation is called a ‘mutation signature.’ Identification of mutation signatures is an important task for elucidation of carcinogenic mechanisms. In previous studies, analyses with statistical approaches (e.g. non-negative matrix factorization and latent Dirichlet allocation) revealed a number of mutation signatures. Nonetheless, strictly speaking, these existing approaches employ an ad hoc method or incorrect approximation to estimate the number of mutation signatures, and the whole picture of mutation signatures is unclear.ResultsIn this study, we present a novel method for estimating the number of mutation signatures—latent Dirichlet allocation with variational Bayes inference (VB-LDA)—where variational lower bounds are utilized for finding a plausible number of mutation patterns. In addition, we performed cluster analyses for estimated mutation signatures to extract novel mutation signatures that appear in multiple primary lesions. In a simulation with artificial data, we confirmed that our method estimated the correct number of mutation signatures. Furthermore, applying our method in combination with clustering procedures for real mutation data revealed many interesting mutation signatures that have not been previously reported.Availability and implementationAll the predicted mutation signatures with clustering results are freely available at http://www.f.waseda.jp/mhamada/MS/index.html. All the C++ source code and python scripts utilized in this study can be downloaded on the Internet (https://github.com/qkirikigaku/MS_LDA).Supplementary information Supplementary data are available at Bioinformatics online.

Remarkable advances in high-throughput sequencing have resulted in rapid data accumulation, and analyzing biological (DNA/RNA/protein) sequences to discover new insights in biology has become more critical and challenging. To tackle this issue, the application of natural language processing (NLP) to biological sequence analysis has received increased attention, because biological sequences are regarded as sentences and k-mers in these sequences as words. Embedding is an essential step in NLP, which converts words into vectors. This transformation is called representation learning and can be applied to biological sequences. Vectorized biological sequences can be used for function and structure estimation, or as inputs for other probabilistic models. Given the importance and growing trend in the application of representation learning in biology, here, we review the existing knowledge in representation learning for biological sequence analysis.

Mutation signatures are defined as the distribution of specific mutations such as activity of AID/APOBEC family proteins. Previous studies have reported numerous signatures, using matrix factorization methods for mutation catalogs. Different mutation signatures are active in different tumor types; hence, signature activity varies greatly among tumor types and becomes sparse. Because of this, many previous methods require dividing mutation catalogs for each tumor type. Here, we propose parallelized latent Dirichlet allocation (PLDA), a novel Bayesian model to simultaneously predict mutation signatures with all mutation catalogs. PLDA is an extended model of latent Dirichlet allocation (LDA), which is one of the methods used for signature prediction. It has parallelized hyperparameters of Dirichlet distributions for LDA, and they represent the sparsity of signature activities for each tumor type, thus facilitating simultaneous analyses. First, we conducted a simulation experiment to compare PLDA with previous methods (including SigProfiler and SignatureAnalyzer) using artificial data and confirmed that PLDA could predict signature structures as accurately as previous methods without searching for the optimal hyperparameters. Next, we applied PLDA to PCAWG (Pan-Cancer Analysis of Whole Genomes) mutation catalogs and obtained a signature set different from the one predicted by SigProfiler. Further, we have shown that the mutation spectrum represented by the predicted signature with PLDA provides a novel interpretability through post-analyses.