Revealing the hidden language of DNA
(English) Genomics has revolutionized in recent years due to the rapid advancements in high-throughput sequencing technologies, leading to an explosion of genomic data. This has opened up new opportunities for utilizing natural language processing (NLP) techniques to analyze and extract knowledge fr...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2024 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/450734 |
| Acceso en línea: | https://hdl.handle.net/2117/450734 https://dx.doi.org/10.5821/dissertation-2117-450734 |
| Access Level: | acceso abierto |
| Palabra clave: | 004 - Informàtica 575 - Genètica general. Citogenètica general. Immunogenètica. Evolució. Filogènia Àrees temàtiques de la UPC::Informàtica Àrees temàtiques de la UPC::Ciències de la salut Àrees temàtiques de la UPC::Enginyeria biomèdica |
| Sumario: | (English) Genomics has revolutionized in recent years due to the rapid advancements in high-throughput sequencing technologies, leading to an explosion of genomic data. This has opened up new opportunities for utilizing natural language processing (NLP) techniques to analyze and extract knowledge from genomic data. However, applying NLP techniques in genomics presents unique challenges due to the differences in genetic data's underlying structure and complexity. State-of-art biological embedding algorithms, such as DNA2VEC, split biological sequences into overlapping sub-sequences of length k, called k-mers, and learn a k-mer-based embedding space. By representing k-mers as continuous vectors, a strong correlation has been observed between the cosine similarity of these embeddings and the global sequence alignment score. This correlation suggests that the k-mer embeddings effectively capture sequence similarities. In other words, if two genes or proteins are embedded closely in the embedding space, they will likely exhibit high sequence similarity. We proposed new embedding methods to go beyond the sequence similarity: untangle 1 (UNT1) and untangle 2 (UNT2). By changing the calculation of the co-occurrences of k-mers in the contextual window, we significantly reduced the correlation to sequence similarity. While capturing fewer semantically similar GO terms, we showed that the novel methods UNT1 and UNT2 provide a broader representation of functional relationships than DNA2VEC. In addition to sequence data, the availability of diverse omics data has witnessed a remarkable surge, with the recognition of its potential to provide complementary insights. In our research, we depart from the conventional use of Singular Value Decomposition (SVD) to represent sequences in a low-dimensional space. Instead, we employ Non-negative Matrix Tri-Factorization (NMTF) to build a directly interpretable biological sequence embedding method. While Singular Value Decomposition (SVD) has successfully produced high-quality embeddings, our choice of NMTF aligns with the goals of explainable AI (XAI) and the need to understand the biological implications of the generated embeddings. NMTF serves as a potent instrument for data fusion, enabling the integration of information from various sources to achieve a deeper knowledge of complex biological systems. By utilizing NMTF, we aim to provide a more comprehensive understanding of the relationships between biological entities, such as k-mer, coding or non-coding regions, and their functional characteristics. We developed a versatile approach for annotating k-mers with genes, which allowed us to investigate the inheritance of annotations across different k-mer lengths. Specifically, we mapped k-mers to genes and vice-versa based on their over-representation in the sequences of the genes. Our study showed that the k-mer clusters are significantly enriched in GO annotations, indicating that the embeddings of k-mers capture functional genome organization. We showed that the same embedding models based on k-mers of DNA not only capture a functional structure of genes, but also effectively distinguish non-coding regions of DNA. We analyzed the non-coding regions of the yeast genome and compared the performance of the models DNA2VEC, UNT1, and UNT2. Our findings suggest that the optimal length of k-mers to classify ncRNAs is shorter than 7. This observation is fundamental since the feasibility of mostly all k-mer-based methods is highly dependent on the volume of the vocabulary increasing exponentially with the k-mer's length. Furthermore, our findings indicate that models utilizing k-mers of length 6 and larger outperform models based on shorter subsequences when predicting chromatin interactions. By emphasizing the importance of longer k-mers, our study provides valuable insights into the potential of utilizing extended sequence motifs to advance our understanding of the structural organization of the genome. |
|---|