Abstract
With different genomes available, unsupervised learning algorithms are essential in learning genome-wide biological insights. Especially, the functional characterization of different genomes is essential for us to understand lives. In this book chapter, we review the state-of-the-art unsupervised learning algorithms for genome informatics from DNA to MicroRNA. DNA (DeoxyriboNucleic Acid) is the basic component of genomes. A significant fraction of DNA regions (transcription factor binding sites) are bound by proteins (transcription factors) to regulate gene expression at different development stages in different tissues. To fully understand genetics, it is necessary of us to apply unsupervised learning algorithms to learn and infer those DNA regions. Here we review several unsupervised learning methods for deciphering the genome-wide patterns of those DNA regions. MicroRNA (miRNA), a class of small endogenous non-coding RNA (RiboNucleic acid) species, regulate gene expression post-transcriptionally by forming imperfect base-pair with the target sites primarily at the 3' untranslated regions of the messenger RNAs. Since the 1993 discovery of the first miRNA let-7 in worms, a vast amount of studies have been dedicated to functionally characterizing the functional impacts of miRNA in a network context to understand complex diseases such as cancer. Here we review several representative unsupervised learning frameworks on inferring miRNA regulatory network by exploiting the static sequence-based information pertinent to the prior knowledge of miRNA targeting and the dynamic information of miRNA activities implicated by the recently available large data compendia, which interrogate genome-wide expression profiles of miRNAs and/or mRNAs across various cell conditions.
Introduction
The central challenge in genome informatics is the "ground truth" problem: while we have sequenced massive amounts of DNA, we lack functional labels for the majority of it. Transcription Factors (TFs) bind to specific DNA regions to regulate gene expression, and small RNA molecules (miRNAs) further tune this expression. However, identifying these binding sites is difficult because they are short, highly variable, and context-dependent. Traditional experimental methods are too slow and expensive to map these elements comprehensively. This review explores how unsupervised learning—ranging from Hidden Markov Models (HMMs) to Nonparametric Bayesian frameworks—can bridge this gap by identifying patterns in raw sequence and expression data without pre-labeled training sets.
Methodology
DNA Motif Discovery and Binding
The "de facto" standard for modeling DNA binding preferences is the Position Weight Matrix (PWM), which assumes each position in a binding site is independent. To account for more complex biological features like dependencies between positions or insertions/deletions (indels), Hidden Markov Models (HMMs) are frequently employed.
Advanced integrative tools like SignalSpider and GEM combine ChIP-Seq signal data with sequence motif discovery to achieve high-resolution binding site predictions. SignalSpider is particularly notable for its linear computational complexity, allowing it to integrate over 100 different ChIP-Seq profiles simultaneously.
miRNA Regulatory Networks
Inferring miRNA regulatory networks involves two main tasks: target prediction and module detection.
- Target Prediction: Tools like PicTar use HMM-based Maximum Likelihood approaches, while TargetScore utilizes a Variational Bayesian-Gaussian Mixture Model (VB-GMM) to integrate sequence scores with experimental expression data.
- Module Detection: Because miRNAs often act synergistically, several frameworks identify regulatory modules:
- GroupMiR: Uses the Indian Buffet Process (IBP) to automatically determine the number of regulatory modules.
- SNMNMF: Employs Sparse Nonnegative Matrix Factorization to decompose expression matrices into interpretable biological components.
- Mirsynergy: A deterministic, two-stage greedy clustering approach designed to find overlapping modules of miRNAs and their target genes.
Results and Discussion
A recurring theme in genomic regulation is that miRNAs and TFs rarely act alone. Models that identify synergistic "modules" are significantly more accurate than those looking for isolated sites.
In comparative tests, algorithm choice significantly impacts performance. Mirsynergy was found to be orders of magnitude faster than SNMNMF (2 hours vs. >1 day) while identifying modules with higher functional enrichment. Similarly, GroupMiR demonstrated remarkable robustness, maintaining over 90% accuracy in predicting miRNA-mRNA memberships even when noise levels reached 80%, significantly outperforming older models like GenMiR++.
Furthermore, Nonnegative Matrix Factorization (NMF) has proven particularly suited for genomics. Unlike Principal Component Analysis (PCA), NMF's non-negativity constraint mirrors biological reality: a gene is either part of a regulatory module or it isn't, avoiding the interpretability issues of negative weights.
Conclusion and Future Directions
While unsupervised learning has provided deep insights into the genome, limitations remain. Most current models treat genomic regions as a "bag of sites," ignoring the 3D spatial arrangement of the genome. Additionally, the computational complexity of some Bayesian methods, such as the $O(N^3)$ complexity of the Indian Buffet Process, limits scalability. Future work must focus on incorporating spatial information and developing dynamic models to track how regulatory networks change in real-time during disease progression or treatment.
Summary updated