Book chapter · 2022
Statistical Relational Learning for Genomics Applications: A State-of-the-Art Review
In Handbook of Machine Learning Applications for Genomics, Studies in Big Data, pp. 31–42, Springer Nature Singapore Pte Ltd. · Published
Summary
What question does this chapter answer?
What statistical relational learning (SRL) models exist, and how can they be applied to the relational, noisy and incomplete data found in genomics?
What did the chapter find?
The chapter reviews the background of SRL (probabilistic graphical models, Bayesian networks, Markov networks) and the main SRL approaches — probabilistic relational models, relational dependency networks and relational Markov networks — and surveys their genomics applications such as annotating genomic sequence elements, handling missing microarray values, identifying drought- and disease-resistance genes in plants, GWAS and mining electronic health records. It identifies the computational complexity of inference, followed by graph size, as the most significant limitations shared by most SRL methods.
Why does it matter?
Genomic data is being generated faster than existing methods can analyze it, and annotating all sequences manually is unfeasible and costly, so the authors argue SRL methods will be essential to automatically annotate sequences and handle large genomic data sets. They expect demand for machine learning methods that can adapt to these big data sets to increase over the next decade.
Key findings
- The chapter explains that statistical relational learning (SRL) combines statistical and probabilistic modeling with structured, relational data representations, making it suited to genomics data that violates the i.i.d. assumption of most machine learning methods.
- The review classifies SRL models into three categories — relational graphical models, latent class models and tensor factorization models — and describes approaches including probabilistic relational models, stochastic logic programs, Bayesian logic programs, relational dependency networks, relational Markov networks and Markov logic networks.
- Probabilistic relational models extend Bayesian networks to relational domains, remove the i.i.d. assumption and have been successfully evaluated on genomic data, according to the review.
- Relational dependency networks can exploit autocorrelation (for example, in the functions of co-located proteins in a cell) and give easy-to-interpret models, while relational Markov networks can represent arbitrary autocorrelation but are impractical because of inefficient parameter estimation.
- The review identifies computational complexity of inference as probably the most significant limitation shared by most SRL methods, followed by graph size proportional to the number of attributes and objects, which limits scalability on realistic genomic datasets.
- Genomics applications of SRL surveyed include annotating genomic sequence elements, modeling missing gene expression microarray values, using inductive logic programming to identify drought-resistant genes in plants, identifying plant disease resistance genes, genome-wide association studies and mining electronic health records for drug discovery.
- The chapter notes that genomics is expected to generate the largest amounts of data by 2025 and that Europe's Million Genomes project, launched in 2018, will create large genomic data sets for which SRL methods will be important.
Source: Biba & Vajjhala (2022), In Handbook of Machine Learning Applications for Genomics, Studies in Big Data, pp. 31–42, Springer Nature Singapore Pte Ltd.. DOI: 10.1007/978-981-16-9158-4_3
Chapter at a glance
| Research question | What is the state of the art in statistical relational learning for genomics? |
|---|---|
| Design | State-of-the-art literature review (book chapter) |
| Data | Published literature on probabilistic graphical models, SRL approaches and genomics applications (66 references) |
| Methods | Narrative review of background models, SRL approaches, and SRL problems and applications in genomics |
| Main result | SRL suits non-i.i.d. relational genomic data, but inference complexity and graph size limit scalability |
| Implication | SRL methods will be essential for handling and annotating large genomic data sets |
| Citation | Biba & Vajjhala (2022) · DOI 10.1007/978-981-16-9158-4_3 |
Abstract
This paper aims to review the state-of-the-art statistical relational learning models (SRL) in genomics. SRL deals with machine learning and data mining in relational domains where observations may be missing, partially observed, and noisy. This chapter introduces a background overview of various models, including probabilistic graphical models, Bayesian networks, dependency networks, Markov networks, first-order logic, and probabilistic inductive logic programming. This chapter also discusses the various statistical relational learning approaches, including probabilistic relational models, stochastic logic programs, Bayesian logic programs, relational dependency networks, relational Markov networks, and Markov logic networks. Finally, the last part of the paper focuses on the practical application of statistical relational learning techniques in genomics. The chapter concludes with a discussion on the limitations of current methods.
Abstract as published in Handbook of Machine Learning Applications for Genomics.
Keywords: Genomic; Artificial intelligence; Machine learning; Probabilistic; Bayesian; Markov; Dependency; Genetics; Bioinformatics
Key terms
- Statistical relational learning (SRL)
- An area of machine learning that combines statistical and probabilistic modeling with languages supporting structured data representations, learning from relational data where observations may be missing, partially observed and noisy.
- Bayesian network
- A directed probabilistic graphical model representing the joint probability distribution over a set of random variables as the product of each variable's probability given its parents.
- Independent and identically distributed (i.i.d.)
- A property of random variables that share the same probability distribution and are mutually independent — an assumption most relational data, including genomic data, violates.
How to cite
Biba, M., & Vajjhala, N. R. (2022). Statistical Relational Learning for Genomics Applications: A State-of-the-Art Review. In S. S. Roy, Y.-H. Taguchi (Eds.), Handbook of Machine Learning Applications for Genomics (pp. 31–42). Springer Nature Singapore Pte Ltd.. https://doi.org/10.1007/978-981-16-9158-4_3
BibTeX
@incollection{biba2022statistical,
title = {Statistical Relational Learning for Genomics Applications: A State-of-the-Art Review},
author = {Biba, Marenglen and Vajjhala, Narasimha Rao},
booktitle = {Handbook of Machine Learning Applications for Genomics},
series = {Studies in Big Data},
editor = {S. S. Roy and Y.-H. Taguchi},
pages = {31--42},
year = {2022},
publisher = {Springer Nature Singapore Pte Ltd.},
doi = {10.1007/978-981-16-9158-4_3},
url = {https://doi.org/10.1007/978-981-16-9158-4_3}
}Related research