Deep Learning Models for Predicting Phenotypic Traits and Diseases from Omics Data

Md. Mohaiminul Islam, Yang Wang, Pingzhao HuInTech 2018 · DOI · OpenAlex

Abstract

Computational analysis of high-throughput omics data, such as gene expressions, copy number alterations and DNA methylation (DNAm), has become popular in disease studies in recent decades because such analyses can be very helpful to predict whether a patient has certain disease or its subtypes. However, due to the high-dimensional nature of the data sets with hundreds of thousands of variables and very small number of samples, traditional machine learning approaches, such as support vector machines (SVMs) and random forests, have limitations to analyze these data efficiently. In this chapter, we reviewed the progress in applying deep learning algorithms to solve some biological questions. The focus is on potential software tools and public data sources for the tasks. Particularly, we show some case studies using deep neural network (DNN) models for classifying molecular subtypes of breast cancer and DNN-based regression models to account for interindividual variation in triglyceride concentrations measured at different visits of peripheral blood samples using DNAm profiles. We show that integration of multi-omics profiles into DNN-based learning methods could improve the prediction of the molecular subtypes of breast cancer. We also demonstrate the superiority of our proposed DNN models over the SVM model for predicting triglyceride concentrations.

Introduction: The Challenge of High-Dimensional Omics Data

Modern high-throughput technologies generate massive "omics" datasets—genomics, transcriptomics, and proteomics—that contain hundreds of thousands of variables. However, these datasets often suffer from a "small $n$, large $p$" problem: they have a very small number of samples (patients) relative to the massive number of features (genes or mutations).

Traditional machine learning (ML) approaches, such as Support Vector Machines (SVM) and Random Forests (RF), struggle with this high dimensionality. They often require labor-intensive, manual "feature engineering" by domain experts to filter out noise. There is a critical need for more robust, automated frameworks that can handle raw, noisy biological data to provide accurate disease classifications and trait predictions.

Overview of the deep learning workflow for biological data analysis.

Deep Learning vs. Conventional Machine Learning

The primary advantage of deep learning (DL) in this context is its ability to perform automatic feature extraction. Instead of researchers deciding which genes are important, the hidden layers of a Deep Neural Network (DNN) automatically learn hierarchical representations of the raw data.

Comparison of characteristics between conventional machine learning and deep learning.

As shown in the comparison above, while conventional ML is easier to interpret and faster to train, DL excels at handling high-throughput data directly and solving problems in an end-to-end fashion.

Methodological Approaches

The authors review and implement several architectures tailored for biological data:

  • Deep Convolutional Neural Networks (DCNN): Applied to 1D genomic data (like Copy Number Alterations). Convolutional layers capture spatial correlations between neighboring genes, respecting biological proximity.
  • Multi-Branch Architectures: Designed to integrate "multi-omics" data by processing different data types (e.g., gene expression and CNA) in parallel branches before concatenating them for a unified prediction.
  • DNN Regression: Fully connected architectures used to predict continuous phenotypic values, such as blood triglyceride levels, from high-dimensional DNA methylation profiles.
  • Regularization: Techniques like Dropout and DropConnect are essential to prevent overfitting on small patient cohorts by ensuring the model learns generalized patterns.

Case Studies and Key Results

The efficacy of these models was tested on several public datasets:

  1. Breast Cancer Subtyping (METABRIC): A DCNN achieved 84.1% accuracy (0.904 AUC), significantly outperforming SVM (76.5%) and Random Forest (82.7%).
  2. Multi-Omics Integration: Combining CNA and gene expression data yielded 79.2% accuracy for molecular subtyping, a substantial improvement over single-source models.
  3. Phenotypic Prediction (GOLDN): For predicting triglyceride concentrations from DNA methylation, the DNN regression model achieved a lower Root Mean Square Error (RMSE) of 88.5 compared to the SVM’s 90.3.

Conclusion and Future Directions

Deep learning offers a powerful solution to the "curse of dimensionality" in omics data. By leveraging multi-omics integration and automated feature extraction, these models provide superior predictive performance for complex clinical tasks. However, challenges remain regarding the "black box" nature of DL models and the computational resources required for training. Future work should focus on improving model interpretability and integrating even more heterogeneous data sources, such as clinical records and medical imaging, to support clinical decision-making.

Summary updated

Abstract

Computational analysis of high-throughput omics data, such as gene expressions, copy number alterations and DNA methylation (DNAm), has become popular in disease studies in recent decades because such analyses can be very helpful to predict whether a patient has certain disease or its subtypes. However, due to the high-dimensional nature of the data sets with hundreds of thousands of variables and very small number of samples, traditional machine learning approaches, such as support vector machines (SVMs) and random forests, have limitations to analyze these data efficiently. In this chapter, we reviewed the progress in applying deep learning algorithms to solve some biological questions. The focus is on potential software tools and public data sources for the tasks. Particularly, we show some case studies using deep neural network (DNN) models for classifying molecular subtypes of breast cancer and DNN-based regression models to account for interindividual variation in triglyceride concentrations measured at different visits of peripheral blood samples using DNAm profiles. We show that integration of multi-omics profiles into DNN-based learning methods could improve the prediction of the molecular subtypes of breast cancer. We also demonstrate the superiority of our proposed DNN models over the SVM model for predicting triglyceride concentrations.
BibTeX
@inproceedings{islam2018deep,
  author = {Md. Mohaiminul Islam and Yang Wang and Pingzhao Hu},
  title = {{Deep Learning Models for Predicting Phenotypic Traits and Diseases from Omics Data}},
  booktitle = {InTech},
  year = {2018},
  doi = {10.5772/intechopen.75311},
}
Figure