Strategies to include prior knowledge in omics analysis with deep neural networks

Kisan Thapa, Meric Kinali, Shichao Pei, Augustin Luna, Özgün BaburPatterns 2025 · DOI

TL;DR

Deep learning models excel at finding patterns in complex biological data, but often overlook valuable prior biological knowledge. This paper reviews strategies for integrating this knowledge into deep learning architectures, particularly for omics analysis, to improve model accuracy, interpretability, and generalizability. The authors categorize these strategies into input transformation, architectural modification, and graph neural networks, offering a roadmap for leveraging biological insights in AI-driven biological discovery.

What Problem Does This Paper Address?

High-throughput molecular profiling technologies (like genomics, transcriptomics, proteomics) generate vast amounts of data, offering a holistic view of biological systems. However, analyzing this high-dimensional, interrelated data presents significant challenges, especially when dealing with limited sample sizes, which can lead to overfitting in machine learning models. While deep learning models are powerful tools for uncovering complex patterns, they typically rely solely on the raw data, ignoring rich, existing biological knowledge about molecular interactions and pathways. This oversight can limit their predictive performance and hinder biological interpretation. The paper addresses the critical need to bridge this gap by systematically reviewing methods that integrate prior biological knowledge into deep learning frameworks for omics analysis.

Methods

The authors categorize the strategies for incorporating prior knowledge into three main approaches:

  • Input Transformation: This involves preprocessing the omics data to incorporate biological knowledge before feeding it into a standard deep learning model.
  • Gene Set Enrichment Scores (GSEA): Methods like GSEA, ssGSEA, and NetPEA transform gene expression data into scores representing the enrichment of predefined gene sets (e.g., pathways). This reduces dimensionality and highlights biologically relevant patterns.
  • Omics-to-Image Conversion: Molecular profiles are converted into 2D images, where gene values are mapped to pixels. This allows the use of Convolutional Neural Networks (CNNs), which are adept at image analysis. Prior knowledge guides the image formation by mapping related genes to close pixels, creating regions that correspond to functional groups. Approaches include functional hierarchy mapping (using ontologies), PCA-guided mapping (using gene sets), and spectral mapping (using interaction networks).
  • Structuring the Neural Network Architecture: This strategy modifies the neural network's structure to mimic the organization of prior biological knowledge.
  • Gene Set Layers: The first hidden layer of a Multi-Layer Perceptron (MLP) or autoencoder can be designed to represent gene sets, with connections reflecting gene set membership. This allows the model to learn gene set-specific representations.
  • Hierarchical Ontologies: The network architecture can mirror hierarchical ontologies, with successive layers representing increasingly abstract biological terms.
  • Gene Layers: An intermediate "gene layer" can be inserted between input and gene set layers to explicitly model individual genes and their relationships to inputs and gene sets, enhancing interpretability.
  • Graph Neural Networks (GNNs): GNNs leverage graph structures (like protein-protein interaction networks, gene regulatory networks) as a guide. Nodes represent biological entities (e.g., genes), and edges represent relationships.
  • Graph Convolutional Networks (GCNs): These models use graph structure to aggregate information from neighboring nodes, learning embeddings that capture both node features and their local network context.
  • Graph Sample and Aggregate (GraphSAGE): A spatial-based GNN that aggregates neighbor information using various functions, scalable for large graphs.
  • Graph Attention Networks (GATs): Incorporate attention mechanisms to assign different weights to neighbors, allowing the model to learn the relative importance of connections. Multi-head attention further enhances this capability.
  • Graph Isomorphism Networks (GINs): Address limitations of GCNs by capturing topological features of nodes, not just aggregated features.

Datasets

The paper discusses the application of these methods to various omics datasets, including gene expression, copy-number alteration (CNV), DNA methylation, and mutations. Specific datasets mentioned include:

  • TCGA (The Cancer Genome Atlas) for lung cancer progression and survival prediction.
  • GDSC (Genomics of Drug Sensitivity in Cancer) dataset for drug sensitivity prediction.
  • Various cancer datasets for survival prediction and subtype classification.

The paper does not introduce new datasets but reviews existing methods applied to publicly available omics data.

Key Insights and Ideas

  • Prior knowledge is not just for regularization; it can fundamentally improve model performance and interpretability. By guiding the learning process, prior knowledge helps models focus on biologically relevant signals, leading to more accurate predictions.
  • Different forms of biological knowledge (gene sets, ontologies, interaction networks) can be integrated using distinct deep learning strategies. This flexibility allows researchers to choose the most appropriate approach based on the available knowledge and the specific biological question.
  • Omics-to-image conversion opens up the powerful domain of CNNs for omics data analysis. This strategy leverages the success of CNNs in image recognition by transforming complex biological data into a visual format.
  • GNNs are a natural fit for integrating network-based biological knowledge. Their ability to learn from graph structures makes them ideal for analyzing interaction networks and inferring relationships.
  • Interpretability is a major benefit of incorporating prior knowledge. By structuring models or transforming inputs based on biological insights, it becomes easier to understand why a model makes a certain prediction, fostering trust and enabling biological discovery.

Key Results

The paper is a review and does not present new experimental results. However, it highlights that the reviewed methods have been successfully applied to:

  • Predicting cancer subtypes and progression.
  • Predicting drug sensitivity and response.
  • Predicting patient survival.
  • Classifying cell types in single-cell transcriptomics.

The key qualitative result is the demonstration that integrating prior biological knowledge consistently improves the generalizability and interpretability of deep learning models in omics analysis, often leading to more robust and biologically meaningful insights compared to models trained solely on raw data.

Limitations and Future Work

  • Incompleteness of Biological Knowledge: The most significant limitation is that our understanding of molecular biology is far from complete. Missing or inaccurate prior knowledge can lead to suboptimal model performance or misinterpretations.
  • Knowledge Overload: Using too much prior knowledge (e.g., an excessive number of gene sets) can paradoxically increase the search space and hurt performance. Selecting relevant prior knowledge is crucial but often a manual and challenging task.
  • Interpretability Caveats: While prior knowledge integration enhances interpretability, it's important to note that deep learning models still learn by optimizing prediction performance. The identified "important" features or pathways should be considered as suggestions for further investigation rather than definitive biological mechanisms.
  • Future Work:
    • Leveraging More Detailed Models: As proteomics data becomes more abundant and detailed models (like directed signaling networks) are developed, future research will likely utilize these richer representations.
  • Single-Cell Omics: The rise of single-cell technologies offers new opportunities to study cell-to-cell heterogeneity and develop more nuanced predictive models.
  • Large Language Models (LLMs): LLMs may aid in curating and building more comprehensive biological knowledge bases, facilitating their integration into deep learning.
  • Heterogeneous Graphs: Developing GNN architectures that can effectively handle heterogeneous graphs (with different node and edge types) is an ongoing area of research.

Summary updated

BibTeX
@article{thapa2025strategies,
  author = {Kisan Thapa and Meric Kinali and Shichao Pei and Augustin Luna and Özgün Babur},
  title = {{Strategies to include prior knowledge in omics analysis with deep neural networks}},
  journal = {Patterns},
  year = {2025},
  doi = {10.1016/j.patter.2025.101235},
}
Figure