Integrating Biological Pathways into Genomic Machine Learning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2146 papers153 conceptsfrozen 2026-10-04

In traditional genomic analysis, researchers often treat a cell's gene expression profile as a flat vector of independent numbers. This "independence assumption" simplifies the math, but it ignores a fundamental biological truth: genes do not act in isolation. They operate within a dense web of regulatory circuits, signaling pathways, and protein-protein interactions.

A growing body of research is moving away from "black-box" models that treat genes as independent features. Instead, researchers are building biologically informed models that bake the known structure of biological networks directly into the architecture of neural networks and machine learning algorithms.

The Problem with the Independence Assumption

When a machine learning model treats 20,000 genes as independent variables, it faces a massive search space with very few samples—often referred to as the "curse of dimensionality." As Onur Dereli and colleagues note in their work on survival analysis, having thousands of features but only hundreds of patients makes models prone to picking up technical noise rather than biological signals.

Furthermore, models that ignore gene-to-gene relationships are difficult to interpret. If a model identifies a "gene signature" for a disease, but those genes have no known functional connection, clinicians cannot easily determine the underlying mechanism. As Syed Haider argues in the development of SIMMS, treating genes as independent features results in predictors that are biologically opaque and fail to identify actionable therapeutic targets.

Encoding Biology into Model Architecture

To solve this, researchers are using "prior knowledge"—existing databases of how genes interact—to constrain how models learn. This shift happens in three primary ways:

1. Graph-Based Constraints

Instead of a fully connected neural network where every input gene can influence every hidden neuron, researchers use partially connected layers. In the GOAE and GONN architectures, neurons are only connected if they share a relationship in the Gene Ontology (GO) database. This forces the model to learn representations that align with known biological processes. Similarly, Xu Shi developed CyNetSVM, which uses protein-protein interaction (PPI) networks to "smooth" the weights of a Support Vector Machine, ensuring that connected genes contribute to the prediction in a coordinated way.

2. Biosemantic Representations

Some frameworks transform 1D expression data into structured graphs before the learning even begins. Xiyue Cao and the team behind scPriorGraph convert raw single-cell data into "biosemantic" graphs. By integrating ligand-receptor interactions and pathway data, the model can identify rare cell types that might look like noise in a standard expression matrix. Jixiang Yu takes this further with scGCOT, using Optimal Transport to simultaneously model cell-cell and gene-gene relationships, capturing the complex "fingerprints" of biological systems.

3. Pathway-Centric Ensembles

Rather than one large model, researchers like Junjie Shen use Bayesian stacking to group genes into functional pathways first. Each pathway gets its own "mini-model," and a "Super Learner" then ensembles these pathway-specific predictions. This approach ensures that the model's final decision is grounded in the activity of specific biological modules rather than isolated, high-variance genes.

Why Prior Knowledge Works

The intuition behind these methods is that biology acts as a regularizer. By restricting the model's hypothesis space to only those patterns that are biologically plausible, we reduce overfitting.

For example, in the NetAP method developed by Jipeng Qiang and John Quackenbush, the algorithm can group two cancer patients together even if they don't share a single mutated gene. Because the model "knows" the underlying interaction network, it can see that the different mutations in both patients hit the same signaling pathway, revealing a shared disease subtype that a standard clustering algorithm would miss.

This structural awareness also improves cross-modal tasks. Ronald Xie and colleagues demonstrated this with BLEEP, which uses contrastive learning to align histology images with transcriptomic profiles. By preserving biological dependencies in the latent space, the model can "impute" molecular data for tissue images with high fidelity.

Moving Toward Mechanistic AI

The ultimate goal of this research direction is to move from "what" to "how." Models like VeloVAE, developed by Yichen Gu and Joshua D. Welch, combine deep learning with the physical constraints of biochemical differential equations. This allows researchers to reconstruct the "hidden" life history of cells, predicting not just their current state, but their future trajectory.

By treating the genome as a structured system rather than a random bag of features, these biologically informed frameworks are turning deep learning from a predictive tool into a discovery engine for the underlying mechanisms of life and disease.

Papers in this version

Showing 60 of 146 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 136 papers
  2. Version 1 · 2026-10-04 · 146 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_gene_independence_assumption_f021ae17,
  title = {gene independence assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/gene-independence-assumption/rdr_f021ae176bcf43dfb510918b08fa31cd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_integrating_biological_pathways_into_gen_f021ae17_v1,
  title = {Integrating Biological Pathways into Genomic Machine Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/integrating-biological-pathways-into-genomic-machine-learning/rdr_f021ae176bcf43dfb510918b08fa31cd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}