Co-evolution based machine-learning for predicting functional interactions between human genes

Doron Stupp, Elad Sharon, Idit Bloch, Marinka Žitnik, Or Zuk, Yuval TabachNature Communications 2021 · DOI

Abstract: Over the next decade, more than a million eukaryotic species are expected to be fully sequenced. This has the potential to improve our understanding of genotype and phenotype crosstalk, gene function and interactions, and answer evolutionary questions. Here, we develop a machine-learning approach for utilizing phylogenetic profiles across 1154 eukaryotic species. This method integrates co-evolution across eukaryotic clades to predict functional interactions between human genes and the context for these interactions. We benchmark our approach showing a 14% performance increase (auROC) compared to previous methods. Using this approach, we predict functional annotations for less studied genes. We focus on DNA repair and verify that 9 of the top 50 predicted genes have been identified elsewhere, with others previously prioritized by high-throughput screens. Overall, our approach enables better annotation of function and functional interactions and facilitates the understanding of evolutionary processes underlying co-evolution.

Introduction

Despite the rapid advancement of the genomic revolution, a significant portion of the human genome remains poorly understood. Approximately 2,000 human genes—collectively referred to as the "Ignorome"—lack functional annotation. Traditional methods for predicting gene function often rely on phylogenetic profiling (PP). The underlying principle of PP is that genes involved in the same biological pathway tend to be co-inherited or co-lost throughout evolution.

However, classical PP methods typically treat the entire eukaryotic tree of life as a single, flat signal. This "one-size-fits-all" approach often fails to capture the nuances of evolution, such as lineage-specific gene losses or the emergence of new functional modules in specific clades. To address these limitations, we developed MLPP (Machine-Learning based Phylogenetic Profiling), a method that shifts the focus from global correlation to clade-specific co-evolution.

Methods

The MLPP framework utilizes a massive dataset of 1,154 eukaryotic proteomes sourced from UniProt and NCBI RefSeq. The core innovation lies in its feature engineering and modeling approach:

  1. Clade-wise Feature Engineering: Instead of a single global correlation score, we calculated the covariance of gene profiles within 49 distinct clades (e.g., Fungi, Nematodes, Mammals). These 49 values serve as the primary features for the machine learning model.
  2. Machine Learning Model: We employed a Random Forest algorithm (specifically LightGBM) to predict functional interactions. The model was trained on high-quality ground truth data from Reactome, KEGG, BioGrid, IntAct, and CORUM.
  3. Contextual Prediction: Beyond predicting simple interactions, the model identifies the biological context (e.g., DNA repair, metabolism, or the immune system) based on Reactome and Gene Ontology (GO) data.
  4. Interpretability: To understand the evolutionary drivers of these predictions, we utilized SHAP values, allowing us to identify which specific clades contribute most to the prediction of different biological pathways.

Key Results and Performance

MLPP demonstrates a significant improvement over traditional phylogenetic profiling methods. Benchmarking against state-of-the-art approaches showed a 14% increase in auROC (Area Under the Receiver Operating Characteristic curve). This performance gain is consistent across a wide variety of biological contexts, ranging from ancient metabolic processes to more recent immune system functions.

Figure 2: Performance benchmarking of MLPP. (A) ROC curves for functional interaction prediction. (B) auROC scores across various biological contexts comparing MLPP and NPP.

In a direct comparison across KEGG pathways, MLPP outperformed traditional methods in 77.5% of cases. The model is particularly effective at capturing signals in complex pathways where traditional global correlation is too noisy to provide clear insights.

Figure 3: Comparative performance across KEGG pathways. (A) Median percentile comparison showing MLPP outperforms NPP in 77.5% of cases. (B-G) Heatmaps illustrating prediction accuracy for specific metabolic pathways.

Insights into the Ignorome

Applying MLPP to the human "Ignorome" led to the prediction of functions for 238 previously uncharacterized genes. A focused analysis on DNA repair identified 50 top candidates. Validation of these results showed that 9 were already present in other databases, while several others were supported by recent high-throughput CRISPR screens.

Our analysis also revealed that different biological pathways rely on distinct evolutionary signals:

  • The "Parasite Signal": Parasitic species, which often undergo massive, modular gene loss, provide high-contrast signals that are exceptionally useful for identifying functionally linked gene modules.
  • Evolutionary Stratification: Ancient pathways (like metabolism) derive their strongest signals from distant clades like Fungi, whereas newer pathways (like the immune system) rely on clades closer to humans, such as Metazoa.

Discussion and Limitations

While MLPP represents a major step forward, it faces challenges with "young genes"—those that appeared recently in evolutionary history (e.g., only in Chordates). These genes lack the deep evolutionary history required to generate a strong co-evolutionary signal. Additionally, as a supervised method, the model may be biased toward well-studied biological processes.

Future work will explore unsupervised machine learning techniques, such as autoencoders, to identify patterns in the Ignorome that do not fit into current functional categories. To facilitate further research, we have provided a public webserver at https://mlpp.cs.huji.ac.il.

Summary updated

Abstract

Over the next decade, more than a million eukaryotic species are expected to be fully sequenced. This has the potential to improve our understanding of genotype and phenotype crosstalk, gene function and interactions, and answer evolutionary questions. Here, we develop a machine-learning approach for utilizing phylogenetic profiles across 1154 eukaryotic species. This method integrates co-evolution across eukaryotic clades to predict functional interactions between human genes and the context for these interactions. We benchmark our approach showing a 14% performance increase (auROC) compared to previous methods. Using this approach, we predict functional annotations for less studied genes. We focus on DNA repair and verify that 9 of the top 50 predicted genes have been identified elsewhere, with others previously prioritized by high-throughput screens. Overall, our approach enables better annotation of function and functional interactions and facilitates the understanding of evolutionary processes underlying co-evolution. The manuscript is accompanied by a webserver available at: https://mlpp.cs.huji.ac.il .
BibTeX
@inproceedings{stupp2021co,
  author = {Doron Stupp and Elad Sharon and Idit Bloch and Marinka Žitnik and Or Zuk and Yuval Tabach},
  title = {{Co-evolution based machine-learning for predicting functional interactions between human genes}},
  booktitle = {Nature Communications},
  year = {2021},
  doi = {10.1038/s41467-021-26792-w},
}
Figure