Rank Projection Trees for Multilevel Neural Network Interpretation

Jonathan Warrell, Hussein Mohsen, Mark GersteinarXiv 2018 · arXiv

Deep neural networks have revolutionized predictive modeling in genomics, yet they often remain "black boxes," making it difficult to extract the biological mechanisms driving their decisions. This paper, originally presented at the NeurIPS 2018 Workshop on Machine Learning for Health, introduces Rank Projection Trees (RPTs)—a multiscale framework designed to interpret neural networks by identifying prioritized, overlapping groups of features rather than isolated inputs.

The Interpretability Challenge in Biology

While standard interpretation methods like saliency maps or gradient-based rankings assign importance scores to individual features (e.g., single genes), biological systems operate through complex, hierarchical interactions known as epistasis. Traditional clustering methods often create rigid, non-overlapping groups that do not reflect the specific patterns a neural network has learned for a given trait.

The authors address this by seeking a method that can identify important feature sets at multiple scales—ranging from individual genes to broad biological pathways—directly relevant to the model's output.

Methods: Rank Projection Trees

The core innovation of this work is the projection of a hierarchical tree structure onto a pre-trained neural network. This process involves four main components:

  1. Mapping Function ($\phi$): A defined mapping connects nodes in a tree $T$ to specific nodes in the neural network $N$.
  2. Ranking Function: Using a chosen scoring metric (such as weights or gradients), the RPT identifies the "Top B" and "Bottom B" nodes at each layer to determine the tree's branching.
  3. Subsets of Influence: By tracing paths from the output down to the input layer, the method isolates subsets of inputs ($S^+$ and $S^-$) that contribute most significantly to the final prediction.
  4. Model Agnosticism: RPT acts as a wrapper, meaning it can be applied to virtually any existing scoring function to add hierarchical context.
Figure 1: The mapping function $\phi$ connects a hierarchical tree structure $T$ to the layers of a neural network $N$.

Experimental Validation

The researchers evaluated RPTs on two high-dimensional genomic datasets:

  • Pan-Cancer Analysis of Whole Genomes (PCAWG): Predicting the co-occurrence of germline and somatic mutations.
  • PsychENCODE: Classifying Schizophrenia versus healthy controls using transcriptomic data from the prefrontal cortex.

To validate the biological relevance of the identified gene sets, the authors used KEGG pathway enrichment and cross-referenced results with GWAS-linked genes and scientific literature citation counts.

Key Findings and Results

The Importance of the Penultimate Layer

The study found that gene groupings identified at the penultimate layer ($L_1$) were often the most trait-relevant. This suggests that as data moves through hidden layers, the network constructs increasingly complex and biologically meaningful "concepts."

Superior Enrichment

In the Schizophrenia model, RPT-derived gene groupings were significantly more enriched with high-confidence GWAS-linked genes compared to randomized trees. As shown in the cumulative distribution function (CDF) plots below, the Rank-projection method consistently outperforms randomized baselines across different layers, with the most pronounced separation often occurring in the layers closer to the output.

Figure 6: CDF plots showing the enrichment of GWAS-linked genes across different layers ($L_0$ to $L_3$). The Rank-projection method (red) shows significantly higher enrichment than the randomized baseline (blue).

Multiscale Discovery

Unlike traditional rigid clustering, RPT allows for overlapping groups, mirroring real-world biology where a single gene may participate in multiple functional pathways. The method successfully extracted statistically significant KEGG pathways that aligned with established disease mechanisms.

Conclusion and Future Directions

Rank Projection Trees provide a robust framework for moving beyond individual feature importance toward a more holistic, hierarchical understanding of neural network decisions. While the current implementation uses a tree structure, future work may explore more complex graph-theoretic approaches, such as Directed Acyclic Graphs (DAGs), to further capture the intricate web of biological interactions. Additionally, the authors suggest that this framework could be extended to other high-dimensional fields like climate modeling or finance where hierarchical feature interactions are prevalent.

Summary updated

Abstract

A variety of methods have been proposed for interpreting nodes in deep neural networks, which typically involve scoring nodes at lower layers with respect to their effects on the output of higher-layer nodes (where lower and higher layers are closer to the input and output layers, respectively). However, we may be interested in picking out a prioritized collection of subsets of the inputs across a range of scales according to their importance for an output node, and not simply a prioritized ranking across the inputs as singletons. Such a situation may arise in biological applications, for instance, where we are interested in epistatic effects between groups of genes in determining a trait of interest. Here, we outline a flexible framework which may be used to generate multiscale network interpretations, using any previously defined scoring function. We demonstrate the ability of our method to pick out biologically important genes and gene sets in the domains of cancer and psychiatric genomics.
BibTeX
@misc{warrell2018rank,
  author = {Jonathan Warrell and Hussein Mohsen and Mark Gerstein},
  title = {{Rank Projection Trees for Multilevel Neural Network Interpretation}},
  year = {2018},
  eprint = {1812.00172},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.00172},
}
Figure