[Paper Review] A mechanistically interpretable neural network for regulatory genomics

Alex M. Tseng, Gökçen Eraslan+2 more|arXiv (Cornell University)|Oct 8, 2024

Gene Regulatory Network AnalysisBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper introduces ARGMINN, a mechanistically interpretable deep neural network for regulatory genomics that directly encodes motifs and their syntactic relationships in learnable weights and activations. By combining a filter regularization to enforce non-redundant motif learning and a modified attention mechanism for instant motif instance detection, ARGMINN enables fully interpretable motif discovery, robustness to sequence variation, and generation of novel functional sequences—all while maintaining full expressivity for complex regulatory syntax.

ABSTRACT

Deep neural networks excel in mapping genomic DNA sequences to associated readouts (e.g., protein-DNA binding). Beyond prediction, the goal of these networks is to reveal to scientists the underlying motifs (and their syntax) which drive genome regulation. Traditional methods that extract motifs from convolutional filters suffer from the uninterpretable dispersion of information across filters and layers. Other methods which rely on importance scores can be unstable and unreliable. Instead, we designed a novel mechanistically interpretable architecture for regulatory genomics, where motifs and their syntax are directly encoded and readable from the learned weights and activations. We provide theoretical and empirical evidence of our architecture's full expressivity, while still being highly interpretable. Through several experiments, we show that our architecture excels in de novo motif discovery and motif instance calling, is robust to variable sequence contexts, and enables fully interpretable generation of novel functional sequences.

Motivation & Objective

• To address the limitations of existing deep learning models in regulatory genomics, where motif discovery is hindered by distributed information across filters and unreliable importance scores.
• To develop a neural network architecture that directly encodes biologically meaningful motifs and their syntactic configurations in interpretable, readable weights and activations.
• To ensure the model maintains full expressivity for complex regulatory syntax while enabling mechanistic interpretability without post-hoc analysis pipelines.
• To enable robust motif discovery and instance calling across variable sequence contexts, including adversarial perturbations.
• To demonstrate the novel capability of generating functionally valid, interpretable DNA sequences based on learned motif patterns.

Proposed method

• Proposes a novel architecture, ARGMINN, which integrates a filter regularization to enforce non-redundant motif learning in the first-layer convolutional filters.
• Introduces a modified self-attention mechanism that directly reveals motif instances and their syntactic relationships (e.g., spacing, orientation) in a single forward pass.
• Uses a loss function combining filter-overlap regularization and L1 sparsity to promote distinct, biologically relevant motif representations.
• Employs TOMTOM to align discovered motifs with known position weight matrices (PWMs), enabling quantitative evaluation of motif interpretability.
• Utilizes multiple oracles—including ARGMINN itself, Borzoi, and fine-tuned Enformer—to validate in silico the functionality of generated sequences.
• Employs a simulated REST dataset with strict syntactic rules (e.g., left-right motif order with 10 bp spacing) to test the model’s ability to learn and represent complex regulatory syntax.

Figure 1: Schematic of the ARGMINN architecture. a) The motif-scanner module produces activations denoting which motifs were found at each position, where activation magnitude reflects match strength. The activations are passed to the syntax builder, which learns higher-order logic between motif ins

Experimental results

Research questions

• RQ1 Can a deep neural network architecture be designed such that motifs and their syntactic relationships are directly readable from learned weights and activations, without relying on post-hoc interpretation?
• RQ2 Does the proposed architecture maintain full expressivity for complex regulatory motifs and syntactic rules, including non-trivial spacing and orientation dependencies?
• RQ3 How robust is the model’s motif discovery and instance calling to natural and adversarial sequence variations?
• RQ4 Can the model generate novel, functional DNA sequences that are interpretable and biologically plausible based on learned motif patterns?
• RQ5 How does the model’s interpretability compare to existing methods in terms of stability, accuracy, and biological relevance?

Key findings

• ARGMINN achieves superior motif discovery and motif-instance/syntax analysis compared to existing methods, with high accuracy in identifying true motifs and their configurations.
• The model demonstrates robustness to sequence perturbations, including dinucleotide shuffling and adversarial edits, maintaining accurate motif detection.
• ARGMINN enables fully interpretable sequence generation: 100 novel sequences were generated per motif pattern, and all were predicted to be functional by multiple independent oracles (ARGMINN, Borzoi, and fine-tuned Enformer).
• In the simulated REST dataset, ARGMINN successfully learned the precise binding syntax (left-right motif order with 10 bp spacing), outperforming baseline models in syntactic fidelity.
• The filter regularization and attention mechanism jointly ensure that motifs are non-redundant and directly interpretable, with no need for complex post-hoc pipelines.
• ARGMINN’s loss-weight robustness was validated across 100 training runs with varying regularization weights, showing consistent motif discovery performance across all settings.

Figure 2: Motif discovery. a) Example of SPI1 motifs discovered by ARGMINN, compared to interpreting the first-layer filters of a standard CNN, using ExplaiNN, and by clustering DeepLIFTShap importance scores using MoDISco. Note that MoDISco combines forward and reverse-complement orientations. b) F