TPCAV: Interpreting deep learning genomics models via concept attribution

This paper introduces TPCAV, a novel concept attribution framework that adapts and enhances the TCAV method with PCA-based decorrelation to provide robust, input-agnostic interpretations of diverse genomics deep learning models, enabling the analysis of both traditional DNA motifs and complex biological features like chromatin states and repetitive elements.

The Gist

Imagine you have a super-smart robot that can read the "instruction manual" of life (our DNA) and predict how a cell will behave. This robot is a Deep Learning Model. It's incredibly powerful, but it's also a "black box." You feed it data, and it gives you an answer, but it doesn't tell you why it made that decision. It's like asking a genius chef why a soup tastes so good, and they just say, "I just know."

Scientists have tried to open this black box before, but they mostly only looked at the raw ingredients (the basic A, C, G, T letters of DNA). They missed the bigger picture, like how the ingredients are stored (chromatin states) or if there are repeated patterns in the recipe (genomic repeats).

This paper introduces a new tool called TPCAV to help us understand what this robot is actually thinking. Here is how it works, using some simple analogies:

1. The Problem: The Robot's "Language" is Too Messy

Imagine the robot's brain is a giant library where every book represents a piece of DNA. But in this library, many books are identical copies, and others are just slight variations of the same story. If you ask the robot, "Which book made you decide this?" it gets confused because the books are all jumbled together.

In technical terms, the data the robot uses is correlated and redundant. It's like trying to find a specific flavor in a smoothie where the strawberries, raspberries, and red dye all taste exactly the same.

2. The Solution: TPCAV (The "Organizer")

The authors created TPCAV (Testing with PCA-projected Concept Activation Vectors). Think of this as a super-organized librarian who steps in to fix the mess.

• The "Concept" Idea: Instead of asking the robot about individual letters, we ask it about concepts. A "concept" is like a theme. For example, "Repetitive Elements" (like a chorus in a song that repeats) or "Chromatin State" (how tightly the DNA is packed, like a rolled-up scroll vs. an open book).
• The "PCA" Magic: The "PCA" part of the name is the librarian's special trick. It takes all those jumbled, duplicate books and sorts them out. It removes the noise and the copies, leaving only the unique, distinct ideas. It's like taking a messy pile of tangled headphones and neatly winding them up so you can see exactly which wire is which.

3. What TPCAV Actually Does

Once the librarian has organized the library, TPCAV does two cool things:

• It checks the "Why": It can tell you, "Hey, the robot made this prediction because it noticed a lot of 'Repetitive Elements' in this section," or "It was influenced by the 'Chromatin State' being loose here." It connects the robot's decision to real-world biological ideas, not just raw code.
• It finds the "Spotlight": It doesn't just say what influenced the robot; it points to where in the DNA that influence happened. It's like a spotlight shining on the specific paragraph in the instruction manual that made the chef decide to add salt.

4. Why This is a Big Deal

Before this, scientists could only understand the robot if it was looking at simple DNA letters. But TPCAV is like a universal translator. It works even when the robot is looking at:

• Complex DNA structures.
• "Foundation models" (robots trained on massive amounts of data, like how humans learn by reading the whole library).
• Signals from the environment (like chemical tags on the DNA).

The Bottom Line

Think of TPCAV as a translator and a detective combined. It translates the robot's complex, messy internal thoughts into clear, human-understandable biological concepts. It helps scientists stop guessing why the robot made a prediction and start understanding the actual biological rules the robot has learned. This allows researchers to discover new ways our genes work, leading to better treatments and a deeper understanding of life itself.

Technical Summary

1. Problem Statement

The interpretation of deep learning models in genomics faces significant limitations with current state-of-the-art methods:

• Input Restriction: Existing feature attribution techniques (e.g., saliency maps, TF-MoDISco) are primarily designed for one-hot encoded DNA sequences. They struggle to interpret models that utilize more complex, general genomic features such as chromatin states, repetitive elements, or tokenized embeddings from foundation models.
• Lack of Global Context: While concept-based methods (like TCAV) offer an input-agnostic framework for global interpretation, they have not been systematically adapted or validated for the specific challenges of genomics deep learning, particularly regarding the high correlation and redundancy inherent in genomic feature embeddings.

2. Methodology: TPCAV

The authors propose TPCAV (Testing with PCA-projected Concept Activation Vectors), an adaptation of the original TCAV method tailored for genomics. The methodology consists of three core technical innovations:

• Concept Activation Vectors (CAVs): The method defines "concepts" (e.g., specific chromatin states, repeat families, or transcription factor motifs) as directions in the model's internal activation space. It trains linear classifiers to distinguish between activations associated with a specific concept versus a baseline.
• PCA-Based Decorrelation: A critical improvement over standard TCAV is the integration of a Principal Component Analysis (PCA) transformation. Genomic embeddings often suffer from high multicollinearity and redundancy. TPCAV applies PCA to decorrelate these embedding features before computing the CAVs, ensuring that the attribution scores reflect distinct biological signals rather than statistical artifacts of correlated inputs.
• Concept-Specific Attribution Maps: The authors introduce a novel strategy to generate input-level attribution maps. This allows the model to highlight specific regions in the input sequence that are most responsible for a prediction based on a specific biological concept, bridging the gap between global concept scores and local feature importance.

3. Key Contributions

• First Application in Genomics: This work represents the first systematic application of concept attribution to interpret deep learning models in the genomics domain.
• Algorithmic Improvement: The development of the PCA-projected variant (TPCAV) specifically addresses the issue of correlated features in genomic embeddings, a problem often overlooked in standard TCAV implementations.
• Input Agnosticism: Unlike traditional methods, TPCAV is not restricted to one-hot DNA. It is designed to interpret:
  • Tokenized foundation models (e.g., DNA language models).
  • Models incorporating multi-modal inputs (e.g., DNA sequence + chromatin signals).
  • Models predicting diverse tasks (e.g., TF binding, gene expression).
• Downstream Utility: The method provides a mechanism to identify representative genomic regions associated with specific concepts, facilitating the discovery of distinct regulatory mechanisms.

4. Results and Evaluation

The authors evaluated TPCAV across a diverse set of genomics models and tasks:

• Benchmarking against TF-MoDISco: On one-hot encoded DNA models for transcription factor (TF) binding prediction, TPCAV demonstrated comparable performance to TF-MoDISco in identifying influential motif features, validating its reliability for standard tasks.
• Generalization to Complex Features: TPCAV successfully interpreted models using repetitive elements and chromatin state annotations, areas where traditional feature attribution methods often fail or provide ambiguous results.
• Foundation Model Interpretation: The approach successfully generalized to interpret features learned by tokenized foundation models, proving its versatility in the era of large-scale genomic pre-training.
• Biological Insight: The method successfully identified representative regions linked to specific biological concepts, offering actionable hypotheses for downstream experimental investigation.

5. Significance

TPCAV represents a paradigm shift in genomic deep learning interpretability. By moving beyond simple sequence-based attribution to a concept-based, input-agnostic framework, it enables researchers to:

1. Understand "Black Box" Models: Decipher how complex models utilize non-sequence data (like chromatin accessibility) to make predictions.
2. Validate Biological Hypotheses: Systematically test whether specific biological concepts (e.g., a specific repeat family) drive model predictions.
3. Future-Proof Interpretation: Provide a robust tool that scales with the evolution of genomic models, from simple CNNs on one-hot DNA to large foundation models with multi-modal inputs.

In summary, TPCAV offers a flexible, robust, and mathematically rigorous complement to existing interpretation techniques, specifically solving the challenges posed by correlated genomic features and diverse input modalities.