🧬 biology

Findings from Sparse Autoencoders for DNA Sequence Models: Motif Detectors, Reading-Frame Features, and the Scarcity of Regulatory Logic

This study demonstrates that while sparse autoencoders effectively extract monosemantic, biologically interpretable features like sequence motifs and reading frames from DNA foundation models, they reveal a significant scarcity of features encoding complex regulatory logic, suggesting current models capture a genomic dictionary rather than a grammar engine.

Original authors: Olivia Denvis

Published 2026-07-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Olivia Denvis

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you have a super-smart robot that has read every book in the world, but instead of human language, it only knows the secret code of life: DNA. This code is written in just four letters (A, C, G, and T), and it tells our bodies how to build cells, fight diseases, and even how to grow. Scientists have built massive computer programs, called "foundation models," that have studied this DNA code so thoroughly that they can predict how genes work better than ever before. But here's the mystery: we know these robots are smart, but we don't know how they think. It's like having a genius who can solve any math problem but refuses to show their work. To peek inside their brains, researchers use a special tool called a "Sparse Autoencoder" (or SAE). Think of an SAE as a high-tech translator that takes the robot's messy, tangled internal thoughts and breaks them down into simple, single ideas. Instead of a chaotic jumble of signals, the SAE tries to find neat, individual "switches" that light up for just one specific thing, like a switch that only turns on when it sees a "start" signal.

The big question scientists are asking is: Do these DNA-reading robots just memorize a dictionary of common words, or have they actually learned the complex grammar rules that govern how life works? In human language, grammar is what lets us understand that "The dog bit the man" is different from "The man bit the dog." In DNA, the "grammar" is the complex arrangement of signals that tells a gene when to turn on, when to stop, and how to interact with other genes. If the robots only have a dictionary, they can recognize words but might miss the deeper meaning. If they have learned the grammar, they truly understand the instructions of life. This paper dives deep into the brains of two of the smartest DNA-reading robots to see if they are just word-memorizers or if they have cracked the code of biological grammar.

The Dictionary vs. The Grammar Engine

In this study, researchers took two very different DNA-reading robots—one that uses a "attention" method (like a human scanning a page) and another that uses a "long-convolution" method (like a sliding window)—and ran them through a special SAE translator. They wanted to see what kind of ideas these robots had stored in their brains.

The results were a bit surprising, and they tell a story of a robot that is excellent at one thing but surprisingly weak at another.

The Robot's "Dictionary" is Amazing
First, the good news: The robots are incredible at recognizing the basic "words" of DNA. When the researchers looked at the SAE's output, they found that the robots had developed very clear, single-purpose switches for specific DNA patterns.

Motif Detectors: About 24% of the robot's active switches were dedicated to spotting specific, famous DNA "words." For example, one switch would light up only when it saw the "start" code (ATG), another only for "stop" codes, and others for specific signals that tell the cell where to cut and paste DNA (splice sites).
Composition Checkers: Another 12% of the switches were like quality control inspectors, checking the local "flavor" of the DNA, such as how much G and C (guanine and cytosine) were in a specific area.
Reading Frame: A small but cool group of switches (about 5%) acted like a rhythm counter. They could tell if the DNA was being read in the correct "triplet" pattern (like counting 1-2-3, 1-2-3) which is essential for making proteins.

The researchers found that these "dictionary" features were much clearer than the robot's raw internal signals. In fact, at the most interesting layer of the robot's brain (around 60% of the way through its processing), 62% of the SAE features could be clearly labeled with a specific meaning, compared to only 18% of the robot's original, messy neurons. It's like taking a blurry photo and suddenly finding that 62% of the pixels are now sharp, recognizable objects.

The Missing "Grammar Engine"
Here is the twist: While the robots are great at recognizing individual words, they seem to have almost no idea how to put those words together into complex sentences. The researchers looked for "regulatory logic"—the complex rules that say things like, "Turn on this gene only if you see Motif A and Motif B, but only if they are spaced exactly 10 letters apart."

The search was disappointing. The researchers found that only 1.4% of the robot's features seemed to do this kind of complex, conditional thinking. Most of the time, when a robot seemed to be reacting to a combination of signals, it turned out to just be reacting to one of the signals strongly, not the combination itself. The robots had a massive, organized dictionary of DNA words, but they hadn't built the grammar engine to understand the rules of the sentence.

Does the Robot Actually Use These Switches?
You might wonder, "If the robot has these switches, does it actually use them to do its job?" To test this, the researchers played a game of "remove and see." They took the specific switches responsible for recognizing splice sites (the cut-and-paste signals) and turned them off. The result? The robot's ability to predict splice sites crashed from a score of 0.89 down to 0.71. When they turned off random switches instead, the robot's performance barely changed. This proved that these "dictionary" features aren't just accidental noise; the robot genuinely relies on them to do its work.

The Same Story, Different Robots
The researchers also checked if this was just a fluke of one specific robot. They compared the "dictionary" of the attention-based robot with the long-convolution robot and a third one. They found that the simple "word" detectors (like the start codon or the TATA box) were almost identical across all three robots. However, the few complex "grammar" features that did exist were totally different in each robot and didn't match up at all. This suggests that while all DNA robots learn the same basic vocabulary, the way they try to handle complex grammar is messy and inconsistent.

The Bottom Line

The study concludes that current DNA foundation models are like brilliant librarians who have memorized every word in the library but haven't quite learned how to write a novel. They have a highly organized, cross-architecture "dictionary" of local DNA patterns—motifs, boundaries, and rhythms—that is far more interpretable than their raw internal signals. However, they struggle to encode the complex "regulatory logic" or grammar that governs how genes interact.

The authors suggest two possibilities: either the robots are doing complex logic, but it's hidden in a way that this specific tool (the SAE) can't see, or the robots are simply leaning heavily on their dictionary and shallow patterns to get the job done. For now, the evidence suggests we have a genomic dictionary, but we are still waiting for the genomic grammar engine.

Technical Summary: Findings from Sparse Autoencoders for DNA Sequence Models

Problem Statement
Genomic foundation models (e.g., Nucleotide Transformer, HyenaDNA) have achieved state-of-the-art performance on tasks such as splice-site recognition and variant effect prediction. However, the internal representations these models learn remain largely uncharted. A central question in mechanistic interpretability is whether these models encode a rich "dictionary" of local sequence patterns (motifs) or if they have learned the complex, conditional, and combinatorial "grammar" required for gene regulation. While Sparse Autoencoders (SAEs) have successfully decomposed language and protein models into human-interpretable features, their application to DNA sequence models has not been systematically explored.

Methodology
The authors conducted a deliberately narrow, focused study to characterize the internal features of two DNA foundation models with contrasting inductive biases:

Models: Nucleotide Transformer v2 500M (NT-500M), an attention-based encoder using 6-mer tokens, and HyenaDNA-medium-160k, a long-convolution decoder using single-nucleotide tokens. DNABERT-2 was used as a cross-architecture anchor.
SAE Training: Top- $K$ Sparse Autoencoders were fitted to the residual stream at a sweep of layers across both models. The dictionary size was set to a 16 $\times$ expansion of the model dimension, with $K$ (active latents) tuned to match model dimensions (e.g., $K=48$ for NT-500M).
Data: Activations were sampled from ~40M tokens of human genome (GRCh38), stratified to ensure representation of coding sequences, promoters, and regulatory regions.
Feature Interpretation Pipeline:
- Motif Recovery: Position Weight Matrices (PWMs) were derived from top-activating contexts and compared against the JASPAR motif database.
- Annotation Alignment: Latent activations were tested against curated genomic tracks (GENCODE splice sites, ENCODE TF peaks, CpG islands, etc.) using AUROC.
- Automatic Labeling: An LLM proposed labels based on top contexts and PWMs, validated by a predicted activation pattern matching held-out data (Spearman $\rho > 0.5$ ).
- Regulatory Logic Test: A conservative test defined "logic-like" features as those requiring a conjunction of motifs or context (e.g., Motif A AND Motif B) that could not be explained by a single additive predictor.
- Causal Ablation: Targeted features were zeroed out in the residual stream to measure the degradation of downstream probes (splice donor, CTCF binding, variant effect).

Key Contributions and Results

1. SAEs Reveal Monosemantic Features Superior to Neurons
SAE features are significantly more interpretable than raw neurons. At the most interpretable layer (Layer 17 of NT-500M):

Confident Labeling: 62% of alive SAE features received a confident label, compared to only 18% of raw neurons.
Genomic Alignment: The fraction of features aligning with a genomic annotation (AUROC > 0.8) rose from 5% (neurons) to 21% (SAEs).
Motif Matching: 27% of SAE features matched a JASPAR motif ( $q < 0.05$ ), versus 6% for neurons.

2. Interpretability is Layer-Dependent
Interpretability follows a distinct "hump" across the network depth, peaking at approximately 0.6 relative depth (Layer 17 for NT-500M, Layer 5 for HyenaDNA).

Early layers capture trivial token identity.
Middle layers contain the most biologically aligned features (splice sites, TF motifs, reading frames).
Final layers become diffuse, task-oriented, and less interpretable, with higher rates of dead latents.

3. The Recovered Vocabulary is Dominated by Local Grammar
The taxonomy of recovered features is dominated by local, largely context-free properties:

Single-Motif Detectors (23.8%): Features firing selectively on start codons (ATG), splice donors/acceptors (GT/AG), poly-A signals, TATA boxes, and specific TF motifs (CTCF, GATA).
Composition Features (12.1%): Tracking GC/CpG content and CpG island membership.
Positional/Boundary Features (8.6%): Detecting exon/intron boundaries and TSS proximity.
Reading-Frame Features (5.2%): A small but striking class detecting the period-3 periodicity of coding sequences (codon usage), which is a computational property rather than a fixed motif.
Repeats (7.4%): Detecting low-complexity regions.

4. Regulatory Logic is Scarce and Weak
The study's primary negative finding is the scarcity of features encoding regulatory logic (combinatorial or context-conditional computation).

Only 1.4% of alive features qualified as "logic-like" under the conditional-selectivity test.
Candidate combinatorial features often dissolved upon scrutiny, reducing to broadened single-motif detectors.
Genuine logic-like features had small interaction effect sizes and were among the least reliably reconstructed.

5. Cross-Model Universality and Causal Relevance

Universality: Simple motif and composition features (e.g., GC content, splice sites) transfer well across architectures (NT-500M, HyenaDNA, DNABERT-2). Combinatorial and reading-frame features do not transfer, suggesting they are model-specific or not robustly isolated.
Causal Ablation: Ablating the top interpretable features for a specific task (e.g., splice donors) caused significant performance drops (e.g., F1 score dropped from 0.89 to 0.71), whereas ablating random features had negligible impact. This confirms the features are causally load-bearing.

Significance and Claims
The paper concludes that current DNA foundation models, when probed via SAEs, recover an organized genomic dictionary rather than a genomic grammar engine.

The models excel at representing a rich catalogue of local sequence patterns and structural regularities (motifs, composition, periodicity).
They appear to encode very little of the conditional, combinatorial logic that governs gene regulation (e.g., motif spacing, orientation, and context-dependence) in a monosemantic, per-layer form.
This finding mirrors recent results in single-cell foundation models, suggesting a broader pattern where biological foundation models capture organized biological knowledge but lack unique regulatory circuitry in their interpretable features.
The authors posit two possibilities: either the regulatory logic exists in a form invisible to single-layer SAEs (distributed, attention-mediated, or non-linear), or current models rely heavily on a motif dictionary and shallow composition. The study emphasizes that reconstruction quality alone is insufficient; grounding in independent labels and causal tests is essential for valid interpretability claims.

The Dictionary vs. The Grammar Engine

The Bottom Line

Technical Summary: Findings from Sparse Autoencoders for DNA Sequence Models

More like this