Interpretability
When do a CNN's first filters learn whole DNA motifs?
Open access · cc by · source: Europe PMC
CNNs only learn whole, readable DNA motifs in their first layer when the architecture stops deeper layers from assembling motifs out of pieces.
Study at a glance
- Design
- Computational / modelling — Systematic architecture sweep of CNNs (varying max-pool size, filter size, filter count) on synthetic sequences with planted motifs, then on in vivo ChIP-seq sequences.
- N
- 25,000 synthetic sequences (split into train/validation/test); in vivo set of 270,382 training and 23,768 test sequences.
- Population
- Synthetic 200-nucleotide DNA sequences with 12 embedded transcription-factor motifs; human genomic sequences from the DeepSEA/ENCODE dataset
- Outcome
- Fraction of first-layer filters matching ground-truth motifs; classification AU-ROC
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
Networks with small first-layer max-pooling learned scattered partial motifs (at best a fraction of 0.127 of filters matched a true motif), while large pooling pushed nearly all filters to learn whole motifs (up to a fraction of 0.980 with a pool size of 25), even though classification accuracy was similar. When pooling was absent, the second layer assembled the whole motifs instead. Matching filters to a motif database was misleading, because many filters matched some database motif without matching the true one. On real data, the distributed-representation networks performed slightly better.
Methodology
The authors planted known transcription-factor motifs into random DNA so the correct answer was known, then trained small convolutional networks that differed only in max-pool size, filter size or filter count. They compared each first-layer filter against the planted motifs to measure how many had learned a whole motif. They then repeated the comparison on real ChIP-seq sequences.
Limitations
The synthetic task assumes binding is fully explained by a single motif-like pattern, which is much simpler than real regulatory DNA. The study only varies architecture and does not test how regularisation or optimisation choices shape representations. For real sequences there is no ground truth, so filter quality could only be compared to known motifs. Visualising a filter also says nothing about how much the network actually relies on it.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Same accuracy, very different internal representations.
What a network's filters represent depends on design choices, not only on the data: with small first-layer pooling only up to 0.127 of filters matched a true motif, versus up to 0.980 with large pooling, at similar accuracy.
Evidence for the claim as stated.
Related papers in this topic
Same topic cluster — not a recommendation engine.