BGPT: Paper Review: TPCAV: Interpreting deep learning genomics models via concept attribution

Explore by Goal

Quick Explanation Copied

TPCAV review (PCA-stabilized concept attribution for genomics)

TPCAV adapts TCAV to genomics by projecting intermediate-layer embeddings through PCA before fitting concept classifiers, aiming to fix TCAV instability caused by correlated/redundant embedding features that can flip concept directions (sign). The paper then applies TPCAV to motif, repeat, and chromatin-accessibility concepts across multiple genomics architectures (one-hot TF binding models, tokenized DNABERT-2 foundation model, and multimodal maxATAC), and introduces a concept-specific attribution-map “dissection” strategy for multi-head models.

Primary source:

Long Explanation

TPCAV: Interpreting deep learning genomics models via concept attribution

Date (paper): Jan 21, 2026 Core idea: PCA-projected TCAV for more stable, genomics-relevant concept influence estimates .

1) What the paper claims (structured)

Problem framing: “Global concept attribution” is underused in genomics; TCAV-like methods exist but correlate/redundant embeddings can yield unintuitive sign errors .
Method change: TPCAV inserts a PCA step to decorrelate intermediate-layer activations before learning concept activation vectors (CAVs) .
Interpretation output: TPCAV ranks concepts using a “motif concept sensitivity” AUC-over-F-score procedure (as described in Results) and provides a concept-specific attribution map for multi-task/multi-head models .
Empirical scope: Evaluations include (i) TF binding prediction with one-hot sequences (Dream Challenge ENCODE models per TF/cell line) (ii) repeats/GC-rich proxies and TF motif concepts (iii) tokenized foundation models via DNABERT-2 fine-tuned for FOXA1 peaks (iv) multimodal maxATAC with explicit open-chromatin concepts, and (v) multi-headed BPNet (mouse ES) with a concept-specific peak dissection workflow .
Comparison points: The paper reports improved reliability vs TCAV for motif sign correctness, and states comparable performance to TF-MoDISco on one-hot TF binding models .

2) Visual map of the algorithmic workflow

Interpretation: This schematic corresponds to the paper’s description: TCAV-style linear concept separation in an intermediate embedding space, with TPCAV’s key modification being PCA decorrelation to mitigate correlated/redundant embedding features .

3) Evidence of the core fix: TCAV sign instability

The paper reports an explicit failure mode in TCAV on a specific example: the CTCF motif concept is found to contribute negatively to CTCF binding predictions in HepG2 cells, despite CTCF binding being “highly positively associated” with the CTCF cognate motif. The authors attribute this to multiple correlated embedding features that encode the motif with both positive and negative effects in different neurons, where TCAV’s linear concept classifier can emphasize negatively contributing correlated features, yielding a concept direction with the wrong sign .

After applying PCA-projected concept activation vectors, the paper states that the same cognate motif concepts are correctly identified as positive influences, increasing reliability relative to TCAV .

4) Concept sensitivity / motif dosage: how the ranking is operationalized

The paper introduces a motif concept sensitivity score defined as the AUC of the motif concept classifier’s F-score across different numbers of motif insertions within concept examples, enabling ranking of motif concepts by model responsiveness to increasing motif dosage .

Note: the plot intentionally avoids inventing F-score values (none are provided in the extracted text). It visualizes only the motif insertion counts mentioned for concept construction/evaluation .

5) Biological interpretation: what kinds of concepts TPCAV supports

5.1 Motifs (cis-regulatory sequence concepts)

The paper reports that TPCAV classifies cognate motif concepts with high F-scores and that the cognate motif ranks highly (often within top 10) for TF binding prediction models; it also notes exceptions where other GC-rich motifs can rank higher, interpreted as possible reliance on GC-content features in the learned model .

5.2 Repeats / annotation-based concepts

For repeat families, the paper reports that a low-complexity GC-rich class is the most frequently classifiable repeat concept across TF binding models, and it interprets this as a potential GC/“open chromatin proxy” effect in sequence-only settings. It also reports a T-rich simple repeat being positively influential in the FOXA1 model, consistent with prior claims that Forkhead TFs preferentially interact with T-rich simple repeats .

The cited mechanistic link to Forkhead/T-rich repeats is supported in the reference list, but the extracted text does not provide DOI coverage for those particular citations beyond what is listed in the provided reference table.

5.3 Chromatin accessibility concepts in multimodal models

In maxATAC models, the paper reports that the matched open-chromatin concept (built by pairing DNase/ATAC signal from peaks with random DNA sequence to separate sequence vs signal) yields the highest concept-classifier F-scores, and that GC-rich concepts shift to negative influence relative to sequence-only models—interpreted as a change in role of GC-content when explicit accessibility information is available .

5.4 Tokenized foundation models

The paper fine-tunes DNABERT-2 to predict FOXA1 ChIP-seq peaks in A549 and then uses TPCAV on activations from the pooler layer to test motif concepts; it also benchmarks multiple foundation models by motif concept classifier success across embedding spaces, reporting DNABERT-2 as outperforming Nucleotide Transformer and HyenaDNA in distinguishability of motif concepts (as framed in embedding linear separability terms) .

6) Critical appraisal (skeptical, evidence-weighted)

6.1 Strength: a plausible, targeted failure mode of TCAV

The paper’s PCA decorrelation motivation is mechanistically coherent: if multiple correlated embedding directions represent the same concept but with mixed signs, a linear classifier can overweight a negatively weighted correlated component, yielding an inverted CAV direction. Adding PCA changes the embedding basis to reduce correlation, aiming to make the hyperplane direction more stable .

6.2 Methodological sensitivity not fully quantified in the extracted text

The extracted Methods mention SVM classifier training, concept example sampling, standardization, and PCA via SVD; however, the reviewable excerpt does not report uncertainty intervals, ablation on PCA rank (how many PCs retained), or cross-seed stability metrics for CAV directions. Without those, it is hard to know whether PCA always reduces variance or sometimes changes concept geometry in ways that trade sign correctness for other forms of interpretive bias .

6.3 Concept-set dependence and “concept leakage” risks

TPCAV is inherently concept-set dependent: results depend on how concept examples are constructed (motif insertions into random genomic backgrounds; open-chromatin examples constructed by pairing signal with random sequence; repeat instances from RepeatMasker). This creates two known interpretability risks: (i) if the random background is not biologically appropriate, the embedding separability may reflect dataset construction artifacts rather than universal motif/annotation mechanisms; (ii) if the model uses correlated proxies (e.g., GC-content), “concept” signals can become partially proxy-based, especially in sequence-only settings where accessibility is absent. The paper explicitly observes GC-related phenomena in ranking and directionality differences across modalities, which supports that proxy use exists—but the excerpt does not show controls that guarantee conceptual purity beyond their construction scheme .

6.4 Comparison with TF-MoDISco is informative but not equivalent

The paper positions TPCAV as comparable to TF-MoDISco for motif interpretation on one-hot models, while also emphasizing a key methodological difference: TF-MoDISco discovers motifs de novo from attribution patterns, whereas TPCAV evaluates influence of known motif concepts defined ahead of time. Therefore, “agreement” between methods can mask different failure modes: a model may reflect the same motif but via different internal representations or may produce negative/positive attribution patterns that are interpreted differently by the clustering mechanism vs the global CAV direction criterion .

7) What would disprove / change the paper’s conclusions?

If PCA-projected concept directions fail to reduce sign errors in TCAV across multiple independent datasets/models (not just the exemplars), the key “stability improvement” claim would weaken .
If concept influence rankings remain dominated by GC-content (or other global sequence statistics) even after modality separation controls, then interpretive claims about specific motif/annotation mechanisms would be less convincing .
If the “concept-specific attribution maps” do not correlate with biologically meaningful subclasses of peaks (the paper reports pathway enrichments via GREAT as support, but extracted text doesn’t include quantitative uncertainty), then the dissection utility would need additional validation .

8) “Known vs inferred vs uncertain” checklist

Category	What the paper reports in the extracted text
Known (from paper)	TPCAV modifies TCAV by adding PCA decorrelation of intermediate embedding activations; it reports improved sign reliability for motif concepts compared with TCAV on a stated TF/cellline example .
Inferred	The PCA step reduces sign errors because correlated/redundant features existed at the chosen layer and PCA reduces correlation in that space .
Uncertain (not shown in excerpt)	How much stability gain generalizes across random seeds, concept sampling choices, PCA rank selection, and layer choice ℓ; the extracted text does not contain those uncertainty quantifications .

9) Practical takeaways for genomics model interpretation

If you see contradictory concept signs from TCAV-like analyses, PCA-decorrelation (TPCAV’s change) is a principled first remedy when embeddings are correlated/redundant .
When interpreting sequence-only models, treat GC-content-driven proxies as a serious confound candidate; the paper explicitly observes GC-rich repeat/motif effects and shows sign reversals when accessibility inputs are added .
For multi-head models, the concept-specific attribution map / dissection strategy can help translate global concept influence into region-level usage heterogeneity (as framed for BPNet peak subsets) .

Author reviews (browse on BGPT)

Feedback:

Updated: April 22, 2026