Pith. sign in

REVIEW 4 major objections 5 minor 72 references

Explainable AI in Genomics: Transcription Factor Binding Site Prediction with Mixture of Experts

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A weighted mixture of three frozen CNN experts predicts transcription-factor binding sites beyond its training factors, and its shift-averaged gradient attributions localize motifs that plain gradients miss.

desk verdict A readable engineering paper whose two headline claims rest on an invalid ANOVA and mostly visual attribution evidence; the core ideas deserve a serious revision rather than a desk reject. read the letter →

arxiv 2507.09754 v2 pith:WB454STM submitted 2025-07-13 cs.LG q-bio.GN

Aakash Tripathi , Ian E. Nielsen , Muhammad Umer , Ravi P. Ramachandran , Ghulam Rasool This is my paper · ORCID
classification cs.LGq-bio.GN
keywords transcriptionfactorbindingsitepredictionmixtureofexpertsconvolutionalneuralnetworksout-of-distributiongeneralizationShiftSmoothattributionmappinggenomicsstatisticalsignificancetesting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Transcription-factor binding site (TFBS) prediction models are usually trained for one transcription factor at a time and struggle on factors they have not seen. This paper tests a mixture-of-experts route: take three pre-trained DeepBind CNNs, freeze them, and let a small gating network learn how to combine their hidden representations for a given input. Their claim is that this ensemble is competitive on the training factors and superior to each individual expert on most out-of-distribution factors, with ANOVA-based statistical support, and that the gating network keeps the experts frozen so new experts could be added without retraining the earlier ones. For interpretability, the paper proposes ShiftSmooth, which averages gradients over circularly shifted copies of an input sequence, and reports that this finds motif nucleotides like the 'G' in 'GATAA' that Vanilla Gradient misses. If the claims hold, the practical payoff is a reusable library of frozen TF-specific experts plus a cheap learned router for new genomes, together with attributions that point at the actual binding motif.

What carries the argument

The load-bearing mechanism is the N:1 mixture of experts: three DeepBind CNN experts, each pre-trained on a different transcription factor (ARID3A, FOXM1, GATA3), are frozen and stripped of their final prediction layer, leaving 32-dimensional embeddings $e_i$. A gating network maps the concatenated embeddings $E=[e_1,e_2,e_3]$ through a linear layer and softmax to weights $\alpha = \operatorname{softmax}(EW_{\mathrm{gate}}+\beta_{\mathrm{gate}})$; the mixture output is $\sum_i \alpha_{:,i}\odot h_i$, where $h_i$ is the hidden representation of expert $i$, and a small classifier converts this into a binding prediction. The gating network is the only part trained, so the ensemble learns to re-weight existing motif detectors rather than retrain them. The second mechanism is ShiftSmooth attribution, $\hat{A}_c(x) = (2N+1)^{-1}\sum_{n=-N}^{N} A_c(x \gg n)\ll n$ with $A_c(x)=\partial S_c(x)/\partial x$ and circular shifts, which averages gradients over shifted viewing windows so that motif attributions are robust to exactly the window-boundary variability that genomic sequences have.

What would settle it

Reconstruct the OOD benchmark with the same six ENCODE factors but report how positive and negative sequences were generated, their lengths, and class balance, and then match dataset difficulty (for example, with motif-free negative sets of equal GC content); if the MoE no longer beats the best single expert on a majority of the six factors, the out-of-distribution claim is falsified. Separately, plant a known motif in synthetic sequences and compare ShiftSmooth and Vanilla Gradient peaks at the motif positions; if ShiftSmooth does not retrieve the planted motif more often, the attribution claim is falsified.

Watch

Extended reading notes

Core claim

The paper claims that a mixture of three frozen CNN experts, each specialized on one JASPAR transcription-factor motif, can be recombined by a learned gating network into a predictor that is not worse than the best single expert on the factors it trained on and is statistically better than every single expert on five of six held-out transcription factors (BCLAF1, CTCF, POLR2A, RBBP5, SAP30, STAT3) from ENCODE. On in-distribution data, the matching expert remains best on its own factor and the MoE is consistently second; on out-of-distribution data, the MoE is best for five factors and second for CTCF, with ANOVA over 30 bootstrap trials giving non-overlapping confidence intervals and very small p-values. The paper further claims that its ShiftSmooth attribution map, unlike Vanilla Gradient, assigns clear positive importance to the guanine in the GATA3 motif 'GATAA' and removes the negative-attribution anomaly that Vanilla Gradient produces, making ShiftSmooth better suited for motif discovery and localization.

Load-bearing premise

The argument rests on the six ENCODE out-of-distribution datasets being directly comparable to one another and to the JASPAR-derived training data, since that is the only basis for reading the MoE's higher AUC as better generalization rather than as easier data.

Editorial extensions

If this is right

  • On out-of-distribution transcription factors, a frozen-expert MoE can beat the best single expert with statistical significance, so a well-chosen gating network is a practical way to reuse existing motif detectors for new genomes.
  • Because the expert weights stay frozen and only the gating network trains, adding a new expert does not invalidate previously learned experts; the same library can serve different tasks by training a new gate.
  • In-distribution results show a specialization trade-off: the expert trained on a factor remains best on that factor, while the MoE is the best overall compromise, which suggests that the MoE protects against worst-case single-expert failures on unseen binding patterns.
  • ShiftSmooth attribution maps are stable under small shifts in the sequence window, which matters because genomic datasets define motif windows somewhat arbitrarily; stable attributions should help locate motifs at base-pair resolution.
  • The paper's statistical protocol of 30 bootstrap trials plus ANOVA provides a template for deciding whether an ensemble's out-of-distribution gain exceeds chance before adopting it in a genomics pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: compare the MoE against an unweighted average of the same three frozen experts; if the simple average performs nearly as well, the gating network's role is mainly variance reduction rather than learned specialization.
  • Because DNA-binding proteins scan motifs in a roughly translation-invariant way, ShiftSmooth's shift-averaging is likely to align better with known motif annotations; this can be quantified by planted-motif recovery on synthetic sequences.
  • The five-of-six OOD result rests on one random draw of six ENCODE factors; checking the result on a larger, stratified set of transcription factors while matching dataset difficulty would show whether the mechanism is ensemble generalization or dataset-specific separability.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a Mixture of Experts (MoE) model for transcription factor binding site (TFBS) prediction, combining three pre-trained CNN experts trained on ARID3A, FOXM1, and GATA3 with a gating network that learns to weight their frozen embeddings. The model is evaluated on the three in-distribution datasets and on six out-of-distribution ENCODE transcription factors (BCLAF1, CTCF, POLR2A, RBBP5, SAP30, STAT3). The authors report AUC scores over 30 bootstrap trials and use one-way ANOVA to claim that the MoE is statistically best on five of the six OOD datasets. The paper also introduces ShiftSmooth, an attribution method that averages gradients over circularly shifted input sequences, and claims it offers superior motif discovery and localization compared to Vanilla Gradients, supported by qualitative visualizations.

Significance. If the central claims were fully supported, the paper would provide a lightweight ensemble approach that generalizes to unseen transcription factors and a simple interpretability method tailored to genomic sequences. The manuscript is transparent about its architecture and gives a formal definition of ShiftSmooth (Eq. 16); it also uses public JASPAR and ENCODE datasets, which aids reproducibility. However, the significance of the contribution is currently limited by the statistical analysis used to support the main performance claim and by the purely qualitative evaluation of the explainability method.

major comments (4)
  1. [III-C2–C3, IV-B] The one-way ANOVA that underpins the claim that the MoE 'is the best for five of the six datasets with statistical significance' (Section IV-B) is computed on 30 bootstrap AUCs per model obtained by resampling with replacement from a single test set (Section III-C2). These bootstrap replicates are not independent: each replicate overlaps heavily with the same test sequences, and the four models' AUCs within a replicate are paired because they are evaluated on the identical resampled data. One-way ANOVA assumes independent groups and independent observations within groups, so both assumptions are violated. No p-values, F-statistics, or effect sizes are reported anywhere; the paper only states that p-values are 'extremely small' and that 95% confidence intervals do not overlap. A valid analysis would use a paired or repeated-measures bootstrap test that accounts for the across-model pairing, and should report effect sizes. As written, the central statistical claim is not supported.
  2. [IV-B, III-C] The out-of-distribution evaluation does not describe how the six ENCODE datasets (BCLAF1, CTCF, POLR2A, RBBP5, SAP30, STAT3) were preprocessed: how positive and negative sequences were obtained, sequence lengths, class balance, and whether peak-calling thresholds vary across factors. If the OOD datasets differ in baseline separability, the observed five-of-six win of the MoE could reflect dataset difficulty rather than ensemble generalization. The authors should report per-dataset construction details and per-dataset AUCs with appropriate uncertainty estimates.
  3. [V] The claim in the Abstract and Section V that ShiftSmooth 'offers superior attribution for motif discovery and localization compared to traditional Vanilla Gradient methods' is supported only by qualitative inspection of one GATA3-containing sequence (Figs. 10–14) and one random sequence (Figs. 15–19). No quantitative faithfulness or localization metric (e.g., recovery of known motifs, or the EvalAttAI evaluation cited in Section II-B1) is reported. The authors should either provide quantitative attribution evaluation or substantially temper the claim.
  4. [III-A, IV] The MoE architecture (Section III-A) is a learned weighted sum of frozen expert embeddings with softmax gating. The paper never compares against simpler baselines such as an unweighted average of expert outputs or a linear combination of expert probabilities. Without such a baseline, it is not established that the gating mechanism, rather than the ensembling of diverse experts, is responsible for the reported OOD improvements. The authors should add an averaging ensemble baseline to the evaluation.
minor comments (5)
  1. [IV-B] In the first paragraph of Section IV-B, 'CLAF1' should be 'BCLAF1'.
  2. [Figure 9] The caption of Figure 9 contains a typo: 'usine' should be 'using'; also the repeated 'ANOV A' spacing is inconsistent.
  3. [III-D] Equation (15) uses bit-shift notation '>>' and '<<' to describe circular shifts; please define the circular shift operation explicitly to avoid confusion with integer bit shifts.
  4. [II-B3] The ShiftSmooth description refers to shifting the 'input image'; in the genomic context, please clarify that the input is a one-hot encoded sequence array.
  5. [Abstract, I] The abstract states that six transcription factors were 'randomly selected' for OOD testing, but the random selection procedure and seed are not described, which hampers reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; MoE OOD result is a held-out benchmark and ShiftSmooth is a defined measurement procedure, with only non-load-bearing self-citations.

full rationale

The central performance claim is not circular. The three expert CNNs are trained on JASPAR-derived ARID3A, FOXM1, and GATA3 data; the MoE gating network is trained on their frozen embeddings from the same in-distribution data; and the six ENCODE transcription factors are entirely held out from training. No OOD label or OOD AUC enters the training objective, so the reported superiority on five of six OOD datasets is a measured generalization result, not a fitted parameter renamed as a prediction. ShiftSmooth (Eq. 16) is likewise a defined measurement procedure: it averages backpropagated gradients over circularly shifted copies of the input, and the qualitative claim that it localizes motifs better than Vanilla Gradient is an empirical evaluation rather than an equation that presupposes its own conclusion. The self-citations [16] and [17] are used only to define attributional robustness and to justify the choice of Vanilla Gradient as a baseline; they are not invoked as a theorem that forces the MoE or ShiftSmooth results. The ANOVA concern raised in the skeptic summary is a statistical-validity issue about dependent bootstrap samples being treated as independent, not a circularity, because the significance claim is not equivalent by construction to the bootstrap input. Therefore no circular step can be exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper's central claims rest on domain assumptions about ChIP-seq data and gradient saliency, a paper-specific assumption that circular shifts of a linear DNA sequence are valid window shifts, and a questionable statistical assumption about bootstrap samples. No fitted constants are used as inputs to a derivation; the hand-chosen values that directly affect results are the number of experts, the number of bootstrap trials, and the unreported ShiftSmooth shift range N.

free parameters (3)
  • ShiftSmooth shift range N = not reported; illustrated as N=2
    The attribution map is an average over 2N+1 shifted sequences (Eq. 16); the experimental value of N is never stated, so the reported ShiftSmooth results depend on an unspecified hand-chosen hyperparameter.
  • Number of experts N_e = 3
    Three pre-trained DeepBind experts (ARID3A, FOXM1, GATA3) were used; results would change with more or fewer experts.
  • Bootstrap trial count B = 30
    AUC distributions and ANOVA are based on 30 bootstrap resamples; this arbitrary count affects statistical power and the width of confidence intervals.
assumptions (5)
  • domain assumption ChIP-seq peak regions used as positive TFBS examples accurately represent true binding sites.
    Standard for TFBS datasets; ENCODE ChIP-seq peaks are treated as ground truth for training and evaluation throughout Sections I, III, and IV.
  • ad hoc to paper Circularly shifting a DNA sequence preserves relevant motif context and is a valid proxy for viewing-window shifts.
    Eq. 15 defines x_n = x >> n as a circular shift with wraparound, and Fig. 3 uses N=2; wraparound creates artificial nucleotide junctions not present in linear DNA, and this is never validated.
  • domain assumption Gradient of the output with respect to input is a faithful measure of nucleotide importance.
    Both Vanilla Gradient and ShiftSmooth rely on this contested assumption; the paper cites Ref. [17] for faithfulness but does not test it on genomic models.
  • domain assumption Bootstrapped AUC samples satisfy the ANOVA assumptions of independence, homoscedasticity, and normality.
    Section III-C2 and Section IV use one-way ANOVA on 30 bootstrap resamples from the same test set; resamples are not independent draws, so this assumption is likely violated.
  • domain assumption The six OOD ENCODE datasets are directly comparable to the JASPAR training datasets and to one another, so AUC differences reflect generalization ability.
    Section IV-B gives no preprocessing details; if OOD datasets differ in difficulty, the reported AUC rankings are confounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable AI in Genomics: Transcription Factor Binding Site Prediction with Mixture of Experts." pith.science (2026). https://pith.science/paper/WB454STM

@misc{pith2026250709754,
  author       = {Pith},
  title        = {Pith review of: Explainable AI in Genomics: Transcription Factor Binding Site Prediction with Mixture of Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WB454STM}},
  note         = {Machine review of arXiv:2507.09754}
}
read the original abstract

Transcription Factor Binding Site (TFBS) prediction is crucial for understanding gene regulation and various biological processes. This study introduces a novel Mixture of Experts (MoE) approach for TFBS prediction, integrating multiple pre-trained Convolutional Neural Network (CNN) models, each specializing in different TFBS patterns. We evaluate the performance of our MoE model against individual expert models on both in-distribution and out-of-distribution (OOD) datasets, using six randomly selected transcription factors (TFs) for OOD testing. Our results demonstrate that the MoE model achieves competitive or superior performance across diverse TF binding sites, particularly excelling in OOD scenarios. The Analysis of Variance (ANOVA) statistical test confirms the significance of these performance differences. Additionally, we introduce ShiftSmooth, a novel attribution mapping technique that provides more robust model interpretability by considering small shifts in input sequences. Through comprehensive explainability analysis, we show that ShiftSmooth offers superior attribution for motif discovery and localization compared to traditional Vanilla Gradient methods. Our work presents an efficient, generalizable, and interpretable solution for TFBS prediction, potentially enabling new discoveries in genome biology and advancing our understanding of transcriptional regulation.

Figures

Figures reproduced from arXiv: 2507.09754 by the authors.

Figure 1
Figure 1. Model training and evaluation process using MoE model and attribution mapping. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Image input into YOLOv7 [33] (left) and image overlaid with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the process for generating ShiftSmooth attribution maps ˆ [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: ROC curves for one trial comparing the performance of the ARID3A, FOXM1, GATA3 expert models and the MoE model on in-distribution TFBS [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Bar plot showing a comparison of the average model AUC scores [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The 95% confidence intervals (shown as boxes) for the model [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: ROC curves comparing the performance of the ARID3A, FOXM1, GATA3 expert models and the MoE model on TFBS prediction across six different [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 9
Figure 9. Figure 9: The 95% confidence intervals (shown as boxes) for the model AUC [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Nucleotide sequence containing GATA3 TFBS. The sequence contains both orientations of the motif including ‘GATAA’ and its inverse, ‘TTATC.’ [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: TFBS prediction using the Vanilla Gradient output from the GATA3 trained expert model for the given sequence in Fig. 10. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: TFBS prediction using the ShiftSmooth Gradient output from the GATA3 trained expert model for the given sequence in Fig. 10. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: TFBS prediction using the Vanilla Gradient output from the MoE model for the given sequence in Fig. 10. [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: TFBS prediction using the ShiftSmooth Gradient output from the MoE model for the given sequence in Fig. 10. [PITH_FULL_IMAGE:figures/full_fig_p013_14.png]
Figure 16
Figure 16. Figure 16: TFBS prediction using the Vanilla Gradient output from the GATA3 trained expert model for the given random sequence in Fig. 15. [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 17
Figure 17. Figure 17: TFBS prediction using the ShiftSmooth Gradient output from the GATA3 trained expert model for the given random sequence in Fig. 15. [PITH_FULL_IMAGE:figures/full_fig_p014_17.png]
Figure 18
Figure 18. Figure 18: TFBS prediction using the Vanilla Gradient output from the MoE model for the given random sequence in Fig. 15. [PITH_FULL_IMAGE:figures/full_fig_p014_18.png]
Figure 19
Figure 19. Figure 19: TFBS prediction using the ShiftSmooth Gradient output from the MoE model for the given random sequence in Fig. 15. [PITH_FULL_IMAGE:figures/full_fig_p014_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 59 canonical work pages

  1. [1]

    Messenger-rna-binding proteins and the messages they carry,

    G. Dreyfuss, V . N. Kim, and N. Kataoka, “Messenger-rna-binding proteins and the messages they carry,” Nature reviews Molecular cell biology, vol. 3, no. 3, pp. 195–205, 2002

    work page 2002

  2. [2]

    Modeling the specificity of protein-dna interactions,

    G. D. Stormo, “Modeling the specificity of protein-dna interactions,” Quantitative biology, vol. 1, no. 2, pp. 115–130, 2013

    work page 2013

  3. [3]

    Transcription factor–dna binding: beyond binding site motifs,

    S. Inukai, K. H. Kock, and M. L. Bulyk, “Transcription factor–dna binding: beyond binding site motifs,” Current Opinion in Genetics & Development , vol. 43, pp. 110–119, 2017. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0959437X1730028X

    work page 2017

  4. [4]

    T. L. Bailey, Discovering Sequence Motifs. Totowa, NJ: Humana Press, 2008, pp. 271–292

    work page 2008

  5. [5]

    Vijayvargiya and P

    S. Vijayvargiya and P. Shukla, Regulatory Motif Identification in Biolog- ical Sequences: An Overview of Computational Methodologies . New Delhi: Springer India, 2013, pp. 111–124

    work page 2013

  6. [6]

    Fast and exact quantification of motif occurrences in biological sequences,

    M. Prosperi, S. Marini, and C. Boucher, “Fast and exact quantification of motif occurrences in biological sequences,” BMC Bioinformatics , vol. 22, no. 1, p. 445, 2021

    work page 2021

  7. [7]

    Deep neural networks identify sequence context features predictive of PREPRINT 15 transcription factor binding,

    A. Zheng, M. Lamkin, H. Zhao, C. Wu, H. Su, and M. Gymrek, “Deep neural networks identify sequence context features predictive of PREPRINT 15 transcription factor binding,” Nature machine intelligence, vol. 3, no. 2, pp. 172–180, 2021

    work page 2021

  8. [8]

    Locating transcription factor binding sites by fully convolutional neural network,

    Q. Zhang, S. Wang, Z. Chen, Y . He, Q. Liu, and D.-S. Huang, “Locating transcription factor binding sites by fully convolutional neural network,” Briefings in bioinformatics , vol. 22, no. 5, p. bbaa435, 2021

    work page 2021

Show all 72 references
  1. [9]

    maxatac: Genome-scale transcription-factor binding prediction from atac-seq with deep neural networks,

    T. A. Cazares, F. W. Rizvi, B. Iyer, X. Chen, M. Kotliar, A. T. Bejjani, J. A. Wayman, O. Donmez, B. Wronowski, S. Parameswaran et al. , “maxatac: Genome-scale transcription-factor binding prediction from atac-seq with deep neural networks,” PLOS Computational Biology , vol. 1...

    2023

  2. [10]

    Predicting the sequence specificities of dna- and rna-binding proteins by deep learning,

    B. Alipanahi, A. Delong, M. T. Weirauch, and B. J. Frey, “Predicting the sequence specificities of dna- and rna-binding proteins by deep learning,” Nature Biotechnology, vol. 33, no. 8, p. 831–838, 2015

    2015

  3. [11]

    Comprehensive evaluation of deep learning architectures for prediction of DNA/RNA sequence binding specificities,

    A. Trabelsi, M. Chaabane, and A. Ben-Hur, “Comprehensive evaluation of deep learning architectures for prediction of DNA/RNA sequence binding specificities,” Bioinformatics, vol. 35, no. 14, pp. i269–i277, 07 2019. [Online]. Available: https://academic.oup.com/bioinformatics/ ...

    2019

  4. [12]

    Modeling genome data using bidirectional lstm,

    N. Tavakoli, “Modeling genome data using bidirectional lstm,” in 2019 IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC), vol. 2, 2019, pp. 183–188

    2019

  5. [13]

    Splice-site identification for exon prediction using bidirectional lstm-rnn approach,

    N. Singh, R. Nath, and D. B. Singh, “Splice-site identification for exon prediction using bidirectional lstm-rnn approach,” Biochemistry and Biophysics Reports , vol. 30, p. 101285, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2405580822000851

    2022

  6. [14]

    Explainable deep neural networks for novel viral genome prediction,

    C. M. Dasari and R. Bhukya, “Explainable deep neural networks for novel viral genome prediction,” Applied Intelligence , vol. 52, no. 3, pp. 3002–3017, 2022. [Online]. Available: https://doi.org/10. 1007/s10489-021-02572-3

    2022

  7. [15]

    DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome,

    Y . Ji, Z. Zhou, H. Liu, and R. V . Davuluri, “DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome,” Bioinformatics, vol. 37, no. 15, pp. 2112–2120, 02 2021. [Online]. Available: https://academic.oup.com/ bioinformatics...

    2021

  8. [16]

    Robust explainability: A tutorial on gradient-based attribution methods for deep neural networks,

    I. E. Nielsen, D. Dera, G. Rasool, R. P. Ramachandran, and N. C. Bouay- naya, “Robust explainability: A tutorial on gradient-based attribution methods for deep neural networks,” IEEE Signal Processing Magazine , vol. 39, no. 4, pp. 73–84, 2022

    2022

  9. [17]

    Evalattai: A holistic approach to evaluating attribution maps in robust and non-robust models,

    I. E. Nielsen, R. P. Ramachandran, N. Bouaynaya, H. M. Fathallah- Shaykh, and G. Rasool, “Evalattai: A holistic approach to evaluating attribution maps in robust and non-robust models,” IEEE Access, vol. 11, pp. 82 556–82 569, 2023

    2023

  10. [18]

    Ancona, E

    M. Ancona, E. Ceolini, C. ¨Oztireli, and M. Gross, Gradient-Based Attribution Methods . Cham: Springer International Publishing, 2019, pp. 169–191

    2019

  11. [19]

    M. A. Carlton and J. L. Devore, Probability with Applications in Engineering, Science, and Technology . Cham, Switzerland: Springer, 2017

    2017

  12. [20]

    Jaspar 2022: The 9th release of the open-access database of transcription factor binding profiles,

    J. A. Castro-Mondragon, R. Riudavets-Puig, I. Rauluseviciute, R. Berhanu&nbsp;Lemma, L. Turchi, R. Blanc-Mathieu, J. Lucas, P. Boddie, A. Khan, N. Manosalva&nbsp;P ´erez, and et al., “Jaspar 2022: The 9th release of the open-access database of transcription factor binding prof...

    2022

  13. [21]

    Loss of bright/arid3a function promotes developmental plasticity,

    G. An, C. A. Miner, J. C. Nixon, P. W. Kincade, J. Bryant, P. W. Tucker, and C. F. Webb, “Loss of bright/arid3a function promotes developmental plasticity,” Stem Cells, vol. 28, no. 9, pp. 1560–1567, 2010

    2010

  14. [22]

    Regulation of the master regulator foxm1 in cancer,

    G.-B. Liao, X.-Z. Li, S. Zeng, C. Liu, S.-M. Yang, L. Yang, C.-J. Hu, and J.-Y . Bai, “Regulation of the master regulator foxm1 in cancer,”Cell Communication and Signaling , vol. 16, no. 1, p. 57, 2018

    2018

  15. [23]

    Gata factor mutations in hematologic disease,

    J. D. Crispino and M. S. Horwitz, “Gata factor mutations in hematologic disease,” Blood, The Journal of the American Society of Hematology , vol. 129, no. 15, pp. 2103–2110, 2017

    2017

  16. [24]

    Acquired and germline predisposition to bone marrow failure: Diagnostic features and clinical implications,

    M. E. Kallen, A. Dulau-Florea, W. Wang, and K. R. Calvo, “Acquired and germline predisposition to bone marrow failure: Diagnostic features and clinical implications,” in Seminars in hematology , vol. 56, no. 1. Elsevier, 2019, pp. 69–82

    2019

  17. [25]

    The encyclopedia of dna elements (encode): Data portal update,

    C. A. Davis, B. C. Hitz, C. A. Sloan, E. T. Chan, J. M. Davidson, I. Gabdank, J. A. Hilton, K. Jain, U. K. Baymuradov, A. K. Narayanan, and et al., “The encyclopedia of dna elements (encode): Data portal update,” Nucleic Acids Research , vol. 46, no. D1, 2017

    2017

  18. [26]

    A user’s guide to the encyclopedia of dna elements (encode),

    E. P. Consortium, “A user’s guide to the encyclopedia of dna elements (encode),” PLoS biology, vol. 9, no. 4, p. e1001046, 2011

    2011

  19. [27]

    Chip-seq: welcome to the new frontier,

    E. R. Mardis, “Chip-seq: welcome to the new frontier,” Nature methods, vol. 4, no. 8, pp. 613–614, 2007

    2007

  20. [28]

    Chip–seq: advantages and challenges of a maturing technol- ogy,

    P. J. Park, “Chip–seq: advantages and challenges of a maturing technol- ogy,” Nature reviews genetics, vol. 10, no. 10, pp. 669–680, 2009

    2009

  21. [29]

    Mixture of experts explained,

    O. Sanseviero, L. Tunstall, P. Schmid, S. Mangrulkar, Y . Belkada, and P. Cuenca, “Mixture of experts explained,” 2023. [Online]. Available: https://huggingface.co/blog/moe

    2023

  22. [30]

    Adaptive mixtures of local experts,

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adaptive mixtures of local experts,” Neural computation, vol. 3, no. 1, pp. 79–87, 1991

    1991

  23. [31]

    Learning factored represen- tations in a deep mixture of experts,

    D. Eigen, M. Ranzato, and I. Sutskever, “Learning factored represen- tations in a deep mixture of experts,” arXiv preprint arXiv:1312.4314 , 2013

    2013 arXiv

  24. [32]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” in International Conference on Learning Representations, 2016

    2016

  25. [33]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 7464–7475

    2023

  26. [34]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 618–626

    2017

  27. [35]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13 . Springer,...

    2014

  28. [36]

    Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence,

    S. Ali, T. Abuhmed, S. El-Sappagh, K. Muhammad, J. M. Alonso- Moral, R. Confalonieri, R. Guidotti, J. Del Ser, N. D ´ıaz-Rodr´ıguez, and F. Herrera, “Explainable artificial intelligence (xai): What we know and what is left to attain trustworthy artificial intelligence,” Inform...

    2023

  29. [37]

    From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable AI,

    M. Nauta, J. Trienes, S. Pathak, E. Nguyen, M. Peters, Y . Schmitt, J. Schl ¨otterer, M. van Keulen, and C. Seifert, “From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable AI,” arXiv preprint arXiv:2201.08164 , 2022. [Online]...

    2022 arXiv

  30. [38]

    Learning important features through propagating activation differences,

    A. Shrikumar, P. Greenside, and A. Kundaje, “Learning important features through propagating activation differences,” in Proceedings of the 34th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 7...

    2017

  31. [39]

    Explainability in transformer models for functional genomics,

    J. Clauwaert, G. Menschaert, and W. Waegeman, “Explainability in transformer models for functional genomics,” Briefings in bioinformat- ics, vol. 22, no. 5, p. bbab060, 2021

    2021

  32. [40]

    Explainability methods for graph convolutional neural networks,

    P. E. Pope, S. Kolouri, M. Rostami, C. E. Martin, and H. Hoffmann, “Explainability methods for graph convolutional neural networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 10 772–10 781

    2019

  33. [41]

    Visualizing the impact of feature attribution baselines,

    P. Sturmfels, S. Lundberg, and S.-I. Lee, “Visualizing the impact of feature attribution baselines,” Distill, vol. 5, no. 1, p. e22, 2020

    2020

  34. [42]

    Transformer interpretability beyond attention visualization,

    H. Chefer, S. Gur, and L. Wolf, “Transformer interpretability beyond attention visualization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 782–791

    2021

  35. [43]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” Advances in neural information processing systems, vol. 30, 2017

    2017

  36. [44]

    Peeking inside the black-box: a survey on explainable artificial intelligence (xai),

    A. Adadi and M. Berrada, “Peeking inside the black-box: a survey on explainable artificial intelligence (xai),” IEEE access, vol. 6, pp. 52 138– 52 160, 2018

    2018

  37. [45]

    Deep inside convolutional networks: visualising image classification models and saliency maps,

    K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep inside convolutional networks: visualising image classification models and saliency maps,” in Proceedings of the International Conference on Learning Represen- tations (ICLR). ICLR, 2014

    2014

  38. [46]

    Transformers in time-series analysis: A tutorial,

    S. Ahmed, I. E. Nielsen, A. Tripathi, S. Siddiqui, R. P. Ramachandran, and G. Rasool, “Transformers in time-series analysis: A tutorial,” Circuits, Systems, and Signal Processing , vol. 42, no. 12, pp. 7433– 7466, 2023

    2023

  39. [47]

    Smoothgrad: removing noise by adding noise,

    D. Smilkov, N. Thorat, B. Kim, F. Vi ´egas, and M. Wattenberg, “Smoothgrad: removing noise by adding noise,” arXiv preprint arXiv:1706.03825, 2017

    2017 arXiv

  40. [48]

    Secure and robust machine learning for healthcare: A survey,

    A. Qayyum, J. Qadir, M. Bilal, and A. Al-Fuqaha, “Secure and robust machine learning for healthcare: A survey,” IEEE Reviews in Biomedical Engineering, vol. 14, pp. 156–180, 2021

    2021

  41. [49]

    A survey on efficient methods for adversarial robustness,

    A. Muhammad and S.-H. Bae, “A survey on efficient methods for adversarial robustness,” IEEE Access , vol. 10, pp. 118 815–118 830, 2022. PREPRINT 16

    2022

  42. [50]

    Recent advances in adversarial training for adversarial robustness,

    T. Bai, J. Luo, J. Zhao, B. Wen, and Q. Wang, “Recent advances in adversarial training for adversarial robustness,” arXiv preprint arXiv:2102.01356, 2021

    2021 arXiv

  43. [51]

    Extended variational inference for propagating uncertainty in convolutional neural networks,

    D. Dera, G. Rasool, and N. Bouaynaya, “Extended variational inference for propagating uncertainty in convolutional neural networks,” in 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 2019, pp. 1–6

    2019

  44. [52]

    Premium-cnn: Propagating uncertainty towards ro- bust convolutional neural networks,

    D. Dera, N. C. Bouaynaya, G. Rasool, R. Shterenberg, and H. M. Fathallah-Shaykh, “Premium-cnn: Propagating uncertainty towards ro- bust convolutional neural networks,” IEEE Transactions on Signal Processing, vol. 69, pp. 4669–4684, 2021

    2021

  45. [53]

    Robustness may be at odds with accuracy,

    D. Tsipras, S. Santurkar, L. Engstrom, A. Turner, and A. Madry, “Robustness may be at odds with accuracy,” in International Conference on Learning Representations , no. 2019, 2019

    2019

  46. [54]

    How explainable are adversarially-robust cnns?

    M. Nourelahi, L. Kotthoff, P. Chen, and A. Nguyen, “How explainable are adversarially-robust cnns?” arXiv preprint arXiv:2205.13042 , 2022

    2022 arXiv

  47. [55]

    Inter- pretable attention model in transcription factor binding site prediction with deep neural networks,

    C. Chen, J. Hou, X. Shi, H. Yang, J. A. Birchler, and J. Cheng, “Inter- pretable attention model in transcription factor binding site prediction with deep neural networks,” bioRxiv, p. 648691, 2019

    2019

  48. [56]

    Predicting effects of noncoding variants with deep learning–based sequence model,

    J. Zhou and O. G. Troyanskaya, “Predicting effects of noncoding variants with deep learning–based sequence model,” Nature methods , vol. 12, no. 10, pp. 931–934, 2015

    2015

  49. [57]

    Mixture-of-experts approach for enhanced drug-target interaction prediction and confidence assessment,

    Y . Lu, S. Lee, S. Kang, and S. Kim, “Mixture-of-experts approach for enhanced drug-target interaction prediction and confidence assessment,” bioRxiv, pp. 2024–08, 2024

    2024

  50. [58]

    Improving language model of human genome for dna–protein binding prediction based on task-specific pre-training,

    H. Luo, W. Shan, C. Chen, P. Ding, and L. Luo, “Improving language model of human genome for dna–protein binding prediction based on task-specific pre-training,” Interdisciplinary Sciences: Computational Life Sciences, vol. 15, no. 1, pp. 32–43, 2023

    2023

  51. [59]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” in International conference on machine learning . PMLR, 2017, pp. 3319–3328

    2017

  52. [60]

    Explainable models using transcription factor binding and epigenome patterns at promoters reveal disease-associated genes and their regulators in the context of cell-types,

    O. Chandra, D. Pramanik, S. Gautam, M. Sharma, N. Dubey, B. Mahato, and V . Kumar, “Explainable models using transcription factor binding and epigenome patterns at promoters reveal disease-associated genes and their regulators in the context of cell-types,” bioRxiv, 2024

    2024

  53. [61]

    The developmental and evolutionary characteristics of transcription factor binding site clustered regions based on an explainable machine learning model,

    Z. Ouyang, F. Liu, W. Li, J. Wang, B. Chen, Y . Zheng, Y . Li, H. Tao, X. Xu, C. Li, Y . Cong, H. Li, X. Bo, and H. Chen, “The developmental and evolutionary characteristics of transcription factor binding site clustered regions based on an explainable machine learning model,”...

  54. [62]

    Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks,

    D. R. Kelley, J. Snoek, and J. L. Rinn, “Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks,” Genome research, vol. 26, no. 7, p. 990—999, July 2016

    2016

  55. [63]

    Deep neural networks for interpreting rna- binding protein target preferences,

    M. Ghanbari and U. Ohler, “Deep neural networks for interpreting rna- binding protein target preferences,” Genome Research, vol. 30, no. 2, pp. 214–226, 2020

    2020

  56. [64]

    Sequential regulatory activity prediction across chromosomes with convolutional neural networks,

    D. R. Kelley, Y . A. Reshef, M. Bileschi, D. Belanger, C. Y . McLean, and J. Snoek, “Sequential regulatory activity prediction across chromosomes with convolutional neural networks,” Genome Res , vol. 28, no. 5, pp. 739–750, 2018

    2018

  57. [65]

    Base-resolution models of transcription factor binding reveal soft motif syntax,

    ˇZ. Avsec, M. Weilert, A. Shrikumar, S. Krueger, A. Alexandari, K. Dalal, R. Fropf, C. McAnany, J. Gagneur, A. Kundaje, and J. Zeitlinger, “Base-resolution models of transcription factor binding reveal soft motif syntax,” bioRxiv, 2020

    2020

  58. [66]

    Edeepssp: Explainable deep neural net- works for exact splice sites prediction,

    S. Amilpur and R. Bhukya, “Edeepssp: Explainable deep neural net- works for exact splice sites prediction,” Journal of Bioinformatics and Computational Biology, vol. 18, no. 04, p. 2050024, 2020

    2020

  59. [67]

    Obtaining genetics insights from deep learning via explainable artificial intelligence,

    G. Novakovsky, N. Dexter, M. W. Libbrecht, W. W. Wasserman, and S. Mostafavi, “Obtaining genetics insights from deep learning via explainable artificial intelligence,” Nature Reviews Genetics , vol. 24, no. 2, pp. 125–137, 2023

    2023

  60. [68]

    Explainn: interpretable and transparent neural networks for genomics,

    G. Novakovsky, O. Fornes, M. Saraswat, S. Mostafavi, and W. W. Wasserman, “Explainn: interpretable and transparent neural networks for genomics,” Genome Biology, vol. 24, no. 1, p. 154, 2023

    2023

  61. [69]

    Optuna: A next- generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next- generation hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , 2019, pp. 2623–2631

    2019

  62. [70]

    On the importance of initialization and momentum in deep learning,

    I. Sutskever, J. Martens, G. Dahl, and G. Hinton, “On the importance of initialization and momentum in deep learning,” in International conference on machine learning . PMLR, 2013, pp. 1139–1147

    2013

  63. [71]

    C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning. New York: Springer, 2006, vol. 4, no. 4

    2006

  64. [2024]

    Available: https://doi.org/10.1093/nar/gkae441

    [Online]. Available: https://doi.org/10.1093/nar/gkae441

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.