One model per experiment is the Kundaje way (model zoo? model herd?) and then does things like backpropagation for sequencing interpretation via DeepLIFT - being ported by AlphaGenome (editorial - need to get this going with things like GenomicsEngland). Has an example with Myc.
May 27, 2026 at 1:47 PM
One model per experiment is the Kundaje way (model zoo? model herd?) and then does things like backpropagation for sequencing interpretation via DeepLIFT - being ported by AlphaGenome (editorial - need to get this going with things like GenomicsEngland). Has an example with Myc.
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
Motivation Characterization of regulatory elements in DNA sequence is a key task in functional genom
www.biorxiv.org
November 21, 2023 at 2:50 AM
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
4) Then, we interrogate each model using DeepLIFT to identify the important seqs for the model. The importance scores highlight known motifs of TFs active in K562, incl GATA, AP-1, SP, ETV/ETS, CEBP. The importance also align across both models, again reconciling the two assays. 8/
August 10, 2026 at 7:00 PM
4) Then, we interrogate each model using DeepLIFT to identify the important seqs for the model. The importance scores highlight known motifs of TFs active in K562, incl GATA, AP-1, SP, ETV/ETS, CEBP. The importance also align across both models, again reconciling the two assays. 8/
3) By tracking the model's operations, we can quantify how much each input was used by the model during prediction.
Excitingly, this recapitulates known TF preferences and isolates the exact model-predicted binding site (vs assumed peak center). (DeepLIFT: doi.org/10.48550/arX...) 15/
Excitingly, this recapitulates known TF preferences and isolates the exact model-predicted binding site (vs assumed peak center). (DeepLIFT: doi.org/10.48550/arX...) 15/
August 6, 2026 at 2:30 AM
3) By tracking the model's operations, we can quantify how much each input was used by the model during prediction.
Excitingly, this recapitulates known TF preferences and isolates the exact model-predicted binding site (vs assumed peak center). (DeepLIFT: doi.org/10.48550/arX...) 15/
Excitingly, this recapitulates known TF preferences and isolates the exact model-predicted binding site (vs assumed peak center). (DeepLIFT: doi.org/10.48550/arX...) 15/
excelling in complex genomic data analysis. Furthermore, interpretability investigations using DeepLIFT and Transcription Factor Motif Discovery from the importance scores algorithm (TF-MoDISco) demonstrate that the derived motifs from our model [7/8 of https://arxiv.org/abs/2505.09883v1]
May 16, 2025 at 6:15 AM
excelling in complex genomic data analysis. Furthermore, interpretability investigations using DeepLIFT and Transcription Factor Motif Discovery from the importance scores algorithm (TF-MoDISco) demonstrate that the derived motifs from our model [7/8 of https://arxiv.org/abs/2505.09883v1]
They conducted a methodological and experimental evaluation of five popular explanation methods ( #Occlusion, #LIME, #GradCAM, #LRP, #DeepLIFT) and ten metrics covering five categories: faithfulness, robustness, localization, complexity, and randomization.
October 27, 2025 at 2:34 PM
They conducted a methodological and experimental evaluation of five popular explanation methods ( #Occlusion, #LIME, #GradCAM, #LRP, #DeepLIFT) and ten metrics covering five categories: faithfulness, robustness, localization, complexity, and randomization.
They a have a stage in their processing pipeline called "explainability" 😉 - seems to consist of 'DeepLIFT' that assigns importance scores to input features -via a chain rule of 'multipliers' to highlight which time-frequency regions in the wavelet spectrogram most influenced each predicted emotion
May 15, 2026 at 10:13 PM
They a have a stage in their processing pipeline called "explainability" 😉 - seems to consist of 'DeepLIFT' that assigns importance scores to input features -via a chain rule of 'multipliers' to highlight which time-frequency regions in the wavelet spectrogram most influenced each predicted emotion
⚡️OSS Project Spotlight⚡️
Captum is an open source library from Meta for model interpretability in PyTorch. It helps developers understand and visualize the decisions made by deep learning models using attribution methods like Integrated Gradients and DeepLIFT.
Learn more: https://captum.ai/
Captum is an open source library from Meta for model interpretability in PyTorch. It helps developers understand and visualize the decisions made by deep learning models using attribution methods like Integrated Gradients and DeepLIFT.
Learn more: https://captum.ai/
Captum · Model Interpretability for PyTorch
Model Interpretability for PyTorch
captum.ai
October 15, 2025 at 5:24 PM
⚡️OSS Project Spotlight⚡️
Captum is an open source library from Meta for model interpretability in PyTorch. It helps developers understand and visualize the decisions made by deep learning models using attribution methods like Integrated Gradients and DeepLIFT.
Learn more: https://captum.ai/
Captum is an open source library from Meta for model interpretability in PyTorch. It helps developers understand and visualize the decisions made by deep learning models using attribution methods like Integrated Gradients and DeepLIFT.
Learn more: https://captum.ai/
CNN classification model of bird vocalizations achieved 94.8% accuracy; XAI techniques (LIME, SHAP, DeepLIFT, Grad-CAM) provide complementary explanations of model predictions, improving trust and interoperability.
Explainability of CNN Based Classification Models for Acoustic Signal
Zubair Faruqui, Mackenzie S. McIntire, Rahul Dubey, Jay McEntee
arxiv.org
September 11, 2025 at 11:18 AM
CNN classification model of bird vocalizations achieved 94.8% accuracy; XAI techniques (LIME, SHAP, DeepLIFT, Grad-CAM) provide complementary explanations of model predictions, improving trust and interoperability.
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
Motivation Characterization of regulatory elements in DNA sequence is a key task in functional genom
www.biorxiv.org
November 21, 2023 at 2:50 AM
DARDN: Identifying transcription factor binding motifs from long DNA sequences using multi-CNNs and DeepLIFT https://www.biorxiv.org/content/10.1101/2023.11.17.567502v1
youtu.be/Z5QQ-Ab5TOY
New YT on the DeepLIFT Explainable AI method. It explores how different baselines can lead to significantly different explanations.
New YT on the DeepLIFT Explainable AI method. It explores how different baselines can lead to significantly different explanations.
DeepLIFT with Python — How Baseline Choice Affects Attributions
YouTube video by A Data Odyssey
youtu.be
March 31, 2026 at 10:52 AM
youtu.be/Z5QQ-Ab5TOY
New YT on the DeepLIFT Explainable AI method. It explores how different baselines can lead to significantly different explanations.
New YT on the DeepLIFT Explainable AI method. It explores how different baselines can lead to significantly different explanations.
My new Explainable ML video of DeepLift is out ☺️
How DeepLIFT Explains Neural Network Predictions | Explainable AI for Computer Vision
YouTube video by A Data Odyssey
youtu.be
March 23, 2026 at 4:05 PM
My new Explainable ML video of DeepLift is out ☺️
I've made some updates to my explainable AI course. All the gradient-based methods now have articles. Videos coming soon :)
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
Explainable AI for Computer Vision: Free Python Course - A Data Odyssey
A free Explainable AI (XAI) course for the theory and Python code: Grad-CAM, SHAP, Occlusion, DeepLIFT and Integrated Gradients and more!
adataodyssey.com
March 10, 2026 at 11:46 AM
I've made some updates to my explainable AI course. All the gradient-based methods now have articles. Videos coming soon :)
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
Cover image from my article on DeepLIFT :)
March 10, 2026 at 11:31 AM
Cover image from my article on DeepLIFT :)
An animation for an upcoming video on DeepLIFT. This is by far the most time-consuming one I've made.
February 26, 2026 at 1:36 PM
An animation for an upcoming video on DeepLIFT. This is by far the most time-consuming one I've made.
Explainable AI (XAI) is about illuminating black-box machine learning models and explaining them in a way that we can understand.
I'm fascinated with this topic. But learning about it was a struggle as there's not much educational content out there. So I made a course. And it's free!
I'm fascinated with this topic. But learning about it was a struggle as there's not much educational content out there. So I made a course. And it's free!
Explainable AI for Computer Vision: Free Python Course - A Data Odyssey
A free course for the theory and Python code for XAI methods including Grad-CAM, SHAP, Occlusion, DeepLIFT, Integrated Gradients and Deconvolution.
adataodyssey.com
February 26, 2026 at 9:11 AM
Explainable AI (XAI) is about illuminating black-box machine learning models and explaining them in a way that we can understand.
I'm fascinated with this topic. But learning about it was a struggle as there's not much educational content out there. So I made a course. And it's free!
I'm fascinated with this topic. But learning about it was a struggle as there's not much educational content out there. So I made a course. And it's free!
Working on the DeepLIFT section of my course
February 6, 2026 at 3:02 PM
Working on the DeepLIFT section of my course
I'm finally working on my course again. Will have some new videos out soon!
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
Explainable AI for Computer Vision: Free Python Course - A Data Odyssey
A free course for the theory and Python code for XAI methods including Grad-CAM, SHAP, Occlusion, DeepLIFT, Integrated Gradients and Deconvolution.
adataodyssey.com
January 23, 2026 at 6:48 AM
I'm finally working on my course again. Will have some new videos out soon!
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
The limitations of raw/vanilla gradients when it comes to explaining model predictions. This is why we need methods like DeepLIFT or Integrated Gradients.
January 21, 2026 at 1:35 PM
The limitations of raw/vanilla gradients when it comes to explaining model predictions. This is why we need methods like DeepLIFT or Integrated Gradients.
Still in a work-in-progress, but you can access my free XAI course below. I'd apprecaite any feedback 😊
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
Explainable AI for Computer Vision: Free Python Course - A Data Odyssey
A free course for the theory and Python code for XAI methods including Grad-CAM, SHAP, Occlusion, DeepLIFT, Integrated Gradients and Deconvolution.
adataodyssey.com
December 8, 2024 at 7:38 PM
Still in a work-in-progress, but you can access my free XAI course below. I'd apprecaite any feedback 😊
adataodyssey.com/xai-for-cv/
adataodyssey.com/xai-for-cv/
Thank you Jacob, that's a huge compliment! Many of the concepts are only there thanks to the inventions by everyone in our field, with really spectacular advances the last few years. Yours and Anshul's foremost :-) Where would anyone be without tfmodisco, deeplift or tangermeme?
April 25, 2025 at 1:37 PM
Thank you Jacob, that's a huge compliment! Many of the concepts are only there thanks to the inventions by everyone in our field, with really spectacular advances the last few years. Yours and Anshul's foremost :-) Where would anyone be without tfmodisco, deeplift or tangermeme?
And then we stratified off-target base edits in non-coding loci based on their predicted consequences on the epigenome. In a case study, an intergenic off-target edit overlaps multiple motifs - our models predict that it specifically disrupts an AP-1 site. Much more in the paper, check out Tong's 🧵!
November 7, 2025 at 6:38 PM
And then we stratified off-target base edits in non-coding loci based on their predicted consequences on the epigenome. In a case study, an intergenic off-target edit overlaps multiple motifs - our models predict that it specifically disrupts an AP-1 site. Much more in the paper, check out Tong's 🧵!
We used model interpretation techniques (DeepLIFT/SHAP & TF-MoDISco) to score the contribution of every nucleotide to accessibility, and discover recurrent patterns of sequences predictive of local chromatin accessibility - and these patterns turned out to primarily resemble TF binding motifs!
May 3, 2025 at 6:27 PM
We used model interpretation techniques (DeepLIFT/SHAP & TF-MoDISco) to score the contribution of every nucleotide to accessibility, and discover recurrent patterns of sequences predictive of local chromatin accessibility - and these patterns turned out to primarily resemble TF binding motifs!
We used model interpretation techniques (DeepLIFT/SHAP & TF-MoDISco) to score the contribution of every nucleotide to accessibility, and discover recurrent patterns of sequences predictive of local chromatin accessibility - and these patterns turned out to primarily resemble TF binding motifs!
May 3, 2025 at 6:02 PM
We used model interpretation techniques (DeepLIFT/SHAP & TF-MoDISco) to score the contribution of every nucleotide to accessibility, and discover recurrent patterns of sequences predictive of local chromatin accessibility - and these patterns turned out to primarily resemble TF binding motifs!
