An interpretable bimodal neural network characterizes the sequence and preexisting chromatin predictors of induced transcription factor binding

Srivastava, Divyanshi; Aydın, Begüm; Mazzoni, Esteban O.; Mahony, Shaun

doi:10.1186/s13059-020-02218-6

Cited by 16 publications

(15 citation statements)

References 69 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We chose neural networks owing to their ability to learn arbitrarily complex predictive sequence patterns ( Kelley et al 2018 ; Fudenberg et al 2020 ; Avsec et al 2021a , b ; Koo et al 2021 ). In particular, hybrid convolutional and recurrent network architectures have successfully been applied to accurately predict TF binding in diverse applications ( Quang and Xie 2016 ; Quang and Xie 2019 ; Srivastava et al 2021 ). The motivation behind these architectures is that convolutional filters can encode binding site motifs and other contiguous sequence features, whereas the recurrent layers can model flexible, higher-order spatial organization of these features.…”

Section: Resultsmentioning

confidence: 99%

Domain-adaptive neural networks improve cross-species prediction of transcription factor binding

et al. 2022

Self Cite

View full text Add to dashboard Cite

The intrinsic DNA sequence preferences and cell-type specific cooperative partners of transcription factors (TFs) are typically highly conserved. Hence, despite the rapid evolutionary turnover of individual TF binding sites, predictive sequence models of cell-type specific genomic occupancy of a TF in one species should generalize to closely matched cell types in a related species. To assess the viability of cross-species TF binding prediction, we train neural networks to discriminate ChIP-seq peak locations from genomic background and evaluate their performance within and across species. Cross-species predictive performance is consistently worse than within-species performance, which we show is caused in part by species-specific repeats. To account for this domain shift, we use an augmented network architecture to automatically discourage learning of training species-specific sequence features. This domain adaptation approach corrects for prediction errors on species-specific repeats and improves overall cross-species model performance. Our results demonstrate that cross-species TF binding prediction is feasible when models account for domain shifts driven by species-specific repeats.

show abstract

Section: Resultsmentioning

confidence: 99%

Domain-adaptive neural networks improve cross-species prediction of transcription factor binding

et al. 2022

Self Cite

View full text Add to dashboard Cite

show abstract

“…It would be important for ML and computational biology researchers to work together to identify additional important steps. For example, many computational biology applications involve training a multi-modal prediction models (e.g., Srivastava et al (2021) incorporates DNA sequence with epigenetic signals to predict TF binding and Chen et al (2020) leverages both cancer histology images and genomic features for survival prediction). An open research question is to define a verification step to check whether explanations properly attributes importance scores to each modality.…”

Section: Discussionmentioning

confidence: 99%

Best Practices for Interpretable Machine Learning in Computational Biology

Chen

Yang

Cui

et al. 2022

Preprint

View full text Add to dashboard Cite

Advances in machine learning (ML) have enabled the development of next-generation prediction models for complex computational biology problems. These developments have spurred the use of interpretable machine learning (IML) to unveil fundamental biological insights through data-driven knowledge discovery. However, in general, standards and guidelines for IML usage in computational biology have not been well-characterized, representing a major gap toward fully realizing the potential of IML. Here, we introduce a workflow on the best practices for using IML methods to perform knowledge discovery which covers verification strategies that bridge data, prediction model, and explanation. We outline a workflow incorporating these verification strategies to increase an IML method's accountability, reliability, and generalizability. We contextualize our proposed workflow in a series of widely applicable computational biology problems. Together, we provide an extensive workflow with important principles for the appropriate use of IML in computational biology, paving the way for a better mechanistic understanding of ML models and advancing the ability to discover novel biological phenomena.

show abstract

“…As for CNN-plus, (i) we just simply encoded DNA sequences and chromatin accessibility signals, but more appropriate methods for dealing with the inputs need to be further explored, e.g, k -mer embedding representing high-order dependencies of nucleotides (48,49), a bimodal neural network for separately handling DNA sequences and chromatin accessibility signals (50); (ii) we just considered a simple CNN architecture composed of three convolutional layers, but advanced DL-based models also need to be further explored, e.g, hybrid neural networks (51,52), transformer architectures (53,54), since they have been proved to be equipped with stronger feature learning ability. Given the massive data being generated by the ENCODE Consortium and other large-scale efforts, there is an excellent opportunity to learn richer representations to more fully understand cell-type-specific and shared binding activities.…”

Section: Discussionmentioning

confidence: 99%

Computational prediction and characterization of cell-type-specific and shared binding sites

Zhang

2022

Preprint

View full text Add to dashboard Cite

Cell-type-specific gene expression is maintained in large part by transcription factors (TFs) selectively binding to distinct sets of sites in different cell types. Recent research works have provided evidence that such cell-type-specific binding is determined by TF’s intrinsic sequence preferences, cooperative interactions with cofactors, cell-type-specific chromatin landscapes, and 3D chromatin interactions. However, computational prediction and characterization of cell-type-specific and shared binding sites is rarely studied. In this paper, we propose two computational approaches for predicting and characterizing cell-type-specific and shared binding sites by integrating multiple types of features, in which one is based on XGBoost and another is based on convolutional neural network (CNN). To validate the performance of our proposed approaches, ChIP-seq datasets of 10 binding factors were collected from the GM12878 (lymphoblastoid) and K562 (erythroleukemic) human hematopoietic cell lines, each of which was further categorized into cell-type-specific (GM12878-specific and K562-specific) and shared binding sites. Then, multiple types of features for these binding sites were integrated to train the XGBoost-based and CNN-based models. Experimental results show that our proposed approaches significantly outperform other competing methods on three classification tasks. To explore the contribution of different features, we performed ablation experiments and feature importance analysis. Consistent with previous studies, we find that chromatin features are major contributors in which chromatin accessibility is the best predictor. Moreover, we identified independent feature contribution for cell-type-specific and shared sites through SHAP values, observing that chromatin features play a main role in the cell-type-specific sites while motif features play a main role in the shared sites. Beyond these observations, we explored the ability of the CNN-based model to predict cell-type-specific and shared binding sites by excluding or including DNase signals, showing that chromatin accessibility significantly improves the prediction performance. Besides, we investigated the generalization ability of our proposed approaches to different binding factors in the same cellular environment or to the same binding factors in the different cellular environments.

show abstract

An interpretable bimodal neural network characterizes the sequence and preexisting chromatin predictors of induced transcription factor binding

Cited by 16 publications

References 69 publications

Domain-adaptive neural networks improve cross-species prediction of transcription factor binding

Domain-adaptive neural networks improve cross-species prediction of transcription factor binding

Best Practices for Interpretable Machine Learning in Computational Biology

Computational prediction and characterization of cell-type-specific and shared binding sites

Contact Info

Product

Resources

About