A Deep Neural Network for Predicting and Engineering Alternative Polyadenylation

Bogard, Nicholas; Linder, Johannes; Rosenberg, Alexander; Seelig, Georg

doi:10.1016/j.cell.2019.04.046

Cited by 169 publications

(280 citation statements)

References 89 publications

Supporting

Mentioning

244

Contrasting

Unclassified

Order By: Relevance

“…The 100 toeholds with the lowest ON/OFF ratio were one-hot encoded and fed as inputs to the static model. Target ON and OFF values of 0.99 and 0.001, respectively, were set and supplied to SeqProp, an open-source python package that enables streamlined development 22 of gradient ascent pipelines for genomic and RNA biology applications 22 . Toehold design constraints were incorporated into the loss function, such that the modified toehold switch contained the conserved sequences and base pairing within the hairpin was preserved.…”

Section: Storm: a Sequence-based Toehold Optimization And Redesign Modelmentioning

confidence: 99%

“…To improve toehold switch design and prediction, we took inspiration from the field of machine learning and deep learning. Machine learning approaches have been applied successfully to synthetic biology 1,19 , as exemplified in a recent study by Yang et al 19 using a 'white box' approach to extract antibiotic mechanisms of action, and in motif finding and DNA sequence prediction tasks [20][21][22] . As typical deep learning approaches require large amounts of training examples, we used a dataset from Angenent-Mari et al (2019) 23 that experimentally characterized 91,534 toehold switches, in both active and repressed states.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Sequence-to-function deep learning frameworks for synthetic biology

Valeri

et al. 2019

Preprint

View full text Add to dashboard Cite

While synthetic biology has revolutionized our approaches to medicine, agriculture, and energy, the design of novel circuit components beyond nature-inspired templates can prove itself challenging without well-established design rules. Toehold switches -programmable nucleic acid sensors -face an analogous prediction and design bottleneck: our limited understanding of how sequence impacts functionality can require expensive, time-consuming screens for effective switches. Here, we introduce the Sequence-based Toehold Optimization and Redesign Model (STORM), a deep learning architecture that applies gradient ascent to re-engineer poorlyperforming toeholds. Based on a dataset of 91,534 toehold switches, we examined convolutional filters and saliency maps of sequences to interpret our sequence-to-function model, identifying hot spots where mutations change toehold effectiveness and features unique to high-performing switches. Our modeling platform provides frameworks for future toehold selection, augmenting our ability to construct potent synthetic circuit components and precision diagnostics, and enabling straightforward translation of this in silico workflow to other circuitries. 3Main Text:

show abstract

Section: Storm: a Sequence-based Toehold Optimization And Redesign Modelmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Sequence-to-function deep learning frameworks for synthetic biology

Valeri

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…Convolutional neural networks (CNNs) in particular can learn the combinatorial patterns embedded within input examples without the need for alignment of examples. Recent studies have begun to take advantage of CNNs to tackle aspects of gene regulation 6 , including models that predict chromatin state [7][8][9] , TF binding 10,11 , polyadenylation 12 , or gene expression 7,13 solely on the basis of DNA (100bp-1Mb) or RNA sequences, with the potential to ferret out relevant motifs.…”

Section: Introductionmentioning

confidence: 99%

Learning immune cell differentiation

Maslova

Ramirez

et al. 2019

Preprint

View full text Add to dashboard Cite

The mammalian genome contains several million cis-regulatory elements, whose differential activity marked by open chromatin determines organogenesis and differentiation. This activity is itself embedded in the DNA sequence, decoded by sequence-specific transcription factors. Leveraging a granular ATAC-seq atlas of chromatin activity across 81 immune cell-types we show that a convolutional neural network ("AI-TAC") can learn to infer cell-type-specific chromatin activity solely from the DNA sequence. AI-TAC does so by rediscovering, with astonishing precision, binding motifs for known regulators, and some unknown ones, mapping them with high concordance to positions validated by ChIP-seq data. AI-TAC also uncovers combinatorial influences, establishing a hierarchy of transcription factors (TFs) and their interactions involved in immunocyte specification, with intriguingly different strategies between lineages. Mouse-trained AI-TAC can parse human DNA, revealing a strikingly similar ranking of influential TFs. Thus, Deep Learning can reveal the regulatory syntax that drives the full differentiative complexity of the immune system.

show abstract

“…Indeed, [24] has recently explored the possibility of using recurrent neural networks for biophysically modeling gene expression in flies. In principle, it should be possible to apply similar strategies to MPRAs performed on genome-wide or random sequence libraries [38,4,39,40,6]. But across all of biology, very few individual cis-regulatory sequences have been characterized to the level that our modeling strategy aims to elucidate.…”

mentioning

confidence: 99%

Biophysical models of cis-regulation as interpretable neural networks

Tareen

Kinney

2019

Preprint

View full text Add to dashboard Cite

The adoption of deep learning techniques in genomics has been hindered by the difficulty of mechanistically interpreting the models that these techniques produce. In recent years, a variety of post-hoc attribution methods have been proposed for addressing this neural network interpretability problem in the context of gene regulation. Here we describe a complementary way to address this problem. Our approach is based on the observation that two large classes of biophysical models for cis-regulatory mechanisms can be expressed as deep neural networks in which nodes and weights have explicit physiochemical interpretations. We also demonstrate how such biophysical networks can be rapidly learned, using modern deep learning frameworks, from the data produced by certain types of massively parallel reporter assays (MPRAs). These results suggest a scalable strategy for using MPRAs to systematically characterize the biophysical basis of gene regulation in a wide range of biological contexts. They also highlight gene regulation as an ideal venue for the development of scientifically interpretable approaches to deep learning. 1 To reduce notational burden, all ∆G values are assume to be in thermal units. At 37 • C, one thermal unit is 1 kBT = 0.62 kcal/mol, where kB is Boltzmann's constant and T is temperature.

show abstract

A Deep Neural Network for Predicting and Engineering Alternative Polyadenylation

Cited by 169 publications

References 89 publications

Sequence-to-function deep learning frameworks for synthetic biology

Sequence-to-function deep learning frameworks for synthetic biology

Learning immune cell differentiation

Biophysical models of cis-regulation as interpretable neural networks

Contact Info

Product

Resources

About