1. What are the contributions mentioned in the paper "Reverse-complement parameter sharing improves deep learning models for genomics" ?
Here, the authors show that conventional deep learning models that do not explicitly model this property can produce substantially different predictions on forward and reverse-complement versions of the same DNA sequence.. The authors present four new convolutional neural network layers that leverage the reverse-complement property of genomic DNA sequence by sharing parameters between forward and reverse-complement representations in the model.. Using experiments on simulated and in vivo transcription factor binding data, the authors show that their proposed architectures lead to improved performance, faster learning and cleaner internal representations compared to conventional architectures trained on the same data.
read more
2. What is the number of parameters in a convolutional layer?
For a standard batch normalization layer following a convolutional layer, the number of parameters is (2 × number_of_input_channels), with one set of γ and one set of β parameters per channel.
read more
3. Why are deep learning models particularly enticing for this problem?
Deep learning models are particularly enticing for this problem because they are capable of inducing hierarchical, predictive patterns of increasing complexity from raw input DNA sequences without relying on explicit featurization (such as featurization into k-mers).
read more
4. How can the authors model the natural RC property of inputs such as DNA sequence?
By introducing four new layers that share weights between forward and RC representations, the authors can enable neural networks to model the natural RC property of inputs such as DNA sequence.
read more
![Fig. 8. Weight sharing in a RC convolutional layer. We depict two filters of size [4,3] that will scan an input with 4 channels. The input channels may be one-hot encoded ACGT (as is the case for the first convolutional layer) or may be the output of a previous RC convolutional layer. The only constraint on the input is that the RC of the channel at index i is present at index n − 1 − i (where n is the number of input channels; note that ACGT satisfies this constraint). The two output filters are individually reverse complemented (length and channel axes are inverted) and then appended as additional channels in the opposite order so as to maintain the property that the reverse complement of the channel at index i is present at n − 1 − i (where n is the number of output channels). This structure allows us to compute reverse complements in downstream layers by simply inverting the length and channel dimensions.](/figures/fig-8-weight-sharing-in-a-rc-convolutional-layer-we-depict-9wyatmvn.webp)
