Reverse-Complement Equivariant Networks for DNA Sequences

Authors: Vincent Mallet, Jean-Philippe Vert

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show experimentally that the new architectures can outperform existing ones. We assess the performance of various equivariant architectures on a set of three binary prediction and four sequence prediction problems used by Zhou et al. [45]
Researcher Affiliation Collaboration Vincent Mallet Structural Bioinformatics Unit, Department of Structural Biology and Chemistry, Institut Pasteur, CNRS UMR3528, C3BI, USR3756 Mines Paris Tech, PSL University, Center for Computational Biology vincent.mallet96@gmail.com Jean-Philippe Vert Google Research, Brain team, Paris jpvert@google.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Finally, we implemented new linear and nonlinear equivariant layers and make all these equivariant layers available in Keras [5] and Pytorch [31].3 code available at https://github.com/Vincentx15/Equi-RC
Open Datasets Yes We assess the performance of various equivariant architectures on a set of three binary prediction and four sequence prediction problems used by Zhou et al. [45] to assess the performance of RCPS networks. The binary classification problems aim to predict if a DNA sequence binds to three transcription factors (TFs), based on genome-wide binarized TF-Ch IP-seq data for Max, Ctcf and Spi1 in the GM12878 lymphoblastoid cell-line [34].
Dataset Splits No The paper mentions using a validation set for hyperparameter tuning, but does not provide specific details on the dataset splits (percentages or counts) in the main text.
Hardware Specification Yes All experiments were run on a single GPU (either a GTX1080 or a RTX6000), with 20 CPU cores.
Software Dependencies No The paper mentions Keras [5] and Pytorch [31] as frameworks used but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We train equivariant models with different combinations of hyperparameters on the training set and assess their performance on the validation set, repeating the process ten times with different random seeds. ... We combine the regular and '+1' dimensions with Re Lu activations and the '-1' dimensions with a tanh activation.