Granger causal inference on DAGs identifies genomic loci regulating transcription

Authors: Alexander P Wu, Rohit Singh, Bonnie Berger

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We applied Gr ID-Net on multimodal single-cell assays that profile chromatin accessibility (ATAC-seq) and gene expression (RNA-seq) in the same cell and show that it dramatically outperforms existing methods for inferring regulatory locus gene links, achieving up to 71% greater agreement with independent population genetics-based estimates.
Researcher Affiliation Academia 1 Computer Science and Artificial Intelligence Laboratory, MIT, Cambridge, MA 02139, USA 2 Department of Mathematics, MIT, Cambridge, MA 02139, USA {rsingh,alexwu,bab}@csail.mit.edu
Pseudocode No The paper describes the model architecture using mathematical equations (Eqn. 4, 5, 6, 7) and textual descriptions but does not provide a formal pseudocode block or algorithm listing.
Open Source Code Yes The code for Gr ID-Net is available at https://github.com/alexw16/gridnet.
Open Datasets Yes We analyzed three single-cell multimodal datasets that characterize a range of dynamic processes, including cancer drug response and cellular differentiation (Cao et al., 2018; Chen et al., 2019; Ma et al., 2020).
Dataset Splits No The paper describes training details like learning rate, epochs, and minibatch size for the Gr ID-Net models. It also mentions comparing full and reduced models using an F-test. However, it does not explicitly define or specify standard train/validation/test splits of the datasets (e.g., percentages or counts for each subset) that are typically used to evaluate model generalization.
Hardware Specification Yes All models were implemented in Py Torch and trained on a single NVIDIA Tesla V100 GPU.
Software Dependencies No The paper states, “All models were implemented in Py Torch” and mentions using “Homer” and “statsmodels package,” but it does not specify version numbers for any of these software dependencies.
Experiment Setup Yes Gr ID-Net models were trained using the Adam optimizer with a learning rate of 0.001 for 20 epochs or until convergence (defined to be the point at which the relative change in the loss function is less than 0.1/|P| across consecutive epochs). A minibatch size of 1024 candidate peak gene pairs was used during training, and trainable parameters in the model were initialized using Glorot initialization (Bengio & Glorot, 2010). All Gr ID-Net models consisted of L = 10 GNN layers; the architectures of the three sub-models ( h(reduced) y , h(full) y , and h(full) x ) were identical but separate.