Learning Identifiable Factorized Causal Representations of Cellular Responses

Authors: Haiyi Mao, Romain Lopez, Kai Liu, Jan-Christian Huetter, David Richmond, Panayiotis Benos, Lin Qiu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Then, we present our implementation of FCR, and empirically demonstrate that it outperforms state-of-the-art baselines in various tasks across four single-cell datasets. To evaluate the efficacy and robustness of the FCR method, we conducted our study on four real single-cell perturbation datasets (Appendix E).
Researcher Affiliation Collaboration 1Genentech 2University of Pittsburgh 3 Stanford University 4 University of Florida
Pseudocode Yes Algorithm 1 Training of FCR
Open Source Code Yes The code is available on Git Hub (https://github.com/Genentech/fcr).
Open Datasets Yes To evaluate the efficacy and robustness of the FCR method, we conducted our study on four real single-cell perturbation datasets (Appendix E). The first of these is the sci Plex dataset (Srivatsan et al., 2020), which provides insights into the impact of several HDAC (Histone Deacetylase) inhibitors on a total of 11,755 cells from three distinct cell lines: A549, K562, and MCF7. The subsequent three datasets are sourced from (Mc Farland et al., 2020).
Dataset Splits Yes We split the data into four datasets: train/validation/test/prediction, following the setup from previous works (Lotfollahi et al., 2023; Wu et al., 2023). First we hold out the 20% of the control cells for the final cellular prediction tasks (prediction). Second, we hold 20% of the rest of data for the task of clustering and statistical test (test). Third, the data excluding the prediction and clustering/test sets are split into training and validation sets with a four-to-one ratio.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., 'Python 3.8, PyTorch 1.9, and CUDA 11.1') or specific versions of solvers/packages used.
Experiment Setup Yes For the sci Plex datasets, the dimensions are as follows: zx is 32, ztx is 64, and zt is 32. Additionally, we set the hyperparameters to ω1 = 3.0, ω2 = 3.0, and ω3 = 5.0, with a batch size of 2046. Additionally, the autoencoder learning rate is set to 3 10 4, the discriminator learning rates are also 3 10 4, and the number of discriminator training steps is 10.