Modelling Cellular Perturbations with the Sparse Additive Mechanism Shift Variational Autoencoder

Authors: Michael Bereket, Theofanis Karaletsos

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate SAMS-VAE both quantitatively and qualitatively on a range of tasks using two popular single cell sequencing datasets.
Researcher Affiliation Collaboration Michael Bereket insitro mbereket@stanford.edu Theofanis Karaletsos insitro* theofanis@karaletsos.com Research supporting this publication conducted while authors were employed at insitro
Pseudocode Yes Algorithm 1 SAMS-VAE generative process
Open Source Code Yes Code availability Our code, which includes implementations of all models and experiment configurations, is available at https://github.com/insitro/sams-vae.
Open Datasets Yes Dataset To assess model generalization to held out samples under individual perturbations, we analyze a subset of the genome-wide CRISPR interference (CRISPRi) perturb-seq dataset from Replogle et al. [17]... We analyze the CRISPR activation (CRISPRa) perturb-seq screen from Norman et al. [15]...
Dataset Splits Yes We randomly sample train, validation, and test splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory, or cloud instance types used for experiments.
Software Dependencies No The paper mentions `scikit-learn [16]` but does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes Each model is trained with a 100 dimensional latent space and MLP encoders and decoders with a single hidden layer of dimension 400 (see Section A.4 for full training details). Based on validation performance and sparsity, a Beta(1, 2) prior was selected for the SVAE+ mask, and a Bern(0.001) prior was selected for SAMS-VAE.