Running PRESCIENT using cloud GPU resources
Training PRESCIENT models using CPU can, in some cases, be time intensive. For this reason, here we provide the path of least resistance to running the whole training pipeline using cloud computing services.This Google Colab notebook provides a free demo of how PRESCIENT can be run with GPU acceleration. PRESCIENT is also available via PyPI for installation and usage on a local workstation, on a compute server, or on other cloud services, including other cloud notebook servers such as Gcloud AI platform and AWS with jupyter. Our example uses Google Colab which offers free access to public GPUs hosted by Google.
Usage and analysis notebooks
Here, we present different examples of various individual pre-processing and analyses steps using PRESCIENT. This will be a growing list of notebooks as we add functionality and update PRESCIENT.
To demonstrate, we use longitudinal scRNA-seq data from Veres et al. 2019 Stage 5 pancreatic
beta-cell differentiation as an example dataset. We present only a demo version of this analysis here for simplicity, please refer to the paper analysis github repo,
for in-depth analyses used in the paper.
To run the following notebooks with the example data, download the following directories containing raw data and trained PRESCIENT models:
Estimating growth rates
PRESCIENT models can incorporate proliferation during training via computing a "growth weight" per cell in a scRNA-seq dataset. This was shown to greatly improve model performance, and we conclude that growth is important for model performance. When lineage tracing is available, empirical growth rates can be computed, however since the primary use case for PRESCIENT is in the absence of lineage tracing, we provide a function for estimating growth from proliferative/death gene signatures. Here, these gene sets are provided in the downloaded data folder and used in the notebook, but you can use any gene set (gst) from common sources like MSigDB.