Getting Started - High-level API

Getting Started - High-level API#

  • Ladder offers a high-level API for running models in settings where there are multiple conditions.

  • It is very easy to get started, as long as you have your AnnData ready! See below for starting out with Patches.

import anndata as ad
import scanpy as sc

from ladder.data import get_data
from ladder.scripts import CrossConditionWorkflow, InterpretableWorkflow # Workflow APIs, read more on which one you'd like for your case!

# Download the data object - skip if you have your data
get_data("Vu")

# Load the anndata object
anndata = ad.read_h5ad("data/vu_2022_ay_wh.h5ad")

## Do optional preprocessing
## IMPORTANT : Model input (anndata.X should be raw counts!!)
anndata.layers["normalized"] = anndata.X

# Find/subset HVGs & swap to raw counts
sc.pp.highly_variable_genes(anndata, n_top_genes=3000, batch_key="sample")
sc.pl.highly_variable_genes(anndata)

anndata = anndata[:, anndata.var["highly_variable"]]
anndata.X = anndata.layers["counts"]

# Initialize workflow object
workflow = InterpretableWorkflow(anndata, verbose=True, random_seed=42)

# Define the condition classes & batch key to prepare the data
factors = ["time", "age", "broad_type"]
workflow.prep_model(factors, batch_key="sample", model_type='Patches', model_args={'ld_normalize' : True})

# Train the model
workflow.run_model(max_epochs=2000, convergence_threshold=1e-5, convergence_window=2000) # Lower the convergence threshold & increase the window if you need a more accurate model, will increase training time