Picasso — Forked

This is a fork of pachterlab/picasso. See the original repository for core documentation. This fork extends Picasso with an updated environment and applies single-cell analysis on NTM (nontuberculous mycobacteria) lung disease data.

---

What I Added

Addition	Description
env/conda_env.yaml	Updated conda environment with Python 3.10 + PyTorch ≥ 2.0
picasso_v2.ipynb	Analysis notebook applying Picasso to NTM scRNA-seq data
figures/	Generated publication-style output figures
cell_type.png	Picasso embedding colored by cell type (see below)

---

Applied Analysis: NTM Lung Disease scRNA-seq

I applied Picasso to single-cell RNA-seq data from NTM (nontuberculous mycobacteria) lung disease samples. Cells are embedded onto flipped elephant-shaped coordinates and colored by cell type identity.

Picasso embedding of NTM lung scRNA-seq data, colored by cell type

Additional figures generated from this analysis:

Figure	Description
figures/picasso_180flipped_elephantcell_type_simplified_v5.pdf	Simplified cell type map
figures/picasso_180flipped_elephantNTM_DE_eff_size_v6.pdf	Differential expression effect sizes
figures/picasso_180flipped_elephantSymptomatic_DA_lfc_v7.pdf	Symptomatic vs. non-symptomatic log fold change

---

Environment Setup (Updated for Python 3.10 + PyTorch 2.0)

The original environment files (Python 3.7) are preserved. I added a modern alternative:

conda env create -f env/conda_env.yaml
conda activate picasso_env

This installs:

• Python 3.10
• PyTorch ≥ 2.0 (with CUDA support)
• NumPy ≥ 1.23, SciPy ≥ 1.10, Pandas ≥ 2.0
• AnnData ≥ 0.9, Matplotlib ≥ 3.7

For the original Python 3.7 environments, see env/env3.7_LINUX.yml or env/env3.7_MACOS.yml.

---

Running with a Checkpoint

The checkpoints/ckpt.pth file contains a pre-trained model state. To load it:

from Picasso import Picasso

pic = Picasso(adata, mapPoints)
pic.fit(checkpt="checkpoints/ckpt.pth")

This resumes training from the saved checkpoint rather than starting from scratch.

---

Original Picasso

Picasso maps points of an input matrix to user-defined, n-dimensional shape coordinates while minimizing reconstruction error via an autoencoder neural network.

Original Picasso example — single-cell data embedded onto elephant coordinates

Examples for running Picasso can be found in examplePicasso.ipynb. The notebook can be run in Google Colab:

Elephant coordinates from Mayer et al. 2010.

---

Citation

If you use Picasso, please cite the original authors:

Pachter Lab. Picasso. GitHub. https://github.com/pachterlab/picasso