pimms-learn 0.5.0

Imputing (MS-based prote-) omics data using self supervised deep learning models.

PIMMS stands for Proteomics Imputation Modeling Mass Spectrometry and is a hommage to our dear British friends who are missing as part of the EU for far too long already (Pimms is a British summer drink).

We published the work in Nature Communications as open access:

Webel, H., Niu, L., Nielsen, A.B. et al.
Imputation of label-free quantitative mass spectrometry-based proteomics data using self-supervised deep learning.
Nat Commun 15, 5405 (2024).
https://doi.org/10.1038/s41467-024-48711-5

We provide new functionality as a python package for simple use (in notebooks) and a workflow for comparsion with other methdos.

For any questions, please open an issue or contact me directly.

Getting started

The models can be used with the scikit-learn interface in the spirit of other scikit-learn imputers. You can try this using our tutorial in colab:

It uses the scikit-learn interface. The PIMMS models in the scikit-learn interface can be executed on the entire data or by specifying a valdiation split for checking training process. In our experiments overfitting wasn't a big issue, but it's easy to check.

Install Python package

For interactive use of the models provided in PIMMS, you can use our python package pimms-learn. The interface is similar to scikit-learn. The package is then availabe as pimmslearn for import in your Python session.

pip install pimms-learn
# import pimmslearn # in your python script

The most basic use for imputation is using a DataFrame.

import numpy as np
import pandas as pd
from pimmslearn.sklearn.ae_transformer import AETransformer
from pimmslearn.sklearn.cf_transformer import CollaborativeFilteringTransformer

fn_intensities = ('https://raw.githubusercontent.com/RasmussenLab/pimms/main/'
                  'project/data/dev_datasets/HeLa_6070/protein_groups_wide_N50.csv')
index_name = 'Sample ID'
column_name = 'protein group'
value_name = 'intensity'

df = pd.read_csv(fn_intensities, index_col=0)
df = np.log2(df + 1)

df.index.name = index_name  # already set
df.columns.name = column_name  # not set due to csv disk file format

# df # see details below to see a preview of the DataFrame

# use the Denoising or Variational Autoencoder
model = AETransformer(
    model='DAE', # or 'VAE'
    hidden_layers=[512,],
    latent_dim=50, # dimension of joint sample and item embedding
    batch_size=10,
)
model.fit(df,
          cuda=False,
          epochs_max=100,
          )
df_imputed = model.transform(df)

# or use the collaborative filtering model
series = df.stack()
series.name = value_name  # ! important
model = CollaborativeFilteringTransformer(
    target_column=value_name,
    sample_column=index_name,
    item_column=column_name,
    n_factors=30, # dimension of separate sample and item embedding
    batch_size = 4096
)
model.fit(series, cuda=False, epochs_max=20)
df_imputed = model.transform(series).unstack()

🔍 see log2 transformed DataFrame

First 10 rows and 10 columns. notice that the indices are named:

Sample ID	AAAS	AACS	AAMDC	AAMP	AAR2	AARS	AARS2	AASDHPPT	AATF	ABCB10
protein group
2019_12_18_14_35_Q-Exactive-HF-X-Orbitrap_6070	28.3493	26.1332	nan	26.7769	27.2478	32.1949	27.1526	27.8721	28.6025	26.1103
2019_12_19_19_48_Q-Exactive-HF-X-Orbitrap_6070	27.6574	25.0186	24.2362	26.2707	27.2107	31.9792	26.5302	28.1915	27.9419	25.7349
2019_12_20_14_15_Q-Exactive-HF-X-Orbitrap_6070	28.3522	23.7405	nan	27.0979	27.3774	32.8845	27.5145	28.4756	28.7709	26.7868
2019_12_27_12_29_Q-Exactive-HF-X-Orbitrap_6070	26.8255	nan	nan	26.2563	nan	31.9264	26.1569	27.6349	27.8508	25.346
2019_12_29_15_06_Q-Exactive-HF-X-Orbitrap_6070	27.4037	26.9485	23.8644	26.9816	26.5198	31.8438	25.3421	27.4164	27.4741	nan
2019_12_29_18_18_Q-Exactive-HF-X-Orbitrap_6070	27.8913	26.481	26.3475	27.8494	26.917	32.2737	nan	27.4041	28.0811	nan
2020_01_02_17_38_Q-Exactive-HF-X-Orbitrap_6070	25.4983	nan	nan	nan	nan	30.2256	nan	23.8013	25.1304	nan
2020_01_03_11_17_Q-Exactive-HF-X-Orbitrap_6070	27.3519	nan	24.4331	25.2752	24.8459	30.9793	nan	24.893	25.3238	nan
2020_01_03_16_58_Q-Exactive-HF-X-Orbitrap_6070	27.6197	25.6238	23.5204	27.1356	25.9713	31.4154	25.3596	25.1191	25.75	nan
2020_01_03_20_10_Q-Exactive-HF-X-Orbitrap_6070	27.2998	nan	25.6604	27.7328	26.8965	31.4546	25.4369	26.8135	26.2008	nan
...

For hints on how to add validation (and potentially test data) to use early stopping, see the tutorial:

PIMMS comparison workflow and differential analysis workflow

The PIMMS comparison workflow is a snakemake workflow that runs the all selected PIMMS models and R-models on a user-provided dataset and compares the results. An example for a publickly available Alzheimer dataset on the protein groups level is re-built regularly and available at: rasmussenlab.org/pimms

It is built on top of

• the Snakefile_v2.smk (v2 of imputation workflow), specified in on configuration
• the Snakefile_ald_comparision workflow for differential analysis

The associated notebooks are index with 01_* for the comparsion workflow and 10_* for the differential analysis workflow. The project folder can be copied separately to any location if the package is installed. It's standalone folder. It's main folders are:

# project folder:
project
│   README.md # see description of notebooks and hints on execution in project folder
|---config # configuration files for experiments ("workflows")
|---data # data for experiments
|---runs # results of experiments
|---src # source code or binaries for some R packges
|---tutorials # some tutorials for libraries used in the project
|---workflow # snakemake workflows

To re-execute the entire workflow locally, have a look at the configuration files for the published Alzheimer workflow:

• config/alzheimer_study/config.yaml
• config/alzheimer_study/comparsion.yaml

To execute that workflow, follow the Setup instructions below and run the following commands in the project folder:

# being in the project folder
snakemake -s workflow/Snakefile_v2.smk --configfile config/alzheimer_study/config.yaml -p -c1 -n # one core/process, dry-run
snakemake -s workflow/Snakefile_v2.smk --configfile config/alzheimer_study/config.yaml -p -c2 # two cores/process, execute
# after imputation workflow, execute the comparison workflow
snakemake -s workflow/Snakefile_ald_comparison.smk --configfile config/alzheimer_study/comparison.yaml -p -c1
# If you want to build the website locally: https://www.rasmussenlab.org/pimms/
pip install .[docs]
pimms-setup-imputation-comparison -f project/runs/alzheimer_study/
pimms-add-diff-comp -f project/runs/alzheimer_study/ -sf_cp project/runs/alzheimer_study/diff_analysis/AD
cd project/runs/alzheimer_study/
sphinx-build -n --keep-going -b html ./ ./_build/
# open ./_build/index.html

Notebooks as scripts using papermill

The above workflow is based on notebooks as scripts, which can then be rendered as html files.'Using jupytext also python percentage script versions are saved.

If you want to run a specific model on your data, you can run notebooks prefixed with 01_, i.e. project/01_*.ipynb after creating hte appropriate data split. Start by cloning the repository.

# navigat to your desired folder
git clone https://github.com/RasmussenLab/pimms.git # get all notebooks
cd project # project folder as pwd
# pip install pimms-learn papermill # if not already installed
papermill 01_0_split_data.ipynb --help-notebook
papermill 01_1_train_vae.ipynb --help-notebook

⚠️ Mistyped argument names won't throw an error when using papermill, but a warning is printed on the console thanks to my contributions:)

Setup workflow and development environment

Either (1) install one big conda environment based on an environment file, or (2) install packages using a mix of conda and pip, or (3) use snakemake separately with rule specific conda environments.

Setup comparison workflow (1)

The core funtionality is available as a standalone software on PyPI under the name pimms-learn. However, running the entire snakemake workflow in enabled using conda (or mamba) and pip to setup an analysis environment. For a detailed description of setting up conda (or mamba), see instructions on setting up a virtual environment.

Download the repository:

git clone https://github.com/RasmussenLab/pimms.git
cd pimms

Using conda (or mamba), install the dependencies and the package in editable mode

# from main folder of repository (containing environment.yml)
conda env create -n pimms -f environment.yml # slower
mamba env create -n pimms -f environment.yml # faster, less then 5mins

If on Mac M1, M2 or having otherwise issue using your accelerator (e.g. GPUs): Install the pytorch dependencies first, then the rest of the environment:

Install pytorch first (2)

⚠️ We currently see issues with some installations on M1 chips. A dependency for one workflow is polars, which causes the issue. This should be fixed now for general use by delayed import of mrmr-selection in njab. If you encounter issues, please open an issue.

Check how to install pytorch for your system here.

• select the version compatible with your cuda version if you have an nvidia gpu or a Mac M-chip.

conda create -n pimms python=3.9 pip
conda activate pimms
# Follow instructions on https://pytorch.org/get-started: 
# CUDA is not available on MacOS, please use default package
# pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
conda install pytorch::pytorch torchvision torchaudio fastai -c pytorch -c fastai -y
pip install pimms-learn
pip install jupyterlab papermill # use run notebook interactively or as a script

cd project
# choose one of the following to test the code
jupyter lab # open 04_1_train_pimms_models.ipynb
papermill 04_1_train_pimms_models.ipynb 04_1_train_pimms_models_test.ipynb # second notebook is output
python 04_1_train_pimms_models.py # just execute the code

Let Snakemake handle installation (3)

If you only want to execute the workflow, you can use snakemake to build the environments for you:

Install snakemake e.g. using the provided snakemake_env.yml file as used in this workflow.

⚠️ Snakefile workflow for imputation v1 only support that atm.

snakemake -p -c1 --configfile config/single_dev_dataset/example/config.yaml --use-conda -n # dry-run
snakemake -p -c1 --configfile config/single_dev_dataset/example/config.yaml --use-conda # execute with one core

Troubleshooting

Trouble shoot your R installation by opening jupyter lab

# in projects folder
jupyter lab # open 01_1_train_NAGuideR.ipynb

Run example on HeLa data

Change to the project folder and see it's README You can subselect models by editing the config file: config.yaml file.

conda activate pimms # activate virtual environment
cd project # go to project folder
pwd # so be in ./pimms/project
snakemake -c1 -p -n # dryrun demo workflow, potentiall add --use-conda
snakemake -c1 -p

The demo will run an example on a small data set of 50 HeLa samples (protein groups):

1. it describes the data and does create the splits based on the example data
   • see 01_0_split_data.ipynb
2. it runs the three semi-supervised models next to some default heuristic methods
   • see 01_1_train_collab.ipynb, 01_1_train_dae.ipynb, 01_1_train_vae.ipynb
3. it creates an comparison
   • see 01_2_performance_plots.ipynb

The results are written to ./pimms/project/runs/example, including html versions of the notebooks for inspection, having the following structure:

│   01_0_split_data.html
│   01_0_split_data.ipynb
│   01_1_train_collab.html
│   01_1_train_collab.ipynb
│   01_1_train_dae.html
│   01_1_train_dae.ipynb
│   01_1_train_vae.html
│   01_1_train_vae.ipynb
│   01_2_performance_plots.html
│   01_2_performance_plots.ipynb
│   data_config.yaml
│   tree_folder.txt
|---data
|---figures
|---metrics
|---models
|---preds

The predictions of the three semi-supervised models can be found under ./pimms/project/runs/example/preds. To combine them with the observed data you can run

# ipython or python session
# be in ./pimms/project
folder_data = 'runs/example/data'
data = pimmslearn.io.datasplits.DataSplits.from_folder(
    folder_data, file_format='pkl')
observed = pd.concat([data.train_X, data.val_y, data.test_y])
# load predictions for missing values of a certain model
model = 'vae'
fpath_pred = f'runs/example/preds/pred_real_na_{model}.csv '
pred = pd.read_csv(fpath_pred, index_col=[0, 1]).squeeze()
df_imputed = pd.concat([observed, pred]).unstack()
# assert no missing values for retained features
assert df_imputed.isna().sum().sum() == 0
df_imputed

:warning: The imputation is simpler if you use the provide scikit-learn Transformer interface (see Tutorial).

Available imputation methods

Packages either are based on this repository, were referenced by NAGuideR or released recently. From the brief description in this table the exact procedure is not always clear.

Method	Package	source	links	name
CF	pimms	pip	paper	Collaborative Filtering
DAE	pimms	pip	paper	Denoising Autoencoder
VAE	pimms	pip	paper	Variational Autoencoder
ZERO	-	-	-	replace NA with 0
MINIMUM	-	-	-	replace NA with global minimum
COLMEDIAN	e1071	CRAN	-	replace NA with column median
ROWMEDIAN	e1071	CRAN	-	replace NA with row median
KNN_IMPUTE	impute	BIOCONDUCTOR	docs	k nearest neighbor imputation
SEQKNN	SeqKnn	tar file	paper	Sequential k- nearest neighbor imputation starts with feature with least missing values and re-use imputed values for not yet imputed features
BPCA	pcaMethods	BIOCONDUCTOR	paper	Bayesian PCA missing value imputation
SVDMETHOD	pcaMethods	BIOCONDUCTOR	paper	replace NA initially with zero, use k most significant eigenvalues using Singular Value Decomposition for imputation until convergence
LLS	pcaMethods	BIOCONDUCTOR	paper	Local least squares imputation of a feature based on k most correlated features
MLE	norm	CRAN		Maximum likelihood estimation
QRILC	imputeLCMD	CRAN	paper	quantile regression imputation of left-censored data, i.e. by random draws from a truncated distribution which parameters were estimated by quantile regression
MINDET	imputeLCMD	CRAN	paper	replace NA with q-quantile minimum in a sample
MINPROB	imputeLCMD	CRAN	paper	replace NA by random draws from q-quantile minimum centered distribution
IRM	VIM	CRAN	paper	iterativ robust model-based imputation (one feature at at time)
IMPSEQ	rrcovNA	CRAN	paper	Sequential imputation of missing values by minimizing the determinant of the covariance matrix with imputed values
IMPSEQROB	rrcovNA	CRAN	paper	Sequential imputation of missing values using robust estimators
MICE-NORM	mice	CRAN	paper	Multivariate Imputation by Chained Equations (MICE) using Bayesian linear regression
MICE-CART	mice	CRAN	paper	Multivariate Imputation by Chained Equations (MICE) using regression trees
TRKNN	-	script	paper	truncation k-nearest neighbor imputation
RF	missForest	CRAN	paper	Random Forest imputation (one feature at a time)
PI	-	-		Downshifted normal distribution (per sample)
GSIMP	-	script	paper	QRILC initialization and iterative Gibbs sampling with generalized linear models (glmnet) - slow
MSIMPUTE	msImpute	BIOCONDUCTOR	paper	Missing at random algorithm using low rank approximation
MSIMPUTE_MNAR	msImpute	BIOCONDUCTOR	paper	Missing not at random algorithm using low rank approximation

DreamAI and GMSimpute are not available for installation on Windows or failed to install.