Ancestry clustering

This tutorial explains how to cluster and classify genomes using 1000 Genomes data with GTM (our ugtm implementation) and t-SNE (sklearn implementation). Data files used in the following tutorial can be downloaded from https://lovingscience.com/ancestries.

Requirements

The following python packages are required:

ugtm
sklearn
altair
matplotlib
numpy
pandas

Files in directory

worldmap_1000G.py: Python script, creates interactive visualization gathering GTM, t-SNE and PCA
runGTM.py: runs GTM (using ugtm package) or t-SNE (using sklearn)
data: directory, contains csv data files
data/dataframe_1000G_noadmixed.csv: csv file, 1000 Genomes project dataframe with corresponding t-SNE and GTM coordinates

Download files

You can download files for ancestry classification using 1000 genomes Phase 3 data from here, which are already formatted for this software. In this tutorial, we will use the following files:

recoded_1000G.noadmixed.mat: 20 populations from 1000 Genomes Project (excluding MXL, ACB, ASW, ITU, STU, GIH)
recoded_1000G.raw.noadmixed.lbls3_3: the corresponding ancestry labels (GBR and CEU were merged to obtain one category for Northern/Western European ancestry)
recoded_1000G.raw.noadmixed.lbls3: the corresponding ancestry labels (GBR and CEU were not merged)
recoded_1000G.raw.noadmixed.ids: the corresponding individual IDs

You can find out how these files were created by clicking here.

Build GTM and t-SNE maps

To build a GTM with parameters [k,m,l,s] = [16,4,0.1,0.3] and 10 principal components, run the following command:

python runGTM.py --model GTM --data recoded_1000G.noadmixed.mat \
--labels recoded_1000G.raw.noadmixed.lbls3 --labeltype discrete \
--out outputname --pca --n_components 10 --regularization 0.1 \
--rbf_width_factor 0.3 --missing --missing_strategy median \
--random_state 8 --ids recoded_1000G.raw.noadmixed.ids

It should be noted that our genotype file has missing values that we are handling with the --missing and --missing strategy options. You should obtain a pdf and an html file. The html file looks like this: 1000G_GTM_20populations.html

To build a t-SNE map, run:

python runGTM.py --model GTM --data recoded_1000G.noadmixed.mat \
--labels recoded_1000G.raw.noadmixed.lbls3 --labeltype discrete \
--out outputname --pca --n_components 10 \
--missing --missing_strategy median \
--random_state 8 --ids recoded_1000G.raw.noadmixed.ids

Click here to access the t-SNE map: 1000G_t-SNE_20populations.html

Evaluation of classification performances in a crossvalidation experiment, compare GTM and linear SVM:

python runGTM.py --model GTM --data recoded_1000G.noadmixed.mat \
--labels recoded_1000G.raw.noadmixed.lbls3_3 --labeltype discrete \
--out outputname --pca --n_components 10 \
--missing --missing_strategy median \
--random_state 8 --crossvalidate

This will give us per-class reports. Default class priors are equiprobable (cf. --prior option), which is generally only OK if classes are balanced. For imbalanced classes, use "--prior estimated" option.

Train on provided data and project a test set onto the map:

The great thing about generative topographic mapping (GTM) is that we can project external test sets on the map without having to re-train the map. The ugtm package also includes some nice functions for classification models and generates posterior probabilities for test set individuals (--test) to belong to a specific class, based on the class labels (--labels) of the training set (--data).

python runGTM.py --model GTM --data recoded_1000G.noadmixed.mat \
--test recoded_1000G_MXL.mat --labels recoded_1000G.raw.noadmixed.lbls3_3 \
--labeltype discrete --out outputname --pca --n_components 10 \
--missing --missing_strategy median \
--random_state 8

This will give us:

predictions for individuals (output_indiv_predictions.csv)
posterior probabilities for each ancestry (output_indiv_probabilities.csv)
posterior probabilities for the whole test set (output_group_probabilities.csv)
a map with projected test set colored in black.

The projection for MXL population (Mexicans) can be visualized here: 1000G_GTM_projection_MXL.html

Addendum 1: map based on AFR superpopulation only

To construct t-SNE and GTM maps based on AFR populations:

Download:
- recoded_1000G.noadmixed.AFR.mat: African (AFR) populations from 1000 Genomes Project (excluding ACB and ASW)
- recoded_1000G.raw.noadmixed.AFR.lbls3: the corresponding ancestry labels
- recoded_1000G.raw.noadmixed.AFR.ids: the corresponding individual IDs

Build GTM and t-SNE:

GTM (cf. output)

python runGTM.py --model GTM --data recoded_1000G.noadmixed.AFR.mat \
--labels recoded_1000G.raw.noadmixed.AFR.lbls3 --labeltype discrete \
--out 1000G_GTM_AFR --pca --n_components 10 \
--regularization 0.1 --rbf_width_factor 0.3 \
--missing --missing_strategy median \
--random_state 8 --ids recoded_1000G.raw.noadmixed.AFR.ids

t-SNE (cf. output):

python runGTM.py --model GTM --data recoded_1000G.noadmixed.AFR.mat \
--labels recoded_1000G.raw.noadmixed.AFR.lbls3 --labeltype discrete \
--out 1000G_t-SNE_AFR --pca --n_components 10 \
--missing --missing_strategy median \
--random_state 8 --ids recoded_1000G.raw.noadmixed.AFR.ids

African subpopulations classification performance:

python runGTM.py --model GTM --data recoded_1000G.noadmixed.AFR.mat \
--labels recoded_1000G.raw.noadmixed.AFR.lbls3 \
--labeltype discrete --out outputname --pca --n_components 10 \
--missing --missing_strategy median \
--random_state 8 --crossvalidate

Addendum 2: Arabidopsis Thaliana geographic visualization

Cf. github repository https://github.com/hagax8/arabidopsis_viz

hagax8 / ancestry_viz