See preliminary Project writeup (0.5 pages, including data source), written prior to the start of work.
My project is an unsupervised approach at identifying patterns of cell types in a 2D layout. It has come to my attention that in the last one or two years, datasets have become available which represent not only the genomic data of sampled cells but also the cells’ location in a 2D surface.
# Make transcriptomes.npy and xy.npy
./preprocess.py
# Create fully-connected graph and get k-nearest-neighbours
./graphs.py- Whiten transcriptome data
- Fully-connected graph: The aforementioned representation comprising both spatial and genomic information will enable me to construct a fully-connected graph, with the edge weight between cell
iand celljcalculated asd(i, j) × c(i, j), wheredrepresents the spatial distance andcrepresents the feature distance between cellsiandj. (For these two distances, I expect to work with Gaussian distance, but given time, I may also experiment with cosine distance or Euclidean distance for either one. Furthermore, I might experiment with variants of the distance calculation stated above so as to favour spatial and feature distance differently.) - Sparsify graph: To sparsify the graph, I will run k-nearest-neighbours and keep only the edges for these neighbourhoods.
- GCN: I intend to train a graph convolution network (GCN) as an autoencoder on the graph data. Being that k-nearest-neighbours will give me a fixed neighbourhood size, my GCN’s aggregator function can be to simply concatenate genomic features of all of the cells in each neighbourhood. (This has shortcomings, but GCN’s are new to me, so I intend this to be at least my first approach. If all proceeds quickly, I may experiment with other aggregator functions.) The embedding produced by the GCN’s encoder will then serve as a new feature representation (replacing the former spatial and feature data) for each cell (or, rather, each neighbourhood, with a neighbourhood being focused on a single cell).
- Spectral clustering: In this new feature space, I will perform spectral clustering. I intend to use the normalized Laplacian to this end, but given time, I may also experiment with a random-walk Lalacian. Moreover, I may experiment with clustering on the feature space in addition to clustering on the eigenvectors of the Laplacians. (GCN’s and spectral clustering are both new to me, so I anticipate the possibility that I will not move quickly enough to compare results with alternative designs.)
- Qualitative inspection: I will perform qualitative examination of neighbourhoods nearest to the cluster centroids, looking for cell-type consistency within clusters and differentiation between clusters.
I plan to apply spectral clustering directly to the graph data (omitting the aforementioned GCN.
- Whiten data
- Fully-connected graph
- Sparse graph
- Spectral clustering
- Qualitative inspection
See preprocess.py
Run run.sh.
Same as "Fully-connected graph": run run.sh
Train the GCN with train.sh. Set your learning rate(s) and epoch counts in LR and EP.
This will save your log to --log_path.
This will save your model to --save_path.
After training, create the embeddings for all the data points with infer-embeddings.sh.
This will save your embeddings to --save_path.
Run python cluster.py.
LBL_CSV=data/class_labels.csv
ID_CSV=data/output_keptCellID.txt
KNN=16
KM=12
DIM=4
XY_NP=data/xy.npy
FEAT_NP=artefacts/nn-embeddings.npy
SAVEDIR=after-gcn
GRAPH=$SAVEDIR/sparse-graph.npy
python run.py \
--no-spatial \
--knn $KNN --km $KM --dim $DIM --lapmode rw \
--embedding "$SAVEDIR/laplacian-embedding-rw.npy" \
--id-csv "$ID_CSV" --lbl-csv "$LBL_CSV" \
--sparse "$GRAPH" \
--gen-np "$FEAT_NP" \
--xy-np "$XY_NP" \
--no-pre