Implementation of the connectome embedding (CE) framework.
Embedding of brain graph or connectome embedding (CE) involves finding a compact vectorized representation of nodes that captures their higher-order topological attributes. CE are obtained using the node2vec algorithm fitted on random walk on a brain graph. The current framework includes a novel approach to align separately learned embeddings to the same latent space. Cepy is tested on Python 3.6, 3.7 and 3.8.
pip install cepy
import cepy as ce
import numpy as np
# Load an adjacency matrix (structural connectivity matrix)
sc_group = ce.get_example('sc_group_matrix')
# Initiate and fit the connectome embedding model
ce_group = ce.CE(permutations = 1, seed=1)
ce_group.fit(sc_group)
# Extract the cosine similarity matrix among all pairwise nodes
cosine_sim = ce_group.similarity()
# Save and load the model
ce_group.save_model('group_ce.json')
ce_loaded = ce.load_model('group_ce.json') # load it
# Load two existing CE models
ce_subject1 = ce.get_example('ce_subject1')
ce_subject2 = ce.get_example('ce_subject2')
# Align the two to the space of the [ce_group]:
ce_subject1_aligned = ce.align(ce_group, ce_subject1)
ce_subject2_aligned = ce.align(ce_group, ce_subject2)
# Extract the node vectorized representations (normalized) for subsequent use - prediction, for example
w_sbject1 = ce_subject1_aligned.weights.get_w_mean(norm = True)
w_sbject2 = ce_subject2_aligned.weights.get_w_mean(norm = True)
A set of example interactive Jupyter notebooks are also available here.
If you find cepy useful for your research, please consider citing the following paper:
Levakov, G., Faskowitz, J., Avidan, G., & Sporns, O. (2021). Mapping individual differences across brain network structure
to function and behavior with connectome embedding. Neuroimage, 242, 118469. Cepy is an open-source software project, and we welcome contributions from anyone. We suggest raising an issue prior to working on a new feature.