Supplementary Data for NIPS Publication: Protein Interface Prediction using Graph Convolutional Networks.

<p>These data sets can be used to re-run the experiments from our paper, Protein Interface Prediction using Graph Convolutional Networks. The data are derived from protein complexes in the docking benchmark dataset v. 5.0. Each file is a python tuple&nbsp;that has been saved using cPickle and compressed using gzip.</p> <p>Links:</p> <p>Paper: https://papers.nips.cc/paper/7231-protein-interface-prediction-using-graph-convolutional-networks</p> <p>Poster:&nbsp;https://zenodo.org/record/1134154</p> <p>Code:&nbsp;https://github.com/fouticus/pipgcn</p> <p>&nbsp;</p> <p><strong>File Descriptions:</strong></p> <p>train.cpkl.gz and test.cpkl.gz have the data formatted for neighborhood based graph convolutions. The diffc_ files are the same data formatted for the diffusion convolutional neural networks that we compare against.&nbsp;</p> <p>train.cpkl.gz is a tuple of length 2:</p> <ul> <li>element 0 is a list of length 175 containing the PDB codes from the docking benchmark dataset</li> <li>element 1 is a list of length 175 containing features for each protein. Each element is a dictionary containing the following keys: <ul> <li>r_vertex: vertex (residue) features for the receptor. numpy array of shape (x, 70) where x is the number of residues in the receptor and 70 is the number of features.</li> <li>l_vertex: vertex (residue) features for the ligand. analogous to above, with shape (y, 70) where y is the number of residues in the ligand.</li> <li>complex_code: PDB code&nbsp;of the complex. matches the list of codes described above.</li> <li>l_edge: edge features for the neighborhood around each residue in the ligand. numpy array of shape (y, 20, 2) where y&nbsp;is defined as above. the second dimension is the edges to the 20 nearest neighboring residues,&nbsp;ordered by decreasing distance. The third dimension allows for two features per edge.&nbsp;</li> <li>r_edge: edge features for the neighborhood around each residue in the receptor. numpy array of shape (x, 20, 2) where x&nbsp;is as above.&nbsp;</li> <li>l_hood_indices: the index of the 20 closest residues to each residue, ordered by decreasing distance. numpy array of shape (y, 20, 1). &quot;Index&quot; means which row in l_vertex gives the vertex features for the closest neighbor, second closest neighbor, etc.&nbsp;</li> <li>r_hood_indices: analogous to above, shape (x, 20, 1).</li> <li>label: 1 or -1 label for each residue pair. numpy array of shape (x*y, 3). Each row looks like (i, j, k) where i is the index of the ligand&nbsp;residue, j is the index of the receptor residue, and k is either -1 (negative example) or 1 (positive example).</li> </ul> </li> </ul> <p>test.cpkl.gz matches the structure of train.cpkl.gz except it has the test set of 55 complexes.&nbsp;</p> <p>Descriptions of the vertex and edge features can be found in Appendix A of &nbsp;<a href="https://mountainscholar.org/handle/10217/185661">this.</a></p> <p>diffc_g2_p2_train.cpkl.gz is a tuple of length 2:</p> <ul> <li>element 0 is a list of the same 175 PDB codes as above.&nbsp;</li> <li>element 1 is a list of features for the 175 complexes. Each element is a dictionary of features with these keys: <ul> <li>r_vertex, l_vertex, complex_code, label: these are the same as described above.&nbsp;</li> <li>&#39;r_power_series&#39;: Stacked diffusion matrices which are powers of the similarity matrix used in the DCNN method. numpy array of shape (x, 2, x) where x&nbsp;is the number of receptor residues. the middle dimension 2 indicates how many &quot;hops&quot; is used for that diffusion (1 vs. 2). In other words, element (i, 0, j) is the similarity after 1 hops between residues i and j. element (i, 1, j) is the similarity after 2 hops.&nbsp;See DCNN paper for details.</li> <li>&#39;l_power_series&#39;: same as above but for the ligand. shape is (y, 2, y).</li> </ul> </li> </ul> <p>diffc_g2_p2_test.cpkl.gz is the same as diffc_g2_p2_train.cpkl.gz but for the 55 test complexes.</p> <p>diff_g2_p5_train.cpkl.gz and diff_g2_p5_test.cpkl.gz are the same as the p2 version above, except that the diffusion matrices have shape (x, 5, x) and (y, 5, y) because one of our comparisons against&nbsp;the DCNN model uses 5 hops instead of just 2.&nbsp;</p> <p>&nbsp;</p> <p>Note: these files were pickled with Python 2.7. If you&#39;re unpickling with Python 3.x you might have to specify encoding as &#39;latin1&#39;.&nbsp;</p> <p>&nbsp;</p> <p>Please direct any questions to:</p> <ul> <li>Alex Fout (fout@colostate.edu)</li> <li>Jonathon Byrd (jonbyrd@colostate.edu)</li> <li>Basir Shariat (basir@cs.colostate.edu</li> <li>Asa Ben-Hur (asa@cs.colostate.edu)</li> </ul>