Algorithm 1. Summary of the preprocessing steps done to each enzyme at training time.

Data: N training enzymes, grid size of l, homothetic transformation ratio λ, p interpolations, probability pflip to flip with respect to each axis.
Input: Raw coordinates contained in PDB files
Output: Volumes of binary voxels representing backbone atoms occupancy
1	foreach of the N enzymes of the training set do
2	Step 1: structural information extraction
3	Extract coordinates of backbone atoms from its PDB file
4	Step 2: holes completion
5	Interpolate consecutive backbone atoms by p new points
6	Step 3: size adjustment
7	Center barycenter S of the coordinates on (0, 0, 0)
8	Homothetic transformation of each point with center S and ratio λ
9	Step 4: enzyme orientation
10	Principal component analysis (PCA) transformation
11	Step 5: random augmentation
12	if True with probability pflip then
13	Flip coordinates with respect to the origin along x—axis
14	if True with probability pflip then
15	Flip coordinates with respect to the origin along y—axis
16	if True with probability pflip then
17	Flip coordinates with respect to the origin along z—axis
18	Step 6: voxelization
19	Center barycenter S of the coordinates on $\left(\frac{l}{2},\frac{l}{2},\frac{l}{2}\right)$
20	Transform coordinate points into binary voxels