Computational method for classifying and predicting ligand docking conformations

1. A computer-implemented method for predicting a conformation of a ligand docked into a protein, the method comprising:

determining one or more poses of the ligand, the poses being representative conformations of the ligand;

extracting features associated with the poses of the ligand;

constructing, based on the extracted features, feature vectors associated with the poses of the ligand;

determining, using a neural network, scores associated with the poses, wherein the determining of the scores associated with the poses comprises at least one convolution of the constructed feature vectors associated with the poses, the convolution of each feature vector comprising a transformation operator and a reduction operator; and

determining a proper conformation for the docked ligand based on the scores.

2. The method of claim 1 , wherein determining one or more poses of the ligand comprises:

dividing the ligand into two or more sections;

anchoring a first section of the two or more ligand sections to a location of the protein;

adding at least one subsequent section of the two or more ligand sections to the first section to form a growing ligand; and

continuing adding subsequent sections to the growing ligand until the ligand is complete.

3. The method of claim 1 , wherein the extracted features comprise features of a first atom and features of the interaction between the first atom and a second atom.

4. The method of claim any of claim 3 , wherein the features of the first atom comprise one or more of an atom type, a radius of the atom, a number of rings in which the atom is included, a size of the ring in which the atom is included, whether the first atom is part of an aromatic ring, or the pairwise potential of the first atom.

5. The method of claim 4 , wherein the pairwise potential of the first atom comprises the sum of the pairwise atom potentials between the first atom and atoms of the protein.

6. The method of claim 3 , wherein the features of the interaction between the first atom and the second atom comprise one or more of a bond type, a distance between the first and second atoms, or whether the first and second atoms are in the same ring.

7. The method of claim 1 , wherein the constructed feature vectors comprise a dense feature vector for each atom of the ligand, wherein each dense feature vector including the features of an atom and features of the interaction between that atom and another atom.

8. The method of claim 1 , wherein determining, using a neural network, scores associated with the poses comprises two convolutions of the feature vectors associated with the poses of the ligand.

9. The method of claim 1 , wherein the transformation operator comprises transforming the feature vectors of neighbor atoms of an atom of interest by a feed-forward linear sub-network.

10. The method of claim 9 , wherein the reduction operator comprises:

aggregating the transformed feature vectors of neighbor atoms of the atom of interest;

applying a commutative reduction function to the aggregated transformed feature vectors of neighbor atoms of an atom of interest to produce a reduced feature map for the atom of interest.

11. The method of claim 9 , wherein the convolution process further comprises an optimization operator that combines the initial feature vector of an atom of interest with the feature vector output after the application of the transformation operator and reduction operator, resulting in a final feature vector for the atom of interest.

12. The method of claim 1 , wherein

the feature vector of each neighbor atom comprises a dense feature vector,

the dense feature vector comprises an atom feature vector for the neighbor atom and an atom pair feature vector for the neighbor atom and the atom of interest, and

the dense feature vector is input into the transformation operator as a concatenation of the atom feature vector and the atom pair feature vector, and

the input is transformed through a fully connected layer and a non-linearity function.

13. The method of claim 1 , wherein determining, using a neural network, scores associated the poses comprises applying a scoring function to the feature vectors of the atoms of the ligand for each pose.

14. The method of claim 13 , wherein the scoring function is a weighted scoring function that applies a weighted vector to the feature vectors of the atoms of the ligand for each pose, wherein the weighted vector is determined by a machine-learning algorithm.

15. The method of claim 14 , wherein the machine-learning algorithm for determining the weight vector is trained on real-world protein structure data.

16. The method of claim 15 , wherein the training of the machine-learning algorithm comprises determining a weight vector {right arrow over (W)}=(w 1 , w 2 , w 3 , . . . , w n ) for the real-world protein structure data, where, when the feature vector for the correct ligand conformation equals (x 1 , x 2 , x 3 , . . . , x n ) and the feature vector for an incorrect ligand conformation is (y 1 , y 2 , y 3 , . . . , y n ), {right arrow over (W)} satisfies the equation (Σ i=1 n w i x i −Σ i=1 n w i y i )>0.

17. The method of claim 1 , wherein determining a proper conformation for the docked ligand based on the scores comprises ranking the scores associated with the ligand poses.

18. The method of claim 1 , wherein determining one or more poses of the ligand comprises:

dividing the ligand into two or more sections;

anchoring a first section of the two or more ligand sections to a first location of the protein;

anchoring a second section of the two or more ligand sections to a second location of the protein, wherein the second location of the protein may be the same or different than the first location of the protein;

extracting features associated with the anchored first section of the ligand and features associated with the anchored second section of the ligand;

constructing, based on the extracted features, feature vectors associated with the anchored first section of the ligand and feature vectors associated with the anchored second section of the ligand;

determining, using a neural network, scores associated with the anchored first section and anchored second section; and

determining a proper anchor for the docked ligand based on the scores.

19. The method of claim 1 , wherein the method further comprises generating a graphical representation of the determined proper conformation for the docked ligand.

20. A computer-implemented method for predicting a conformation of a ligand docked into a protein, the method comprising:

dividing the ligand into two or more sections;

anchoring a first section of the two or more ligand sections to a location of the protein;

extracting features associated with the anchored first section;

constructing, based on the extracted features, feature vectors associated with the anchored first section;

determining, using a neural network, a score associated with the anchored first section, wherein the determining of the score comprises at least one convolution of the constructed feature vectors, the convolution of each feature vector comprising a transformation operator and a reduction operator; and

determining a proper conformation for the docked ligand based on the score.

21. A computer-implemented method for predicting an anchor section of a ligand docked into a protein, the method comprising:

dividing the ligand into two or more sections;

anchoring a first section of the two or more ligand sections to a first location of the protein;

anchoring a second section of the two or more ligand sections to a second location of the protein, wherein the second location of the protein may be the same or different than the first location of the protein;

extracting features associated with the anchored first section of the ligand and features associated with the anchored second section of the ligand;

constructing, based on the extracted features, feature vectors associated with the anchored first section of the ligand and feature vectors associated with the anchored second section of the ligand;

determining, using a neural network, scores associated with the anchored first section and anchored second section, wherein the determining of the scores comprises at least one convolution of the constructed feature vectors, the convolution of each feature vector comprising a transformation operator and a reduction operator; and

determining a proper anchor for the docked ligand based on the scores.

22. The method of claim 21 , wherein the method further comprises generating a graphical representation of the anchor section of the ligand.

23. A computer-implemented method for predicting a conformation of a potential growing ligand section docked into a protein, the method comprising:

dividing the ligand into a plurality of sections;

creating a first growing ligand section by:

anchoring a first section of the plurality of sections to a first location of the protein; and

adding a second section of the plurality of sections to the first section in a first conformation;

creating a second growing ligand section by:

anchoring a first section of the plurality of sections to a first location of the protein; and

adding a second section of the plurality of sections to the first section in a second conformation;

extracting features associated with the first growing ligand section and features associated with the second growing ligand section;

constructing, based on the extracted features, feature vectors associated with the first growing ligand section and feature vectors associated with the second growing ligand section;

determining, using a neural network, scores associated with the first growing ligand section and the second growing ligand section, wherein the determining of the scores comprises at least one convolution of the constructed feature vectors, the convolution of each feature vector comprising a transformation operator and a reduction operator; and

determining a potential growing ligand section for the docked ligand based on the scores.

24. The method of claim 23 , wherein the method further comprises generating a graphical representation of the potential growing ligand section.