2018AffinitySummary

Pinning Model

Can be found in feat_flex/ligand_web

Goal: Given ligand coordinates and a set of possible coordinates for each atom, select coordinates for each atom that preserves the shape of the coordinates

Benefits: Learns shape preservation for either rigid ligands or conformers (faster than trying all possible combinations)

General description: Runs a 3d convolution in the ligand space and then takes those features and runs a 3d convolution in the possible coordinate space. Can repeat by taking those features and running another 3d convolution in the ligand space and do more convolutions in a similar fashion.

Variations:

1. Using a feed-forward network or a RNN,
2. Using 2-6 convolutions in the pattern stated above,
3. Using Gradient Descent Optimizer or Adam Optimizer,
4. Matching a transformed version of crystal ligand or conformer,
5. Choosing from 2 poses to 10 poses
6. Using weighted sums instead of concatenations (for
7. Using batches of sizes 1, 10, 30, 100
8. Random pins or random grids
9. Use grey map pins
10. Use top 2 or 10 beads from bead model as pins
11. Did preliminary tests with just triangles, squares and icosahedrons

Top statistics:

Model	SeqTest Accuracy	SeqTest RMSD	SeqTest AUC
Feed-forward, 5 convolutions, crystal transformed, 2 poses, gradient descent	0.8000001	0.70817494	~~
Feed-forward, 2 convolutions, conformer, 2 poses, gradient descent, 2.4A grids	0.9751578	2.1780472	0.9963048
Feed-forward, 2 convolutions, conformer, 10 poses, gradient descent, 2.4A grids	0.84553695	6.642	0.9758124
Feed-forward, 2 convolutions, conformer, 2 poses, gradient descent, 0.6A grids	0.98905164	0.69357973	0.9986486

More statistics can be found here.

Higher Definition Model

Can be found in high_def/metrics_tm

Goal: Give protein and ligand atoms and coordinates, predicts grid location that corresponds to the crystal ligand location for each atom

Benefits: Reduces amount of memory because doesn't need to calculate features for all possible grid locations

General description: Creates a large grid over the binding site. Then runs a 3d convolution over the ligand and another 3d convolution over the grid space and performs an outer product. Can repeat by taking smaller grids from selected grids in the larger grids

Variations:

1. 1-3 iterations (10 divisions, 2 divisions, 2 divisions)
2. Ranging the space from -25A to +25A or -12A to +12A
3. Using cross entropy loss or noise-contrastive estimation loss
4. Compared with just 20 divisions (equivalent to 2 iterations)
5. Combine with previous pinning model

Statistics:

Model	SeqTest AUC
1 layer cross entropy	0.7612224
2 layer cross entropy	0.7345741, 0.49980348
3 layer cross entropy	0.79310304, 0.5035895, 0.500349
1 layer nce	0.80953294
2 layer nce	0.80062085, 0.57625467
3 layer nce	0.8110932, 0.5955195, 0.5003045
20 divisions	0.65434194

Bi-convolution

Can be found in high_def/bi_conv

Goal: Given relative ligand coordinates, pin coordinates, and pin features, learn features from two different spaces

Benefits: The features are invariant to ordering of the coordinates representing each of the spaces

General description: Performs a 3D convolution over the space of the original ligand atoms with pin features and then performs a second convolution with the new features over the space of the pins coordinates

Variations

1. Random grids (2.4A, 1.2A, .6A) for pins
2. Random coordinates for pins
3. Grey map for pins
4. 2 different concatenations of ligand features
5. Cross entropy loss or swap loss

Top statistics:

Model	SeqTest Accuracy	SeqTest RMSD	SeqTest AUC
Basic bi_conv on 2.4A grids	0.8329366	7.4538136	0.9742352
Initial ligand concatenation on 2.4A grids	0.8538086	6.0495987	0.97240996
No grids with 2nd ligand concatenation	0.46637124	13.798717	0.8164979
0.6A grids	0.9107372	5.179973	0.985856
No grids 2 sets of bi_conv	0.9323402	3.0136988	0.9805218
Grey map swap	0.35714287	3.4412975	3.4412975

Fixing beads

Can be found in high_def/fix_bead

Goal: Beads and ligand vector representations should be invariant to switching beads

Variations:

1. 1-3 convolutions
2. Model35 (euclidean distance) or Model38 (cosine distance)

Statistics:

Model	GoodBeadAcc	SeqBeadAcc	GoodBindAcc	SeqBindAcc
1 Euclidean	0.9789973	0.9841414	0.973965	0.9860906
2 Euclidean	0.98313314	0.9812303	0.98346484	0.9756007
3 Euclidean	0.9790819	0.98118925	0.98620284	0.95882887

For full statistics, click here and here.