PreliminaryTest.md

Models

integer square:

Model	Explanation	Accuracy
conv_rand_inp	Used random shape matching	0.0122875
conv_basic	Used convolutions to create features	0.52933125
conv_no_loc	Used no coordinate information	0.533325
conv_less_mem	Used independent convolutions to reduce memory usage	0.43043125
rnn_conv	Used RNN instead of feed-forward net	0.56195625
rnn_no_loc	Used RNN and no coordinate information	0.53053125
birnn	Used bidirectional rnn instead of feed-forward net	0.542495

float square:

Model	Explanation	Accuracy
conv_basic2	Used convolutions to create features	0.358625
conv_no_loc2	Used no coordinate information	0.360305
conv_basic_k15	Used kernel size of 15 instead of 21	0.362722
conv_basic_k25	Used kernel size of 25 instead of 21	0.37927225
conv_basic_pix5	Used pixel size of .5 instead of .7	0.361384
conv_basic_pix9	Used pixel size of .9 instead of .7	0.41228775
adam_conv_basic	Used adam optimizer as opposed to gradient descent	0.3422775
add_conv_basic	Instead of concatenation, used weighted sum of pin combination	0.3561075
add2_conv_basic	Instead of concatenation, used weighted sum of lig feats and pin coords	0.35749
add3_conv_basic	Instead of concatenation, used weighted sum of pin feats and relative lig distances	0.356545
add12_conv_basic	Combination of 1 and 2	0.3499175
add13_conv_basic	Combination of 1 and 3	0.3566525
batch_conv_basic	Batches in groups of 10	0.5996925
batch100_conv_basic	Batches in groups of 100	0.7342025
single_batch_conv_basic	Combined batches of 10 with min	0.5141025
single_max_batch_conv_basic	Combined batches of 10 with max	0.19069
batch_adam_conv_basic	Batches in 10 with adam optimizer	0.5742675
more_conv_basic	Added another cross convolution to a total of 3 convolutions	0.40216
rnn_conv2	Used RNN instead of feed-forward net	0.40108
rnn_no_loc2	Used RNN and no coordinate information	0.3743025
birnn2	Used bidirectional rnn instead of feed-forward net	0.22496
lstm_basic	Used LSTM for feature creation instead of convolutions	0.142

icosahedron:

Model	Explanation	Accuracy
conv_basic_iso	Used convolutions to create features	0.4063
conv_no_loc_iso	Used no coordinate information	0.4024625
rnn_conv_iso	Used RNN instead of feed-forward net	0.444475
rnn_no_loc_iso	Used RNN and no coordinate information	0.42755
birnn_iso	Used bidirectional rnn instead of feed-forward net	0.14515975

original_square:

• rnn_conv: 0.56195625
• conv_rand_inp: 0.0122875
• conv_basic: 0.52933125
• conv_less_mem: 0.43043125
• conv_no_loc: 0.533325
• rnn_no_loc: 0.53053125
• birnn: 0.542495 (50 hidden)

adv_square:

• conv_basic: 0.358625
• birnn: 0.22496
• rnn_conv: 0.40108
• rnn_no_loc: 0.3743025
• conv_no_loc: 0.360305

icos:

• conv_basic: 0.4063
• rnn_no_loc: 0.42755
• birnn: 0.14515975
• conv_no_loc: 0.4024625
• rnn_conv: 0.444475

add_weights:

• 1: 0.3561075
• 2: 0.35749
• 3: 0.356545
• 1&2: 0.3499175
• 1&3: 0.3566525

more models:

• more_conv: 0.40216
• adam_conv: 0.3422775
• lstm: 0.142
• rnn_batch_pix9: 0.83133

batches:

• batch10_rnn: 0.8852425
• batch10: 0.5996925
• adam_batch10: 0.5742675
• single_batch10: 0.5141025
• single_max_batch: 0.19069
• batch100: 0.7342025

pixels and kernels:

• k15: 0.362722
• k25: 0.37927225
• pix5: 0.361384
• pix9: 0.41228775

Model explanations

Note: all models are used with float squares unless explicitly stated

Model	Explanation	Accuracy
10_5_conv_basic	Starts with 10 poses, goes to 5 poses, and then finally chooses a pose
adam_conv_basic	Used adam optimizer as opposed to gradient descent	0.3422775
add_conv_basic	Instead of concatenation, used weighted sum of pin combination	0.3561075
add2_conv_basic	Instead of concatenation, used weighted sum of lig feats and pin coords	0.35749
add3_conv_basic	Instead of concatenation, used weighted sum of pin feats and relative lig distances	0.356545
add12_conv_basic	Combination of 1 and 2	0.3499175
add13_conv_basic	Combination of 1 and 3	0.3566525
add23_conv_basic	Combination of 2 and 3	too big
add123_conv_basic	Combination of 1, 2, and 3	too big
batch_conv_basic	Batches in groups of 10	0.5996925
batch100_conv_basic	Batches in groups of 100	0.7342025
batch1000_conv_basic	Batches in groups of 1000	too big
batch_adam_conv_basic	Batches in 10 with adam optimizer	0.5742675
birnn	Used bidirectional rnn instead of feed-forward net (integer squares)	0.542495
birnn2	Used bidirectional rnn instead of feed-forward net (float squares)	0.22496
birnn_iso	Used bidirectional rnn instead of feed-forward net (icosahedrons)	0.14515975
conv_basic	Used convolutions to create features (integer squares)	0.52933125
conv_basic2	Used convolutions to create features (float squares)	0.358625
conv_basic_iso	Used convolutions to create features (icosahedrons)	0.4063
conv_basic_k15	Used kernel size of 15 instead of 21	0.362722
conv_basic_k25	Used kernel size of 25 instead of 21	0.37927225
conv_basic_pix5	Used pixel size of .5 instead of .7	0.361384
conv_basic_pix9	Used pixel size of .9 instead of .7	0.41228775
conv_less_mem	Used independent convolutions to reduce memory usage (integer squares)	0.43043125
conv_no_loc	Used no coordinate information (integer squares)	0.533325
conv_no_loc2	Used no coordinate information (float squares)	0.360305
conv_no_loc_iso	Used no coordinate information (icosahedrons)	0.4024625
conv_rand_inp	Used random shape matching (integer squares)	0.0122875
lstm_basic	Used LSTM for feature creation instead of convolutions	0.142
more_conv_basic	Added another cross convolution to a total of 3 convolutions	0.40216
rnn_conv	Used RNN instead of feed-forward net (integer squares)	0.56195625
rnn_conv2	Used RNN instead of feed-forward net (float squares)	0.40108
rnn_conv_iso	Used RNN instead of feed-forward net (icosahedrons)	0.444475
rnn_no_loc	Used RNN and no coordinate information (integer squares)	0.53053125
rnn_no_loc2	Used RNN and no coordinate information (float squares)	0.3743025
rnn_no_loc_iso	Used RNN and no coordinate information (icosahedrons)	0.42755
single_batch_conv_basic	Combined batches of 10 with min	0.5141025
single_max_batch_conv_basic	Combined batches of 10 with max	0.19069

ligand:

• conv: train-0.211336666667, test-0.292093762416 (could run for longer)
• adam: train-0.24562, test-0.306924516532 (also could run for longer)

Preliminary Runs

Model	Explanation	Train Accuracy	Test Accuracy	Average RMSD
ligand_adam	adam optimizer instead of gradient descent	0.24562	0.306924516532
ligand_adam3	adam and 3 convolutions	0.576979742173
ligand_adam6	adam and 6 convolutions	0.167024174328
ligand_conv	2 convolutions	0.211336666667	0.292093762416	13.436826244990032
ligand_conv3	3 convolutions	0.376234096692
ligand_conv4	4 convolutions	too big
ligand_conv5	5 convolutions	0.396642808453
ligand_conv6	6 convolutions	0.0
ligand_rnn	rnn instead of feed forward
ligand_rnn3_adam_pix9	rnn, adam, 3 convoluations, pixel .9

Full Runs on xstream

Learning to find correct pose given original ligand coordinates and multiple random poses for each atom

Model	Explanation	Train Accuracy	Test Accuracy	Train Average RMSD	Test Average RMSD	Last Batch Ran
0, fwd2	feed-forward net with 2 convolutions and 2 poses, gradient descent optimizer	0.303333333333	.54	2.2899334	3.044272	83583
1, fwd3	feed-forward net with 3 convolutions and 2 poses, gradient descent optimizer	0.238888888889	0.493333333333	3.4184973	2.4580858	71899
2, fwd4	feed-forward net with 4 convolutions and 2 poses, gradient descent optimizer	0.261111111111	0.518888888889	3.2393553	2.30844	80299
3, fwd5	feed-forward net with 5 convolutions and 2 poses, gradient descent optimizer	0.434444444444	0.637777777778	1.9871486	1.4491253	69603
4, fwd6	feed-forward net with 6 convolutions and 2 poses, gradient descent optimizer	0.373333333333	0.577777777778	2.380698	1.8441339	69730
5, fwd_adam2	feed-forward net with 2 convolutions and 2 poses, adam optimizer	0.29	0.487777777778	3.1777396	2.4393547	66699
6, fwd_adam3	feed-forward net with 3 convolutions and 2 poses, adam optimizer	0.318888888889	0.511111111111	2.9273443	2.4509358	66785
7, fwd_adam4	feed-forward net with 4 convolutions and 2 poses, adam optimizer	0.498888888889	0.693333333333	1.7167966	1.2009867	65166
8, fwd_adam5	feed-forward net with 5 convolutions and 2 poses, adam optimizer	0.231111111111	0.537777777778	3.395176	2.3522975	62458
9, fwd_adam6	feed-forward net with 6 convolutions and 2 poses, adam optimizer	0.337777777778	0.501111111111	2.8802297	2.4606843	71999
10, rnn2	rnn with 2 convolutions and 2 poses, gradient descent optimizer	0.306666666667	0.517777777778	2.955842	2.2691903	34099
11, rnn3	rnn with 3 convolutions and 2 poses, gradient descent optimizer	0.463333333333	0.618888888889	1.839397	1.5545195	35341
12, rnn4	rnn with 4 convolutions and 2 poses, gradient descent optimizer	0.453333333333	0.602222222222	1.9027276	1.630737	34980
13, rnn5	rnn with 5 convolutions and 2 poses, gradient descent optimizer	0.351111111111	0.558888888889	2.7551725	2.0155606	32377
14, rnn6	rnn with 6 convolutions and 2 poses, gradient descent optimizer	0.363333333333	0.563333333333	2.4736986	2.0263968	33516
15, rnn_adam2	rnn with 2 convolutions and 2 poses, adam optimizer	0.332222222222	0.585555555556	2.7406085	1.950406	80611
16, rnn_adam3	rnn with 3 convolutions and 2 poses, adam optimizer	0.433333333333	0.614444444444	2.1726098	1.6305083	33255