janos-code/deepdta-tf

Tensorflow implementation of "DeepDTA deep drug-target binding affinity prediction"

DeepDTA-tf

Tensorflow implementation of "DeepDTA deep drug-target binding affinity prediction"

envs

• python 3.6
• tensorflow 1.12.0

data

• kiba: 10 cross validation
• newkd: all for training
• drugbank: new drug, new target, new pair 5 cross validation
• FDA, Merck, XJ, YS, ZDC: just for prediction

data format

• cross validation: ligands.csv, proteins.csv, inter.csv
• all training: ligands.csv, proteins.csv, pairs.csv

benchmark

kiba as training set, predict new data(FDA & Merck) and calc AUC

Config files

• data_xxx.cfg

[model]
data_path = path of training data
path = path for saving the model

[cv]
cv_num = 5
# 1: new pair, 2: new target, 3: new drug
problem_type = 1

• model.cfg

[model]
params...

• predict.cfg

Notice: change model.path & data.prediction for prediction
[model]
params...
path = ../../model/newkd

[data]
path = ../../data/
prediction = FDA,Merck,XJ,YS,ZDC

How to run

• training

cd src/deepdta or cd/src/deepdti
# cross validation
python cv.py ../../config/model.cfg ../../config/data_kiba.cfg
# all for training
python main.py ../../config/model.cfg ../../config/data_newkd.cfg
# special conditions for training
python special.py ../../config/model.cfg ../../config/data_drugbank.cfg

• prediction

cd src/deepdta or cd/src/deepdti
python predict.py ../../config/predict.cfg

• DUD-E is different from other datasets, so I wrote the file separately.

# cross validation
python cv_dude.py ../../config/model.cfg ../../config/data_dude.cfg
# predict dude using trained DrugBank model
python predict_dude.py ../../config/predict.cfg

Model file

path	description
drugbank-newdrug	new drug 5 cv
drugbank-newtarget	new target 5 cv
drugbank-newpair	new pair 5 cv
drugbank-ecfp	ecfp + seq all training
drugbank-smiles	smiles + seq all training
kiba	smiles + seq all training & 10 cv
newkd	smiles + seq all training

Results

• AUCs/AUPRs(smiles + seq)

dataset used in training	FDA	Merck	FDA + Merck
kiba	0.5220/0.0984	0.3677/0.3513	0.3722/0.0801
newkd	0.5149/0.0481	0.3396/0.1494	0.6389/0.1108
DB201707	0.5257/0.0413	0.7510/0.3879	0.5720/0.0643
DB201707-del-P11388	0.6015/0.0922	0.5323/0.1938	0.5462/0.0880
DUD-E	0.5637/0.0510	0.6698/0.3256	0.6068/0.0828

• AUCs(fingerprints + seq)

dataset used in training	FDA	Merck	FDA + Merck
DB201707	0.3585	0.4719	0.4324

DrugBank 201707 Result

1. CNN (smiles + seq)

• AUC 5 cross validation

type	1	2	3	4	5
new target	0.798	0.842	0.801	0.786	0.801
new pair	0.904	0.900	0.900	0.895	0.897
new drug	0.844	0.810	0.787	0.802	0.831

• AUPR 5 cross validation

type	1	2	3	4	5
new target	0.461	0.507	0.483	0.381	0.412
new pair	0.663	0.635	0.644	0.628	0.651
new drug	0.501	0.479	0.405	0.421	0.409

2. CNN (smiles + seq) vs ECFPCNN (fingerprints + seq)

• AUC 5 cross validation

methods	new target	new drug	new pair
CNN	0.844	0.872	0.924
ECFPCNN	0.837	0.857	0.912

• AUPR 5 cross validation

methods	new target	new drug	new pair
CNN	0.488	0.563	0.720
ECFPCNN	0.524	0.551	0.702

Drop P11388 col, other as training, P11388 col as test

P11388 col: 539

model	AUC	AUPR
CNN	0.992	0.659
ECFPCNN	0.827	0.047

DUD-E as training, P11388 col as test

model	AUC	AUPR
CNN	0.490	0.015

DUD-E Result

• AUC 5 cross validation

type	1	2	3	4	5
new target	0.917	0.905	0.892	0.896	0.917

• AUPR 5 cross validation

type	1	2	3	4	5
new target	0.569	0.502	0.457	0.495	0.554

NeoData Result

1. (708, 1512)

• AUC 10 cross validation

type	1	2	3	4	5	6	7	8	9	10
new target	0.897	0.895	0.873	0.916	0.899	0.835	0.912	0.918	0.786	0.874
new pair	0.927	0.929	0.920	0.951	0.949	0.944	0.950	0.936	0.936	0.949
new drug	0.894	0.880	0.873	0.874	0.880	0.902	0.879	0.916	0.886	0.894

• AUPR 10 cross validation

type	1	2	3	4	5	6	7	8	9	10
new target	0.676	0.677	0.597	0.658	0.686	0.373	0.707	0.713	0.195	0.521
new pair	0.721	0.744	0.724	0.756	0.729	0.724	0.769	0.763	0.719	0.774
new drug	0.707	0.639	0.506	0.559	0.525	0.670	0.560	0.687	0.618	0.703

• AUC 5 cross validation

type	1	2	3	4	5
new target	0.880	0.858	0.894	0.910	0.815
new pair	0.921	0.924	0.944	0.941	0.925
new drug	0.888	0.870	0.880	0.902	0.888

• AUPR 5 cross validation

type	1	2	3	4	5
new target	0.632	0.591	0.609	0.633	0.355
new pair	0.719	0.702	0.751	0.731	0.696
new drug	0.622	0.505	0.543	0.560	0.648

2. (549, 424) del none rows & cols

• AUC 10 cross validation

type	1	2	3	4	5	6	7	8	9	10
new target	0.862	0.862	0.800	0.877	0.796	0.845	0.852	0.865	0.797	0.849
new pair	0.891	0.885	0.902	0.923	0.927	0.915	0.925	0.908	0.909	0.914
new drug	0.753	0.807	0.804	0.893	0.729	0.763	0.822	0.805	0.820	0.854

• AUPR 10 cross validation

type	1	2	3	4	5	6	7	8	9	10
new target	0.580	0.527	0.502	0.607	0.468	0.439	0.496	0.529	0.417	0.478
new pair	0.667	0.699	0.651	0.663	0.715	0.684	0.725	0.697	0.667	0.714
new drug	0.385	0.530	0.444	0.700	0.362	0.436	0.486	0.433	0.519	0.403

• AUC 5 cross validation

type	1	2	3	4	5
new target	0.837	0.826	0.815	0.824	0.775
new pair	0.874	0.890	0.913	0.900	0.894
new drug	0.734	0.833	0.800	0.804	0.780

• AUPR 5 cross validation

type	1	2	3	4	5
new target	0.504	0.605	0.449	0.485	0.349
new pair	0.653	0.680	0.676	0.654	0.659
new drug	0.386	0.524	0.412	0.446	0.326

ToDo List:

folds_neo_708_1512: (708, 1512) 5 cv split folds: (549, 424) 5 cv split

(549, 424): DeepDTA & NeoDTI 5 cv in new d, t, p

neo t neo d neo p