dpm
ChemAxon UGM Budapest 2012
Contribution of 2D, 3D structural features of drug molecules in the prediction of Drug Profile Matching Ágnes Peragovics Department of Biochemistry Eötvös Loránd University Budapest, Hungary
DRUG PROFILE MATCHING (DPM) METHOD
Interaction Profile (IP) generation
Drug
Effect Profile (EP) generation
drugs
drugs
Interaction Profile matrix A set of random proteins
Effect Profile matrix
HIERARCHICAL EFFECT DATABASE
Main effect
Subcategories ACE inhibitor Calcium channel blocker Renin blocking agent Angiotensin II receptor antagonist
Antihypertensive agent
Neutral endopeptidase inhibitor Sodium depletion diuretic Sympathetic blocker Vasodilator Diuretic
DRUG PROFILE MATCHING (DPM) METHOD
Interaction Profile matrix
Effect Profile matrix
Effect Probability matrix
PREDICTION LIST – ACE INHIBITORS Drug name Probability Captopril 1.000 Cilazapril 1.000 Ramipril 1.000 Methacycline 1.000 Trandolapril 1.000 Quinapril 1.000 Benazepril 1.000 Moexipril 1.000 Rescinnamine 1.000 Fosinopril 0.999 Rolitetracycline 0.999 Deserpidine 0.997 Enalapril 0.997 Lisinopril 0.996 Buprenorphine 0.991 Spirapril 0.990 Dasatinib 0.988 Telmisartan 0.985 Diatrizoate 0.977
True Positive (TP)
False Positive (FP)
False Negative (FN)
True Negative (TN)
ROC CURVES: MEASURE OF ACCURACY
TPR
AUC=1.0
AUC=0.50
FPR
TRADITIONAL 2D AND 3D SIMILARITY SEARCHES
2D descriptors
3D descriptors
ChemAxon chemical fingerprints Tanimoto similarity
ChemAxon Flexible 3D alignment alignment scores
2D similarity matrix
3D similarity matrix
Screening protocol for a given effect: 1. Select an active molecule as a query structure 2. Sort the database molecules by decreasing similarity 3. Calculate AUC to measure accuracy 4. Repeat steps 1-3 for each active molecule of a given effect 5. Calculate average AUC and standard deviation
RESULTS OF DRUG CLASSIFICATION
DPM
2D
3D
DPM
2D
3D
RIGID STRUCTURAL CATEGORIES
Benzodiazepines DPM
2D
3D
1.00
0.96±0.04
0.95±0.03
Glucocorticoids
DPM
2D
3D
1.00
0.99±0.01
1.00±0.00
STRUCTURALLY DIVERSE EFFECTS
Anti-inflammatory agent
DPM
2D
3D
0.93
0.52±0.11
0.50±0.03
Anxiolytic agent DPM
2D
3D
0.94
0.59±0.08
0.60±0.11
Anti-glaucoma agent DPM
2D
3D
0.96
0.49±0.04
0.51±0.04
STRUCTURALLY DIVERSE EFFECTS Fast sodium channel inhibitor
GABA agent
NMDA receptor Calcium channel antagonist agent
Antiepileptic agent DPM
2D
3D
0.97
0.55±0.07
0.50±0.11
Antiepileptic agent. NOS
CONCLUSIONS
1. Rigid structural categories can be handled effectively by 2D and 3D searches. The more complex DPM method is not required in these cases. 2. The main strength of DPM is to screen wide effect categories containing structurally diverse molecules.
Possible explanation: interaction to non-target protein sites
biologically more relevant rare conformers?
THANK YOU FOR YOUR ATTENTION!
Department of Psychiatry and Psychotherapy Semmelweis University
Department of Biochemistry Eötvös Loránd University
Printnet Ltd.
Pál Czobor István Bitter Gábor Csukly László Tombor
András Málnási-Csizmadia Zoltán Simon Balázs Jelinek Csaba Hetényi Margit Vigh-Smeller Anna Rauscher László Végner Zhenhui Yang Gergely Zahoránszky-Kőhalmi
Péter Hári Zoltán Brandhuber Domonkos Nagy Máté Marót Barna Bíró
PREDICTIVE POWER ON EXTERNAL DATA?
Random splitting experiment Training set
Test set
Modified 2D and 3D search • same query molecules as in DPM training set • search performed on the same test sets
RESULTS OF RANDOM SPLITTING DPM
2D
3D
DPM
2D
3D
RESULTS OF RANDOM SPLITTING
RESULTS OF RANDOM SPLITTING
TEN-FOLD CROSS-VALIDATION
ROBUSTNESS VALUES FOR MAIN EFFECTS AND SUBGROUPS
Contribution of 2D, 3D structural features of drug molecules in the prediction of Drug Profile Matching Ágnes Peragovics Department of Biochemistry Eötvös Loránd University Budapest, Hungary
DRUG PROFILE MATCHING (DPM) METHOD
Interaction Profile (IP) generation
Drug
Effect Profile (EP) generation
drugs
drugs
Interaction Profile matrix A set of random proteins
Effect Profile matrix
HIERARCHICAL EFFECT DATABASE
Main effect
Subcategories ACE inhibitor Calcium channel blocker Renin blocking agent Angiotensin II receptor antagonist
Antihypertensive agent
Neutral endopeptidase inhibitor Sodium depletion diuretic Sympathetic blocker Vasodilator Diuretic
DRUG PROFILE MATCHING (DPM) METHOD
Interaction Profile matrix
Effect Profile matrix
Effect Probability matrix
PREDICTION LIST – ACE INHIBITORS Drug name Probability Captopril 1.000 Cilazapril 1.000 Ramipril 1.000 Methacycline 1.000 Trandolapril 1.000 Quinapril 1.000 Benazepril 1.000 Moexipril 1.000 Rescinnamine 1.000 Fosinopril 0.999 Rolitetracycline 0.999 Deserpidine 0.997 Enalapril 0.997 Lisinopril 0.996 Buprenorphine 0.991 Spirapril 0.990 Dasatinib 0.988 Telmisartan 0.985 Diatrizoate 0.977
True Positive (TP)
False Positive (FP)
False Negative (FN)
True Negative (TN)
ROC CURVES: MEASURE OF ACCURACY
TPR
AUC=1.0
AUC=0.50
FPR
TRADITIONAL 2D AND 3D SIMILARITY SEARCHES
2D descriptors
3D descriptors
ChemAxon chemical fingerprints Tanimoto similarity
ChemAxon Flexible 3D alignment alignment scores
2D similarity matrix
3D similarity matrix
Screening protocol for a given effect: 1. Select an active molecule as a query structure 2. Sort the database molecules by decreasing similarity 3. Calculate AUC to measure accuracy 4. Repeat steps 1-3 for each active molecule of a given effect 5. Calculate average AUC and standard deviation
RESULTS OF DRUG CLASSIFICATION
DPM
2D
3D
DPM
2D
3D
RIGID STRUCTURAL CATEGORIES
Benzodiazepines DPM
2D
3D
1.00
0.96±0.04
0.95±0.03
Glucocorticoids
DPM
2D
3D
1.00
0.99±0.01
1.00±0.00
STRUCTURALLY DIVERSE EFFECTS
Anti-inflammatory agent
DPM
2D
3D
0.93
0.52±0.11
0.50±0.03
Anxiolytic agent DPM
2D
3D
0.94
0.59±0.08
0.60±0.11
Anti-glaucoma agent DPM
2D
3D
0.96
0.49±0.04
0.51±0.04
STRUCTURALLY DIVERSE EFFECTS Fast sodium channel inhibitor
GABA agent
NMDA receptor Calcium channel antagonist agent
Antiepileptic agent DPM
2D
3D
0.97
0.55±0.07
0.50±0.11
Antiepileptic agent. NOS
CONCLUSIONS
1. Rigid structural categories can be handled effectively by 2D and 3D searches. The more complex DPM method is not required in these cases. 2. The main strength of DPM is to screen wide effect categories containing structurally diverse molecules.
Possible explanation: interaction to non-target protein sites
biologically more relevant rare conformers?
THANK YOU FOR YOUR ATTENTION!
Department of Psychiatry and Psychotherapy Semmelweis University
Department of Biochemistry Eötvös Loránd University
Printnet Ltd.
Pál Czobor István Bitter Gábor Csukly László Tombor
András Málnási-Csizmadia Zoltán Simon Balázs Jelinek Csaba Hetényi Margit Vigh-Smeller Anna Rauscher László Végner Zhenhui Yang Gergely Zahoránszky-Kőhalmi
Péter Hári Zoltán Brandhuber Domonkos Nagy Máté Marót Barna Bíró
PREDICTIVE POWER ON EXTERNAL DATA?
Random splitting experiment Training set
Test set
Modified 2D and 3D search • same query molecules as in DPM training set • search performed on the same test sets
RESULTS OF RANDOM SPLITTING DPM
2D
3D
DPM
2D
3D
RESULTS OF RANDOM SPLITTING
RESULTS OF RANDOM SPLITTING
TEN-FOLD CROSS-VALIDATION
ROBUSTNESS VALUES FOR MAIN EFFECTS AND SUBGROUPS
