Geometric filters for protein–ligand complexes based on phenomenological molecular models

Molecular docking is a widely used method of computer-aided drug design capable of accurate prediction of protein-ligand complex conformations. However, scoring functions used to estimate free energy of binding still lack accuracy. Aim. Development of computationally simple and rapid algorithms for...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Datum:	2013
Hauptverfasser:	Sudakov, O.O., Balinskyi, O.M., Platonov, M.O., Kovalskyy, D.B.
Format:	Artikel
Sprache:	English
Veröffentlicht:	Інститут молекулярної біології і генетики НАН України 2013
Schriftenreihe:	Вiopolymers and Cell
Schlagworte:	Bioinformatics
Online Zugang:	https://nasplib.isofts.kiev.ua/handle/123456789/153215
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!
Назва журналу:	Digital Library of Periodicals of National Academy of Sciences of Ukraine
Zitieren:	Geometric filters for protein–ligand complexes based on phenomenological molecular models / O.O. Sudakov, O.M. Balinskyi, M.O. Platonov, D.B. Kovalskyy // Вiopolymers and Cell. — 2013. — Т. 29, №. 5. — С. 418-423. — Бібліогр.: 9 назв. — англ.

Institution

Digital Library of Periodicals of National Academy of Sciences of Ukraine

id	nasplib_isofts_kiev_ua-123456789-153215
record_format	dspace
spelling	nasplib_isofts_kiev_ua-123456789-1532152025-02-09T23:07:59Z Geometric filters for protein–ligand complexes based on phenomenological molecular models Геометричні фільтри для комплексів білок–ліганд на основі феноменологічних молекулярних моделей Геометрические фильтры для комплексов белок–лиганд на основе феноменологических молекулярных моделей Sudakov, O.O. Balinskyi, O.M. Platonov, M.O. Kovalskyy, D.B. Bioinformatics Molecular docking is a widely used method of computer-aided drug design capable of accurate prediction of protein-ligand complex conformations. However, scoring functions used to estimate free energy of binding still lack accuracy. Aim. Development of computationally simple and rapid algorithms for ranking ligands based on docking results. Methods. Computational filters utilizing geometry of protein-ligand complex were designed. Efficiency of the filters was verified in a cross-docking study with QXP/Flo software using crystal structures of human serine proteases thrombin (F2) and factor Xa (F10) and two corresponding sets of known selective inhibitors. Results. Evaluation of filtering results in terms of ROC curves with varying filter threshold value has shown their efficiency. However, none of the filters outperformed QXP/Flo built-in scoring function Pi . Nevertheless, usage of the filters with optimized set of thresholds in combination with Pi achieved significant improvement in performance of ligand selection when compared to usage of Pi alone. Conclusions. The proposed geometric filters can be used as a complementary to traditional scoring functions in order to optimize ligand search performance and decrease usage of computational and human resources. Молекулярний докінг є широко застосовуваним обчислювальним методом пошуку лігандів біомолекул, здатним до достатньо точного передбачення конформацій комплексів білок–ліганд. У той же час скоринговим функціям, що використовують для оцінки сили зв’язування, бракує точності. Мета. Розробка обчислювально простих та швидких алгоритмів для вибору потенційних лігандів з комплексів, отриманих у результаті докінгу. Методи. Створено обчислювальні фільтри, засновані на геометричних співвідношеннях у комплексі білок–ліганд, ефективність яких перевірено крос-докінговим дослідженням із застосуванням кристалічних структур людських серинових протеаз тромбіна (F2) і фактора 10а (F10), а також двох відповідних наборів відомих селективних інгібіторів за допомогою програмного забезпечення QXP/Flo. Результати. Оцінено результати застосування фільтрів у термінах ROC-кривих із змінними пороговими значеннями та показано їхню ефективність. Проте жоден з фільтрів не перевершив за ефективністю вбудовану скорингову функцію Pi програми QXP/ Flo. Тим не менш, використання фільтрів з оптимізованими пороговими значеннями у комбінації з Pi дозволило значно збільшити ефективність порівняно із застосуванням лише Pi. Висновки. Розроблені геометричні фільтри можуть слугувати доповненням до традиційних скорингових функцій для оптимізації пошуку лігандів і зменшення залучення обчислювальних та люд- ських ресурсів. Молекулярный докинг – широко используемый вычислительный метод поиска лигандов биомолекул, способный довольно точно предсказывать конформацию комплекса белок–лиганд. В то же время скоринговые функции, используемые для оценки силы связывания, недостаточно точны. Цель. Разработка вычислительно простых и быстрых алгоритмов для выбора потенциальных лигандов из комплексов, полученных в результате докинга. Методы. Созданы вычислительные фильтры на основе геометрических соотношений в комплексе белок–лиганд, эффективность которых проверена кросс-докинговым исследованием c применением кристаллических структур человеческих сериновых протеаз тромбина (F2) и фактора 10а (F10), а также двух соответствующих наборов известных селективных ингибиторов с помощью программного обеспечения QXP/Flo. Результаты. Оценены результаты применения фильтров в терминах ROC-кривых с переменными пороговыми значениями и показана их эффективность. Однако ни один из фильтров не превзошел по эффективности встроенную скоринговую функцию Pi программы QXP/Flo. Тем не менее, использование фильтров с оптимизированными пороговыми значениями в комбинации с Pi позволило существенно увеличить эффективность в сравнении с применением только Pi. Выводы. Разработанные геометрические фильтры могут служить дополнением к традиционным скоринговым функциям для оптимизации поиска лигандов и уменьшения привлечения вычислительных и человеческих ресурсов. 2013 Article Geometric filters for protein–ligand complexes based on phenomenological molecular models / O.O. Sudakov, O.M. Balinskyi, M.O. Platonov, D.B. Kovalskyy // Вiopolymers and Cell. — 2013. — Т. 29, №. 5. — С. 418-423. — Бібліогр.: 9 назв. — англ. 0233-7657 DOI: http://dx.doi.org/10.7124/bc.000833 https://nasplib.isofts.kiev.ua/handle/123456789/153215 577.322.23 en Вiopolymers and Cell application/pdf Інститут молекулярної біології і генетики НАН України
institution	Digital Library of Periodicals of National Academy of Sciences of Ukraine
collection	DSpace DC
language	English
topic	Bioinformatics Bioinformatics
spellingShingle	Bioinformatics Bioinformatics Sudakov, O.O. Balinskyi, O.M. Platonov, M.O. Kovalskyy, D.B. Geometric filters for protein–ligand complexes based on phenomenological molecular models Вiopolymers and Cell
description	Molecular docking is a widely used method of computer-aided drug design capable of accurate prediction of protein-ligand complex conformations. However, scoring functions used to estimate free energy of binding still lack accuracy. Aim. Development of computationally simple and rapid algorithms for ranking ligands based on docking results. Methods. Computational filters utilizing geometry of protein-ligand complex were designed. Efficiency of the filters was verified in a cross-docking study with QXP/Flo software using crystal structures of human serine proteases thrombin (F2) and factor Xa (F10) and two corresponding sets of known selective inhibitors. Results. Evaluation of filtering results in terms of ROC curves with varying filter threshold value has shown their efficiency. However, none of the filters outperformed QXP/Flo built-in scoring function Pi . Nevertheless, usage of the filters with optimized set of thresholds in combination with Pi achieved significant improvement in performance of ligand selection when compared to usage of Pi alone. Conclusions. The proposed geometric filters can be used as a complementary to traditional scoring functions in order to optimize ligand search performance and decrease usage of computational and human resources.
format	Article
author	Sudakov, O.O. Balinskyi, O.M. Platonov, M.O. Kovalskyy, D.B.
author_facet	Sudakov, O.O. Balinskyi, O.M. Platonov, M.O. Kovalskyy, D.B.
author_sort	Sudakov, O.O.
title	Geometric filters for protein–ligand complexes based on phenomenological molecular models
title_short	Geometric filters for protein–ligand complexes based on phenomenological molecular models
title_full	Geometric filters for protein–ligand complexes based on phenomenological molecular models
title_fullStr	Geometric filters for protein–ligand complexes based on phenomenological molecular models
title_full_unstemmed	Geometric filters for protein–ligand complexes based on phenomenological molecular models
title_sort	geometric filters for protein–ligand complexes based on phenomenological molecular models
publisher	Інститут молекулярної біології і генетики НАН України
publishDate	2013
topic_facet	Bioinformatics
url	https://nasplib.isofts.kiev.ua/handle/123456789/153215
citation_txt	Geometric filters for protein–ligand complexes based on phenomenological molecular models / O.O. Sudakov, O.M. Balinskyi, M.O. Platonov, D.B. Kovalskyy // Вiopolymers and Cell. — 2013. — Т. 29, №. 5. — С. 418-423. — Бібліогр.: 9 назв. — англ.
series	Вiopolymers and Cell
work_keys_str_mv	AT sudakovoo geometricfiltersforproteinligandcomplexesbasedonphenomenologicalmolecularmodels AT balinskyiom geometricfiltersforproteinligandcomplexesbasedonphenomenologicalmolecularmodels AT platonovmo geometricfiltersforproteinligandcomplexesbasedonphenomenologicalmolecularmodels AT kovalskyydb geometricfiltersforproteinligandcomplexesbasedonphenomenologicalmolecularmodels AT sudakovoo geometričnífílʹtridlâkompleksívbíloklígandnaosnovífenomenologíčnihmolekulârnihmodelei AT balinskyiom geometričnífílʹtridlâkompleksívbíloklígandnaosnovífenomenologíčnihmolekulârnihmodelei AT platonovmo geometričnífílʹtridlâkompleksívbíloklígandnaosnovífenomenologíčnihmolekulârnihmodelei AT kovalskyydb geometričnífílʹtridlâkompleksívbíloklígandnaosnovífenomenologíčnihmolekulârnihmodelei AT sudakovoo geometričeskiefilʹtrydlâkompleksovbelokligandnaosnovefenomenologičeskihmolekulârnyhmodelei AT balinskyiom geometričeskiefilʹtrydlâkompleksovbelokligandnaosnovefenomenologičeskihmolekulârnyhmodelei AT platonovmo geometričeskiefilʹtrydlâkompleksovbelokligandnaosnovefenomenologičeskihmolekulârnyhmodelei AT kovalskyydb geometričeskiefilʹtrydlâkompleksovbelokligandnaosnovefenomenologičeskihmolekulârnyhmodelei
first_indexed	2025-12-01T14:53:19Z
last_indexed	2025-12-01T14:53:19Z
_version_	1850318052684464128
fulltext	BIOINFORMATICS UDC 577.322.23 Geometric filters for protein–ligand complexes based on phenomenological molecular models O. O. Sudakov1, O. M. Balinskyi1, M. O. Platonov2, D. B. Kovalskyy3 1Taras Shevchenko National University of Kyiv 64, Volodymyrs’ka Str., Kyiv, Ukraine, 01601 2Institute of Molecular Biology and Genetics, NAS of Ukraine 150, Akademika Zabolotnoho Str., Kyiv, Ukraine, 03680 3Department of Biochemistry, University of Texas Health Science Center 7703 Floyd Curl Drive, San Antonio, TX 78229-3900, USA saa@univ.kiev.ua Molecular docking is a widely used method of computer-aided drug design capable of accurate prediction of protein-ligand complex conformations. However, scoring functions used to estimate free energy of binding still lack accuracy. Aim. Development of computationally simple and rapid algorithms for ranking ligands based on docking results. Methods. Computational filters utilizing geometry of protein-ligand complex were designed. Ef- ficiency of the filters was verified in a cross-docking study with QXP/Flo software using crystal structures of hu- man serine proteases thrombin (F2) and factor Xa (F10) and two corresponding sets of known selective inhibi- tors. Results. Evaluation of filtering results in terms of ROC curves with varying filter threshold value has shown their efficiency. However, none of the filters outperformed QXP/Flo built-in scoring function Pi . Nevertheless, usage of the filters with optimized set of thresholds in combination with P i achieved significant improvement in performance of ligand selection when compared to usage of P i alone. Conclusions. The proposed geometric fil- ters can be used as a complementary to traditional scoring functions in order to optimize ligand search perfor- mance and decrease usage of computational and human resources. Keywords: drug design, molecular modeling, docking, scoring, geometric filtering. Introduction. Nowadays, computer-aided drug design is a widely used technique. It is mostly based on mo- lecular docking and scoring approach [1]. Docking is the procedure of protein (target) and small molecule (li- gand) complexes geometry optimization aimed at fin- ding the global energy minimum of the system. Accor- ding to thermodynamics, the most likely configuration of the complex corresponds to the Gibbs free energy mi- nimum. Energy is usually estimated using certain force field model and its minimization is performed by va- rious methods. Typically, a large collection of small molecules is docked against the protein active site. For each optimized complex, different characteristics (sco- res) are calculated to estimate binding free energy. Li- gands with the highest scores are filtered for further testing in biophysical, biochemical and/or cell-based screening assays. Although there are many publicly available and commercial tools for molecular docking and filters for scoring, some problems still exist. While docking usual- ly can provide adequate results for optimized geometry prediction, scoring is a tricky thing and requires human intrusion like visual inspection of three-dimensional molecular complex structures by a drug discovery ex- pert [2]. As the mechanisms of intermolecular interaction in a protein–ligand complex exceed the classical mecha- nics limits, accurate prediction of binding energy needs quantum mechanical calculations, which boost require- ments for memory size and floating point calculations speed by orders of magnitude. Furthermore, flexibility of both protein and ligand molecules causes an increase 418 ISSN 0233–7657. Biopolymers and Cell. 2013. Vol. 29. N 5. P. 418–423 doi: 10.7124/bc.000833 � Institute of Molecular Biology and Genetics, NAS of Ukraine, 2013 of freedom degrees of a typical system up to hundreds or thousands for simulations with implicit solvent mo- dels and even tens of thousands in case of explicit solvent. As a result of these complications, precise virtual screening of large collections of small molecules be- comes practically impossible, which forced us to use simplified models with empirical scoring algorithms. In this paper, we introduce geometric filters, which are designed to select protein-ligand complexes from the database of molecular docking results. The filters use the molecular geometry of protein–ligand complex as a main filtering characteristic as opposed to appro- ximated potentials of inter-atomic interaction or other loosely defined and computationally expensive func- tions. The main idea behind this approach is based on the fact that molecular docking can predict the molecular geometry stationary point rather accurately, which has been proved by numerous X-ray structural analysis experiments [3]. All mentioned above makes the propo- sed filters robust and quick for interactive usage. Materials and methods. In this study, four types of geometry-based filters were introduced: nearest atom filter (NA), center of mass filter (COM), out coef- ficient filter (OUT) and hydrogen bond filter (HB). Des- cription of the filters is provided below. Nearest atom filter finds atom of the ligand that is the nearest to the given atom of protein in the current complex. Ligand passes the filter if this distance is less than the specified value: min , min l l pr r R � � � � where l is ligand atom index, p is the given protein atom index, rn � is position of the n-th atom, Rmin is the specified minimal distance. This filter has complexity of O(n) and can select ligands that are partially located close to the given atom of the protein active site and evidently screens it from solvent. Filtering results may be modified by considering only ligand atoms of cer- tain type. Center of mass filter finds the distance from ligand center of mass to the given protein atom. Ligand passes the filter if this distance is less than the specified value: r m m R l l l l l � � � � min , where mn is mass of the n-th atom. This filter can select ligands that are located close to the specified atom of the protein active site and evidently screen it. Out coefficient filter calculates numerical characteri- stic which approximates the probability of destroying the given protein-ligand complex. The following model is used. The complex will be destroyed if ligand binds to some external molecule and bonds with protein are des- troyed. The probability of this event, P, may be appro- ximated as P N N e l � , where Ne is average number of ligand atoms that may bind to the external molecule and Nl is total number of ligand atoms. The less P, the more stable the complex . Number Ne is estimated as N pe k e k N l � � , where k is ligand atom index and pk e is the probability that k-th ligand atom will bind to the external molecule. Probability pk e depends on the number of protein atoms that bind to the k-th ligand atom and shield it from outside. Probability of shielding may be described by the Markov field model with the Gibbs distribution: p e k e nk� � , where nk is average number of protein atoms which shield the k-th ligand atom. Number nk may be estima- ted in the same way: n b k p k p N p � � , where bp k is the probability that p-th protein atom binds to the k-th ligand atom. Probability bp k in turn also may be described by Markov field model: b e k p r r R l p kp� � � � � , 419 GEOMETRIC FILTERS FOR PROTEIN–LIGAND COMPLEXES where Rkp is characteristic length of chemical bond bet- ween k-th ligand atom and p-th protein atom. The final expression for P is thus P e N r r R P N k N l l p kp pl � � � � � � � � � ��exp . This filter requires O(n2) operations. To simplify the case, Rkp is defined to be the same for all atom pairs and equal to 1.1 C. In this case, the exact value influ- ences only the value of filtering function but not its be- havior. Hydrogen bonds filter calculates estimated number of hydrogen bonds between ligand and protein atoms. Each hydrogen bond is characterized by strength coefficient that may be estimated as p Ae e e H r r R r r R A H H A H H� � � � � � � � � � 1 2 cos ,� where r A1 � is position of the 1-st acceptor, r A 2 � is posi- tion of the second acceptor, r H � is position of hydrogen atom, RH is characteristic length of hydrogen bond and � is the bond angle. To optimize the filters parameters and verify their effectiveness, we conducted a cross-docking study. Hu- man serine proteases thrombin (gene F2) and factor Xa (gene F10) were selected as targets. X-ray crystal struc- tures of the protein catalytic sites were retrieved from RCSB Protein Data Bank [4], entries 1oyt and 1f0s res- pectively. Two sets of selective small molecule inhibi- tors containing 244 compounds for thrombin and 331 compounds for factor Xa were retrieved from MDDR database [5]. After generation of stereoisomers and ioni- zation using LigPrep software [6], compounds were do- cked into the three-dimensional protein active site struc- ture using QXP/Flo software [7] with 100 steps of SDOCK+ routine. 10 lowest energy complex structures were selected for each compound structure, which re- sulted in a total of 10,580 and 7,410 complexes for thrombin and factor Xa inhibitor sets respectively. Fil- ters were applied to the complexes, and their perfor- mance was evaluated in terms of receiver operating cha- racteristics (ROC). Description of all filters is provided in Table 1. Ar- bitrary atoms of thrombin active site which were selec- ted for the nearest atom filters and center of mass filter are shown in Fig 1. Results and discussion. To evaluate efficiency of the filters, we conducted a virtual screening study whe- re two sets of small-molecule inhibitors of two serine proteases were docked against two sets of their selecti- ve inhibitors. This resulted in a set of protein–ligand complexes with both «native» and «wrong» inhibitors. For each filter, a receiver operating characteristic (ROC) was built. For each of the two proteins, the compounds which have at least one protein-ligand complex passed through a filter were considered positives. Out of positi- ves, essentially, the compounds from one protein’s inhi- bitor set were considered true positives (TP), while com- pounds from another protein’s inhibitor set were consi- dered false positives (FP). Additionally, a ROC was built for the docking software QXP/Flo+ built-in scoring function Pi (Fig. 2). For the two protein crystal struc- tures, we compare only filters which are independent of arbitrary protein atom selection: OUT, HB and Pi. It is clear from the ROCs, that the filters can be used efficiently for selection of inhibitors, except the hydro- gen bond filter in case of factor Xa. However, the QXP/ Flo built-in scoring function outperforms any of the pro- posed filters. Next, we focused on selection of thrombin inhi- bitors with introduction of atom-specific filters NA and 420 SUDAKOV O. O. ET AL. Filter ID Filter name Protein atom Pi QXP/Flo built-in scoring function Pi – OUT Out coefficient – HB Hydrogen bonds – NA182 Nearest atom Leu99 CG NA314 Nearest atom Asp189 OD1 NA63 Nearest atom Tyr60 CD1 COM Center of mass Gly216 HN Table 1 Summary of all filters applied in the study. Residue numbering according to the crystal structure, PDB ID 1oyt COM. ROCs for all filters for this case are provided in Fig. 3. As mentioned in Materials and methods, the num- ber of compounds in the two sets is different. Further- more, as 10 complexes were generated for every stereo- isomer and every possible ionization state, different compounds also have different number of complexes in the docking output. As a result, the number of thrombin inhibitor complexes (true positives) is about 25 % higher than that of factor Xa inhibitors (false positives). To investigate an impact of this inequality, in addition to obvious compound-based random guess ROC (TPR = FPR), a complex-based random guess ROC curve was built (Fig. 3). At this point, all complexes had equ- 421 GEOMETRIC FILTERS FOR PROTEIN–LIGAND COMPLEXES Fig. 1. Three-dimensional structure of human thrombin catalytic site in complex with small- molecule inhibitor retrieved from PDB entry 1oyt. Atoms selected for filtering are shown: 1 – atom 314, Asp189 OD; 2 – center of mass atom, Gly216 NH; 3 – atom 182, Leu99 CG; 4 – atom 314, Tyr60 CD1 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 FPR T P R 1 2 3 4 5 6 7 Fig. 2. Receiver operating characteristics for thrombin and factor Xa, for the filters which are independent of protein atom selection: 1 – FXa PI; 2 – Thrombin PI; 3 – Thrombin HB; 4 – Thrombin OUT; 5 – FXa OUT; 6 – Random guess; 7 – FXa HB 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 FPR T P R 1 2 3 9 8 7 6 5 4 Fig. 3. Receiver operating characteristics for different filters and their combination. ROC curve for random selection of protein-ligand comp- lexes is included; C – Combination; 1 – PI; 2 – HB; 3 – COM; 4 – NA314; 5 – OUT; 6 – NA182; 7 – Random guess (complex based); 8 – NA63; 9 – Random guess (compound based) 422 SUDAKOV O. O. ET AL. al probability to pass a «random guess filter», and this probability was considered the filter parameter, vary- ing along the ROC curve. As one can see from the Fig. 3, all ROC curves are located above the compound-based random guess line, which proves the filters efficiency. Furthermore, all of them are located above complex-based random guess curve, except those for the nearest atom filters NA182 and NA63, which are only effective in certain ranges. However, none of the filters outperformed the built-in scoring function Pi. DesPi te that, use of Pi alone would not be a proper choice. Really, let us consider that in a tyPi cal docking setup, we screen a set of about 50,000 compounds to obtain a docking library of no more than 5,000 compounds, which are going to be tested experi- mentally. As we do not expect more than a few percent of true binders in the initial set, size of the docking lib- rary can be estimated as size of initial set multiplied by false positive ratio (FPR), which in this case should be 10 % at maximum. As one can see from the ROCs, even the best-performing at this FPR filters, Pi and HB, give only about 40 % of true positives, which is generally not acceptable as it means loss of more than half of po- tentially active compounds at the very first stage of drug development. To address this issue, we carried out mul- tiple filtering, in which all protein-ligand complexes were sequentially conducted through all 7 filters, inclu- ding the built-in scoring, thus applying logical conjunc- tion to the filter conditions. In this computation, both TPR and FPR are the functions of 7 variables, which are filter cut-off values. The values of TPR and FPR were samp- led in a broad range of filter parameters to optimize fil- tering performance. The resulting ROC data points are plotted in Fig. 3, and filter cut-off values for them are provided in Table 2. Conclusions. The proposed geometric filters for protein–ligand complexes have shown their efficiency for selection of specific inhibitors in a cross-docking study for serine proteases thrombin and factor Xa. How- ever, their efficiency in terms of receiver operating cha- racteristics is lower than that of QXP/Flo+ native sco- ring function. Nevertheless, the filters can significantly improve virtual screening performance when used in combination with the scoring function. When compa- red to usage of the scoring function alone, target-spe- cific tuning of filtering parameters achieved an incre- ase of TPR from 40 % to 80 % at 10 % FPR, and from 30 % to 65 % at 5 % FPR. Acknowledgements. Software development and computations were performed using Ukrainian Natio- nal Grid Infrastructure and computing cluster of Taras Shevchenko National University of Kyiv [8, 9]. Î. Î. Ñóäàêîâ, Î. Ì. Áàë³íñüêèé, Ì. Î. Ïëàòîíîâ, Ä. Á. Êîâàëüñüêèé Ãåîìåòðè÷í³ ô³ëüòðè äëÿ êîìïëåêñ³â á³ëîê–ë³ãàíä íà îñíîâ³ ôåíîìåíîëîã³÷íèõ ìîëåêóëÿðíèõ ìîäåëåé Ðåçþìå Ìîëåêóëÿðíèé äîê³íã º øèðîêî çàñòîñîâóâàíèì îá÷èñëþâàëüíèì ìåòîäîì ïîøóêó ë³ãàíä³â á³îìîëåêóë, çäàòíèì äî äîñòàòíüî òî÷- íîãî ïåðåäáà÷åííÿ êîíôîðìàö³é êîìïëåêñ³â á³ëîê–ë³ãàíä. Ó òîé æå ÷àñ ñêîðèíãîâèì ôóíêö³ÿì, ùî âèêîðèñòîâóþòü äëÿ îö³íêè ñèëè çâ’ÿçóâàííÿ, áðàêóº òî÷íîñò³. Ìåòà. Ðîçðîáêà îá÷èñëþ- âàëüíî ïðîñòèõ òà øâèäêèõ àëãîðèòì³â äëÿ âèáîðó ïîòåíö³éíèõ ë³ãàíä³â ç êîìïëåêñ³â, îòðèìàíèõ ó ðåçóëüòàò³ äîê³íãó. Ìåòîäè. Ñòâîðåíî îá÷èñëþâàëüí³ ô³ëüòðè, çàñíîâàí³ íà ãåîìåòðè÷íèõ ñï³ââ³äíîøåííÿõ ó êîìïëåêñ³ á³ëîê–ë³ãàíä, åôåêòèâí³ñòü ÿêèõ ïå- ðåâ³ðåíî êðîñ-äîê³íãîâèì äîñë³äæåííÿì ³ç çàñòîñóâàííÿì êðèñ- òàë³÷íèõ ñòðóêòóð ëþäñüêèõ ñåðèíîâèõ ïðîòåàç òðîìá³íà (F2) ³ ôàêòîðà 10à (F10), à òàêîæ äâîõ â³äïîâ³äíèõ íàáîð³â â³äîìèõ ñå- ëåêòèâíèõ ³íã³á³òîð³â çà äîïîìîãîþ ïðîãðàìíîãî çàáåçïå÷åííÿ QXP/Flo. Ðåçóëüòàòè. Îö³íåíî ðåçóëüòàòè çàñòîñóâàííÿ ô³ëüò- ð³â ó òåðì³íàõ ROC-êðèâèõ ³ç çì³ííèìè ïîðîãîâèìè çíà÷åííÿìè òà ïîêàçàíî ¿õíþ åôåêòèâí³ñòü. Ïðîòå æîäåí ç ô³ëüòð³â íå ïå- TPR FPR Pi OUT HB NA182 NA314 NA63 COM 0.926 0.242 3.1 0.77 1.0 4.6 9.1 4.8 5.8 0.869 0.160 3.4 0.75 1.0 4.7 9.8 4.7 5.8 0.803 0.094 3.5 0.73 0.9 4.7 8.9 4.6 6.3 0.721 0.057 3.7 0.75 0.9 4.7 8.4 4.5 6.5 0.631 0.048 3.7 0.75 0.9 4.7 8.1 4.3 6.4 Table 2 Best-scoring filter cut-off value combinations. Values for nearest atom and center of mass filters are in angstroms (C) ðåâåðøèâ çà åôåêòèâí³ñòþ âáóäîâàíó ñêîðèíãîâó ôóíêö³þ Pi ïðî- ãðàìè QXP/ Flo. Òèì íå ìåíø, âèêîðèñòàííÿ ô³ëüòð³â ç îïòèì³- çîâàíèìè ïîðîãîâèìè çíà÷åííÿìè ó êîìá³íàö³¿ ç Pi äîçâîëèëî çíà÷- íî çá³ëüøèòè åôåêòèâí³ñòü ïîð³âíÿíî ³ç çàñòîñóâàííÿì ëèøå Pi. Âèñíîâêè. Ðîçðîáëåí³ ãåîìåòðè÷í³ ô³ëüòðè ìîæóòü ñëóãóâàòè äîïîâíåííÿì äî òðàäèö³éíèõ ñêîðèíãîâèõ ôóíêö³é äëÿ îïòèì³çà- ö³¿ ïîøóêó ë³ãàíä³â ³ çìåíøåííÿ çàëó÷åííÿ îá÷èñëþâàëüíèõ òà ëþä- ñüêèõ ðåñóðñ³â. Êëþ÷îâ³ ñëîâà: êîìï’þòåðíà ðîçðîáêà ë³ê³â, ìîëåêóëÿðíå ìî- äåëþâàííÿ, äîê³íã, ñêîðèíãîâà ôóíêö³ÿ, ãåîìåòðè÷í³ ô³ëüòðè. À. À. Ñóäàêîâ, À. Ì. Áàëèíñêèé, Ì. Î. Ïëàòîíîâ, Ä. Á. Êîâàëüñêèé Ãåîìåòðè÷åñêèå ôèëüòðû äëÿ êîìïëåêñîâ áåëîê–ëèãàíä íà îñíîâå ôåíîìåíîëîãè÷åñêèõ ìîëåêóëÿðíûõ ìîäåëåé Ðåçþìå Ìîëåêóëÿðíûé äîêèíã – øèðîêî èñïîëüçóåìûé âû÷èñëèòåëüíûé ìåòîä ïîèñêà ëèãàíäîâ áèîìîëåêóë, ñïîñîáíûé äîâîëüíî òî÷íî ïðåäñêàçûâàòü êîíôîðìàöèþ êîìïëåêñà áåëîê–ëèãàíä. Â òî æå âðåìÿ ñêîðèíãîâûå ôóíêöèè, èñïîëüçóåìûå äëÿ îöåíêè ñèëû ñâÿ- çûâàíèÿ, íåäîñòàòî÷íî òî÷íû. Öåëü. Ðàçðàáîòêà âû÷èñëèòåëü- íî ïðîñòûõ è áûñòðûõ àëãîðèòìîâ äëÿ âûáîðà ïîòåíöèàëüíûõ ëèãàíäîâ èç êîìïëåêñîâ, ïîëó÷åííûõ â ðåçóëüòàòå äîêèíãà. Ìå- òîäû. Ñîçäàíû âû÷èñëèòåëüíûå ôèëüòðû íà îñíîâå ãåîìåòðè- ÷åñêèõ ñîîòíîøåíèé â êîìïëåêñå áåëîê–ëèãàíä, ýôôåêòèâíîñòü êîòîðûõ ïðîâåðåíà êðîññ-äîêèíãîâûì èññëåäîâàíèåì ñ ïðèìåíå- íèåì êðèñòàëëè÷åñêèõ ñòðóêòóð ÷åëîâå÷åñêèõ ñåðèíîâûõ ïðîòå- àç òðîìáèíà (F2) è ôàêòîðà 10à (F10), à òàêæå äâóõ ñîîòâåò- ñòâóþùèõ íàáîðîâ èçâåñòíûõ ñåëåêòèâíûõ èíãèáèòîðîâ ñ ïîìî- ùüþ ïðîãðàììíîãî îáåñïå÷åíèÿ QXP/Flo. Ðåçóëüòàòû. Îöåíåíû ðåçóëüòàòû ïðèìåíåíèÿ ôèëüòðîâ â òåðìèíàõ ROC-êðèâûõ ñ ïå- ðåìåííûìè ïîðîãîâûìè çíà÷åíèÿìè è ïîêàçàíà èõ ýôôåêòèâ- íîñòü. Îäíàêî íè îäèí èç ôèëüòðîâ íå ïðåâçîøåë ïî ýôôåêòèâ- íîñòè âñòðîåííóþ ñêîðèíãîâóþ ôóíêöèþ Pi ïðîãðàììû QXP/Flo. Òåì íå ìåíåå, èñïîëüçîâàíèå ôèëüòðîâ ñ îïòèìèçèðîâàííûìè ïî- ðîãîâûìè çíà÷åíèÿìè â êîìáèíàöèè ñ Pi ïîçâîëèëî ñóùåñòâåííî óâåëè÷èòü ýôôåêòèâíîñòü â ñðàâíåíèè ñ ïðèìåíåíèåì òîëüêî Pi. Âûâîäû. Ðàçðàáîòàííûå ãåîìåòðè÷åñêèå ôèëüòðû ìîãóò ñëó- æèòü äîïîëíåíèåì ê òðàäèöèîííûì ñêîðèíãîâûì ôóíêöèÿì äëÿ îïòèìèçàöèè ïîèñêà ëèãàíäîâ è óìåíüøåíèÿ ïðèâëå÷åíèÿ âû÷èñ- ëèòåëüíûõ è ÷åëîâå÷åñêèõ ðåñóðñîâ. Êëþ÷åâûå ñëîâà: êîìïüþòåðíàÿ ðàçðàáîòêà ëåêàðñòâ, ìîëå- êóëÿðíîå ìîäåëèðîâàíèå, äîêèíã, ñêîðèíãîâàÿ ôóíêöèÿ, ãåîìåòðè- ÷åñêèå ôèëüòðû. REFERENCES 1. Hurmach V. V., Balinskyi O. M., Platonov M. O., Borysko P. O., Prylutskyy Y. I. Molecular docking method involving SH2-do- mains // Biotechnology.–2012.–5, N 2.–P. 31–40. 2. Cole J. C., Murray C. W., Nissink J. W., Taylor R. D., Taylor R. Comparing protein-ligand docking programs is difficult // Proteins.–2005.–60, N 3.–P. 325–332. 3. Wang R., Lu Y., Wang S. Comparative evaluation of 11 scoring functions for molecular docking // J. Med. Chem.–2003.–46, N 12.–P. 2287–2303. 4. Berman H. M., Westbrook J., Feng Z., Gilliland G., Bhat T. N., Weissig H., Shindyalov I. N., Bourne P. E. The protein data bank // Nucleic Acids Res.–2000.–28, N 1.–P. 235–242. 5. Schuffenhauer A., Zimmermann J., Stoop R., van der Vyver J.J., Lecchini S., Jacoby E. An ontology for pharmaceutical ligands and its application for in silico screening and library design // J. Chem. Inf. Comput. Sci.–2002.–42, N 4.–P. 947–955. 6. LigPrep User manual, Version 2.3.–New York: Schrodinger, 2009.–116 p. 7. Mcmartin C., Bohacek R. S. QXP: Powerful, raPi d computer algorithms for structure-based drug design // J. Comput. Aided Mol. Des.–1997.–11, N 4.–P. 333–344. 8. Zynovyev M., Svistunov S., Sudakov O., Boyko Y. Ukrainian Grid Infrastructure: Practical Experience // Proc. 4-th IEEE Work- shop IDAACS 2007 (September 6–8, 2007, Dortmund, Germa- ny).–Dortmund, 2007.–P. 165–169. doi:10.1109/IDAACS. 2007.4488397 9. Salnikov A. O., Sliusar I. A., Sudakov O. O., Savytskyi O. V., Kor- nlyuk A. I. MolDynGrid Virtual Laboratory as a part of Ukrai- nian Academic Grid infrastructure // Proc. IEEE International Workshop on Intelligent Data Acquisition and Advanced Com- puting Systems: Technology and Applications (21–23 Septem- ber 2009, Rende (Cosenza), Italy.–Rende, 2009.–P. 237–240. Received 02.04.13 423 GEOMETRIC FILTERS FOR PROTEIN–LIGAND COMPLEXES

Geometric filters for protein–ligand complexes based on phenomenological molecular models

Institution

Ähnliche Einträge