Prediction of Post-translational modification sites and Multi-domain protein structure

Transcription

1 Prediction of Post-translational modification sites and Multi-domain protein structure Robert Newman, Ph.D. Department of Biology Dukka KC, Ph.D. Department of Computational Science and Engineering North Carolina A&T State University Greensboro, NC USA

2 A Brief Overview of Cell Signaling The Akt/PKB Signaling Pathway EMD Calbiochem

3 The Human Kinome Manning et al. (2002)

4 The Human Kinome Manning et al. (2002)

5 The Human Phosphorylome CELLULAR PROTEINS Manning et al. (2002)

6 The Human Phosphorylome CELLULAR PROTEINS Manning et al. (2002)

7 The Human Phosphorylome CELLULAR PROTEINS Manning et al. (2002)

8 The Human Phosphorylome CELLULAR PROTEINS Known P sites: >100,000 Known KSRs: <2,000 Phosite PLUS (Hornbeck et al. (2012))

9

10 Crosstalk Between Cellular Signaling Pathways ROS Signals H2O2 NO O2 Src Akt1 Adapted from Bhattacharyya (2014)

11 Central Theme of the lab Central Theme Elucidation of Protein sequence Structure Function Evolution relationship using various computational approaches Past Projects From sequence to structure Clique based algorithm for protein side-chain packing Refinement of protein structure (KC et al. 2002) From Sequence/structure to function Improvement of protein functional site prediction using phylogenetic motifs and the MINER software (KC, Livesay 2009) From structure to evolution Internal symmetry detection in protein structure (NAR,2014) Evolution of symmetric proteins

12 Outline of the Talk Protein post-translation modification Site prediction Feature Extraction from Protein Sequences (FEPS) Protein Phosphorylation Site Prediction (RF-Phos) Protein Hydroxylation Site Prediction (RF-hydroxysite)

13 Protein Post-translation modification site prediction Post-translational modification (PTM): Chemical modifications occurring in proteins after they are synthesized by ribosomes ( translating mrna into polypeptide chains) PTM increases the functional diversity of the proteome and influence almost all aspects of cell biology. Some PTMS: phosphorylation, hydroxylation, glycosylation, methylation, etc. Identifying and understanding PTMs (as well as sites) is critical in the study of cell biology.

14 Protein Post-translation modification site prediction Discuss our Random Forest based approaches for two of the PTMs site prediction approach RF-Phos: Phosphorylation site prediction approach (Ismail et.al, 2016) RF-Hydroxysite: Hydroxylation site prediction approach (Ismail et al. 2016) Critical thing is extracting the features (numeric values) from protein sequences for Machine Learning Based Approaches

15 FEPS: A tool for feature extraction from protein sequences Ismail, Smith, KC, SubmiLed.

16 FEPS: A tool for feature extraction from protein sequences - Motivation Existing methods: PseAAC (Chou et al, 2008): Computes only one types of features. PROFEAT(Zhu et al. 2011): (10 features) SPiCE (Reinders et al. 2014) ) (18 features) A new tool is required: more features Flexible Easy-to-use With different output formats

17 FEPS: Feature Extraction From Protein Sequence Developed FEPS: web-based bioinformatics tool for sequence-based protein feature extraction available at bcb.ncat.edu/features/ (In Submission) Developed under linux with python + django + JavaScript + mysql. Uses 46 published methods; 6 use any one of the 544 physicochemical properties of amino acids so it makes total of 3304 of different features. Four feature types accept user-defined amino acid indices that make the number of features unlimited. 17

18 Types of Features in FEPS 18

19 FEPS: A tool for feature extraction from protein sequences bcb.ncat.edu/features

20 FEPS: A tool for feature extraction from protein sequences 1- Input Input can be a single files or multiple fasta files Browse to input files Input files

21 FEPS: A tool for feature extraction from protein sequences 2- Feature groups Select one feature type at a time Seven types each with a group of methods Select a method for the list

22 FEPS: A tool for feature extraction from protein sequences 3- Feature type options Tweak feature options

23 FEPS: A tool for feature extraction from protein sequences 3- Physicochemical properties AAindex database Select an amino acid property out of 544

24 FEPS: A tool for feature extraction from protein sequences 3- User-defined indices Amino acid properties User-defined indices

25 FEPS: A tool for feature extraction from protein sequences - Motivation 4- Output options Four file formats are available File formats

26 RF-Phos: Random Forest based protein phosphorylation Site prediction RF-Phos: A Novel General Phosphorylation Site Prediction Tool Based on Random Forest. Hamid D. Ismail, Ahoi Jones, Jung H. Kim, Robert H. Newman, Dukka B. KC., Biomed Res. Intl, ID: ,2016.

27 Phosphorylation Protein synthesis DNA RNA Protein Phosphorylation Addition of a phosphate group to an amino acid residue (Ser, Thr, Tyr) 27

28 Phosphorylation One third of the proteins in the human proteome are substrates for phosphorylation Regulatory functions: Enzymes and receptors are switched off and on Conformational change in structure of enzymes and receptors (activated and deactivated) Plays critical roles in cell cycle, growth, apoptosis and signal transduction pathways. One of the most important post-translational modification sites. Hence, identification of Phosphorylation sites is very important. 28

29 Random Forest: An ensemble classifier Original training data Random Forest (RF) [Breiman et al., 1984] (1) Random bootstrap samples (with replacement) (2) Build decision trees (3) Combine trees

30 Existing methods for Phosphorylation Site Prediction NetPhosK[Blom et al., 2004], GPS2.1 [Xue et al., 2008], NetPhos[Blom et al., 1999], PPRED [Biswas et al., 2010], Musite[Gao et al., 2010], and PhosphoSVM[Dou et al. 2014]. Methods Technique Year NetPhosK ANN 2005 GPS 2.1 PSSM/ GA 2005 NetPhos ANN 2005 Musite SVM 2010 PhosphoSVM SVM 2014 A new method with better performance is highly required.

31 RF-Phos: Random Forest based protein phosphorylation Site prediction Annotated protein sequences with known phosphorylated Ser/Thr/Tyr residues were obtained from P.ELM database [Dinkel et al, 2011]. Annotated fasta format sequence >O S CHK1 86 T CDK2 99 Y ab1 MAQSTATSPDGGTTFEHLWSSLEPDSTYFDLPQSSRG NNEVVGGTDSSMDVFHLEGMTTSVMAQFNLLSSTM DQMSSRAASASPYTPEHAASVPTHSPYAQPSSTF

32 Phosphorylation Site Prediction Redundant sequences (with similarity > 30%) were removed with CD-HIT [Weizhong et al., 2006]. Sliding windows of size 7, 9, 11, 13, 15, 17, 19, 21 were generated from the sequences XXXXXXXXXXXXXXXXX Positive: The windows with experimentally characterized (or +ve) S/T/Y in the middle. XXXX[S/T/Y]XXXX Negative: The windows with -ve S/T/Y in the middle. XXXX[S/T/Y]XXXX 32

33 Phosphorylation Site Prediction: Benchmark Dataset Redundant sequence windows were removed from both +ve windows and ve windows. The remaining windows were used to construct the random forest model. Residue Positive windows (size 9) Negative Before After Used S 20,577 1,554 1,543 T 5, Y 2,

34 Feature Extraction with FEPS Eight types of features were extracted i- Shannon entropy (1 feature) ii- Relative entropy (1feature) iii- Information gain (1 feature) iv- Accessible surface area (9 features) v- Overlapping Properties (90 features) vi- Average cumulative hydrophobicity (4 features) vii- Sequence Features (180 features) viii- Composition, Transition and Distribution (147 features) 34

35 RF- Phos: Overall Algorithm

36 Comparison Metrics (Accuracy) (Precision) (Specificity) (Sensitivity) (F1-score) (Matthew s correlation coefficient)

37 RF- Phos: Random Forest based Predictor for Protein Phosphorylation Sites Comparing RF-Phos to the other methods (10-fold cross validation) Methods Tyrosine AUC Sen (%) Sp (%) MCC NetPhosK GPS Swaminathan NetPhos PPRED Musite PhosphoSVM RF- Phos

38 Phosphorylation Site Prediction Top 10 One of the salient feature of Random Forest is that you can annotate the importance of features 38

39 RF- Phos: Random Forest based Predictor for Protein Phosphorylation Sites

40 RF- Phos: Random Forest based Predictor for Protein Phosphorylation Sites Job ID Download link Predicted sites Annotated sequence

41 RF-Phos: Random Forest based Predictor for Protein Phosphorylation Sites Output screen Output text file

42 RF-Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Ismail, Newman and KC, RF-Hydroxysite: a random forest based predictor for hydroxylation sites, Molecular Biosystems, 12, , 2016.

43 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Addition of OH (Hydroxyl group) Hydroxylase Proline Hydroxyproline Hydroxylase Lysine Hydroxylysine Essential elements of collagen and connective tissues Instability is associated with cancers

44 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Existing methods and motivation add reference SVM-based [Hu et al, 2010], ihyd-pseaac[xu et al, 2014], PredHydroxy [Shi, S.P., et al., 2015] Accuracy Methodology Method P K Feature Algorithm SVM- based PSSM SVM ihyd- PseAAC PseAAC LDA PredHydroxy PWAAC + HQI SVM The previous studies found that the evolutionary and physicochemical information are important. There is still room for improvement

45 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Position weight amino acid composition (PWAA) High quality physicochemical indices (HQI) Type I entropy Overlapping properties (OP) Average cumulative hydrophobicity Protein disordered region Type II entropy Feature Extraction using FEPS

46 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Feature selection (selected ü ) Features P K Position weight amino acid (PWAA) û û Propensity (HQI1) ü û Solvent accessibility (HQI2) û û Alpha-helix frequency (HQI3) ü ü Crystallographic waters (HQI4) ü ü Amino acid composition of MEM (HQI5) ü û Composition of AA in intracellular (HQI6) û û Conformational preference (HQI7) ü û Partition energies (HQI8) ü û Type I entropy (ENTI) ü û Overlapping properties (OP) û û Average cumulative hydrophobicity (ACH) ü ü Protein disordered region (PDR) û û ENTII (Type II entropy) û û

47 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Evaluation results (Jackknife) for different window sizes (lysine and proline) 0.80 K P K P K P K P K P K P K P Accu Prec Sens Spec F1sc MCC Residues Window size

48 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Evaluation results (Independent sample) K P K P K P K P K P K P K P Accu Prec Sens Spec F1sc MCC AUC

49 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Evaluation results Side-by-side comparison (Independent sample window size=15) PredHydroxy RF-Hydroxysite Metrics P L P L Accuracy Sensitivity Specificity MCC

50 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites RF-Hydroxysite

51 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites

52 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Job id Download link Annotated sequence

53 RF- Hydroxysite: Random Forest based Predictor for Protein Hydroxylation Sites Text file

54 Acknowledgements NCAT Startup Fund NSF Beacon Grant