Detecting unfolded regions in protein sequences. Anne Poupon Génomique Structurale de la Levure IBBMC Université Paris-Sud / CNRS France

Transcription

1 Detecting unfolded regions in protein sequences Anne Poupon Génomique Structurale de la Levure IBBMC Université Paris-Sud / CNRS France

2 Large proteins and complexes: a domain approach Structural studies large proteins difficult to produce in sufficient quantities difficult to crystallize (inter domain mobility) Domains are the evolutionary units of proteins discovery of new domains discovery of new structures/functions Domain approach rewarding for the study of complexes combined with electron microscopy

3 objectives bioinformatics tools for the automatic detection of domains identification linker regions genetic engineering for creation of meaningful fragment libraries screening for soluble folded fragments apply to large proteins involved in DNA remodeling/repair multiprotein complexes

4 Unfolded fragments Induced conformational change Linker region between two structural domains

5 Unfolded fragments Induced conformational change Linker region between two structural domains

6 Unfolded fragments and crystallography gene cloning expression solubility crystals Most prediction methods predict regions that might be unfolded under certain c We want to predict regions that are always unfolded

7 E Q E G N A T A I F F L P D G K L H L N E L F E L G N A T A I F L P D E G K Q H L N E L I R S TKF LENEDR F F LPDEGKLQHLENELTHD I GNAT I G N A I F F L P D G H L L E A T K L E N E Q Principles and limits of HCA

8 Principles and limits of HCA Folded Unfolded Folded Requires a lot of manpower and expertise Not automated No statistics

9 Principles and limits of HCA «Why» do we see the unfolded regions on the HCA diagram? Because the «textures» of folded and unfolded regions are different The amino acid composition is different The hydrophobic clusters are different These differences have to be quantified and exploited

10 Data set Swisprot sequence PDB sequence Linker set (L) N-ter fragment Internal fragment C-ter fragment residues residues Set U Prelink chains aa Disopred 75 chains 459 aa

11 Data set PDB seq PDB seq 2 PDB seq 3 PDB seq 3 Only sequences that are unfolded in all homolog PDB sequences are retai

12

13 Amino acid compositions G N A T I F F L P D E G K L Q H L E N E L T H D I I T K F L E N E D R R S aa Probabilities of occurence of this 2 aa sequence in linker (PL) and structured (PS) regions n! PL = n V! n I!... n G! pl V n v pl I n I... pl G n G n! PS = n V! n I!... n G! ps V n v ps I n I... ps G n G

14 I R S TKF LENEDR F F LPDEGKLQHLENELTHD I GNAT I Cluster distance Sequence Differences in hydrophobic clusters

15 Computed parameters G N A T I F F L P D E G K L Q H L E N E L T H D I I T K F L E N E D R R S aa For this residue Ratio Cluster distance Product R = PL/PS d P = R.d These 3 parameters are ploted for all residues in a protein sequence

16 Calibrating on PDB ratio FP : false positives real FP : the regions that are unfolded under certain conditions are removed > aa Rule FP real FP R R R

17 Calibrating on PDB ratio R product P 5 > aa 5<ratio< product> > aa FP Rule Rule 2 5 R R 3 R 2 R P P real FP 6

18 Calibrating on PDB Length < >3 total N-ter Fragments Nb fragments Predicted (%) C-ter Fragments Nb fragments Predicted (%) Internal fragments Nbfragments

19 Calibrating on PDB 3 ratio R product P 5 Rule 3 : P 3 > aa 5<ratio< product> > aa > aa produit>3 Length < >3 total Nbfragments Rules Rules

20 Testing on CASP5 Sensitivity = TP/(TP +FN) Specificity = TN/(TN+FP) Group Sensitivity (%) Sensibility (%) Prelink

21 Testing on Yeast proteins Nb proteins Nb with linker predicted False positives False negative Unfolded 4 7 Structured 6 Soluble not crystallized Chorismate synthase (saccharomyces cerevisiae)

22 Yeast DNA repair protein SET SET 8 Not expressed Not soluble Soluble Crystallized (structure)

23 Limitations Ratio Ratio Ratio Sequence Sequence P922 P523 P Sequence How to account for the «shape» of the curve?

24 Perspectives Not all parameters have been optimized: window length thresholds More parameters are needed neighborhood The rules have been determined empirically We should use statistical learning algorythms We should also try tree learning algorithms

25 Thanks to... Karen Coeytaux Herman van Tilbeurgh Joël Janin Julie Bernauer Sophie Cheruel... and the rest of the Yeast Structural Genomics team.