Prediction of protein disorder Zsuzsanna Dosztányi MTA-ELTE Momentum Bioinformatics Group Department of Biochemistry Eotvos Lorand University, Budapest, Hungary

Protein Structure/Function Paradigm Dominant view: 3D structure is a prerequisite for protein function

But…. Heat stability Protease sensitivity Failed attempts to crystallize Lack of NMR signals Increased molecular volume “Freaky” sequences …

IDPs Intrinsically disordered proteins/regions (IDPs/IDRs) Do not adopt a well-defined structure in isolation under native-like conditions Highly flexible ensembles Functional proteins

p53 tumor suppressor transactivation DNA-binding tetramerization regulation Wells et al. PNAS 2008; 105: 5762 TAD DBDTDRD Disordered region

Bioinformatics of protein disorder Part 1 Prediction of protein disorder  Databases  Prediction of protein disorder Part 2 Biology of disordered proteins  Prediction of functional regions within IDPs

Datasets Ordered proteins in the PDB  over structures  few 1000s folds Some structures in the PDB classify as disordered! only adopt a well-defined structure in complex in crystals, with cofactors, proteins, … Disorder in the PDB  Missing electron density regions from the PDB  NMR structures with large structural variations  Less than 10% of all positions  Usually short (<10 residues), often at the termini

Disprot Current release: 6.02 Release date: 05/24/2013 Number of proteins: 694 Number of disordered regions: 1539 Experimentally verified disordered proteins collected from literature (X-ray, NMR, CD, proteolysis, SAXS, heat stability, gel filtration, …)

Additional databases Combining experiments and predictions Genome level annotations  MobiDB:  D2P2:  IDEAL:

Sequence properties of disordered proteins Amino acid compositional bias High proportion of polar and charged amino acids (Gln, Ser, Pro, Glu, Lys) Low proportion of bulky, hydrophobhic amino acids (Val, Leu, Ile, Met, Phe, Trp, Tyr) Low sequence complexity Signature sequences identifying disordered proteins Protein disorder is encoded in the amino acid sequence

Amino acid compositions He et al. Cell Res. 2009; 19: 929

Prediction methods for protein disorder Based on amino acid propensity scales or on simplified biophysical models GlobPlot, FoldIndex, FoldUnfold, IUPred, UCON Machine learning approaches PONDR VL-XT, VL3, VSL2; Disopred; POODLE S and L ; DisEMBL; DisPSSMP; PrDOS, DisPro, OnD-CRF, POODLE-W, RONN Over 50 methods

1.Amino acid propensity scale GlobPlot Compare the tendency of amino acids:  to be in coil (irregular) structure.  to be in regular secondary structure elements Linding (2003) NAR 31, 3701

GlobPlot From position specific predictions Where are the ordered domains? Longer disordered segments? Noise vs. real data

GlobPlot: downhill regions correspond to putative domains (GlobDom) up-hill regions correspond to predicted protein disorder

Globular proteins Large entropy penalty Large number of inter-residue contacts

2. Physical principles IUPred If a residue cannot form enough favorable interactions within its sequential environment, it will not adopt a well defined structure it will be disordered Dosztanyi (2005) JMB 347, 827  Based on an energy estimation method  Parameters calculated from statistics of globular proteins  No training on disordered proteins

IUPred The algorithm: …PSVEPPLSQETFSDLWKLLPENNVLSPLPSQAMDDLMLSPDDIEQWFTEDPGPDEAPRMPEAAPRVAPAPAAPTPAA... Based only on the composition of environment of D’s we try to predict if it is in a disordered region or not: Amino acid composition of environ- ment: A – 10% C – 0% D – 12 % E – 10 % F – 2 % etc… Estimate the interaction energy between the residue and its environment Decide the probability of the residue being disordered based on this

3. Machine learning approaches INPUTOUTPUT.ATVQLSMIWQSTR..ATVQLSMIWQSTR. D O

DISOPRED2: …..AMDDLMLSPDDIEQWFTED….. Assign label: D or O D O F(inp) SVM with linear kernel Ward (2004) JMB 337, 635

DISOPRED2 Cutoff value!

PONDR VSL2 Differences in short and long disorder  amino acid composition  methods trained on one type of dataset tested on other dataset resulted in lower efficiencies PONDR VSL2: separate predictors for short and long disorder combined length independent predictions Peng (2006) BMC Bioinformatics 7, 208

4. Metaservers: Sequence Disorder prediction methods PONDR VLXT PONDR VL3 PONDR VSL2 IUPred FoldIndex TopIDP Prediction ANN Meta-predictor Xue et al. Biochem Biophys Acta. 2010; 180: 996

Disordered regions and secondary structure Coil is an ordered, irregular structural element Disordered proteins usually do not contain stable secondary structural elements  (e.g. by CD) They can contain transient secondary structure elements  (by NMR) Pure random coil never occurs Use secondary structure predictions methods for disordered proteins with extreme caution Long segments without predicted secondary structure may indicate proteins disorder (NORsnet)

Accuracy True positive: Disordered residues are predicted as disordered False positive: Ordered residues predicted as disordered True negative: Ordered residues predicted as ordered False negative: Disordered residues predicted as ordered 75-90%

Prediction of protein disorder Disordered residues can be predicted from the amino acid sequence  ~ 80% at the residue level Methods can be specific to certain type of disorder  accordingly, accuracies vary depending on datasets Predictions are based on binary classification of disorder

Heterogeneity in protein disorder Flexible loop RC-like Compact Transient structures

Modularity in proteins Many proteins contains multiple domains Composed of ordered and disordered segments Average length of a PDB chain is < 300 Average length of a human proteins ~ 500 Average length of cancer-related proteins > 900 Structural properties of full length proteins …

Practical