1. 2014 Using machine learning to predict binding sites in proteins Jenelle Bray Stanford University October 10, 2014 #GHC14 2014

2. Protein Function  Proteins are biological molecules that: −Catalyze metabolic reactions −Replicate DNA −Transport molecules −Respond to stimuli

3. 2014 Protein Structure

4. 2014 Protein Binding Sites UC Davis ChemWiki

5. 2014 Goal: Predict where ATP Binds  Adenosine triphosphate (ATP) is the primary energy currency of the cell  Transports chemical energy within cells for most reactions that require energy in the cell

6. 2014 ATP Model Based on FEATURE  Builds 3D models of local environment around a protein site given training sets  Calculates chemical properties at varying radial distances from site, and creates a vector containing values of each property in each radial volume  Constructs Naïve Bayes model by comparing distribution of feature vectors between positive and negative sites Liang MP, Banatao DR, Klein TE, Brutlag DL, Altman RB. "WebFEATURE: an interactive web tool for identifying and visualizing functional sites on macromolecular structures." Nucleic Acids Res. 2003 Jul 1;31(13):3324-7. Wei L, Altman RB. Recognizing protein binding sites using statistical descriptions of their 3D environments. Pac Symp Biocomput. 1998:497-508.

7. 2014 Extending the Use of FEATURE  So far, FEATURE only used to predict a protein functional site or a single ion binding site – never a whole small molecule (ligand)  Want to combine FEATURE models to create an overall model to predict ATP binding

8. 2014 Training Set  All PDBs (experimental 3D protein structure files) with ATP bound clustered by 30% sequence similarity, and one protein in each cluster used as positive training set – leads to 190 proteins  For negative training set, proteins with ligands not containing any part of ATP selected, then also clustered by 30% similarity – leads to 3345 proteins  Leave 20% out of training data for validation

9. 2014 Combining Atomic Models  Build individual FEATURE models for 3 atoms in each section of ATP  Need to combine the 9 atomic models to give one overall molecular model  Train a logistic regression model with the atomic FEATURE scores as features

10. 2014 ATP Docking  Want to train model on ATP poses that can actually fit in a binding pocket  For positive proteins, calculate FEATURE score for each of 9 atoms in crystal structure ATP  For negative, use Vina Autodock to dock 1000 ATP poses into a protein  Do this for random sample of negatives equal to number of positive proteins

11. 2014 Choosing ATP Poses for Training  For each negative protein, calculate FEATURE scores of the nine atoms for all 1000 ATP poses, then choose pose with highest sum of (normalized) individual scores −Ensures model can distinguish good ATP poses in non-ATP binding proteins from those in real ATP-binding proteins

12. 2014 Logistic Regression Model  Build logistic regression model with the 9 individual atomic FEATURE scores for each protein in training set

13. 2014 Model Validation  Dock 1000 poses into all training proteins (positive and negative)  Use logistic regression model to score and rank every pose, and choose highest scoring pose for each protein Validation AUC = 0.83 Compares favorably to dock energy (physics based model) with AUC = 0.74

14. 2014 ATP Binding Prediction for a Protein Kinase

15. 2014 Acknowledgments  Russ Altman for supporting the research  Altman group  LinkedIn for sending me to GHC

16. 2014 Got Feedback? Rate and Review the session using the GHC Mobile App To download visit www.gracehopper.org