ProteinShop: A Tool for Protein Structure Prediction and Modeling Silvia Crivelli Computational Research Division Lawrence Berkeley National Laboratory

The Protein Structure Prediction Problem To determine how proteins, the building blocks of living cells, fold themselves into three-dimensional shapes that define the role they play in life.

Importance of Protein Structure Prediction The shape of a protein determines its function. Knowledge of structure is used in many ways: –Drug design –Design of synthetic proteins –Re-engineering defective proteins Genome projects are providing sequences for many proteins whose structure will need to be determined.

Protein Structures ProGlyLeuSer Proteins consist of a long chain of amino acids, the primary structure N OH R H N O H R H N OH R H N O H R H N O H R H N O H R H N O H R H N O H R H Side chain H-bond Backbone Amino acid

Protein Structures ProGlyLeuSer Proteins consist of a long chain of amino acids, the primary structure The constituent amino acids may encourage hydrogen bonding that form regular structures, called secondary structures The secondary structures fold together to form a compact 3-dimensional shape, called the tertiary structure  -helix  -sheet

The problem can be formulated as a global minimization problem, as it is assumed that the tertiary structure occurs at the global minimum of the free energy function of the primary sequence Ab Initio Approach Our Goal: To provide an approach that relies more on physical principles than on information from known proteins

Ab Initio Method Tertiary structure is believed to minimize potential energy: Min V MM (x) where x = atom coordinates Difficulties: Proposed energy function may not match nature O(e n 2 ) local minima Very large parameter space e.g., modestly sized protein 100 amino acids ~ 1,600 atoms ~ 4,800 variables

The Search Algorithm Given the amino acid sequence of a protein, find the global minimum of the free energy function. Generate Starting Configurations Global Optimization Phase 1Phase 2

Secondary Structure Predictions in Phase 1 SKIGIDGFGRIGRLVLRAALSCGAQ CBBBBBCCCAAAAAAACCCBBBBBC Sequence: Type: Weight: Sequence: Servers predict secondary structure likely to be in a target protein based on a large database of known proteins.

Matching the predicted strands is a combinatorial problem Which strands are paired? Which orientation? ? ? ? parallel anti-parallel Which residues are paired? odd even

There are n!2 n-2 possible n-stranded motifs 96 motifs for n=4 960 motifs for n=5 It takes weeks to create some of these configurations using constrained local minimizations! Distribution of Beta Sheets in Proteins with Applications to Structure Prediction Ruckzinski, Kooperberg, Bonneau, and Baker, Proteins 48,2002

CASP4 Competition Fourth community-wide experiment on the Critical Assessment of Techniques for Protein Structure Prediction (2000) Our group predicted 8 proteins Largest protein had 240 aa Most complex fold had 2 β-strands

ProteinShop Interactive tool for protein manipulation Designed to quickly create initial configurations It takes weeks to create a number of configurations using constrained minimizations It takes a few hours to create the same configurations with ProteinShop

Phase 1 with ProteinShop Phase 1 Amino Acid Sequence Phase 2 Initial Configurations Final Configuration 2 nd ary Structure Prediction Geometry Generation Structure Sequence Direct Manipulation Pre-configuration Initial Configurations ProteinShop takes minutes

CASP4 Competition (before ProteinShop) CASP5 Competition (with ProteinShop) Our group predicted 20 proteins Largest protein had 417 aa Most complex fold had 13 β-strands Our group predicted 8 proteins Largest protein had 240 aa Most complex fold had 2 β-strands

Phase 2 Phase 1 Amino Acid Sequence Phase2: Global Optimization Initial Configurations Final Configuration Subspace Selection Initial Configurations Subspace Optimization Candidate Selection Final Configuration Takes months to converge using hundreds of processors on Seaborg!

Phase 2 with ProteinShop Phase 1 Amino Acid Sequence Phase2: Global Optimization Initial Configurations Final Configuration Subspace Selection Initial Configurations Subspace Optimization Candidate Selection Final Configuration Monitoring System Direct Manipulation Steering System Will reduce computation time

Monitoring System Monitor progress of overall optimization/each optimization process

Monitoring System Monitor progress of overall optimization/each optimization process Alert user to important events during optimization A sudden drop in internal energy A group of processes getting stuck Test new heuristics for expanding nodes of the tree

Steering System Change configurations during optimization to account for developments not anticipated during Phase 1 Manipulate proteins that don’t seem to be realistic or that are stuck in a local minimum Allow pruning of the optimization tree Assign multiple processes to a configuration that just had a drop in internal energy Assign stuck processes to other configurations

Plans for the Future  Use of the monitoring and steering features to develop and test a new method for protein structure prediction  Compete in CASP6 (Critical Assessment of Techniques for Protein Structure Prediction)  Expand and enhance ProteinShop

O. Kreylos, N. Max, B. Hamann, S. Crivelli, and W. Bethel. Interactive Protein Manipulation, Winner of the Best Application Award IEEE Visualization 2003, Seattle. ProteinShop Available to academic and non-profit organizations proteinshop.lbl.gov