Bridging Molecular Timescales with MELD and Blue Waters Alberto - - PowerPoint PPT Presentation

Bridging Molecular Timescales with MELD and Blue Waters

Alberto Perez

We need to know protein structures to make new drugs

DNA Protein sequence Protein structure Rational Drug Design

We need to know protein structures to make new drugs

DNA Protein sequence Protein structure Rational Drug Design EASY Expensive Time consuming Not always possible HARD

Key challenge: develop computational tools to predict protein structures from sequence

A grand challenge in structural biology is predicting the 3D structure of a protein given the sequence

Nguyen, H., Maier, J., Huang, H., Perrone, V. & Simmerling, C.

• J. Am. Chem. Soc. 136, 13959–13962 (2014).

6 months of continuous sampling is not enough for even a simple protein

Molecular Modeling is a computational grand challenge

Perez, A., Morrone, J. A. & Dill, K. WIREs Comput Mol Sci 125, e1309 (2017).

Computational brute force will not solve these grand challenges

When will we be able to fold larger proteins?

We developed MELD to scale to larger systems

Perez, A., Morrone, J. A., Simmerling, C. & Dill, K. Curr. Opin. Struct. Biol. 36, 25–31 (2016). Perez, A., MacCallum, J. L., Coutsias, E. A. & Dill, K.. J. Chem. Phys. 143, 243143 (2015).

MD is the basis of our method

force field

• Sparse
• Ambiguous
• Noisy

+ data

I lost my keys in the beach

MELD uses a Bayesian inference approach to incorporate data into simulations

MacCallum*, Perez*, & Dill, Proc. Natl. Acad. Sci. U.S.A. 112, 6985–6990 (2015). force field

We use Hamiltonian Replica Exchange to enhance sampling Low Temperature High Temperature / / Strong Restraints Vanishing Restraints

95 residues 67 residues 68 residues

MELD performed high accuracy blind predictions of 3D structure

Å Å Å

Top cluster experiment prediction

Perez et al. Science Advances (2016)

Blue Waters is key for CASP — the structure prediction competition

• 3 months — daily new targets
• 200 competing groups and methods
• Hundreds of proteins
• Strict deadlines (some as short as 5-7 days)
• We are the only physics-based methodology in

CASP

BW’s team help indispensable during CASP

• 30 GPU nodes per protein
• Sparse communication between nodes
• Helping with compilation of the OpenMM/MELD

plugin

Beyond folding — binding and pathways

1 2 3 4 5 6 7 8 9 10 lllll rrrrr rrlll rrrll rllll rrrrl lrrrr llrrr lllrr llllr

11 12 13 14 15 16 17 18 19 20 21 22

llrll rrlrr rlllr lrrrl lrlll rrrlr lrrll rrllr rlrrr lllrl rllrr llrrl

• 1. Perez, A., Sittel, F., Stock, G. & Dill, K. J Chem Theory Comput 14, 2109–2116 (2018).
• 2. Morrone, J. A. et al. J Chem Theory Comput 13, 863–869 (2017).
• 3. Morrone, J. A., Perez, A., MacCallum, J. & Dill, K. J Chem Theory Comput 13, 870–876 (2017).

1 2 3 4 5 6 7 8 9 10 lllll rrrrr rrlll rrrll rllll rrrrl lrrrr llrrr lllrr llllr

Don’t miss these posters for more details!

Protein Folding of Nonthreadables

James Robertson, Alberto Perez, Ken Dill Protein Structure Prediction Remains Important and Challenging MELD is an Accelerator for Molecular Simulations

• MELD uses temperature and Hamiltonian

replica exchange molecular dynamics (MD) to enhance conformational sampling and give free energies

• MELD simulations run on GPU-accelerated

supercomputers like NCSA Blue Waters

MELD Folds Nonthreadable Proteins

• MELD populations are predictor of folding
• MELD is limited by force field deficiencies

• Threading methods predict 86% of human

protein structures, but many proteins are nonthreadable

• MELD folds proteins fast, is physics-

based, and not limited like threading

• Can MELD fold nonthreadable proteins?

Emiliano Brini James Robertson

Thanks!

alberto.perez17@gmail.com