Revision as of 23:10, 20 October 2019

TU Darmstadt

Modeling

Introduction

In synthetic biology, theoretical models are often used to gain insights, predict and improve experiments. In our project we are modifying Virus-like particles (VLPs) by attaching proteins to the surface of the P22 capsid through a linker. The linking is catalyzed using the enzyme Sortase A7M, which is a calcium independent mutant of the wild type Sortase A from Staphylococcus aureus. We performed modeling to predict the unknown structure of the Sortase A7M, to improve the linker between proteins and therefore optimizing the modification efficiency of our platform.
Two different modeling approaches were used to determine the structure of Sortase A7M. We compared machine learning approaches to traditional comparative, Monte-Carlo based modeling methods. The results were evaluated using an energy-scoring function and molecular dynamics (MD) simulations. The most promising Sortase A7M structures were used to perform a docking simulation to screen for optimal linkers.

In silico modeling and simulation of proteins requires a 3D structure, which can be obtained from the RCSB Protein Data Bank. However, if no 3D structures are annotated, as it is the case with sortase A7M, the structure has to be determined by other means. The structure prediction of sortase A7M was done using two different approaches.

Deep Learning

Background

Machine Learning is a class of algorithms that aim to determine a function between two datasets. This is commonly done by presenting the algorithm with training data as well as a scoring function to measure its success at processing the input data. During training a feedback loop is used to allow the algorithm to automatically find a function to fit the data. In contrast, classical algorithms are often hardcoded to solve a specific problem and only allow for limited flexibility.

A neural network consists of neurons, which are commonly referred to as nodes. They process input using weights, which are adjusted during its training. Nodes in neural networks are linked together: One neuron processes the inputs of other neurons, loosely mimicking the structure of biological brains. While one usually has a fixed amount of input and output neurons limited by the data one wishes to classify, adding layers of hidden neurons can improve the classification. This is often referred to as deep learning and has led to revolutions in applications like speech and image recognition.

Using Machine Learning to predict protein structures has many advantages compared to conventional methods especially for iGEM teams who often only have limited access to resources. After training a neural network, which is a computationally expensive process and often done in centralized data centers, it can be used to predict the structure of a wide variety of proteins. ^[1] Using pretrained models, novel protein structures can be obtained within seconds ^[2] compared to conventional methods taking several hours or days. ^[3]

Until earlier this year the use of Machine Learning in the prediction of protein structures has been restricted to applications within human-written algorithms. ^[2] AlQuarishi demonstrated a complete deep learning approach that is able to make predictions within 1-2 Å of other approaches ^[2] , while only using a fraction of the computational power. This enables accurate structural prediction with less powerful as well as less expensive hardware and thus significantly reduces the cost of structural modeling.

Procedure

We used AlQuarashi’s approach in combination with his pretrained model, which was trained on the Proteinnet database containing all structures released prior to the start of CASP12 (12th Critical Assessment of Techniques for Protein Structure Prediction – 2016). The results were tested against the CASP12 datasets and reached distance root-mean-square deviation (RMSD) values between 10 and 13 Å. The RMSD is defined as root-mean-square deviation of all atom positions compared to a template structure. It is defined as: $$ RMSD = \sqrt{\sum_i^N \left((||v_t - v_i||)^2\right)},$$ where v_i is a vector of all All proteins in the CASP datasets were not published until after the competition and thus represent an assessment with only little bias. ^[4] We used these pretrained datasets to make structural predictions for our Sortase A7M. The predicted structure was then relaxed in a Molecular Dynamics Simulation using GROMACS.

In the following, the specific steps for obtaining a tertiary structure predicted by AlQuarashi’s model are listed.

We used the amino acid sequence of the Sortase A7M in the FASTA format to predict the tertiary structure of the amino acid backbone using AlQuarishi’s Tensor Flow implementation of his end-to-end differentiable learning of protein structure with the pretrained preCASP Proteinnet database. The Output file was a .tertiary file which contains a sequential 3x3 Matrix with atomic coordinates from each amino acid backbone starting at the N-Terminus.
As the standard format for protein structure information is the PDB file format, we wrote a python script to combine the structural information from the FASTA and .tertiary files into a PDB file. For ease of use we used the Biotite Python Module.
Using Rosetta's fixed backbone design program 'fixbb' with the 'hpatch', the optimal position of the side-chains was determined and added to the PDB file. The fixed backbone tool adds the corresponding side-chains and optimizes their conformation. The Hpatch database ensures that hydrophilic side-chains are to be preferred on the surface of the protein as our sortase is present in an aqueous environment.

Results

Animation 1: The raw PDB File converted from the .tertiary file.

Animation 2: The PDB-File after Step 3.

For analysis the Strucure was viewed in Pymol . As can be seen in the pictures below, no secondary structures could be recognized by Pymol. Thus, a Ramachandran Plot was used to evaluate the dihedral angles of the backbone. It was found that the angles do not match with the typical angles for α-helices and β-sheets.

Animation 3: The cartoon view in Pymol.

Figure 1: Ramachandran plot of the predicted structure.

During training the predictions in AlQuarashi’s Model were optimized for their RMSD which is the root-mean-square deviation of the distance between the atoms of the prediction and reference structure. Thus, even though the predictions are expected to have a similar shape to the physical structure, they may not be in the energy minimum. Hence, we applied a GROMACS molecular dynamics in order to relax the structure obtained by AlQuarashi’s deep learning model.