Modelling

To assist Wet Lab in selecting candidates for surface display, AlphaFold was used to generate 3D models of the fusion proteins. Structural alignment on PyMOL allowed for the investigation of conformational changes. Finally, the stability of the proteins were analyzed through molecular dynamics simulations with GROMACS.

Introduction

The Wet Lab team of meduCA intended to leverage surface display mechanism to display the constructed fusion proteins, which consist of surface protein and a specific carbonic anhydrase, on Caulobacter crescentus, Escherichia coli and Cyanobacteria (UTEX).

Surface protein used:

• Vibrio—Colwellia—Bradyrhizobium—Shewanella repeat protein (VCBS)
• Ice Nucleation Protein N-terminal (INPN)
• S-layer protein RsaA (RsaA) We utilize AlphaFold to predict the structure of the fusion proteins, and investigate the conformational changes of the structures and their stability under different environmental conditions for in-silico analysis and selecting the candidates for cell display.

Background

Carbonic Anhydrase (CA) are enzymes that catalyze the reversible hydration of carbon dioxide [1] . Some of the CA candidates for surface display investigated were of interest to Wet Lab, while others were chosen as a result of Bioinformatics . When a surface protein is fused with a CA, the resulting fusion protein could exhibit conformational changes on the enzymatic part. This is a critical consideration for Wet Lab because CA activity is required for microbial induced calcium carbonate precipitation (MICP) when forming biocement. Therefore, results from this project will aid Wet Lab in deciding which surface protein and CA strain to include in the fusion protein for the overall success of MICP. Conformational changes could impact enzyme activity, which influences the rate of calcium carbonate precipitation and the strength and durability of the resulting biocement.

Cell surface display is a technique in which functional proteins are anchored to the outer membrane, enabling their direct presentation on the cell surface, providing better access to external environment for calcification [2] . Heterologous proteins can be expressed on the cell surface, allowing direct interaction with the environment and facilitates functional assays without cell lysis [3] . The N-terminal domain of INPN and VCBS domains from certain Gram negative bacteria can be used to anchor enzymes to the outer membrane, enabling surface display [4] .

Construct of surface protein

• Cyanobacteria: VCBS N-terminal + CA + VCBS C-terminal
• E. coli: INPN + CA
• C. crescentus : RsaA N-terminal + CA + RsaA C-terminal

Prediction of protein structures was done using AlphaFold. [5] is a deep learning based tool developed by Google DeepMind that predicts the 3D structure of proteins from their amino acid sequences. Through machine learning, this artificial intelligence was trained off the Protein Data Bank (PDB), which is a database that links specific proteins to experimentally determined and publicly available macromolecular structures. Traditionally, determining protein structures required experimental methods like X-ray crystallography or cryo-EM which are extremely costly [6]. AlphaFold instead uses neural networks to infer how residues interact and fold, producing highly accurate models from protein sequences. The quality of the predicted structure is evaluated by predicted Local Distance Difference Test (pLDDT) score, which estimates the confidence of the predicted residue position on a scale of 0 to 100 [7].

Structural alignment was conducted and evaluated on predicted structures using Python Molecular Visualizer ([8] ). PyMOL is a molecular visualization tool available both as a software and as a Python library, allowing researchers to interactively explore structures or script analyses. This graphical program allows alignment to be conducted both manually and programmatic use, allowing for consistent re-runs for different input structures. In PyMOL, structural alignment is a method for comparing two or more protein structures to assess their similarity through computing a root mean square deviation (RMSD) minimization. PyMOL calculates the best alignment and report the RMSD between corresponding backbone atoms by superimposing one structure onto another, measured in Ångströms (Å). A lower RMSD, with values below 2 Å, generally considered high quality for backbone alignment and indicates higher structural similarity [9] . This process is especially useful for validating AlphaFold predictions against experimentally determined structures and analyzing conformational changes.

$RMSD = \sqrt{\frac{1}{N}\sum^N_{i=1}(x_i-\hat{x_i})^2}$

where:

$x_i$ = Backbone atom coordinates of predicted protein structure

$\hat{x_i}$ = Backbone atom coordinates of reference protein structure

These results will be useful for Wet Lab to consider when selecting candidates for surface display, as RMSD highlights which fusion proteins are predicted to fold with lower conformational change, potentially indicating a better preservation of enzymatic activity.

GROningen MAchine for Chemical Simulations ([10]) is a versatile package to perform molecular dynamics (MD). GROMACS numerically integrates Newton’s equations of motion for systems comprising millions of particles. This project employed MD simulations using GROMACS to investigate how the structural stability of the fusion proteins respond to various pH values (4, 6, 7, 9), analyzing structural fluctuations and compactness.

The two metrics recorded for each fusion protein is

• Root-Mean-Square Deviation (RMSD) RMSD of the backbone was computed relative to the reference structure. This metric tracks how far the protein diverges from its starting conformation during the simulation. Higher values of RMSD suggest less structural stability.

• Radius of Gyration (Rg) The radius of gyration is a measure of how compact or extended the molecule is over time. A more stable Rg implies a more tightly folded, compact structure, whereas large fluctuations in Rg indicates unfolding of the structure. $R_g = \sqrt{\frac{\sum_i m_i |r_i - r_{com}|^2}{\sum_i m_i}}$ where: $m_i$ = Mass of atom $i$ $r_i$ = Position vector of atom $r_{com}$ = center of mass of the structure

Aim

This project aims to develop two pipelines that are reusable for other teams.

• The structural analysis and validation of fusion proteins for cell surface display using AlphaFold.
• The molecular dynamics of the fusion proteins at various pH values (pH of 4, 6, 7, 9) to investigate protein stability using GROMACs.

Methods

The aim of this pipeline is to be reusable for future teams, runnable on high performance computing (HPC), and automate batch AlphaFold and structural alignment.

High-performance computing cluster is a system made up of many powerful computers working together. Tasks can be scheduled to use specific computing resources, which is important for running simulations that are too demanding for a regular personal computer. By distributing the workload across multiple computers, HPC makes it possible to obtain results more efficiently while work with larger datasets.

Graphics processing units (GPU) acceleration is important because they are optimized for parallel processing, which makes them particularly effective for calculations used in molecular dynamics. Offloading intensive tasks to GPUs significantly reduces simulation time and increases overall performance, enabling longer simulations with greater efficiency.

Preliminary AlphaFold Run

AlphaFold architecture involves querying the input sequence against database and evaluating large neural networks, both of which requires massive computing resources. We wanted to perform an initial run of AlphaFold and examine the quality of the predicted structures, so we chose to use Google DeepMind’s AlphaFold web server, which allows users to upload the .fasta files containing the protein sequence and perform AlphaFold prediction in the cloud without the need to manage the computational workflow themselves. We tried using both DNA sequence and amino acid sequence, which are both acceptable input formats of AlphaFold.

The predicted CA structure had mostly high confidence pLDDT scores, while the fusion protein had regions with lower scores. The predicted CA also agreed with the experimentally solved structure on Protein Data Bank base on the RMSD from structural alignment. This is as expected since AlphaFold is trained from the experimentally solved structures on Protein Data Bank.

Interestingly, we also found that the quality scores are different depending on whether the input is in DNA of amino acid form. Since protein folding prediction models are optimized for amino acid sequences and using them avoids translation-related inconsistencies, amino acid sequences were used as input to ensure consistent and reliable results moving forward.

Batch AlphaFold Structure Prediction Pipeline

While Google DeepMind’s AlphaFold web server allows simple and accessible AlphaFold runs, it can only process one protein sequence at a time. Therefore, to perform batch protein structure prediction, it is necessary to create a pipeline that utilizes computing resources and allow multiple AlphaFold runs simultaneously. We have decided to use Nextflow, which is an workflow management system commonly used in bioinformatics. Initially, we wanted to adapt an existing AlphaFold pipeline from nf-core/proteinfold. nf-core is a community that consists of a curated collection of Nextflow pipelines. However, to perform AlphaFold on high-performance computing cluster, the database need to be laid out in a specific manner depending on the version of AlphaFold. Proteinfold expects a different database layout, which we could not manage to set up on the HPC cluster since the source files were not provided. To ensure maximum compatibility and reproducibility, we have chosen to directly use the more well-defined version AlphaFold2.

For a single sequence, AlphaFold produce five predicted structures. They are always named by ranked_0.pdb, ranked_1.pdb, …, ranked_4.pdb, and in decreasing order of confidence. We only examine ranked_0.pdb for downstream analysis since it has the highest confidence.

Structural Alignment Pipeline

To examine whether the CA in the predicted fusion protein structure exhibit conformational changes that could affect their enzymatic activity, we wanted to perform structural alignment and evaluate the results. Originally, we used PyMOL’s graphical user interface to test a few predicted strutcures. We then built a pipeline with Nextflow that can execute multiple structural alignment given a list of fusion proteins structures, CA structures form PDB, and optionally, the surface protein structures, and write the RMSD output to text file. We executed this pipeline locally, as opposed to on HPC, as structural alignment requires far less compute than structural prediction. The outputs are summarized in the results and discussion section.

Molecular Dynamics Simulation Pipeline

AlphaFold is an excellent tool for predicting the static structure of proteins, but it does not provide information on how protein stability or activity may vary under different environmental conditions. Because the bacteria surface display the CAs are embedded in a biobrick that contains martian regolith, and we do not know the local pH, we need to consider the effects of pH on enzymatic activity. Changes in pH can alter the protonation states of key active-site residues, such as histidines and acidic side chains, potentially affecting hydrogen bonding networks and electrostatic interactions, which in turn may influence enzymatic activity. We decided to use molecular dynamics (MD) simulations with the software GROMACS to investigate the effects of pH on the structural stability of the CAs.

GROMACS doesn’t natively support pH-dependent MD simulation, but we can tackle this by protonating ionizable residues prior to simulation to approximate specific pH conditions. We found a web-based tool H++ that can generate the .pdb files of proteins at specific pH. [11] However, H++‘s generated .pdb files are designed to be compatible with Assisted Model Building with Energy Refinement (AMBER), another software commonly used for MD simulation. Specifically, we used tleap from AmberTools to solvate and ionize the protein system and generate AMBER topology (.prmtop) and coordinate (.inpcrd) files. [12] These were then converted into GROMACS-compatible formats (.gro and .top) using the Python library ParmEd. We selected GROMACS over AMBER because it is highly optimized for efficiency and scalability, making it an ideal choice for large-scale simulations [10].

Specifications

We created reusable batch AlphaFold, structural alignment and molecular dynamics simulation pipeline. All the necessary files in this section are provided in our GitLab repository. This section contains the technical specification.

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in RFC 2119.

AlphaFold2 Pipeline

• You MUST have Apptainer, Nextflow (we used V25.04.7) and installed in HPC environment.
• You MUST modified the paths in the provided filenextflow.config to reflect the path of your directories on HPC cluster.
• The input files MUST all have .fasta extension, and be located in params.fasta_dir . They should all be in DNA or amino acid format for consistent AlphaFold outputs.
• this process SHOULD produce the outputs in params.alphafold_outdir They will appear as individual folders with the fasta file’s name. Each folder will contain five .pdb files named by ranked_0.pdb to ranked_4.pdb , ranked by the confidence of prediction in descending order. There will also be five .json files which corresponds to the .pdb files that stores the per-residue pLDDT scores. We used the predicted structure ranked_0.pdb with the highest confidence for structural alignment.

Structural Alignment Pipeline

• You MAY want to visually inspect the pLDDT score of the AlphaFold predicted structure. You can do this by opening ranked_0.pdb in the PyMOL Graphical User Interface and executing the following command:

spectrum b, blue_red, minimum=0, maximum=100

This colors the structure based on the pLDDT score assigned to each residue. Red indictates high confidence and blue indictates low confidence.

• The folder layout MUST be in the following format. Each folder within structures/MUST contain one fusion protein, and reference structures to perform the alignment. The fusion protein must be named ranked_0.pdb Our Nextflow script loops through each folder and align the fusion protein to all other .pdb files in the folder.

├─ pymol.nf
├─ pymol.config
└─ structures/
   ├─ HpCA_INPN/
   │   └─ ranked_0.pdb
   │   └─ INPN.pdb
   │   └─ HpCA.pdb
   └─ SazCA_VCBS/
       └─ ranked_0.pdb
       └─ c_domain.pdb
       └─ n_domain.pdb
       └─ SazCA.pdb

• To run the script, execute this command in the parent folder:

nextflow pymol.nf -c pymol.config

• You SHOULD expect to find a file alignment_rmsds.txt in each of the folder. In example above, that will be HpCA_INPN/ and SazCA_VCBS/ An example of the content of this text file is provided below:

RMSD (fusion vs SazCA.pdb): 0.2152739316225052
RMSD (fusion vs n_domain.pdb): 2.418569564819336
RMSD (fusion vs c_domain.pdb): 17.096477508544922

Molecular Dynamics Pipeline

• You MUST use the following parameters when processing a structure on H++, the rest of the parameters can be kept as default.
  • Water Model: TIP3P
  • Add counter-ions: Neutralize using Na+/Cl-
• You SHOULD change the parameter “The pdb structure will be protonated assuming pH of” to the desired pH at which you want the protein to be protonated.
• After H++ finishes processing the structure, you MUST click on “VIEW_RESULTS” and download the file labeled: “AMBER-compatible PDB (PQR) structure in predicted protonation state.”, which is the input for the molecular dynamics script.

maestro workflows

The structural prediction, alignment, and molecular dynamics pipelines were converted into maestro workflows for redistribution, improved reliability, and correctness. To learn more, visit the maestro page, or the repositories on GitLab [structural analysis] [molecular dynamics].

Results and Discussion

PyMOL alignment results of the fusion proteins can be observed in Table 1. The name of each fusion protein indicates the strain of carbonic anhydrase (CA) and the associated surface protein. The strains of CA used were CA from Burkholderia henselae (BhCA), CA I from Brucella suis (BtCAI), CA II from Brucella suis (BtCAII), CA from Helicobacter pylori (HpCA), and CA from Sulfurihydrogenibium azorense (SazCA). The CA fusion proteins were displayed using three chassis systems: UTEX cyanobacteria with VCBS anchors, E. coli BL21 with INPN-based display, and Caulobacter crescentus CB2A with its native RsaA S-layer.

Fusion Protein	Alignment against CA	Alignment against surface protein	Alignment against N-terminal of surface protein	Alignment against C-terminal of surface protein
BhCA_VCBS	17.751	/	2.433	13.949
BtCAI_VCBS	0.165	/	0.461	6.714
BtCAII_VCBS	0.535	/	0.339	27.038
HpCA_VCBS	0.718	/	2.240	8.435
SazCA_VCBS	0.215	/	2.419	17.010
BhCA_INPN	8.466	20.761	20.761	/
BtCAII_INPN	11.995	27.222	27.222	/
HpCA_INPN	4.210	9.586	9.586	/
SazCA_INPN	3.316	20.310	20.310	/
BhCA_RsaA	11.310	/	1.771	/
BtCAII_RsaA	10.357	/	1.491	/
HpCA_RsaA	45.313	/	1.444	/
SazCA_RsaA	51.307	/	1.541	/
BtCA_BL21_RsaA	27.805	/	11.380	/
BhCA_RsaA_secreted	/	/	/	0.089
BtCAII_RsaA_secreted	/	/	/	1.041
SazCA_RsaA_secreted	/	/	/	0.100
HpCA_RsaA_secreted	/	/	/	9.861

Table 1. Root mean square deviation of structural alignment results when aligning fusion proteins against their CA and surface proteins in PyMOL. Fusion protein are named by [CA name]_[surface protein].

The AlphaFold predicted fusion proteins were imported into PyMOL, and structural alignment was performed to assess where the CA in the fusion proteins has undergone significant conformational change. The results of the structural alignment pipeline are presented in Table 1.

Wet Lab explained the surface protein is expected to undergo conformational change when fused with a CA because it serves as a scaffold in surface display. This explains why the RMSD values are higher than 2 Å for almost all surface protein alignments. Nevertheless, these results remain relevant to the project when observing the conformational change of the surface protein.

The results from Figure 3 shows that VCBS fused with BtCAI produced the lowest RMSD compared to the alignments investigated. This suggests that this fusion protein exhibits the least conformational change, and be more likely to be catalytically efficient when surface displayed. When using BtCAII in the fusion protein, the alignment against C terminal of surface protein produced the highest RMSD value and BhCA showed the highest RMSD value for alignment against CA. Therefore, Wet Lab is advised against using BtCAII and BhCA for surface display because it could potentially have an effect of enzymatic activity.

The results from Figure 4 shows that HpCA and INPN would have the lowest overall RMSD and BtCAII or BhCA produced the best results for fusion proteins with RsaA according to Figure 5.

Since these results are predictions, the conformational changes may not affect protein function. However, the results are worth noting for Wet Lab because the abnormally high RMSD values of fusion proteins with RsaA could signify incompatibility between the CA and the surface protein, potentially disrupting the function of the CA. The selected fusion protein candidates will be tested by Wet Lab for successful surface display and enzymatic activity.

In Figures 6 and 7, the RMSD profiles indicate that structural fluctuations are highest at pH 4 for both SazCA and HpCA. At higher pH values (6, 7, and 9), the RMSD values for SazCA show reduced fluctuations, while HpCA reaches a stable plateau between 0.10 and 0.12 nm. These results suggest that both SazCA and HpCA maintain greater structural stability and remain closer to their initial conformations under less acidic conditions.

Interestingly, BtCAII displays its greatest RMSD fluctuation at pH 6, followed by a noticeable drop around 0.9 ns. This apparent instability may reflect incomplete equilibration or insufficient simulation time to reach a steady state, rather than an inherent loss of structural stability. Furthermore, there was less fluctuation measured for the radius of gyration in Figure 11 at pH 7 compared to Figures 9 and 10 at pH 4 and 6 respectively, indicating a more compact and conformationally stable ensemble at higher pH values.

Conclusion and Future Directions

Since enzymatic activity requires the right conformation, the CA in the fusion protein is likely active if their structure remains relatively unchanged. These results provide a basis for estimating the probability of successful surface display, guiding the Wet Lab in selecting which strain of CA and which structural protein are most likely to generate functional constructs.

AlphaFold predicts the proteins as static structures, but our CAs will be deposited in martian regolith, which could have varying local pH. The results from MD simulations allowed for the exploration of the stability of the CAs under different pH conditions, providing insight to Wet Lab regarding the optimal pH range for CA activity in martian regolith.

In the future, docking studies could be performed in addition to MD simulation to examine ligand-enzyme interactions under different pH conditions. By accounting for changes in protonation states and protein conformation, these studies can provide insights into how pH affects binding affinity, stability, and specificity of the CA. Such information would help guide the design of experiments and the selection of optimal conditions for functional assays.