1. Introduction

We conducted molecular modeling analysis on the single-chain antibody fragment variable (scFv) employed in this project to investigate the interaction between the scFv and the carcinoembryonic antigen (CEA) present on the surface of tumor cells. We employed molecular docking and corresponding calculations to predict the structure of the antibody and its affinity for CEACAM5.

To assess the practical value of our antibody, we thoroughly examined the exploitability of scFv, including aggregation and stability. Traditionally, researchers have reduced the immunogenicity of mouse antibodies by grafting the CDR region from mouse antibodies onto the human variable region framework, thus creating humanized antibodies¹ . Using the original sequence as a reference, we utilized modeling techniques to identify the key amino acid sites involved in antigen binding. Through virtual mutation of these sites, we aimed to enhance the affinity of the anti-CEA scFv and enable the engineered bacteria to exhibit stronger tumor cell targeting ability. In general, the protein molecular modeling process encompasses structural prediction of Lpp-OmpA-scFv and molecular docking with CEACAM5, followed by the modification of the antibody proteins based on an understanding of their complex structure (Figure 1).

2. Structure prediction of Lpp-OmpA-scFv

Background

Since the structure of the Lpp-OmpA-scFv is unknown, our first step is to predict the structure of the antibody. AlphaFold2 (AF2), an advanced deep learning model, has achieved unprecedented performance in predicting the structure of single-chain proteins². UCSF ChimeraX is the excellent molecular visualization tool launched by Resource for Biocomputing, Visualization, and Informatics (RBVI) following UCSF Chimera. We attempted to open the AlphaFold tool in ChimeraX and use ColabFold to make new protein predictions³.

Methodology

We obtained the amino acid sequences of the Lpp-OmpA-scFv protein in a FASTA file. Additionally, we opened the AlphaFold panel in ChimeraX and copied the protein's amino acid sequence into the designated box.

In the Options, we selected "Energy-minimize predicted structures", "Trim fetched structure to the aligned structure sequence" and "Use PDB templates when predicting structures". We clicked on the predict button to initiate the execution. Selecting all options may cause longer processing time, but it yields more accurate predictions. The 3D structure of the protein was predicted using ColabFold, a free computational environment provided by Google.

Results

Upon analyzing the prediction results, we downloaded the optimal prediction (Figure 2.A).We assessed the reliability of each segment of the predicted structure by analyzing the additional generated images(Figure 2.B, C, D). The AlphaFold prediction provided expected position error values for each residue pair (X.Y),showing the predicted position error at residue X when aligned with residue Y in the true structure. These residue-residue "predicted aligned error" values can be visualized with an error plot (Figure 2.B).Additionally, a sequence coverage plot was generated to examine the number of similar sequences found at different positions in the Lpp-OmpA-scFv (Figure 2.C).The predicted structures include atomic coordinates and confidence estimates for each residue, with scores ranging from 0 to 100. Higher scores indicate higher confidence. This confidence measure is called pLDDT and corresponds to the model's predicted per-residue scores on the IDDT-Ca metric (Figure 2.D).

Analysis

To predict the transmembrane protein regions of the aforementioned structure, we incorporated an implicit membrane into the protein structure. Additionally, we modified the membrane properties by utilizing the Analyze Transmembrane Proteins tools in Discovery Studio. The Hidden Markov Model (HMM) was employed to predict the transmembrane helices based on the amino acid sequence of the protein. Subsequently, a hidden membrane consisting of two parallel planes was introduced to the protein structure (Figure 3). The placement of the membrane was determined by optimizing the simplified solvation energy. If there was a significant charge difference between protein residues located outside the membrane, adjustments were made to the membrane.

3. Molecular docking

Background

We employed molecular docking to investigate the interaction between the antigenic protein and the scFv. Carcinoembryonic antigen related cell adhesion molecule 5(CEACAM5), also known as CEA or CD66e, belongs to the carcinoembryonic antigen family⁴. ZDOCK is a protein interaction docking program based on fast Fourier transform. It is primarily utilized to explore all potential binding modes by translating and rotating two proteins in space, and subsequently assess each binding model through an energy-based scoring function. In ZDOCK version 3.0.2, IFACE statistical potential energy, structural complementarity, and electrostatic scoring functions are employed. ZDOCK 3.0.2 was employed to dock CEACAM5 and Lpp-OmpA-scFv proteins, and the most optimal docking result was selected upon completion. The PyMol V2.4.0 software was utilized to label and present the binding sites of the docking complex..

Methodology

The structure of the human CEACAM5 protein was searched for in the PDB database (https://www.rcsb.org/) and subsequently downloaded. The Lpp-OmpA-scFv prediction from the previous step was optimally prepared. We uploaded both protein structures and performed the ZDOCK 3.0.2 prediction. Following this, we downloaded and installed PyMol V2.4.0 software. Upon completion of the molecular docking, we opened the resulting complex structure file with the best prediction using PyMol V2.4.0 software. To delete all solvent molecules and display sticks within 5 angstroms of scFv or CEACAM5, we entered the command below into the command line.

remove solvent
PyMOL>bg_color white
PyMOL>show sticks, byres CEACAM5 within 5 of scFv
PyMOL>show sticks, byres scFv within 5 of CEACAM5

Afterward, we conducted a search and selection process to identify residues that establish contact bonds, which were then utilized for labeling and displaying the binding sites of the docking complex. We particularly emphasized amino acids capable of forming hydrogen bonds on the interaction surface. Following that, we uploaded the docking complexes to PDBePISA in order to analyze the interaction domains and surfaces of the proteins involved.

Results

Amino acids capable of forming hydrogen bonds in the docking molecules are shown in the picture (Figure 4). The gray dotted line is the amino acid residue interaction hydrogen bond, the capital letter is the abbreviation of the corresponding amino acid, and the number is the position information of the amino acid in the protein sequence. For example, S362 is serine 362 in CEACAM5 protein. G130 is the glycine 130 in Lpp-OmpA-scFv protein. Upload the best predicted protein structure in PDBePISA, analyze the interaction between the two proteins and download the summary of the interaction surface (Table 1).

Notes:(A)Results of molecular docking, Lpp-OmpA-scFv is shown in green and CEACAM5 is shown in orange. (B)Schematic representation of the molecular docking surface. Among them, amino acids that can form hydrogen bonds are shown.

4. Developmental analysis of antibodies

Background

High concentrations of antibodies often result in their aggregation, which can diminish their activity and potentially induce an immune response⁵. The calculation of antibody aggregation trend is an indicator to measure the tendency of protein surface amino acid aggregation⁶. Identifying the protein surface specify regions that are prone to aggregation allows us to use targeted mutation to engineer proteins with higher stability. The aggregation of scFv proteins is also an important factor affecting their developability. We calculated Developability Indices (DI) (Fv) scores for assessing the exploitability of scFv.

Methodology

We use the apply forcefield command in the change forcefield tools to apply a CHARMm forcefield to the Lpp-OmpA-scFv protein structure. Then we calculated the solvent accessible surface area (SAA) for the side-chain of each residue in Discovery Studio. In the Tools Explorer, we launched the Macromolecules and Predict Protein Aggregation Tools panel, click ‘Calculate Aggregation Scores’. In the parameter setting interface that appears, we chose the proteins for the calculation and set the Cutoff Radius to 5,7,10. Per atom aggregation propensity score is calculated as the ratio of the actual side-chain SAA to the SAA of side-chain atoms in a fully exposed residue of the same type, multiplied by the residue hydrophobicity, for all residues within the Cutoff Radius of each atom. We use the Developability Index (DI) to analysis anti-CEA scFv protein, which is a measure used to rank the aggregation propensity of related proteins based on their hydrophobic and total charge properties. The DI is calculated from the aggregation propensity score (APS) minus the weighted squared total of the charge as described by the following formula:

DI = [APS] - β x [total charge]²

[where APS is the positive aggregation score and is calculated as the summation of all positive atomic aggregation scores.]

Results

The color is based on the protein aggregation trend score of the Cutoff radius=10. Atoms with a high protein aggregation trend score will be shown in red, while atoms with a low protein aggregation trend score will be shown in blue (Figure 5.C). In addition, a line chart and a point chart are automatically created (Figure 5.A, Figure 5.B). In the line chart, the amino acid sequence number was taken as the horizontal coordinate, and the protein aggregation trend score was taken as the vertical coordinate. In the point chart, the sequence number of protein aggregation sites was taken as the horizontal coordinate, and the score of protein aggregation trend was taken as the vertical coordinate. Site 1 is shown in red in Figure.5C, and we click the arrow on the right side of the Display Style at the top of the window to better show the details of the amino acids in Figure 5.D. Because anti-CEA scFv protein is an antibody with variable regions, the APS is calculated for these to generate DI(Fv) scores (Table 2).

Notes: (A). Line chart of protein aggregation tendency score. (B).Point chart of aggregation tendency score.(C).Solvent surfaces stained according to protein aggregation tendency scores.(D).Show detail of site 1.

5. Redesign of antibody molecules

① Humanization of anti-CEA scFv by CDR Grafting

Background

Humanization is the process of engineering an antibody that retains the antigen-binding specificity and affinity of a non-human antibody, but exhibits low human immunogenicity and does not compromise stability⁷. Humanized scFv can mitigate the human immune response against mouse antibodies. Humanized scFv can reduce human anti-mouse antibody response. This project we aimed to humanize the heavy chain variable (V_H) and light chain variable (V_L) regions of the anti-CEA scFv to develop a humanized scFv exhibiting minimal immunogenicity while maintaining high binding affinity for CEACAM5.

Methodology

The FASTA file of anti-CEA-scFv (without Lpp-OmpA) obtained in the previous steps were prepared. Open the FASTA file in Discovery Studio and load the sequences to annotate in a Sequence Window. Before humanizing scFv, we first identified and annotated domains and Complementarity Determining Regions (CDR) of antibody sequences. Several Hidden Markov Models (HMM) are precomputed for the protocol and used to scan the input protein sequence to identify the antibody domains. Then we selected the sequence, clicked on "Predict Humanizing Mutations" and set the target scores parameter to 0.9. We set both the "Germline and Frequent Residues" and "Machine Learning Prediction" parameters to “True” and started the prediction program. After completion of the run, we checked the residue substitution alignment to contain the various mutations suggested by the scheme, and calculated the mutation stability energy of the mutations.

Results

First of all, we predicted sequence properties of anti-CEA scFv protein in Discovery Studio. The tool is perfect to predict the antibody domain, the features can be filtered to retain only those that are within the specified antibody domain or complementarity determining regions (CDRs). The characteristic prediction results of anti-CEA scFv protein sequences are shown in Figure 6.

Notes: The feature and antibody domains and CDRs showing in the graphic. The color of light chain CDR is magenta, the color of heavy chain CDR is red.

Predicting humanized antibody structure requires residues in the framework region of the mutation-variable domain, which can be obtained from the antibody sequence annotation information in our previous step, making these residues more similar to the types found in human antibodies (Figure 7). The interactive residue analysis report for each variable domain contains a legend that defines the colors and styles that represent the residue features in the table. There is a detail legend in the interactive report, which we display by hovering the cursor over each residue cell. The interactive residue analysis report for each variable field contains a legend that defines the color and style to represent the residue characteristics in the table (Table 3, Table 4).

Notes: The first sequence is a query annotated using the normal antibody annotation style, and the second sequence is also a query sequence, but cursor residues are indicated by blue caret, hotspots are orange, and the combined Kabat and IMGT CDR regions are magenta or red, depending on the domain. The next three sequences show various mutations in hot spot residues not excluded from substitutions.

The residue substitution ratio contains a variety of mutations suggested by the scheme, of which the following three types are commonly generated, Best_Single_Mutations, Frequent_Residue_Substitutions and Germline_Substitutions (Figure 8). Then we analyzed the developability index (DI) for the three humanized mutation types, where we found that Best_Single_Mutations model have better development performance.

Notes: The protocol contains recommendations for the three types of humanized mutations. (A) Best_Single_Mutations sequence: The germ line sequence, the type of residue found in the ML model, or the frequency > 5%, produces the most stable structure; (B) Frequent_Residue_Substitutions sequence: Where appropriate, a residue of the type most commonly found in that location in human antibodies is replaced; (C) Germline_Substitutions sequence: Germline_substitutions sequence: Mutate residues in appropriate places to find a type in the best-matched human germline sequence.

② Virtual amino acid mutations increase the binding affinity of antibodies

Background

Site-specific amino acid mutation of protein is often used in antibody design. Virtual amino acid mutation based on interaction can be performed on antibody-antigen complex through Discovery Studio software to improve the binding affinity of antibody.

Methodology

Firstly, we uploaded the result structure file of the molecular docking. Subsequently, the protein complex was subjected to the CHARMm force field, and amino acid residues within a 3 Å radius around the ligand were selected. The Single Mutations mode was chosen, and all the previously selected amino acids were mutated to ALA. We used the MODELER program to mutate to specified types and optimize the mutated residues, as well as any surrounding residues within the selected radius. We ranked the results of mutational energy to identify the key amino acid residues that affect antibody affinity. These key amino acids were then mutated to the remaining 19 standard amino acids. These mutations were ranked in order of mutation energy, from lowest to highest.

Additionally, we calculated the mutation energy to evaluate the effect of mutations on protein stability. The difference in free energy of folding between the mutant structures and the wild-type proteins (ΔΔGmut) represented the energy effect of each mutation on protein stability.

ΔΔG_mut = ΔΔG_{folding(mutant)} - ΔΔG_{folding(wild type)}
ΔΔG_folding = ΔG_fld - ΔG_unf

[The ΔG_fld and ΔG_unf terms represent the free energies of folded and denatured states]

Results

By sorting the results of single mutations in mutation energy, only 5 mutations with the highest and lowest mutation energy were displayed. Mutations in the last five index in the table (LEU196, ARG177, ILE193, TYR190, ASN194) can lead to a decrease in affinity, so these five amino acids are key amino acids for the interaction between antibodies and antigens(Figure 9, Table 6). We marked the positions of these key amino acids in the protein structure (Figure 10). Amino acids in the antibody that resulted in decreased antibody affinity after mutation to ALA are highlighted in red.

Notes: Amino acid residues within 3 Å around the CEACAM5 were selected. Stabilizing mutation energy is less than -0.5 kcal/mol; neutral mutation energy is between -0.5 to 0.5 kcal/mol; destabilizing mutation energy is greater than 0.5 kcal/mol.

Notes: The table in the Summary section of the report lists up to ten mutations with the five lowest energy mutations and five highest energy mutations. The ranking of the mutations was sorted from lowest to highest energy mutations.

Notes: Molecule labeled in yellow represent anti-CEA scFv; Molecule labeled in blue represent CEACAM5. Amino acids in the antibody that resulted in decreased antibody affinity after mutation to ALA are highlighted in red, whereas amino acids with increased antibody affinity after mutation are highlighted in magenta.

In order to obtain more detailed information on how mutations enhance antibody affinity, we performed saturation mutagenesis on the five amino acids mentioned above. From the results (Table 7, Figure 11), we found that the amino acid site ASN195 is crucial for the affinity of anti-CEA scFv. When ASN195 becomes TRP, GLN, PHE, GLU and HIS, the affinity of anti-CEA scFv can be significantly improved.

By calculation of mutations on protein stability, we found that, except for ASN195 mutated to GLU and GLN, several other mutations with increased affinity did not affect antibody stability (Table 8).

Notes: The table in the Summary section of the report lists up to ten mutations with the five lowest energy mutations and five highest energy mutations. The ranking of the mutations are sorted from lowest to highest energy mutations

Notes: The table in the Summary section of the report lists up to ten mutations with the five lowest energy mutations and five highest energy mutations. The ranking of the mutations are sorted from lowest to highest energy mutations.

Introduction

Initially, when engineered bacteria expressing scFv was introduced to tumor cells, the bacteria exhibited random distribution alongside the cells. Over time, however, the movement of engineered bacteria expressing scFv became irregular. The experimental findings indicated that the position of the tumor cells remained relatively stable. Additionally, after discussing with the team members of the wet experiment, it was concluded that considering the death and division of bacteria and cells in the model is unnecessary in short time. There was minimal change observed in the number of engineered bacteria expressing scFv and tumor cells within a short duration compared to the baseline.

We have obtained the following data by consulting information and calculation: 24-well plate with one well bottom area of 2cm², and 50,000 tumor cells were distributed more uniformly in the 24-well plate, and on average, a single tumor cell occupies an area of 4 × 10⁸ μm². Engineered bacteria expressing scFv exhibits irregular movement and has a certain probability of entering cells. We learned from literature review that the radius of tumor cells and normal cells is approximately 8 μm⁸.

To determine the true distance between engineered bacteria expressing scFv and tumor cells, we randomly generated the horizontal and vertical coordinates of engineered bacteria expressing scFv using a modeling approach. cells number and infected time are deduced from distance. We assumed that engineered bacteria expressing scFv moves irregularly and forms an area over time. Considering the differential invasion rate of engineered bacteria expressing scFv towards normal cells and tumor cells, we can calculate the required time for each engineered bacterium expressing scFv to enter tumor cells by estimating the spreading rate of these bacteria. Based on the model results, we generated a comparative graph depicting the quantity of engineered bacteria expressing scFv infecting normal cells and tumor cells during a certain period of infection time.

Model Establishment

Determination of parameters

Based on the calculations performed by the wet lab experimental members, the initial number of normal and tumor cells seeded in each well of the 24-well plate was close to 50 000. Referring to the modeling of HZAU-China 2018, we learned that the affinity ratio of attenuated Salmonella to infect tumor cells and normal cells was about 1:10, so we used this ratio as the infection affinity of engineered bacteria without scFv expression.

To model the killing effect of the engineered bacteria on tumor cells during co-culture, the number of tumor cells and normal cells dying after a period of infection were predicted. Among them, we assumed that 20 attenuated Salmonella typhimurium needed to enter the tumor cells before the cells would rupture and die, based on our observations and estimation in wet experiments. Moreover, if a cell undergoes breakdown and dies, the attenuated Salmonella  present inside the cell will also perish. These assumptions derived from the experimental group suggest that when a certain quantity of tumor cells is invaded by attenuated Salmonella, cellular rupture occurs and we can estimate the number of dead cells during the co-culture period.

In wet lab, we performed several experiments with engineered bacteria expressing scFv and infecting tumor cells (see Experiments for the further). We observed the experimental results by fluorescence microscopy and attempted to calculate and count the number of cells and bacteria using ImageJ software. We calculated that the affinity of the scFv-expressing strain to infect tumor cells was about 5.5 times higher than that of the original strain without scFv expression. This data collected by the experimental group will be used in our modeling.

Results and Analysis

The figures (Figure 12.A) presented below show the comparison of our engineered Salmonella infestation in normal cells and tumor cells. From the results of our modeling, over an extended period,  attenuated Salmonella with scFv demonstrated more than two dozen times greater infestation in BGC-823 compared to HEK293T (Figure 12.A). Within a five-hour timeframe, over 80% of BGC-823 cells were infected by attenuated Salmonella, which aligns with the data obtained from the experimental group. From our modeling results, it was observed that the number of dead cells didn't change much at the beginning for a while（Figure 12.B）. At 50 minutes, there was a surge in the number of dead BGC-823 cells and not much change in the number of HEK293T cells. However, at 150 minutes, the number of dead HEK293T cells started to increase.

Similarly, we constructed a difference map of attenuated Salmonella  invading tumor cells and normal cells over time from the affinity enhancement data validated in experimental groups. Furthermore, the anti-CEA scFv expression was modeled to evaluate the anti-CEA scFv cytotoxicity against tumor cells. The results showed that attenuated Salmonella with scFv infected tumor cells more easily, resulting in a higher rate of tumor cell death (Figure 12.C & D).

Changes of tumor cells infected by attenuated Salmonella typhimurium over time were displayed in the following figures:

1 KETTLEBOROUGH C A, SALDANHA J, HEATH V J, et al. Humanization of a mouse monoclonal antibody by CDR–grafting: the importance of framework residues on loop conformation [J]. Protein engineering, design and selection, 1991, 4(7): 773-83.
2 JUMPER J, EVANS R, PRITZEL A, et al. Highly accurate protein structure prediction with AlphaFold [J]. Nature, 2021, 596(7873): 583-9.
3 MIRDITA M, SCHüTZE K, MORIWAKI Y, et al. ColabFold: making protein folding accessible to all [J]. Nature methods, 2022, 19(6): 679-82.
4 TAHERI M, SARAGOVI U, FUKS A, et al. Self recognition in the Ig superfamily: identification of precise subdomains in carcinoembryonic antigen required for intercellular adhesion [J]. Journal of Biological Chemistry, 2000, 275(35): 26935-43.
5 GRETCH D, SUTER M, STINSKI M. The use of biotinylated monoclonal antibodies and streptavidin affinity chromatography to isolate herpesvirus hydrophobic proteins or glycoproteins [J]. Analytical biochemistry, 1987, 163(1): 270-7.
6 THIAGARAJAN G, SEMPLE A, JAMES J K, et al. A comparison of biophysical characterization techniques in predicting monoclonal antibody stability; proceedings of the MAbs, F, 2016 [C]. Taylor & Francis.
7 SAFDARI Y, FARAJNIA S, ASGHARZADEH M, et al. Antibody humanization methods–a review and update [J]. Biotechnology and Genetic Engineering Reviews, 2013, 29(2): 175-86.
8 Takeishi, N., Imai, Y., Yamaguchi, T. & Ishikawa, T. Flow of a circulating tumor cell and red blood cells in microvessels. Physical Review E 92, 063011 (2015).