Overview

The project we have developed is a comprehensive automated monoclonal antibody drug design system. Currently, we have 55 species covered by our heavy chain design and 9 commonly used species covered by our light chain design. Inspired by the AI tools related to humanization of murine antibodies, we have innovated and extended them to more species. Our project starts with the other sources of antibody sequences against target species antigen (such as mouse-derived antibodies), automatically generates antibody sequences that may have better efficacy and lower immunogenicity, and performs comprehensive scoring.

Users can enter an antibody data based on existing experimental results, or select the antibody data with a specific number for a specific species from the antibody database we provide. At the same time, the user is required to provide the name of the target species for which the antibody is desired. Our model will provide the user with multiple potential antibody data. These antibody data preserve the specific binding ability of the original sequence to a particular antigen, while minimizing the likelihood of eliciting an immune response from the target species. We will also provide multiple scoring metrics for these antibody sequences as well as the original sequence.

Core Contributions

1. Database construction: Through independent collection and screening of antibody data, a database covering multiple species was constructed.

2. Model application: The Deep One Class model in image anomaly detection was used to score antibody data by species, solving the problem of sample data imbalance.

3. Structural scoring interface: A structural scoring interface was introduced to call the IgFold model and use TM-Score and RMSD to score the similarity of protein PDB files.

In addition to the immunogenicity-related species specification scores at the core of the project, our generation and evaluation can be divided into two main parts:

In the sequence generation part, we synthesized the characteristics of the input antibody data as well as the statistical patterns of the antibody data of the target species, and adopted the strategy of conservative mutation to determine the amino acids with higher probability at each site, and then generated a library of potential antibody data sequences by means of permutations and combinations.

In the structure scoring part, we folded the original sequences and the generated potential sequences through the IGFOLD model to determine whether the generated sequences can be folded into a complete structure, and visualized the differences between the structures of the original sequences and those of the generated sequences.

Result

Here is an example of what our software would get.

First, in the sequence generation phase, we obtained a large amount of potential sequence space.

We further put these sequences into the model for sequence scoring.

Further, we will perform structural scoring. First of all, from the pdb file, the sequence was successfully folded into the typical structure of the antibody, and it was relatively close to the original CDR region structure. In order to better measure the similarity, we used the structure score, which uses numbers to better characterize the score of the structure.

Validation

1. Sequence validation:

Because existing tools can only provide humanization scores, we adapted our model for multi-species scoring to score human sequences and validated it by comparison with existing tools. In order to verify that our generated sequences are plausible and to validate the reliability of our species specification scores, we analyzed the generated sequences in terms of available humanization tools and antibody structures, respectively.

We attempted to validate the results of our project in conjunction with existing antibody humanization tools.

The first core of the validation is whether the sequences generated based on statistical regularity and conserved mutations satisfy the mutability and conservatism of the existing humanization tools in the humanization process sites.

>

We chose a murine antibody sequence with a complete heavy chain sequence of:

            
              EVMLVESGGGLVMPGGSLKLSCAASGFTFSNYAMSWVRQIPEKRLEWVATISIGGHFTYFPDSVKGRFTISRDNAKNTLYLRMSSLRSEDTAMYYCVRHEGYGRPYFDYWGQGTTLTVSS
            
          

We used the current mainstream antibody humanization tool BioPhi for humanization and obtained the following results.

The possible amino acids for each site of the FR sequence we generated are as follows, and we set a maximum of two possible amino acids for each site in order to limit the size of the generated sequence space as well as to ensure as many conserved mutations as possible.

We compare and analyze the original sequence, the humanization results given by BioPhi, and the humanization results from our project.

In the above figure, the first row shows the FR fragment of the original sequence, the second row shows the humanization results given by BioPhi, and the third and fourth rows show the humanization results given by us. The reddish colored sites in the second row of the sequence are the sites that BioPhi considered necessary to mutate in the original antibody sequence, and the reddish colored sites in the third and fourth rows of the sequence are the sites that BioPhi and we considered necessary to mutate in our results.

The comparison shows that our model offers more possibilities compared to BioPhi, and the permutations can be combined in a way that gives both data more in line with the human antibody pattern and results that are as conservative as possible in terms of mutations, and this conservatism will be reflected in the sequence scoring of our model.

Comparing the CDR region, BioPhi tends to keep the CDR region of the mouse-derived antibody, and the red position of the humanized results it gives maintains the invariance of the CDR part of the sequence as much as possible, which is the same idea as that of our antibody sequence generation to ensure that the antibody clock still maintains its proper specific recognition and binding effect after speciation.

In addition, analyzing from the scoring point of view, BioPhi gave the result of 0.69 OASis identity for the original sequence and 0.81 OASis identity score for the humanized one.

Our model, on the other hand, although scoring the overall sequence higher, still improves the score of the original sequence from 0.81 to 0.99.

Because the details of different model scoring are slightly different, but from the point of view of the score enhancement, our model gives a more reasonable result.

In the following, we validated the light chain results, and we selected a special mouse-derived antibody in order to form a distinction with the multiple mutations in the heavy chain validation, and to validate the accuracy of our model from multiple perspectives.

This antibody has already shown extremely high human origin in the BioPhi test results, and its probability of human origin is basically 100%.

After humanization, BioPhi did not make any mutations to the original sequence.

Our model gives just two possible outcomes for this sequence and one of them is the same as the original sequence, the first two data in the figure below are the outcomes we give with ratings of 0.96 and 0.95, and the last sequence is the original sequence, which also has a rating of 0.95.

Our model only gives a mutation at one site, i.e., the position of Kabat L100 is mutated from G to Q. Three of the five sequences with the highest multiple human homology given by BioPhi maintain the same results as ours at and around this site.

The second core of the validation is to determine the reasonableness of the speciation scores of the antibody sequences, e.g., to determine the reasonableness between the ordering of the scores of each sequence. Considering that the existing mainstream websites only give one humanization result, while our model gives users more possibilities, we would like to compare the scoring results of our model with the scoring results of BioPhi for the similarity of the ordering.

Considering that our model can output up to 999 results, we randomly selected several sequences with different scores for comparison and analysis.

The sequences we chose are given in the figure below, the first four are the generated results given by us, the yellow marker rows are the humanized results given by BioPhi, the blue marker rows are the original sequences, and the right hand side is the sequence scores given by our model.

The results of the BioPhi website ratings will be given below, one at a time, in the order in which they appear in the chart above.

Comparative analysis shows that the results we give are better than those given by BioPhi in terms of OASis Identity, OASis Percentile and Germline Content scores.

Moreover, the ordering of the scores given by us is consistent with the ordering of the scores given by BioPhi.

This confirms the reasonableness of our sequence generation results and sequence scoring results.

2. Structural validation

In the structure scoring section, we first fold the original sequence and the generated potential sequence through the IgFold model, determine whether the generated sequence can be folded into a complete structure, and visualize the difference between the structure of the original sequence and the structure of the generated sequence.

After obtaining the PDB files, we compare the similarity of the PDBs of the original and generated sequences by two commonly used PDB similarity comparison methods, TM-Score and RMSD, to give a reasonable structural score.

It can be seen that the structures of the CDR areas of the three PDB files are relatively similar.

The degree of structural similarity between the two generated sequences is greater. The following are structural comparisons of the original sequence and the generated sequence.

TM-Score is above 0.99 and RMSD is about 0.2. Because the output sequence still maintains structural similarity with the original sequence in the CDR region, this means that most of the antigen-binding region is maintained, and the antigen-antibody binding ability is still high.

This means that the sequence we output can not only fold into the correct antibody structure, but also has guaranteed antigen-antibody binding ability.