An artificial intelligence-based bispecific antibody design method

Abstract

This invention discloses an artificial intelligence-based method and system for designing bispecific antibodies, belonging to the fields of bioinformatics and antibody engineering technology. This invention uses the interface region between the parental monoclonal antibody and the monoantigen as the basis for determining key amino acid sites, and introduces virtual mutations to construct a candidate bispecific antibody sequence library. The candidate sequences are then input into an artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody and analyze antibody-antibody interactions. Through multi-round structure prediction and quantitative scoring, this invention achieves a systematic evaluation of the stability of the bispecific antibody and its binding ability to the two antigens, simulating actual bispecific antibody action scenarios and improving the reliability of the design results. This invention quantifies complex intermolecular interactions into a computable scoring system, enabling high-throughput virtual screening of bispecific antibodies, significantly shortening the design cycle, reducing R&D costs, and effectively improving the success rate of bispecific antibody design.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and antibody engineering technology, specifically to a method for predicting and designing bispecific antibody structures based on an artificial intelligence model. Background Technology

[0002] Bispecific antibodies (BsAbs) are a class of engineered antibodies capable of simultaneously recognizing two different antigens or epitopes, showing broad application prospects in areas such as tumor immunotherapy and autoimmune diseases. Compared to natural monoclonal antibodies, the design and development of bispecific antibodies face more complex structural requirements: not only must the two different antigen-binding arms be correctly paired and light chain mismatches be avoided, but the bispecific antibody must also ensure good spatial conformational stability and binding activity when binding to two antigens simultaneously. Currently, the screening and optimization of bispecific antibodies mainly rely on experimental methods, such as hybridoma screening and phage display library screening. However, these traditional methods suffer from problems such as long cycles, low throughput, and high costs, making it difficult to meet the rapidly evolving needs of antibody drug development.

[0003] In recent years, the development of AI-powered structure prediction tools, such as AlphaFold, has provided a new technical approach for antibody structure modeling. Existing methods typically assess the folding stability of bispecific antibodies based on the structure prediction of monoclonal antibody-antigen complexes. However, the actual application scenarios of bispecific antibodies involve the simultaneous binding of two antigens, and their spatial conformation depends not only on the antibody's own folding but also on the steric hindrance and interaction forces during the simultaneous binding of two antigens. Current technologies lack specific design methods for the "bispecific antibody-dual antigen" complex scenario, making it difficult to accurately simulate the real state of bispecific antibodies binding to two antigens simultaneously, resulting in low design success rates and a heavy burden on subsequent experimental screening. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based bispecific antibody design method and system. By introducing key amino acid mutations, multi-scenario structure prediction, and interface analysis, it solves the technical problems of low success rate and difficulty in simulating real binding scenarios in existing bispecific antibody designs.

[0005] The objective of this invention can be achieved through the following technical solutions: An artificial intelligence-based bispecific antibody design system includes an input module, a mutation module, a first prediction module, a scoring module, a second prediction module, and an analysis module. The input module is used to obtain the sequence of the bispecific antibody to be designed and the sequence information of its corresponding first antigen and second antigen; The determination module is used to determine rules based on preset key amino acid sites, which affect subsequent scoring details; the key amino acid sites are determined based on the interface region where the parental monoclonal antibody and the monoantigen bind. The first prediction module is used to input the bispecific antibody sequences before and after mutation into an artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself and analyze the interaction force data between antibodies within the bispecific antibody; the interaction force data includes hydrogen bonds, charge interactions, hydrophobic interactions, and contact area. The scoring module is used to construct a bispecific antibody self-stability scoring system based on the interaction force data obtained from the first prediction module, and to screen and obtain candidate bispecific antibodies; The second prediction module is used to input the screened candidate bispecific antibodies and the first antigen and the second antigen into the artificial intelligence structure prediction model to predict the structure of the bispecific antibody-first antigen complex and the structure of the bispecific antibody-second antigen complex. The analysis module is used to perform docking interface analysis on the predicted complex structure, including hydrogen bonds, charge interactions, hydrophobic interactions, and binding area at the antigen-antibody binding interface, and outputs the final screening results.

[0006] Preferably, the rules for determining the key amino acid sites include: obtaining the complex structure of the parental monoclonal antibody and the corresponding antigen, and marking amino acid residues whose distance between the antibody side and the antigen side is less than a preset threshold as key amino acid sites.

[0007] Preferably, the first prediction module is further configured to predict the structure of a biantibody-biantigen complex in which the biantibody binds to both the first antigen and the second antigen simultaneously.

[0008] Preferably, the scoring system constructed by the scoring module has the following scoring principle: the smaller the antibody-antibody interaction force within the bispecific antibody, the higher the stability of the bispecific antibody pairing and the higher the score.

[0009] Preferably, when the analysis module performs interface analysis, it also sorts the candidate bispecific antibodies by combining the binding free energy estimation.

[0010] Preferably, the artificial intelligence structure prediction model includes AlphaFold3 or its equivalent model.

[0011] To address this issue, the present invention also discloses an artificial intelligence-based method for designing bispecific antibodies, comprising the following steps: S1: Obtain the sequence of the bispecific antibody to be designed and the sequence information of its corresponding first antigen and second antigen; S2: Based on the interface region where the parental monoclonal antibody binds to the monoantigen, key amino acid sites are determined, which affect the scoring details; S3: Input the bispecific antibody sequences before and after mutation into the artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself and analyze the antibody-antibody interaction forces within the bispecific antibody; S4: Based on the interaction force data obtained in step S3, construct a bispecific antibody self-stability scoring system and screen out candidate bispecific antibodies; S5: Input the screened candidate bispecific antibodies, the first antigen, and the second antigen into the artificial intelligence structure prediction model to predict the structure of the bispecific antibody-first antigen complex and the structure of the bispecific antibody-second antigen complex. S6: Perform docking interface analysis on the predicted complex structure, including hydrogen bonds, charge interactions, hydrophobic interactions, and binding area at the antigen-antibody binding interface, and output the final screening results.

[0012] Preferably, step S3 further includes: predicting the structure of the biantibody-biantigen complex that binds to both the first antigen and the second antigen simultaneously.

[0013] Preferably, in step S4, the scoring principle of the scoring system is: the smaller the antibody-antibody interaction force within the bispecific antibody, the higher the score.

[0014] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention uses the interface region where the parental monoclonal antibody binds to the single antigen as the basis for determining the key amino acid sites, introduces virtual mutation to construct a candidate library, and combines multiple rounds of structural prediction to achieve a systematic evaluation of the stability of the bispecific antibody itself and its ability to bind to the bispecific antigen. By predicting and scoring the structure of the bispecific antibody itself through the first prediction module, candidate sequences with high pairing stability are screened out, reducing the blindness of the subsequent experimental screening. By using the second prediction module to predict and analyze the interface of bispecific antibody-monoantigen complexes and bispecific antibody-bispecific antigen complexes, the actual action scenarios of bispecific antibodies were simulated, improving the reliability of the design results. This invention quantifies complex intermolecular interactions into a computable scoring system, enabling high-throughput virtual screening of bispecific antibodies, significantly shortening the design cycle, reducing R&D costs, and effectively improving the design success rate of bispecific antibodies. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a block diagram of a bispecific antibody design system based on artificial intelligence according to the present invention.

[0017] Figure 2 This is a flowchart of an artificial intelligence-based bispecific antibody design method according to the present invention. Detailed Implementation

[0018] Example 1: A method for designing bispecific antibodies based on artificial intelligence This embodiment provides an artificial intelligence-based bispecific antibody design method, including the following steps.

[0019] Step 1: Input sequence Obtain the bispecific antibody sequence to be designed. The bispecific antibody sequence includes a first binding domain against a first antigen and a second binding domain against a second antigen. The first antigen and the second antigen are different antigens or different epitopes of the same antigen.

[0020] Specifically, in this embodiment, the first antigen is tumor-associated antigen A, and the second antigen is immune checkpoint molecule B. The bispecific antibody sequence is constructed based on two parental monoclonal antibodies (parental antibody 1 and parental antibody 2), with parental antibody 1 specifically binding to antigen A and parental antibody 2 specifically binding to antigen B.

[0021] Simultaneously, the amino acid sequences or structural information of the first and second antigens are obtained for subsequent complex structure prediction.

[0022] Step 2: Identify key amino acid sites Determination of key amino acid sites Two parental monoclonal antibodies (hereinafter referred to as "parental antibodies") corresponding to the target bispecific antibody are obtained, and their structures are predicted or determined based on known crystal structure data to identify the interface regions where each parental antibody binds to the antigen.

[0023] The interface region is defined as: in the antibody-antigen complex structure, the amino acid residues whose distance between the antibody side and the antigen side is less than a preset threshold. In this embodiment, the preset threshold is set to 4.5. The amino acid residues on the antibody side that satisfy this distance condition are marked as key amino acid sites.

[0024] If the crystal structure of the parent antibody-antigen complex is known, it can be obtained directly from the protein structure database (PDB); if the crystal structure is not known, an artificial intelligence structure prediction model (such as AlphaFold3, RoseTTAFold, etc.) is used to predict the complex structure, and then interface analysis is performed based on the predicted structure.

[0025] 2.2 Classification of key amino acid sites The identified key amino acid sites are classified, including but not limited to the following types: Key sites located in the CDR region of the heavy chain: core residues involved in direct antigen binding; Key sites located in the CDR region of the light chain: residues involved in antigen binding or stabilizing the heavy chain conformation; Key sites located in the framework region but involved in antigen binding: residues that, although not in the CDR region, interact with the antigen; Key sites located in the antibody constant region that may affect bispecific antibody pairing: such as the CH1-CL interface, heavy chain hinge region, and other residues that may affect the correct pairing of bispecific antibodies.

[0026] Step 3: First round of structural prediction – prediction of the structure of the dual-antibody compound The candidate bispecific antibody sequences (including the original sequence and the mutant sequence) constructed in step 2 are input into the artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself.

[0027] In this embodiment, the artificial intelligence structure prediction model used is AlphaFold3, and its default parameters are used for structure prediction. For each candidate sequence, five structure models are predicted, and the model with the highest confidence is selected for subsequent analysis.

[0028] The predicted 3D structure of the bispecific antibody was analyzed, and the interaction force data between antibodies within the bispecific antibody was extracted, including: Hydrogen bonds: Count the number and bond length of hydrogen bonds formed between the two chains of the bispecific antibody; Charge interaction: Calculate the electrostatic potential distribution and charged residue pairing in the interface region; Hydrophobic interaction: Calculate the area of ​​hydrophobic interaction at the interface; Antibody-antibody contact area: The change in solvent-accessible surface area (ΔASA) between the two chains is calculated as a quantitative indicator of the contact area.

[0029] In this embodiment, the get_area function of PyMOL and the Contact Map analysis tool are used to calculate the above-mentioned interaction data.

[0030] Step 4: Construct a bispecific antibody self-stability scoring system Based on the interaction data obtained in step 3, a scoring system is established to evaluate the stability of antibody self-pairing. The scoring principle is: the smaller the antibody-antibody interaction force within the bispecific antibody (i.e., the lower the risk of mismatch), the higher the score.

[0031] In this embodiment, the scoring system is constructed using a weighted summation method, with the specific formula as follows: Score = w1·S_hbond + w2·S_charge + w3·S_hydro + w4·S_contact in: S_hbond is the hydrogen bond score, which is normalized based on the number of hydrogen bonds. The more hydrogen bonds there are, the stronger the interaction, and the lower the score. S_charge is the charge interaction score, which is scored based on the complementarity of charged residues at the interface. The stronger the electrostatic complementarity, the stronger the interaction, and the lower the score. S_hydro represents the hydrophobic interaction score, which is based on the area of ​​the hydrophobic interaction. The larger the hydrophobic area, the stronger the interaction, and the lower the score. S_contact is the contact area score, which is based on the ΔASA value. The larger the contact area, the stronger the interaction, and the lower the score. w1, w2, w3, and w4 are weighting coefficients that satisfy w1 + w2 + w3 + w4 = 1. In this embodiment, based on the contribution of each interaction to the stability of the dual resistance, w1 = 0.3, w2 = 0.2, w3 = 0.3, and w4 = 0.2 are set.

[0032] Each sub-score is normalized to the [0,1] interval. The higher the final score, the better the stability of the bispecific antibody pairing.

[0033] Based on the scoring results, the candidate sequences are sorted, and the top N candidate bispecific antibodies with the highest scores are selected to proceed to the next round of analysis. In this embodiment, N is set to 10.

[0034] Step 5: Obtain the antigen structure Obtain the 3D structures of the first and second antigens. If the antigen structure is known, obtain it directly from the protein structure database (PDB); if no structure is known, use an artificial intelligence structure prediction model (such as AlphaFold3) to predict the structure.

[0035] In this embodiment, the PDB ID of the first antigen A is 8Z8L, and the PDB ID of the second antigen B is 2OCW. The structure file is downloaded directly from the PDB.

[0036] Step 6: Second round of structure prediction – Bispecific antibody-monoantigen complex structure prediction The candidate bispecific antibodies obtained in step 4 are input into the artificial intelligence structure prediction model along with the first antigen and the second antigen, respectively, to predict the structure of the bispecific antibody-first antigen complex and the structure of the bispecific antibody-second antigen complex.

[0037] In this embodiment, the complex prediction mode of AlphaFold3 was used to predict the complex structure of each candidate bispecific antibody with antigen A and antigen B. Five structural models were obtained for each complex prediction, and the model with the highest confidence was selected for subsequent analysis.

[0038] Step 7: Third round of structure prediction (optional) – Bispecific antibody-dual antigen complex structure prediction To further simulate the actual action of bispecific antibodies, candidate bispecific antibodies were simultaneously input into an artificial intelligence structure prediction model along with the first and second antigens to predict the 3D structure of the bispecific antibody-bispecific antigen complex.

[0039] In this embodiment, AlphaFold3 is used to predict the structure of the ternary complex of the bispecific antibody simultaneously binding to antigen A and antigen B. This step can be used to verify whether the bispecific antibody exhibits steric hindrance or conformational conflict.

[0040] Step 8: Interface Analysis The docking interface of the complex structures predicted in steps 6 and 7 was analyzed using molecular visualization software. In this embodiment, PyMOL was selected for the following analysis: Antigen-antibody binding interface analysis Hydrogen bonds: The number and length of hydrogen bonds formed at the antigen-antibody interface are statistically analyzed. The more hydrogen bonds and the shorter the bond length, the stronger the binding. Charge interaction: Analyze the distribution of charged residues in the interface region; the stronger the electrostatic complementarity, the higher the binding strength. Hydrophobic interaction: The hydrophobic interaction area of ​​the interface region is calculated. The larger the hydrophobic area, the stronger the bonding strength. Binding area: Calculates the change in solvent-accessible surface area (ΔASA) before and after antigen-antibody binding. The larger the binding area, the stronger the interaction.

[0041] Combined with free energy estimation The binding free energy of the complex structure can be estimated using the PyMOL plugin or third-party tools (such as FoldX or Rosetta). The more negative ΔG is, the more stable the binding.

[0042] Spatial steric hindrance analysis For the biantibody-biantigen complex structure predicted in step 7, analyze whether there is steric hindrance or conformational conflict when the two antigens bind simultaneously.

[0043] Step 9: Final Stability Verification and Screening The candidate bispecific antibodies are comprehensively evaluated and ranked based on the combined results of the scoring system in step 4 and the interface analysis in step 8.

[0044] Evaluation indicators include: Bis-antibody stability score (from step 4): the higher the score, the better; Antigen A binding strength: comprehensively evaluated based on indicators such as the number of hydrogen bonds, binding area, and binding free energy analyzed in step 8; Antigen B binding strength: Same as above; Dual antigen binding capacity: For the complex predicted in step 7, assess whether there is steric hindrance.

[0045] Screening principles: Prioritize retaining the top 30% of candidate sequences based on the intrinsic stability score of bispecific antibodies; Based on this, sequences that have good binding activity with both antigen A and antigen B are screened. Sequences with significant steric hindrance in bispecific antibody-bispecific antigen complexes were excluded.

[0046] In this embodiment, three final candidate sequences were selected from 10 candidate bispecific antibodies as preferred candidates for subsequent experimental verification.

[0047] Example 2: An Artificial Intelligence-Based Bispecific Antibody Design System This embodiment provides an artificial intelligence-based bispecific antibody design system for implementing the method described in Embodiment 1. The system includes the following modules: Input module: Used to obtain the sequence or structural information of the bispecific antibody to be designed and its corresponding first and second antigens.

[0048] Mutation module: Used to introduce virtual mutations into antibody sequences according to preset key amino acid site determination rules, constructing a candidate bispecific antibody sequence library. The key amino acid sites are determined based on the interface region between the parental monoclonal antibody and the single antigen.

[0049] The first prediction module is used to input the bispecific antibody sequences before and after mutation into an artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself and analyze the interaction force data between antibodies within the bispecific antibody; the interaction force data includes hydrogen bonds, charge interactions, hydrophobic interactions, and contact area.

[0050] Scoring module: Used to construct a bispecific antibody self-stability scoring system based on the interaction force data obtained from the first prediction module, and to screen and obtain candidate bispecific antibodies.

[0051] The second prediction module is used to input the screened candidate bispecific antibodies and the first and second antigens into the artificial intelligence structure prediction model to predict the structure of the bispecific antibody-first antigen complex and the bispecific antibody-second antigen complex.

[0052] Analysis module: Used to perform docking interface analysis on the predicted complex structure, including hydrogen bonds, charge interactions, hydrophobic interactions and binding area at the antigen-antibody binding interface, and output the final screening results.

[0053] Visualization module: used to display the predicted 3D structures of the bispecific antibodies and their complexes, as well as the interface analysis results.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based bispecific antibody design system, characterized in that, include: The input module is used to obtain the sequence of the bispecific antibody to be designed and the sequence information of its corresponding first antigen and second antigen; The mutation module is used to introduce virtual mutations into the bispecific antibody sequence according to a preset key amino acid site determination rule to construct a candidate bispecific antibody sequence library; the key amino acid sites are determined based on the interface region between the parental monoclonal antibody and the monoantigen. The first prediction module is used to input the candidate bispecific antibody sequence into an artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself and analyze the interaction force data between antibodies within the bispecific antibody. The scoring module is used to construct a bispecific antibody self-stability scoring system based on the interaction force data obtained by the first prediction module, and to screen and obtain candidate bispecific antibodies; The second prediction module is used to input the screened candidate bispecific antibodies and the first antigen and the second antigen into the artificial intelligence structure prediction model to predict the structure of the bispecific antibody-first antigen complex and the structure of the bispecific antibody-second antigen complex. The analysis module is used to perform docking interface analysis on the predicted complex structure and output the final screening results.

2. The artificial intelligence-based bispecific antibody design system according to claim 1, characterized in that, The rules for determining the key amino acid sites include: Obtain the complex structure of the parental monoclonal antibody and the corresponding antigen, and mark the amino acid residues on the antibody side and the antigen side with a distance of less than a preset threshold as key amino acid sites.

3. The artificial intelligence-based bispecific antibody design system according to claim 2, characterized in that, The preset threshold is 4.

5. .

4. The artificial intelligence-based bispecific antibody design system according to claim 1, characterized in that, The interaction force data analyzed by the first prediction module includes: Number and length of hydrogen bonds, charge interaction, hydrophobic interaction area, and contact area between antibodies.

5. The artificial intelligence-based bispecific antibody design system according to claim 1, characterized in that, The scoring module constructs a bispecific antibody self-stability scoring system, and its scoring principle is as follows: The weaker the antibody-antibody interaction within the bispecific antibody, the higher the stability of the bispecific antibody pairing and the higher the score.

6. The artificial intelligence-based bispecific antibody design system according to claim 5, characterized in that, The calculation formula for the scoring system is: Score = w1·S_hbond + w2·S_charge + w3·S_hydro + w4·S_contact Where S_hbond is the hydrogen bond score, S_charge is the charge interaction score, S_hydro is the hydrophobic interaction score, and S_contact is the contact area score; w1, w2, w3, and w4 are preset weighting coefficients, and w1 + w2 + w3 + w4 = 1.

7. The artificial intelligence-based bispecific antibody design system according to claim 1, characterized in that, The second prediction module is also used for: The selected candidate bispecific antibodies are simultaneously input into an artificial intelligence structure prediction model along with the first and second antigens to predict the structure of the bispecific antibody-bispecific antigen complex.

8. The artificial intelligence-based bispecific antibody design system according to claim 1, characterized in that, The interface analysis performed by the analysis module includes: Estimation of hydrogen bonds, charge interactions, hydrophobic interactions, binding area, and binding free energy at the antigen-antibody binding interface.

9. A method for designing bispecific antibodies based on artificial intelligence, characterized in that, Includes the following steps: S1: Obtain the sequence of the bispecific antibody to be designed and the sequence information of its corresponding first antigen and second antigen; S2: Based on the interface region where the parental monoclonal antibody binds to the single antigen, key amino acid sites are determined, and virtual mutations are introduced to construct a candidate bispecific antibody sequence library; S3: Input the candidate bispecific antibody sequence into the artificial intelligence structure prediction model to predict the 3D structure of the bispecific antibody itself and analyze the antibody-antibody interaction forces within the bispecific antibody; S4: Based on the interaction force data obtained in step S3, construct a bispecific antibody self-stability scoring system and screen out candidate bispecific antibodies; S5: Input the screened candidate bispecific antibodies, the first antigen, and the second antigen into the artificial intelligence structure prediction model to predict the structure of the bispecific antibody-first antigen complex and the structure of the bispecific antibody-second antigen complex. S6: Perform docking interface analysis on the predicted complex structure and output the final screening results.

10. The method for designing bispecific antibodies based on artificial intelligence according to claim 9, characterized in that, The interaction force data analyzed in step S3 includes the number and length of hydrogen bonds, charge interactions, hydrophobic interaction area, and antibody-antibody contact area; the scoring system in step S4 is constructed according to the system described in claim 5 or 6; the docking interface analysis in step S6 includes the hydrogen bonds, charge interactions, hydrophobic interactions, binding area, and binding free energy estimation of the antigen-antibody binding interface.

Images