Automatic calculation method and system for binding free energy of biological macromolecules based on structure prediction, antibody sequence clustering and lie-fep graph optimization

Abstract

The application discloses a kind of biological macromolecule binding free energy automatic calculation method and system based on structure prediction, antibody sequence clustering and LIE-FEP graph optimization, build Campaign project container and BinderObject single-molecule bilayer data architecture, real-time writeback of whole-process file, task state, quality control information realizes data traceability;Through double-strategy clustering compression computing space, generate classification FEP edge according to mutation type, only carry out structure prediction to target subset, reduce the power consumption.Relying on HPC, balanced MD, neFEP and LIE energy collection are automatically executed in stages, and free energy data with error weight is obtained;OptimizationGraph is constructed to complete LIE and FEP global joint optimization, correct thermodynamic cycle deviation, and output free energy and confidence interval through Bootstrap resampling.This method unifies two sets of free energy system, improves antibody affinity prediction accuracy and automation level, and can be used for antibody high-throughput screening and affinity maturation research and development.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to an automated method and system for calculating the binding free energy of biomolecules based on structure prediction, antibody sequence clustering, and LIE-FEP diagram optimization. Background Technology

[0002] In the development of biopharmaceuticals such as antibodies, nanobodies, Fab, scFv, and bispecific binding molecules, the binding affinity between candidate molecules and antigens is a core evaluation indicator for affinity maturation, high-throughput virtual screening, and molecular modification decisions. Current affinity assessments fall into two technical routes: wet experimental detection and computer simulation. Wet experiments such as SPR, BLI, ELISA, and cell function binding assays have low throughput, take months, and are costly, making them unsuitable for initial screening of tens of thousands of mutant candidate molecules. Computational simulation methods such as molecular dynamics, free energy perturbation (FEP), and linear interaction energy (LIE) can achieve low-cost, high-throughput virtual assessment and have become the mainstream auxiliary means in early antibody development.

[0003] With the widespread adoption of high-precision protein structure prediction tools such as AlphaFold, ESMfold, and RosettaFold, the batch acquisition of three-dimensional structures of antibody-antigen complexes has become a reality, theoretically supporting parallel simulations of large-scale candidate molecules. However, the existing computational workflow suffers from multi-dimensional systemic defects: Data fragmentation and the lack of a unified data carrier throughout the entire process make traceability and matching difficult. In existing molecular simulation workflows, antibody light and heavy chain sequences, IMGT standardized numbers, CDR mutation sites, PDB 3D structures, molecular dynamics force field parameters, FEP relative free energy, and LIE absolute free energy results are scattered and stored in independent scripts, folders, and spreadsheets, without a centralized, unified data object for each project. In practical use, problems such as candidate molecule ID confusion, sequence-to-3D structure mismatch, duplicate configuration of simulation parameters, and inability to trace calculation results back to the original input are prone to occur. The entire process relies on manual verification and correction, which is costly and prone to errors.

[0004] FEP candidate pairing relies on manual screening, resulting in combinatorial explosion and significant computational waste. FEP alchemy is only applicable to highly similar candidate molecule pairs with few CDR mutations, no amino acid insertions or deletions, and no drastic charge changes. If FEP calculations are performed on all candidates pairwise, an exponential number of combinations will occur, leading to low reliability for most pairing predictions and rendering the computational investment worthless. Existing tools rely solely on engineers manually verifying the number, type, and insertion / deletion characteristics of mutations. Screening criteria vary from person to person and are inconsistent, making automated batch pre-screening impossible and unsuitable for processing hundreds of candidate molecules. Furthermore, existing tools are only designed for small molecule ligands and lack specific clustering screening logic targeting antibody CDR region mutation characteristics.

[0005] The inability to automatically distinguish complex mutation candidate pairs results in poor simulation adaptability. For antibody pairings with CDR charged mutations, amino acid insertions / deletions, and low sequence similarity, direct simulation using conventional FEP can easily lead to problems such as imbalance of system electroneutrality, severe topological mutations causing non-convergence of alchemical conversions, and a surge in free energy errors. Existing automated processes cannot automatically identify such special pairings and match them with dedicated processing solutions such as double-box neutralization and substitution calculations. They can only be manually eliminated or the simulation script can be modified individually, resulting in insufficient automation.

[0006] The LIE absolute free energy and FEP relative free energy systems are thermodynamically disconnected, leading to conflicting candidate rankings. The traditional LIE model outputs the absolute binding free energy of a single molecule, while FEP outputs the difference in relative free energy between two molecules, and the two calculation processes run independently. First, the affinity rankings of the same candidate molecule obtained through LIE and FEP networks are inconsistent, making it impossible to establish a unified priority. Second, the FEP network diagram contains multiple closed loops, and there is no unified mechanism to distribute and correct the thermodynamic errors caused by these loops. Third, isolated candidate molecules without direct FEP connections cannot be included in the global unified ranking, and many mutants can only be judged based on LIE results with limited accuracy, reducing the reliability of the assessment.

[0007] Traditional LIE models have large prediction biases and lack molecular-specific correction mechanisms. The classic standard LIE model only sets a set of globally uniform weight coefficients to synchronously scale the bound and free state energy terms, which cannot distinguish the differential contributions of van der Waals and electrostatic interactions in the bound and free states. At the same time, it does not set independent correction terms for different antibody mutants, and the systematic bias caused by the structural, charge, and conformational differences brought about by CDR mutations cannot be eliminated, resulting in high prediction errors.

[0008] The entire molecular simulation process involves many manual operations and has weak scalability. From structural file parsing, force field allocation, solvent box construction, neutralization and salting, energy minimization, equilibrium simulation, production trajectory sampling, non-equilibrium FEP alchemy conversion to LIE energy extraction and BAR error analysis, the existing process relies on manual script writing and debugging for more than ten steps. The calculation cycle for a single batch of candidates can be as long as several weeks, which cannot meet the high-throughput and rapid iteration screening needs in the early stages of antibody discovery.

[0009] The prediction results only output single-point free energy values, without any quantitative indicators of reliability. Existing free energy calculation tools only output a single bound free energy prediction value, and cannot quantify the confidence interval of the prediction results. Researchers cannot distinguish between high-confidence and low-confidence candidate molecules, wet experimental verification is blind, and the effective positive molecule detection rate is low.

[0010] Existing publicly available technologies can only achieve one function in sequence clustering, small molecule FEP graph construction, basic LIE calculation, and single-step automated molecular simulation. There is no integrated and complete solution for antibody macromolecules, including centralized data management, CDR dual-strategy clustering pre-screening, automatic FEP edge classification and labeling, extended LIE model, LIE-FEP thermodynamic graph joint optimization, and full-process automated simulation. It cannot simultaneously solve multiple industry pain points such as data fragmentation, computing power redundancy, conflict between two sets of free energy results, poor adaptation of complex mutations, cumbersome manual operation, and lack of confidence assessment in prediction. Summary of the Invention

[0011] To address the aforementioned problems, the purpose of this invention is to provide an automated calculation method and system for the binding free energy of biomacromolecules based on structure prediction, antibody sequence clustering, and LIE-FEP diagram optimization. This reduces errors caused by manual processing, improves reusability between different projects, enhances the consistency and success rate of simulation settings, and improves the thermodynamic consistency of the free energy network and the coverage of antibody candidate priority decision-making.

[0012] To address the aforementioned technical problems, the first technical solution adopted by this invention is as follows: An automated calculation method for the binding free energy of biomolecules based on sequence clustering and LIE-FEP graph optimization, characterized in that it is applicable to the affinity prediction of binding molecules such as antibodies, nanobodies, Fab, and scFv to antigens, employing a two-layer data architecture of a project-level Campaign global data container and a single-molecule BinderObject, with all calculation files, task status, and quality control information written back to the Campaign to achieve data traceability, including the following steps: S1. Data initialization and sequence standardization: Create a Campaign object to carry all the data of the project, and generate a unique BinderObject for each input combined with the molecular sequence; standardize the sequence according to the IMGT numbering rules, identify the CDR mutation site, mutation number and mutation type, and store the chain type, number and mutation characteristics into the corresponding BinderObject and classify it into the Campaign. S2. Clustering, FEP Graph Construction and Structure Preprocessing: Extract CDR mutation features from each BinderObject, perform mutation threshold dense clustering and CDR similarity density clustering in parallel, merge to obtain independent candidate clusters, and label isolated molecules as singleton clusters; determine the mutation type only for molecular pairs within the same cluster, distinguish and generate regular FEP edges, charge mutation special FEPs, and insertion / deletion replacement edges, and write back the cluster ID, FEP edge list, edge attributes, and quality control labels to the Campaign; select only the subset of molecules with FEP edges to generate FASTA, call the structure prediction tool to obtain the antigen-antibody complex structure, complete the GROMACS format conversion, and store the structure path and processing status in the Campaign; S3, HPC phased automated simulation and energy harvesting: Based on the structure and simulation parameters stored in the Campaign, GROMACS input files and SLURM scheduling scripts are automatically generated, and high-performance cluster tasks are submitted in three steps: First, a balanced MD simulation job is submitted, and the trajectory, energy file, and verification log are written back to the Campaign upon completion; Second, a non-equilibrium free energy perturbation job for the neFEP is submitted based on the balanced conformation, outputting the raw ΔΔG data and recording the storage path; Third, the simulation trajectory is parsed in batches, and the extended LIE is run to calculate the interaction energy of the binding state and free state corresponding to each BinderObject; LIE energy, FEP relative free energy ΔΔG, standard error SE, job ID, run failure / rerun flag, and quality control label are all written back to the Campaign. S4. Joint Thermodynamic Optimization of the Optimization Graph: An optimization graph is constructed by reading LIE node data and error-weighted FEP edge data from the Campaign. The LIE energy stored in BinderObject is used as the node prior, and ΔG and SE are used as undirected edge attributes. A loss function is constructed based on thermodynamic conservation constraints, and the global parameters and molecule-specific offset terms of the LIE model are jointly optimized and expanded. Thermodynamic loop closures in the network graph are identified and loop closure corrections are performed to eliminate thermodynamic inconsistencies. Through Bootstrap resampling error propagation, the optimized free energy ΔG and confidence interval CI of each candidate molecule are output. The LIE model calculation formula is as follows: ΔG LIE,i =α bound •E VDW，bound,i -α free•E VDW，free,i +β bound •E ELEC，bound,i -β free •E ELEC，free,i +γ+f i Where α bound α free β bound β free γ are globally fixed parameters, f i This is a candidate molecule-specific offset term; S5. Results Output and Complete Project Archiving: Based on the full data output within the Campaign, standardized results are output, including optimized ΔG, candidate ranking, confidence interval, clustering grouping, ΔG / SE / processing method of each FEP edge, and a wet experiment Top-N candidate priority list; sequences, structure files, HPC job records, simulation trajectories, raw energy data, and optimization result data packages are all archived to the Campaign, achieving full-link traceability of molecules, FEP edges, simulation tasks, and final results.

[0013] In the above technical solution, the mutation threshold dense clustering rule in S2 is as follows: if the total number of CDR mutations of two candidate molecules is less than a preset threshold, an undirected connection edge is established, and the connected components are used as independent candidate clusters; the CDR similarity density clustering adopts the DBSCAN algorithm, and a multi-dimensional feature vector is constructed based on the sequence similarity of 6 CDR regions to complete the clustering.

[0014] In the above technical solution, the FEP edge determination rules in S2 specifically include: if a molecular pair has an amino acid insertion / deletion mutation, it is marked as a substitution treatment edge and does not participate in the neFEP calculation; if there is no insertion or deletion but there is a charge mutation, it is marked as a double-box neutralized special FEP edge; if there is no insertion or deletion and no charge mutation, and the total number of mutations is less than the threshold, a regular FEP edge is generated; if the total number of mutations exceeds the threshold, no FEP edge is constructed, and the evaluation is completed solely by LIE and graph propagation.

[0015] In the above technical solution, the three-stage HPC tasks in S3 are executed sequentially: balanced MD simulation, neFEP free energy perturbation, and LIE energy batch acquisition. All tasks are scheduled to be parallelized on a high-performance cluster via SLURM scripts. FEP edges that fail in simulation are automatically marked, and reruns or weight reduction processing in subsequent optimization stages are supported.

[0016] In the above technical solution, the FEP calculation adopts non-equilibrium free energy perturbation, including: extracting one representative conformation frame every 150-250 ps from the equilibrium simulation trajectory, with a total sampling frame count of 40-60 frames; the single conformation alchemical transformation time is 60-100 ps, ​​the lambda value is linearly transformed from 0 to 1, and the van der Waals and Coulomb interactions enable the soft-core potential to eliminate endpoint singularities; the BAR algorithm is used to calculate the pairwise relative binding free energy ΔΔG and standard error.

[0017] In the above technical solution, the S4 loop closure correction operation includes: traversing all thermodynamic loops within the OptimizationGraph, allocating loop deviations with the BAR standard error SE corresponding to each FEP edge as the weight, correcting the ΔΔG value of each edge, and achieving global thermodynamic self-consistency.

[0018] In the above technical solution, the Bootstrap resampling time in S4 is no less than 100 times. By perturbing the FEP side error and repeatedly optimizing the LIE input energy, a 95% confidence interval is obtained to identify high uncertainty mutation candidates.

[0019] In the above technical solutions, the antibody sequence standardization numbering adopts the IMGT, Chothia, Kabat or AHo numbering system, and the site alignment and CDR region identification are completed by the ANACRI or AbNumber tools.

[0020] To achieve the above objectives, the second technical solution of the present invention is: an automated calculation system for the binding free energy of biomolecules based on structure prediction, antibody sequence clustering, and LIE-FEP diagram optimization, characterized in that it is used to execute the calculation method described in any one of claims 1 to 8, the system comprising: a data integration layer, deploying a centralized data management module for Campaign, maintaining the mapping relationship between candidate molecule BinderObject and BinderModel, and uniformly storing, verifying, and archiving full data of sequences, structures, mutations, simulations, and free energies; and a sequence structure preprocessing module, integrating an antibody standardization numbering submodule, structure analysis, and force field allocation. The system includes submodules for sequence-structure matching verification, clustering and FEP graph construction (including mutation feature extraction, dual-strategy clustering, FEP edge classification labeling, and batch generation of simulation tasks), free energy calculation and graph optimization (including fully automated molecular dynamics pipeline, non-equilibrium FEP and BAR correction, LIE energy extraction, OptimizationGraph global optimization, and bootstrap uncertainty analysis), and output modules for affinity ranking, FEP network graphs, clustering reports, confidence intervals, and standardized campaign archives.

[0021] In summary, the advantages of using the technical solution of this invention compared to traditional technical means are as follows: This invention manages sequence, structure, and simulation data uniformly through a campaign central object, enabling candidate antibodies to have traceable data relationships from input, clustering, graph construction, simulation to result output, reducing errors caused by manual processing, and improving reusability between different projects.

[0022] This invention automatically identifies a more suitable set of candidate pairs for pairwise comparisons by clustering based on the number of CDR mutations and region similarity, before performing costly FEP computations. Compared to computationally calculating each candidate pairwise, this approach significantly reduces invalid or unreliable FEP edges, allowing computational resources to be focused on candidate pairs most likely to produce reliable ΔΔG.

[0023] This invention incorporates insertions / deletions, charged mutations, mutation locations, and physical properties into the simulation plan through FEP graph construction and edge labeling mechanisms. This enables the system to automatically distinguish between regular FEP edges, edges requiring double-box or special processing, and candidate pairs that are not suitable for direct FEP, thereby improving the consistency and success rate of simulation setup.

[0024] This invention combines the single-molecule coverage capability of LIE with the high-precision relative free energy capability of FEP, and jointly optimizes ΔG and ΔΔG in the same graphical model through OptimizationGraph. This allows candidate molecules lacking direct FEP comparisons to be incorporated into the overall ranking through graph propagation and LIE node information. This approach reduces ranking conflicts when LIE and FEP are used independently, and improves the thermodynamic consistency of the free energy network through cyclic closed-loop correction.

[0025] The automated simulation process of this invention can compress the traditional, lengthy manual setup and iteration process for assessing binding free energy into a more scalable computational pipeline, supporting rapid and prioritized screening of a large number of candidate molecules in antibody discovery activities. According to the validation description in the technical solution, this hybrid method can achieve an average absolute error of approximately 2 kcal / mol relative to experimental binding data, and is expected to shorten the binding affinity assessment cycle from several months to several weeks.

[0026] This invention also expands the applicability of candidate comparison. For candidate molecules that cannot be directly compared stably using conventional FEP, such as those with charged mutations, insertions / deletions, or different epitope relationships, the system can utilize special edge markers, LIE node estimation, and graph optimization integration mechanisms for indirect comparison or unified ranking, thereby improving the coverage of antibody candidate priority decision-making. Attached Figure Description

[0027] The foregoing and other objects, features and advantages of the present invention will become apparent from the following detailed description taken in conjunction with the accompanying drawings.

[0028] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart illustrating the process of creating a structured Campaign object from a sequence. Figure 3 Flowchart for generating simulated inputs and running them step-by-step on HPC; Figure 4 This is a flowchart from LIE / FEP to the final ΔG ranking. Detailed Implementation

[0029] Based on the preferred embodiments of the present invention, and through the following description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

[0030] The invention will be further described with reference to the following figures: S1 Data Initialization and Sequence Normalization A project-level global data container was built to uniformly store all candidate molecules, structure files, simulation tasks, energy data, optimization results, task logs, and quality control labels for the project. 120 FASTA sequences of antibodies to be evaluated were imported, and a unique BinderObject was generated for each antibody sequence, with all data specific to each molecule encapsulated within the corresponding BinderObject. Sequence standardization was performed using the IMGT numbering system and the ANARCI tool, automatically identifying the six CDR regions of the heavy and light chains. The mutation sites, number of mutations, and mutation types (common amino acid mutations, charge mutations, insertion / deletion mutations) for each sequence were statistically analyzed. Chain information, IMGT site numbers, and CDR mutation characteristics were all written to the corresponding BinderObject and synchronously archived to persistent storage in the Campaign object.

[0031] S2 Clustering, FEP Map Construction, and Complex Structure Preprocessing A dual-strategy clustering method is used to divide candidate clusters. All BinderObjects are traversed to read CDR mutation features, and two types of clustering are performed in parallel: the first is a dense clustering method based on the CDR mutation threshold, where a mutation threshold of 5 is set. If the total number of CDR mutations between two molecules is ≤5, an undirected connection is established, and connected components are treated as independent candidate clusters; unconnected molecules are marked as singleton clusters. The second is a density clustering method based on the CDR six-region similarity (DBSCAN), where the CDR similarity is used to construct a multi-dimensional feature vector and then standardized before clustering. The results of the two clustering methods are merged to obtain non-overlapping computational clusters. No FEP pairing is performed between clusters, significantly reducing the number of computational combinations. In this embodiment, 120 molecules are ultimately divided into 8 candidate clusters, reducing the theoretical number of FEP pairings from 7140 to 150.

[0032] The system automatically generates classification labels for FEP edges, enumerating all molecule pairs within each cluster and classifying them into three categories according to preset rules: Molecular pairs with amino acid insertions / deletions are marked as substitution treatment edges and do not participate in neFEP simulation; edges without insertions / deletions but with charge changes due to mutations are marked as double-box neutralization special FEP edges; edges with no insertions / deletions, no charge mutations, and a total mutation count ≤ 5 are generated as regular FEP edges; edges with a total mutation count greater than 5 are not constructed, and evaluation is completed solely based on LIE energy and graph propagation. All cluster IDs, molecule pairing information, FEP edge types, mutation attributes, and quality control labels are written back to the campaign.

[0033] Subset-oriented structure prediction and preprocessing only selects a subset of molecules with FEP pairings to generate FASTA files. AlphaFold-Multimer is then used to batch predict the 3D structure of antigen-antibody complexes. The predicted structures are automatically converted into GROMACS-recognizable topology files, and solvent box construction and ion neutralization preprocessing are completed. All structure file storage paths, modeling completion status, and structure quality control results are stored in the Campaign, eliminating the need to batch model all 120 molecules and saving computational resources.

[0034] S3 HPC Phased Automated Simulation and LIE / FEP Energy Harvesting Based on the complex structure, simulated force field, duration, and scheduling parameters stored within the Campaign, the program automatically generates the GROMACS topology input file and the SLURM cluster scheduling script, submitting the high-performance computing task in three steps. All task execution records, file paths, and running status are written back to the Campaign in real time. The first step is to submit the equilibrium MD simulation HPC job: perform an energy minimization and NPT-constrained equilibrium simulation for 200 ps for each complex system. After the simulation is completed, write the trajectory file, energy output file, running log, and verification checkpoints into the corresponding fields of the Campaign. The second step is to submit the neFEP non-equilibrium free energy perturbation task: extract the equilibrium MD stable conformation as the initial structure, perform bidirectional alchemical transformation simulation on all regular FEP edges and special FEP edges, calculate the relative free energy ΔΔG and standard error SE for each pair of pairs using the BAR algorithm, and write back the original trajectory, ΔΔG, SE values, task ID, and failure flags to the Campaign; if the simulation fails due to an error, it is automatically marked and supports one-click rerun or weight reduction during the optimization stage; The third step involves batch collecting trajectories and performing extended LIE calculations: all equilibrium MD production trajectories are read, and the van der Waals and electrostatic interactions of the complex-bound state, free antigen, and free antibody are extracted respectively. These are then substituted into the extended LIE model to calculate the absolute binding free energy ΔG_LIE,i of a single molecule. The LIE energy value corresponding to each group of molecules is written to its own BinderObject and simultaneously summarized to the Campaign for unified management.

[0035] S4 OptimizationGraph: Global Thermodynamic Joint Optimization The campaign batch reads the LIE prior energy, error-weighted FEP edges ΔG and SE stored in all BinderObjects, and constructs an OptimizationGraph thermodynamic optimization model: with candidate molecules as nodes, and the initial value of each node is the internal ΔG of the BinderObject. LIE,i The standard error SE is calculated using FEP pairings as undirected edges and edge weights of BAR. The thermodynamic conservation relation ΔG is applied. FEP(i,j) =ΔG j -ΔG i A multi-objective loss function is constructed, and a nonlinear least squares algorithm is used to simultaneously and globally optimize and extend the LIE model while fixing the parameter α. bound α free β bound β free The algorithm calculates γ and the molecular-specific offset correction term f_i for each BinderObject. It iterates through all thermodynamic cycles in the graph, assigning cycle deviations based on the edge SE as weights, and performs cycle closure correction to eliminate thermodynamic inconsistencies in the network graph. A Bootstrap perturbation resampling method is used for 100 iterations, with random perturbations of FEP edge errors and LIE input energies for repeated global optimization. The 95% confidence interval of the combined free energy ΔG for each candidate molecule after optimization is statistically analyzed, and high-uncertainty mutant molecules are labeled.

[0036] S5 Results Output and Full Project Archiving Based on the complete storage of sequences, structures, simulation tasks, raw energies, and optimization data within the Campaign, standardized result data packages are output in batches. These packages include a globally optimized thermodynamically consistent ranking of candidate binding free energies, confidence intervals for each molecule, its cluster group, ΔG, SE, mutation treatment type, and cycle error contribution for each FEP edge. A Top-N candidate list for wet experimental validation is also output, along with mutation modification suggestions and weight reduction tips for high-risk unstable molecules. Finally, all original FASTA sequences, complex structure files, SLURM job scripts, MD / neFEP trajectories, raw energies of LIE and FEP, optimization calculation records, and quality control markers are archived into the Campaign object, enabling full traceability of any candidate molecule, any FEP edge, and any simulation task.

[0037] Implementation effect verification The optimized Pearson correlation coefficient (R) of the combined free energy and SPR in vitro affinity detection results in this embodiment is 0.85, with a mean absolute error of 1.8 kcal / mol. The prediction accuracy is significantly higher than that of the single LIE model and the manual segmented FEP calculation scheme. Among the 20 candidate molecules screened according to the output priority, 14 molecules have better affinity than the parent molecules, and the positive screening rate in wet experiments is 70%. At the same time, by relying on clustering to compress the computational space, subset-oriented modeling, and HPC automated scheduling, the overall computational power consumption is reduced by more than 70%, and the amount of manual operation is greatly reduced.

[0038] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any equivalent substitutions or fine-tunings of the simulation duration, clustering threshold, resampling times, and structure prediction tools under the technical concept of the present invention shall fall within the protection scope of the present invention.

[0039] 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 automated method for calculating the binding free energy of biomolecules based on sequence clustering and LIE-FEP graph optimization, characterized in that, It is suitable for predicting the affinity between antibodies, nanobodies, Fab, scFv and other binding molecules and antigens. It adopts a two-layer data architecture of project-level Campaign global data container and single-molecule BinderObject. The entire process of calculation files, task status and quality control information are written back to the Campaign to achieve data traceability, including the following steps: S1. Data initialization and sequence standardization: Create a Campaign object to carry all the data of the project, and generate a unique BinderObject for each input combined with the molecular sequence; standardize the sequence according to the IMGT numbering rules, identify the CDR mutation site, mutation number and mutation type, and store the chain type, number and mutation characteristics into the corresponding BinderObject and classify it into the Campaign. S2. Clustering, FEP Graph Construction and Structure Preprocessing: Extract CDR mutation features from each BinderObject, perform mutation threshold dense clustering and CDR similarity density clustering in parallel, merge to obtain independent candidate clusters, and label isolated molecules as singleton clusters; determine the mutation type only for molecular pairs within the same cluster, distinguish and generate regular FEP edges, charge mutation special FEPs, and insertion / deletion replacement edges, and write back the cluster ID, FEP edge list, edge attributes, and quality control labels to the Campaign; select only the subset of molecules with FEP edges to generate FASTA, call the structure prediction tool to obtain the antigen-antibody complex structure, complete the GROMACS format conversion, and store the structure path and processing status in the Campaign; S3, HPC phased automated simulation and energy harvesting: Based on the structure and simulation parameters stored in the Campaign, GROMACS input files and SLURM scheduling scripts are automatically generated, and high-performance cluster tasks are submitted in three steps: First, a balanced MD simulation job is submitted, and the trajectory, energy file, and verification log are written back to the Campaign upon completion; Second, a non-equilibrium free energy perturbation job for the neFEP is submitted based on the balanced conformation, outputting the raw ΔΔG data and recording the storage path; Third, the simulation trajectory is parsed in batches, and the extended LIE is run to calculate the interaction energy of the binding state and free state corresponding to each BinderObject; LIE energy, FEP relative free energy ΔΔG, standard error SE, job ID, run failure / rerun flag, and quality control label are all written back to the Campaign. S4. Joint Thermodynamic Optimization of the Optimization Graph: An optimization graph is constructed by reading LIE node data and error-weighted FEP edge data from the Campaign. The LIE energy stored in BinderObject is used as the node prior, and ΔG and SE are used as undirected edge attributes. A loss function is constructed based on thermodynamic conservation constraints, and the global parameters and molecule-specific offset terms of the LIE model are jointly optimized and expanded. Thermodynamic loop closures in the network graph are identified and loop closure corrections are performed to eliminate thermodynamic inconsistencies. Through Bootstrap resampling error propagation, the optimized free energy ΔG and confidence interval CI of each candidate molecule are output. The LIE model calculation formula is as follows: ΔG LIE,i =α bound •E VDW，bound,i -α free •E VDW，free,i +β bound •E ELEC，bound,i -β free •E ELEC，free,i +γ+f i Where α bound α free β bound β free γ are globally fixed parameters, f i This is a candidate molecule-specific offset term; S5. Results Output and Complete Project Archiving: Based on the full data output within the Campaign, standardized results are output, including optimized ΔG, candidate ranking, confidence interval, clustering grouping, ΔG / SE / processing method of each FEP edge, and a wet experiment Top-N candidate priority list; sequences, structure files, HPC job records, simulation trajectories, raw energy data, and optimization result data packages are all archived to the Campaign, achieving full-link traceability of molecules, FEP edges, simulation tasks, and final results.

2. The method according to claim 1, characterized in that, In step S2, the mutation threshold dense clustering rule is as follows: if the total number of CDR mutations between two candidate molecules is less than a preset threshold, an undirected connection edge is established, and the connected components are used as independent candidate clusters; the CDR similarity density clustering adopts the DBSCAN algorithm, and a multi-dimensional feature vector is constructed based on the sequence similarity of 6 CDR regions to complete the clustering.

3. The method according to claim 1, characterized in that, The specific rules for determining FEP edges in step S2 include: if a molecule pair has an amino acid insertion / deletion mutation, it is marked as a substitution treatment edge and does not participate in the neFEP calculation; if there is no insertion or deletion but there is a charge mutation, it is marked as a double-box neutralized special FEP edge; if there is no insertion or deletion and no charge mutation, and the total number of mutations is less than the threshold, a regular FEP edge is generated; if the total number of mutations exceeds the threshold, no FEP edge is constructed, and the evaluation is completed solely by LIE and graph propagation.

4. The method according to claim 1, characterized in that, In S3, the three-stage HPC task is executed sequentially: balanced MD simulation, neFEP free energy perturbation, and LIE energy batch acquisition. All tasks are scheduled to be parallelized on a high-performance cluster via SLURM scripts. FEP edges that fail in simulation are automatically marked, and reruns or subsequent optimization stages are supported for weight reduction processing.

5. The calculation method according to claim 4, characterized in that, The FEP calculation employs non-equilibrium free energy perturbation, including: extracting one representative conformation frame every 150–250 ps from the equilibrium simulation trajectory, with a total sampling frame count of 40–60 frames; the single-conformation alchemical transformation time is 60–100 ps, ​​the lambda value is linearly transformed from 0 to 1, and the van der Waals and Coulomb interactions utilize a soft-core potential to eliminate endpoint singularities; the BAR algorithm is used to calculate the paired relative binding free energy ΔΔG and standard error.

6. The method according to claim 1, characterized in that, Step S4, the cyclic closure correction operation, includes: traversing all thermodynamic cycles within the OptimizationGraph, allocating cycle deviations with the BAR standard error SE corresponding to each FEP edge as the weight, correcting the ΔΔG value of each edge, and achieving global thermodynamic self-consistency.

7. The method according to claim 1, characterized in that, In step S4, the Bootstrap resampling is performed no less than 100 times. By perturbing the FEP side error and repeatedly optimizing the LIE input energy, a 95% confidence interval is obtained to identify high-uncertainty mutation candidates.

8. The calculation method according to claim 1, characterized in that, Antibody sequence standardization numbering adopts the IMGT, Chothia, Kabat or AHo numbering system, and site alignment and CDR region identification are completed by ANACRI or AbNumber tools.

9. An automated calculation system for the binding free energy of biomolecules based on structure prediction, antibody sequence clustering, and LIE-FEP diagram optimization, characterized in that, The system, used to execute the calculation method according to any one of claims 1 to 8, comprises: a data integration layer, deploying a centralized data management module for Campaign, maintaining the mapping relationship between candidate molecule BinderObject and BinderModel, and uniformly storing, verifying, and archiving full data of sequences, structures, mutations, simulations, and free energies; a sequence structure preprocessing module, integrating an antibody standardization numbering submodule, a structure analysis and force field allocation submodule, and a sequence-structure matching verification submodule; a clustering and FEP graph construction engine, including a mutation feature extraction submodule, a dual-strategy clustering submodule, an FEP edge classification labeling submodule, and a simulation task batch generation submodule; a free energy calculation and graph optimization engine, including a fully automated molecular dynamics pipeline submodule, a non-equilibrium FEP and BAR correction submodule, a LIE energy extraction submodule, an OptimizationGraph global optimization submodule, and a bootstrap uncertainty analysis submodule; and a result output module, used to output affinity ranking, FEP network graph, clustering report, confidence interval, and standardized Campaign archive file.

Images