Sorting model-based PBAE polymer screening method
By using a sorting model-based approach, PBAE polymer structures are automatically assembled and ranked according to their relative properties. This solves the problem of integrating multi-source data, enabling efficient and accurate screening of PBAE polymers and improving screening efficiency and precision.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to quickly and accurately screen for the optimal PBAE polymers for specific cells, and multi-source heterogeneous data are difficult to integrate for training robust predictive models.
A ranking model-based approach is adopted. By constructing a dataset, extracting features, and training the model, polymer structures are automatically assembled using chemical reaction rules. Combined with molecular features and experimental conditions, relative performance ranking is performed to screen out target candidate polymers.
It significantly improved the model's prediction accuracy and generalization ability, enhanced screening efficiency and practicality, and enabled high-throughput, standardized screening of PBAE polymers.
Smart Images
Figure CN121963885A_ABST
Abstract
Description
A PBAE polymer screening method based on a ranking model Technical Field
[0001] This invention relates to the field of machine learning technology, and in particular to a PBAE polymer screening method based on a ranking model. Background Technology
[0002] Poly(β-amino ester) (PBAE) is a polymer system widely used for the delivery of various gene therapy payloads. It can efficiently encapsulate nucleic acids (such as DNA and RNA) through electrostatic interactions and promote cell uptake and lysosomal escape through mechanisms such as the proton sponge effect, thereby successfully delivering genes into cells. However, the transfection efficiency of PBAE is highly cell type specific, that is, for different types of cells, PBAE polymers with different structures need to be matched to achieve the best delivery effect (see reference [1]). Therefore, rapidly and accurately screening the optimal PBAE polymer for specific cells is a key step in gene therapy research and development.
[0003] Although high-throughput synthesis and screening technologies such as combinatorial chemistry (see reference [2]) and barcode (see reference [3]) have been used to discover new polymer structures, it is still a tedious, time-consuming and costly task to conduct one-by-one experimental verification of a large number of candidate polymers in different cell lines, which cannot meet the needs of rapid research and development.
[0004] To accelerate the screening process, researchers have attempted to use machine learning models to predict polymer properties. However, the amount of effective experimental data generated by a single research team is typically limited (often in the hundreds to thousands), insufficient to train a robust and reliable predictive model. Furthermore, data from publications by different research teams exhibit systematic differences due to variations in experimental units, measurement methods, environmental conditions, and evaluation criteria. This heterogeneity makes directly merging data from different sources to train traditional regression models impractical, as the models may confuse performance differences between polymer structures and experimental conditions. This creates data silos, preventing the effective integration and utilization of a large amount of existing research data. Therefore, there is an urgent need to develop a method that can integrate and utilize multi-source heterogeneous data to achieve efficient and accurate screening of PBAE polymers.
[0005] References: [1] J. Sunshine, JJ Green, KP Mahon, et al. Small-Molecule End-Groups of Linear Polymer Determine Cell-type Gene-Delivery Efficacy. AdvancedMaterials, 2009, 21(48): 4947-4951[2] JJ Green, R. Langer, DG Anderson. A Combinatorial PolymerLibrary Approach Yields Insight into Nonviral Gene Delivery. Accounts ofChemical Research, 2008, 41(6): 749-759[3] JE Dahlman, KJ Kauffman, Y. Xing, et al. Barcoded nanoparticles for high throughput in vivo discovery of targeted therapeutics. Proceedings of the National Academy of Sciences, 2017, 114(8): 2060-2065 Summary of the Invention
[0006] This invention proposes a PBAE polymer screening method based on a ranking model, which solves the problems of scattered sources, inconsistent standards, and difficulty in directly using existing PBAE polymer experimental data for machine learning.
[0007] The technical solution of this invention is implemented as follows:
[0008] This invention provides a PBAE polymer screening method based on a ranking model, comprising the following steps:
[0009] Dataset construction: Based on multi-source PBAE polymer experimental data, data from the same literature and obtained in the same experimental unit are divided into the same group. From the group with sufficient data, several polymer samples with experimentally verified transfection efficiency are selected to construct the validation set, and the remaining polymer samples are used to construct the training set.
[0010] Feature extraction: For each PBAE polymer in the dataset, based on the molecular structure information of its backbone monomers, side chain monomers, and blocking agents, a complete structural expression of the PBAE polymer is automatically generated through chemical reaction rules; molecular features are extracted based on the structural expression, and corresponding experimental condition features are extracted in conjunction with it.
[0011] Model training: Using the molecular features and experimental conditions of the training set as input, and the grouping information of the training set as group labels, a ranking model based on a list-level learning paradigm is used for training to learn the relative performance ranking of polymer samples within the same group; the model hyperparameters are fined using the validation set to obtain the trained ranking model.
[0012] New polymer screening: For new PBAE polymer combinations, the feature extraction process is executed, and the extracted molecular features are combined with the selected experimental condition features and input into the trained ranking model to obtain the predicted ranking score. Target candidate polymers are then selected based on the ranking score.
[0013] Specifically, the data sources and experimental conditions include at least one of the following: literature sources, experimental units, cell types, and nucleic acid types.
[0014] Specifically, the group with sufficient data refers to a group in which the number of polymer samples within the same group is greater than a set threshold.
[0015] Specifically, the method for automatically assembling and generating the structural expression of a complete PBAE polymer through chemical reaction rules includes the following steps: identifying primary amine groups on backbone monomers and acrylate groups on side-chain monomers as polymerization reaction sites based on the RDKit tool; generating a repeating unit structural expression of the polymer through Michael addition reaction; and reacting the repeating unit structure with a blocking agent at a given degree of polymerization to generate the SMILES expression of the complete PBAE polymer.
[0016] Specifically, the molecular features include: RDKit molecular descriptors for backbone monomers, side-chain monomers, and blocking agents; polymer fingerprints calculated based on the polymer repeating unit structure; and LogP values calculated based on the complete polymer SMILES expression.
[0017] Specifically, the method for tuning model hyperparameters using the validation set includes the following steps: 1) Using the LGBMRanker model from the LightGBM toolkit as the ranking model for training and evaluation; 2) Calculating the TopK precision of the ranking model on each group of the validation set, and taking the average of the TopK precision of each group as the overall evaluation metric, where K is a positive integer; 3) Using the automatic hyperparameter optimization tool Optuna, automatically updating the model hyperparameters according to the optimization direction of the overall evaluation metric, and retraining the model based on the updated hyperparameters; 4) Repeating the training, evaluation, and update operations until the overall evaluation metric no longer improves in N consecutive iterations, at which point the current hyperparameters are determined to be the optimal hyperparameters; 5) Training the ranking model with the optimal hyperparameters using the full data of the training and validation sets to obtain the final ranking model.
[0018] Specifically, the experimental conditions include: cell type, nucleic acid type, molar ratio of backbone monomer to side chain monomer, nucleic acid weight, weight ratio of polymer to nucleic acid, and degree of polymerization.
[0019] Furthermore, the selected experimental condition features are key feature combinations determined from multiple experimental condition features through feature importance analysis, including the degree of polymerization and the molar ratio of nucleic acid to polymer.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention transforms the absolute prediction problem of PBAE polymer performance into a relative ranking problem within a group, and uses a ranking model based on list-level learning for training, which effectively overcomes the heterogeneity problem caused by inconsistent units and conditions of experimental data from different literature sources; it enables scattered and heterogeneous multi-source data to be directly integrated into a unified training set, significantly expanding the available data scale and diversity of the model, fundamentally alleviating the bottleneck of data scarcity, and thus improving the prediction accuracy and generalization ability of the model; (2) The present invention constructs a validation set by adopting the optimal screening and partitioning strategy, that is, selecting the sample with the best known transfection efficiency from the experimental group with sufficient data as the validation benchmark, and using TopK accuracy as the core evaluation index for model tuning, creating a training and evaluation environment that closely matches the actual virtual screening target; it directly simulates the accurate selection from the candidate library. The core task of finding the optimal polymer makes the model optimization process strictly aligned with the final application goal, which significantly improves the practicality and reliability of the model in actual screening scenarios; (3) By integrating automated chemical reaction rules based on RDKit, this invention realizes the automatic assembly from monomer SMILES to complete PBAE polymer structure, and constructs a multi-dimensional feature automatic extraction process covering molecular descriptors, polymer fingerprints, LogP and key experimental parameters, which completely replaces the tedious and error-prone manual operation, realizes high-throughput and standardized digital characterization of massive polymer candidate structures, and thus greatly improves the efficiency of the overall screening process; (4) In the final new polymer prediction stage, this invention identifies key experimental parameters through feature importance analysis, and only combines and predicts these key features and molecular features, rather than enumerating all feature combinations. This optimization strategy greatly reduces unnecessary computational overhead. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 is a flowchart of a PBAE polymer screening method based on a sorting model according to the present invention.
[0023] Figure 2 is a schematic diagram of the process of constructing the dataset in this embodiment.
[0024] Figure 3 is a flowchart of the automatic assembly and generation of the complete PBAE polymer structure expression in this embodiment.
[0025] Figure 4 is a flowchart of the training and optimization process of the sorting model in this embodiment.
[0026] Figure 5 is a schematic diagram of the TopK index of the ranking model in this embodiment on the validation set.
[0027] Figure 6 shows the difference in ranking score distribution between the training set and the validation set for the ranking model in this embodiment.
[0028] Figure 7 is a flowchart of the new polymer combination prediction process in this embodiment. Detailed Implementation
[0029] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0030] Referring to Figure 1, this invention provides a PBAE polymer screening method based on a ranking model, comprising the following steps:
[0031] S1, Constructing the dataset: Based on multi-source PBAE polymer experimental data, data from the same literature and obtained in the same experimental unit are divided into the same group. From the group with sufficient data, several polymer samples with experimentally verified transfection efficiency are selected to construct the validation set, and the remaining polymer samples are used to construct the training set.
[0032] Specifically, the data sources and experimental conditions include at least one of the following: literature sources, experimental units, cell types, and nucleic acid types.
[0033] Specifically, the group with sufficient data refers to a group in which the number of polymer samples within the same group is greater than a set threshold. In this embodiment, the set threshold is 150.
[0034] As shown in Figure 2, in a preferred embodiment of the present invention, more than 30 PBAE polymer-related articles were collected by searching databases such as Web of Science, and more than 3,000 experimental data were manually extracted. From each group with sufficient data, the top 60 polymer samples in terms of transfection efficiency were selected to form the validation set. Groups that meet this condition include, for example, the JHGBM-551, ARPE-19, and COS-7 cell lines, so the total number of samples in the validation set can be 180. The remaining data and data from other groups with a sample size of less than 150 together constitute the training set.
[0035] S2, Feature Extraction: For each PBAE polymer in the dataset, based on the molecular structure information of its backbone monomers, side chain monomers, and blocking agents, the complete structural expression of the PBAE polymer is automatically generated through chemical reaction rules; molecular features are extracted based on the structural expression, and corresponding experimental condition features are extracted in association.
[0036] Specifically, the method for automatically assembling and generating the complete structural expression of the PBAE polymer through chemical reaction rules includes the following steps:
[0037] The RDKit tool was used to identify primary amine groups on backbone monomers and acrylate groups on side-chain monomers as polymerization reaction sites.
[0038] The structural expression of repeating polymer units was generated through Michael addition reaction;
[0039] At a given degree of polymerization, the repeating unit structure is reacted with a blocking agent to generate the SMILES expression of the complete PBAE polymer.
[0040] As shown in Figure 3, the automatic assembly process can automatically determine the reaction sites based on the input SMILES strings of backbone monomers, side chain monomers and blocking agents, generate a dedicated PSMILES expression to describe the repeating unit of the polymer, and a complete polymer SMILES expression at a specified degree of polymerization, without the need for manual drawing.
[0041] Specifically, as shown in Figure 3, the molecular features include:
[0042] RDKit molecular descriptors for backbone monomers, side-chain monomers, and blocking agents;
[0043] Polymer fingerprints calculated based on polymer repeating unit structures;
[0044] The LogP value is calculated based on the complete polymer SMILES expression.
[0045] The RDKit molecular descriptors may include molecular weight, number of heavy atoms, number of rotatable bonds, number of hydrogen bond donors / acceptors, and topological polar surface area, etc., with a total of 651 dimensions for the backbone, side chains, and blocking agents (each descriptor is 217 dimensions). The polymer fingerprint is a 600-dimensional polyBERT fingerprint calculated by PSMILES for the polymer formed by backbone monomers and side chain monomers. The LogP value is the lipid-water partition coefficient calculated based on the SMILES expression. In one embodiment, the LogP of repeating units at degrees of polymerization of 4 and 16 and the LogP of repeating units plus blocking agents were calculated, resulting in a total of 4-dimensional features. In summary, the total molecular feature dimensions of a single polymer can reach 1255 dimensions.
[0046] S3, Model Training: Using the molecular features and experimental condition features of the training set as input, and the grouping information of the training set as grouping labels, a ranking model based on a list-level learning paradigm is used for training to learn the relative performance ranking of polymer samples within the same group; the model hyperparameters are fine-tuned using the validation set to obtain the trained ranking model.
[0047] Specifically, the method of tuning model hyperparameters using a validation set includes the following steps:
[0048] The LGBMRanker model from the LightGBM toolkit is used as the ranking model for training and evaluation.
[0049] Calculate the TopK precision of the ranking model on each group of the validation set, and take the average of the TopK precision of each group as the overall evaluation metric, where K is a positive integer;
[0050] The automatic hyperparameter optimization tool Optuna is used to automatically update the model hyperparameters according to the optimization direction of the overall evaluation metric, and the model is retrained based on the updated hyperparameters.
[0051] The training, evaluation, and update operations are performed repeatedly until the overall evaluation metric no longer improves in N consecutive iterations, at which point the current hyperparameter is determined to be the optimal hyperparameter.
[0052] The ranking model with optimal hyperparameters is trained using the full dataset of the training and validation sets to obtain the final ranking model.
[0053] In one embodiment of the present invention, as shown in Figure 4, the validation set groups are the JHGBM-551, ARPE-19, and COS-7 cell line groups mentioned in step S1. The precision (Pr) of each group in the Top50 (K=50) ranking results is calculated using the formula Pr = TP / n, where TP is the number of true optimal polymer samples in the predicted top 50 ranking results, and n is 50. The average of the Top50 precision Pr of the three groups is fed back to Optuna for hyperparameter optimization. Optimization objectives include key parameters such as learning rate and maximum tree depth. When the average precision no longer improves after multiple iterations, the optimal hyperparameters are determined. Subsequently, all training and validation set data are merged, and the LGBMRanker model is retrained with the optimal hyperparameters to obtain the ranking model used for final prediction.
[0054] Figure 5 shows the model precision for different K values under the optimal hyperparameters. When K=1 (i.e., taking the highest-ranked sample in each group), the Top1 precision is 1, indicating that the model can accurately locate the optimal PBAE polymer in each group. When K=5, the average precision of the three groups reaches 0.73, proving that the candidate samples ranked higher in the model ranking results have extremely high accuracy and can effectively support subsequent screening tasks.
[0055] The differences in model ranking scores between the training and validation sets are shown in Figure 6 (Train represents the training set, and Test represents the validation set). In Figure 6, (a) shows the distribution difference curves of model ranking scores between the training and validation sets in the JHGBM-551 group, (b) shows the distribution difference curves of model ranking scores between the training and validation sets in the ARPE-19 group, and (c) shows the distribution difference curves of model ranking scores between the training and validation sets in the COS-7 group. Both groups show significant distribution characteristics. The overall score of the validation set is slightly higher than that of the training set, which is consistent with expectations—because the validation set samples are all the best PBAE polymer data validated in each cell line. The success ratio in the figure is defined as the percentage of samples in the validation set whose ranking score exceeds the highest score in the training set (for example, 14 / 60 means that out of 60 validation set samples, 14 samples have a ranking score higher than the highest score in the training set). The results show that some validation set samples in each group have ranking scores that exceed the best level in the training set, proving that the model has the ability to screen out better PBAE polymers from candidate samples.
[0056] S4, New Polymer Screening: For new PBAE polymer combinations, the feature extraction process is executed, and the extracted molecular features are combined with the selected experimental condition features and input into the trained ranking model to obtain the predicted ranking score. Target candidate polymers are then screened based on the ranking score.
[0057] Specifically, as shown in Figure 3, the experimental conditions include: 14-dimensional cell type encoded by OneHot, 2-dimensional nucleic acid type encoded by OneHot, molar ratio of backbone to side chain monomers, nucleic acid weight, polymer to nucleic acid weight ratio, and degree of polymerization data after logarithmic transformation. If no degree of polymerization data is available, the degree of polymerization is calculated backward from the molecular weight described in the literature. If neither degree of polymerization nor molecular weight data is available, 16 is used as the default padding.
[0058] Furthermore, the selected experimental condition features are key feature combinations determined from multiple experimental condition features through feature importance analysis, including the degree of polymerization and the molar ratio of nucleic acid to polymer.
[0059] In one embodiment, as shown in Figure 7, a virtual screening library containing 140,000 unverified new PBAE polymer combinations is first constructed. Feature importance analysis is performed on the model to determine the molar ratio of nucleic acids to polymers and the degree of polymerization as key features. Subsequently, the key features are discretized and combined: the molar ratio is selected at four levels (30, 69, 90, and 160); the degree of polymerization is selected at three levels (16, 26, and 34); and the remaining non-key experimental condition features use the parameters with the highest frequency in the training set as default values. The molecular characteristics of the new polymers are combined with the above different key features and input into the final ranking model for prediction. A ranking score threshold (e.g., 7.5) is set, and polymer combinations with scores higher than this threshold are selected as high-potential target candidate polymers for subsequent experimental verification.
[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A PBAE polymer screening method based on a ranking model, characterized in that, The process includes the following steps: Dataset construction: Based on multi-source PBAE polymer experimental data, data from the same literature and obtained in the same experimental unit are grouped together. From the group with sufficient data, several polymer samples with experimentally verified high transfection efficiency are selected to construct a validation set. The remaining polymer samples are used to construct a training set. Feature extraction: For each PBAE polymer in the dataset, based on the molecular structure information of its backbone monomers, side chain monomers, and blocking agents, a complete structural expression of the PBAE polymer is automatically generated through chemical reaction rules. Molecular features are extracted based on the structural expression, and corresponding experimental condition features are also extracted. Model Training: Using molecular features and experimental condition features of the training set as input, and grouping information of the training set as group labels, a ranking model based on a list-level learning paradigm is used for training to learn the relative performance ranking of polymer samples within the same group; the model hyperparameters are fine-tuned using a validation set to obtain a trained ranking model; New Polymer Screening: For new PBAE polymer combinations, the feature extraction process described above is executed, and the extracted molecular features are combined with the selected experimental condition features and input into the trained ranking model to obtain a predicted ranking score. Target candidate polymers are then screened based on the ranking score.
2. The PBAE polymer screening method based on a ranking model as described in claim 1, characterized in that, The data sources and experimental conditions include at least one of the following: literature sources, experimental units, cell types, and nucleic acid types.
3. The PBAE polymer screening method based on a ranking model as described in claim 1, characterized in that, The group with sufficient data refers to a group in which the number of polymer samples within the same group is greater than a set threshold.
4. The PBAE polymer screening method based on a ranking model as described in claim 1, characterized in that, The method for automatically assembling and generating the structural expression of a complete PBAE polymer using chemical reaction rules includes the following steps: identifying primary amine groups on backbone monomers and acrylate groups on side-chain monomers as polymerization reaction sites based on the RDKit tool; generating a repeating unit structural expression of the polymer through Michael addition reaction; and reacting the repeating unit structure with a blocking agent at a given degree of polymerization to generate the SMILES expression of the complete PBAE polymer.
5. The PBAE polymer screening method based on a ranking model as described in claim 4, characterized in that, The molecular features include: RDKit molecular descriptors for backbone monomers, side-chain monomers, and blocking agents; polymer fingerprints calculated based on the polymer repeating unit structure; and LogP values calculated based on the complete polymer SMILES expression.
6. The PBAE polymer screening method based on a ranking model as described in claim 1, characterized in that, The method for tuning model hyperparameters using a validation set includes the following steps:
1. Using the LGBMRanker model from the LightGBM toolkit as the ranking model for training and evaluation; 2. Calculating the TopK precision of the ranking model on each group of the validation set, and taking the average of the TopK precision of each group as the overall evaluation metric, where K is a positive integer; 3. Using the automatic hyperparameter optimization tool Optuna, automatically updating the model hyperparameters according to the optimization direction of the overall evaluation metric, and retraining the model based on the updated hyperparameters; 4. Repeating the training, evaluation, and update operations until the overall evaluation metric no longer improves in N consecutive iterations, at which point the current hyperparameters are determined to be the optimal hyperparameters; 5. Training the ranking model with the optimal hyperparameters using the full data of the training and validation sets to obtain the final ranking model.
7. The PBAE polymer screening method based on a ranking model as described in claim 1, characterized in that, The experimental conditions include: cell type, nucleic acid type, molar ratio of backbone monomer to side chain monomer, nucleic acid weight, weight ratio of polymer to nucleic acid, and degree of polymerization.
8. The PBAE polymer screening method based on a ranking model as described in claim 7, characterized in that, The selected experimental condition features are key feature combinations determined from multiple experimental condition features through feature importance analysis, including the degree of polymerization and the molar ratio of nucleic acid to polymer.