Method for locating key amino acid sites and predicting optimal amino acids based on model explainability tool and matrix inversion

Abstract

The application discloses a method for positioning key amino acid points and predicting optimal amino acids based on model interpretability tools and matrix inversion transformation. In the protein large language model, the model interpretability tools in the field of computer vision are introduced, which can effectively identify the discrimination area and key information of the model, establish the connection between the classification result and the input to identify the most critical amino acid point, and introduce PointMLP to infer the microenvironment around the key amino acid point in the form of point cloud to predict the optimal amino acid. The application can greatly reduce the training cost of the model and efficiently screen out beneficial mutants, and the stability effective modification rate is 31%-80%.

Description

Technical Field

[0001] This invention belongs to the field of protein engineering and relates to a method for locating key amino acid sites and predicting the optimal amino acid based on model interpretability tools and matrix inverse transformation. Background Technology

[0002] Protein engineering is the process of enhancing existing properties or acquiring new properties by altering the amino acid sequence composition. Directed evolution, through techniques such as error-prone PCR and site-specific saturation mutagenesis, can achieve specific screening for certain protein properties. Directed evolution mainly involves two steps: mutation and screening. This process typically requires constructing a library of random mutant sequences and iteratively screening for high-performance mutants. However, the large size of these random mutant libraries leads to low high-throughput screening efficiency. Constructing high-quality sequence mutant libraries and finding efficient methods for screening mutant libraries are currently urgent areas for breakthroughs in directed evolution.

[0003] Current protein modification has shifted from traditional and complex directed evolution to more efficient computationally assisted and semi-rational design, with the core strategy of "precise site selection followed by targeted mutation." In practical applications, researchers typically combine protein 3D structure analysis with strategies such as CASTing to identify physical hotspots like catalytic pockets or flexible regions. Simultaneously, they utilize bioinformatics tools such as multiple sequence alignment (MSA) to extract high-frequency mutation sites or ancestral sequence features from the vast number of sequences generated through natural evolution. In recent years, artificial intelligence technology has further revolutionized the site selection paradigm. Deep learning-based protein language models (such as ESM2, Evolutionary Scale Modeling) can directly predict zero-sample mutation effects. This early site selection strategy, integrating structure, evolution, and underlying AI algorithms, significantly reduces the size of experimental mutation libraries and substantially improves the success rate of modifying the function and stability of target proteins. Summary of the Invention

[0004] The purpose of this invention is to provide a method for locating key amino acid sites and predicting optimal amino acids based on model interpretability tools and matrix inverse transformation. Based on the unique characteristics of protein sequence, structure, and function as a unified whole, this invention establishes a method—CASPE (Critical Amino Acids Streamline Protein Evolution)—to find the optimal amino acid that effectively enhances certain protein properties (such as heat resistance, acid resistance, and alkali resistance) in the current microenvironment, thereby reducing model training load and locating key amino acids.

[0005] The technical solution for achieving the objective of this invention is as follows:

[0006] The method for locating key amino acid sites based on model interpretability tools and matrix inverse transformation includes the following steps:

[0007] S1. Based on the protein language model, proteins are converted into feature vectors (embeddings), and proteins are classified to train a protein classification model.

[0008] S2. Analyze the protein classification model based on model interpretability tools, locate key amino acid sites related to protein characteristics, and obtain the contribution value of each position in the feature vector through model interpretability tools.

[0009] S3. Convert the contribution value of the feature vector into the contribution value of the amino acid sequence, sort them according to the amino acid contribution value, select the top 5-10 sites as core sites, and select 1-2 sites before and after each core site, which are the key amino acid sites.

[0010] A method for locating key amino acid sites and predicting the optimal amino acid based on model interpretability tools and inverse matrix transformation includes the following steps:

[0011] S1. Based on the protein language model, proteins are converted into feature vectors, and proteins are classified to train a protein classification model.

[0012] S2. Analyze the protein classification model based on model interpretability tools, locate key amino acid sites related to protein characteristics, and obtain the contribution value of each position in the feature vector through model interpretability tools.

[0013] S3. Convert the contribution value of the feature vector into the contribution value of the amino acid sequence, sort them according to the amino acid contribution value, select the first 5-10 sites as core sites, and select 1-2 sites before and after each core site, which are the key amino acid sites.

[0014] S4. Predict the three-dimensional structure of the protein sequence using AlphaFold or ESMFold. Based on the three-dimensional structure, obtain the amino acid microenvironment around key amino acid sites. Use point clouds to describe the amino acid microenvironment. During data sampling, focus on the central amino acid. x-axis direction A three-dimensional coordinate system is constructed along the y-axis, and the amino acid microenvironment is quantitatively described using (x, y, z, value), in the form of: Save it as an h5 file;

[0015] S5. Input the point cloud data in the h5 file into the PointMLP framework for training, calculate the cross-entropy loss, and use the Adam optimizer to optimize the model training process. The learning rate is 0.001 and decreases with the number of training iterations. When the loss tends to stabilize, stop training to obtain the optimal amino acid prediction platform APCNet.

[0016] S6. For the validation protein sequence, the key amino acid sites obtained in S2 and the amino acid microenvironment obtained in S4 are input into the optimal amino acid prediction platform APCNet to predict the probability of 20 amino acids existing in the current amino acid microenvironment. The amino acid with the highest probability is selected as the optimal amino acid.

[0017] Furthermore, S1 specifically involves: converting protein sequences into feature vectors based on the ESM2 large language model, connecting the MLP to ESM2 to classify the protein sequences, adding BatchNorm1d, ReLU, and Dropout layers to the classification layer to prevent model overfitting, and training to obtain a protein classification model.

[0018] Furthermore, in S1, protein sequences are obtained from open-source protein databases and tagged with protein characteristic labels. Protein characteristics include, but are not limited to, heat resistance, acid resistance, alkali resistance, and organic solvent resistance. Specifically, for example, protein sequences adapted to bacterial optimal growth temperatures above 60°C are tagged with strong heat stability; protein sequences adapted to bacterial optimal pH below 4 are tagged with strong acid resistance; and protein sequences adapted to bacterial optimal pH above 8 are tagged with strong alkali resistance.

[0019] Furthermore, in S2, the model interpretability tool is Grad-CAM, and the contribution value at each position in the feature vector is set to... , The calculation formula is:

[0020] (1)

[0021] in, This represents the activation value of the element in the i-th row and j-th column of the embeddings when the model propagates forward to the linear layer. This represents the gradient of the classification weights with respect to the element in the i-th row and j-th column of the embeddings during backpropagation of the model to the linear layer. and After multiplication, the ReLU function is used to filter out the positions that have a positive impact on classification.

[0022] Furthermore, in S3, the contribution value of the amino acid sequence is... , The calculation formula is:

[0023] (2)

[0024] in, It is the contribution matrix of embeddings. It is the inverse of the attention weight matrix of the i-th transformer layer. This indicates that the mean of the contribution values ​​in the j-th column is used as the contribution of the j-th element of the amino acid. Further, in S4, the PDB file obtained from the three-dimensional structure prediction of the protein sequence by AlphaFold or ESMFold is converted into a PQR file using PDB2PPQR. The amino acid microenvironment information surrounding key amino acid sites is obtained from the PQR file and converted into point cloud data, with the central amino acid serving as the data tag.

[0025] Furthermore, in S4, the amino acid microenvironment includes the atomic microenvironment, atomic charge, and solvent accessible surface area (SASA). The atomic charge is sampled using BioPython, the atomic charge is calculated using Gaussian and Multiwfn, and the SASA is calculated using FreeSASA.

[0026] Furthermore, in S4, the intermediate encoding layer of the PointMLP framework is used to construct the optimal amino acid prediction platform APCNet, with the number of point cloud inputs ranging from 256 to 1536.

[0027] Furthermore, in S4, L2 regularization is introduced to prevent overfitting.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] (1) Traditional protein big language models guide enzyme modification by inputting the entire protein sequence into the model for prediction. Due to the black-box nature of deep learning models, the relationship between the modification points obtained by researchers and the actual performance to be modified is unknown. This invention introduces Grad-CAM, a model interpretability tool from the field of computer vision, into the field of protein big language models for the first time. It can effectively identify the discriminative regions of the model and key information. First, ESM2 is used to expand the amino acid sequence into two-dimensional data similar to the image. Based on Grad-CAM, the relationship between amino acids and protein characteristics is obtained, and the connection between the classification results and the input is established to identify the most critical amino acid points.

[0030] (2) This invention introduces PointMLP to infer the microenvironment surrounding key amino acid sites in the form of point clouds, discarding non-critical data, which can greatly reduce the training load of the model to obtain a lightweight model and predict the optimal amino acid. This strategy can greatly reduce the training cost of the model and efficiently screen out beneficial mutants, with a stability and effective modification rate of 31%~80%. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the amino acid microenvironment data sampling process;

[0032] Figure 2 (a) Description of the process of deep learning guiding protein evolution in the literature and (b) Schematic diagram of the CASPE strategy, where DB is the database, CAS is the key amino acid site, and APCNet is the amino acid point cloud classification network.

[0033] Figure 3 The diagram shows (a) the process of fine-tuning ESM2 and locating key amino acid sites based on Grad-CAM, and (b) the process of acquiring the amino acid microenvironment and training APCNet. Modules 1-4 are the same module.

[0034] Figure 4 This is a schematic diagram of the thermostability modification of cellulase based on CASPET, where (ab, gh) represents the experimental results of BG, (cd, ij) represents the experimental results of EG, (ef, kl) represents the experimental results of CBHI, (a / c / e) represents the mutation sites of BG, EG and CBHI screened based on CASPET, (g / i / k) represents the heatmap of the mutation sites of BG, EG and CBHI screened based on CASPET, and (b / d / f, h / j / l) represents the thermostability experimental results of the mutants of BG, EG and CBHI screened based on CASPET.

[0035] Figure 5 This diagram illustrates the pH stability modification of beta-glucosidase based on CASPEA. (a / c / d) represents the transferability verification of CASPEA sequences related to acid tolerance, (b / e / f) represents the transferability verification of CASPEA sequences related to alkali tolerance, (ab) is a heatmap of BG mutation sites screened based on the CASPEA framework, and (cf) represents the mutant performance of BG screened based on the CASPEA framework. Detailed Implementation

[0036] The technical solution of the present invention will now be clearly and completely described with reference to specific embodiments and accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] The method for locating key amino acid sites and predicting optimal amino acids based on Grad-CAM and inverse matrix transformation of this invention is defined as CASPE, and specifically includes:

[0038] (1) Classify protein sequences according to ESM2. Protein sequences were obtained from open-source protein databases and labeled with protein characteristics, including but not limited to heat resistance, acid resistance, and alkali resistance. Based on ESM2, the protein sequences were converted into embeddings, and MLP was used as a classifier to classify the embeddings. To prevent overfitting, BatchNorm1d, ReLU, and Dropout layers were added to the classification layer to train the protein classification model. Since one dimension of the embeddings is related to the length of the input protein sequence, zreo padding was used to expand it by 1600 amino acids. After one-dimensional expansion, it was input into the subsequent MLP classifier.

[0039] (2) Based on the protein classification model, Grad-CAM is used to identify key amino acids related to protein characteristics. During the model inference process, each position in the amino acid sequence has a different contribution value to the final classification weight. The larger the contribution value, the closer the relationship between the position and the protein characteristics. In the ESM2 encoding stage, the amino acid sequence is processed by the attention matrix of multiple transform layers to obtain embeddings. Therefore, the contribution value of the embeddings can be converted into the contribution value of the amino acid sequence through the inverse transformation of the attention matrix, thereby obtaining the key amino acid positions. Specifically, in order to locate the key amino acids, the key data (weight file) in the model inference process is used for reverse inference, and the contribution value of each position on the amino acid sequence is deduced from the classification weight value. The whole process is mainly divided into two steps. First, the contribution value of each position in the embeddings is obtained through Grad-CAM. , The calculation formula is:

[0040] (1)

[0041] in, This represents the activation value of the element in the i-th row and j-th column of the embeddings when the model propagates forward to the linear layer. This represents the gradient of the classification weights with respect to the element in the i-th row and j-th column of the embeddings during backpropagation of the model to the linear layer. and After multiplication, the ReLU function is used to filter out the positions that have a positive impact on classification.

[0042] Then the contribution values ​​of embeddings are converted into the contribution values ​​of amino acid sequences. , The calculation formula is:

[0043] (2)

[0044] in, It is the contribution value of the amino acid sequence. It is the contribution matrix of embeddings. It is the inverse of the attention weight matrix of the i-th transformer layer. This means that the mean of the contribution values ​​in the j-th column is used as the contribution of the j-th element of the amino acid.

[0045] Finally, based on Sort the sites and select the top 5-10 sites as key amino acid sites.

[0046] (3) The three-dimensional structure of the protein sequence is predicted using AlphaFold or ESMFold, and the amino acid microenvironment surrounding key amino acid sites is obtained based on the three-dimensional structure. The amino acid microenvironment includes the atomic microenvironment composed of C, H, O, N, and S, SASA, and atomic charge. Among them, C, H, O, N, and S represent the three-dimensional structural information of amino acids, SASA and atomic charge represent the physicochemical properties of amino acids, and RESP charge is currently the most suitable charge for molecular simulation of flexible small molecules. Figure 1 To address the issues of training accuracy and speed, point cloud data is introduced to describe the amino acid microenvironment. Point cloud data is a disordered collection with fixed dimensions and variable quantity. The shape of the object represented by the point cloud should possess rotation and translation invariance, which aligns with the spatial structure characteristics of proteins. Due to the variable quantity of point cloud data, the difference in the number of coordinate points in the amino acid microenvironment caused by different sampling locations can be effectively resolved, redundancy in 3D mesh data is removed, and data storage and model loading are facilitated. BioPython is used to sample atomic charges, Gaussian and Multiwfn are used to calculate atomic charges, and FreeSASA is used to calculate SASA. During the point cloud data sampling process, the central amino acid is used as the reference point. x-axis direction A three-dimensional coordinate system is constructed along the y-axis, and the amino acid microenvironment is quantitatively described using (x, y, z, value), in the form of: And save it as an h5 file.

[0047] S4. PointMLP is introduced to train point cloud data, constructing the optimal amino acid prediction platform APCNet. For the loss function, cross-entropy is used as the classifier's loss function, and L2 regularization is introduced to prevent overfitting. The model training result can reach a maximum of 88%. Specifically, the point cloud data from the h5 file is input into the PointMLP framework for training. Cross-entropy loss is calculated, and the Adam optimizer is used to optimize the model training process. The learning rate is 0.001 and decreases with the number of training iterations. Training stops when the loss tends to plateau, resulting in the optimal amino acid prediction platform APCNet.

[0048] S5. For the validation protein sequence, the key amino acid sites obtained in S2 and the amino acid microenvironment obtained in S3 are input into the optimal amino acid prediction platform APCNet in S4. The probability of 20 amino acids existing in the current amino acid microenvironment is predicted, and the amino acid with the highest probability is selected as the optimal amino acid.

[0049] Example 1

[0050] Based on Grad-CAM and inverse matrix transformation, a method for locating key amino acid sites related to thermal stability and predicting the optimal amino acid sequence is defined as CASPET (Thermal Stability Related CASPE). First, protein sequences related to protein thermal stability are input into the model. To obtain data most relevant to protein thermal stability, Grad-CAM is used to extract key sites from the protein sequences. Figure 3 a) to achieve data condensation. In the field of computer vision, Grad-CAM has certain limitations in locating complex objects, making it difficult to accurately delineate the boundaries of complex objects. Considering the ambiguity of Grad-CAM localization, the site before / after the key localization point is selected for subsequent model training. The amino acid microenvironment around the key amino acid site is obtained as input data for subsequent model training ( Figure 3 b). The microenvironment surrounding the key amino acid sites is extracted and input into pointMLP for training. Cross-entropy loss is calculated, and the Adam optimizer is used to optimize the model training process. The learning rate is 0.001 and decreases with the number of training iterations. Training is stopped when the loss tends to stabilize, resulting in the optimal amino acid prediction platform APCNet.

[0051] The significance of using the Grad-CAM method to extract key amino acid sequences lies in two aspects: First, it allows data most relevant to protein thermostability to be used for later model training, reducing computational load while improving model accuracy. Second, the data obtained in this way fundamentally breaks free from the constraints of individual protein sequences and focuses on the amino acid microenvironment with specific characteristics. Since deep learning datasets are divided into training and testing sets, if the entire amino acid sequence is used as the original training data, and later modifications encounter amino acid sequences that are the same as or similar to those in the training set, the model will produce the same prediction results as the input data during training. However, by selecting the few amino acids that contribute most to protein thermostability using Grad-CAM, theoretically, the model is related to specific amino acids for protein thermostability, rather than a model trained on a set of certain amino acid sequences, which is beneficial for later protein stability modifications.

[0052] First, CASPET was used to guide the modification of cellulase, such as... Figure 4 As shown in g, 4i, and 4k, the mutation sites of BG, EG, and CBHI were first located based on Grad-CAM. 39, 24, and 37 key amino acid sites were selected from the three sequences, respectively. The amino acid microenvironment of these sites was then input into APCNet to predict the optimal amino acids, ultimately obtaining the modification sites for BG, EG, and CBHI. Protein expression and enzyme activity assays were performed on the mutants and their wild-types. The results are shown below. Figure 4 As shown in h, 4j, and 4l, the beneficial mutants of the three enzymes are: BG, N227A, H229F, D240N, D306A, Q319A, P321A; EG, A45N, T91I, G92S, S149N, V150L, ​​A234T; CBH1, Y78F, N188T, I203T, Q312S, Q313L. Among these, BG yielded 6 beneficial mutants out of 10 mutants (60%), EG yielded 6 out of 14 mutants (43%), and CBH1 yielded 5 out of 16 mutants (31%). Figure 4 b / d / f).

[0053] Example 2

[0054] Based on Grad-CAM and inverse matrix transformation, a method for locating key amino acid sites related to pH stability and predicting optimal amino acids was developed. The entire process framework was defined as CASPEA (pH stability-related CAPSE). First, protein sequences related to pH stability were obtained from the literature (Rong, H., et al., AcidBasePred: a protein acid-basetolerance prediction platform based on deep learning[J], Chin. J.Biotechnol.,2024, 40: 4670-4681). Protein sequences adapted to bacterial optimal pH below 4 were tagged as acid-resistant, and those adapted to bacterial optimal pH above 8 were tagged as alkali-resistant. The database retrieved corresponding protein sequences based on the microorganism's optimal pH. After filtering sequences by length (greater than 100 and less than 1000), a total of 84,303 acid-tolerant sequences and 49,912 alkali-tolerant sequences were obtained. The training, validation, and test sets were divided into 8:1:1 ratios of acid-resistant and alkali-resistant sequences, respectively. The training method was the same as in Example 1, training models related to protein acid and alkali stability. The validation set accuracies for the acid and alkali stability models were 67.25% and 70.349%, respectively. 26 and 37 sites were located using Grad-CAM, respectively, and then screened using APCNet (APCNetAA: acid-resistant APCNet; APCNetAL: alkali-resistant APCNet), resulting in 10 mutation sites for subsequent wet experiments. Figure 5 (ab). The amino acid sites obtained from acid and alkali resistance modifications of the protein are D14G, I90L, T120Q, D236G, A239G, D240N, L241Q, A266G, E267Q, and E351D; D14G, A42Q, F43Y, C44G, D68E, A239S, L241E, A266G, E351D, and L379Q. When the same sequence was modified for acid and alkali resistance respectively, the model predicted three repeat sites: D14G, A266G, and E351D. The wet test results are as follows: Figure 5 As shown in cf, the beneficial mutants obtained from acid and alkali resistance modifications are D14G, I90L, T120Q, D236G, A239G, D240N, A266G, and E267Q; and D14G, D68E, L241E, and E351D, respectively. Based on BG's acid and alkali resistance modifications, the beneficial mutants accounted for 80% and 40% of the total mutants, respectively.

Claims

1. A method for locating key amino acid sites based on model interpretability tools and matrix inverse transformation, characterized in that, Includes the following steps: S1. Based on the protein language model, proteins are converted into feature vectors, and proteins are classified to train a protein classification model. S2. Analyze the protein classification model based on model interpretability tools, locate key amino acid sites related to protein characteristics, and obtain the contribution value of each position in the feature vector through model interpretability tools. S3. Convert the contribution value of the feature vector into the contribution value of the amino acid sequence, sort them according to the amino acid contribution value, select the top 5-10 sites as core sites, and select 1-2 sites before and after each core site, which are the key amino acid sites.

2. A method for locating key amino acid sites and predicting the optimal amino acid based on model interpretability tools and matrix inverse transformation, characterized in that: Includes the following steps: S1. Based on the protein language model, proteins are converted into feature vectors, and proteins are classified to train a protein classification model. S2. Analyze the protein classification model based on model interpretability tools, locate key amino acid sites related to protein characteristics, and obtain the contribution value of each position in the feature vector through model interpretability tools. S3. Convert the contribution value of the feature vector into the contribution value of the amino acid sequence, sort them according to the amino acid contribution value, select the first 5-10 sites as core sites, and select 1-2 sites before and after each core site, which are the key amino acid sites. S4. Predict the three-dimensional structure of the protein sequence using AlphaFold or ESMFold. Based on the three-dimensional structure, obtain the amino acid microenvironment around key amino acid sites. Use point clouds to describe the amino acid microenvironment. During data sampling, focus on the central amino acid. x-axis direction A three-dimensional coordinate system is constructed along the y-axis, and the amino acid microenvironment is quantitatively described using (x, y, z, value), in the form of: Save it as an h5 file; S5. Input the point cloud data in the h5 file into the PointMLP framework for training, calculate the cross-entropy loss, and use the Adam optimizer to optimize the model training process. The learning rate is 0.001 and decreases with the number of training iterations. When the loss tends to stabilize, stop training to obtain the optimal amino acid prediction platform APCNet. S6. For the validation protein sequence, the key amino acid sites obtained in S2 and the amino acid microenvironment obtained in S4 are input into the optimal amino acid prediction platform APCNet to predict the probability of 20 amino acids existing in the current amino acid microenvironment. The amino acid with the highest probability is selected as the optimal amino acid.

3. The method according to claim 1 or 2, characterized in that, S1 specifically involves: converting protein sequences into feature vectors based on the ESM2 large language model, connecting the MLP to ESM2 to classify the protein sequences, adding BatchNorm1d, ReLU and Dropout layers to the classification layer to prevent overfitting, and training to obtain a protein classification model.

4. The method according to claim 1 or 2, characterized in that, In S1, protein sequences are obtained from open-source protein databases and tagged with protein characteristics.

5. The method according to claim 4, characterized in that, Proteins possess one or more of the following properties: heat resistance, acid resistance, alkali resistance, and organic solvent resistance.

6. The method according to claim 5, characterized in that, Protein sequences that are compatible with the optimal growth temperature of bacteria above 60℃ are labeled as having strong thermal stability; protein sequences that are compatible with the optimal pH of bacteria below 4 are labeled as having strong acid resistance; and protein sequences that are compatible with the optimal pH of bacteria above 8 are labeled as having strong alkali resistance.

7. The method according to claim 1 or 2, characterized in that, In S2, the model interpretability tool is Grad-CAM, and the contribution value at each position in the feature vector is set to... , The calculation formula is: （1） in, This represents the activation value of the element in the i-th row and j-th column of the embeddings when the model propagates forward to the linear layer. This represents the gradient of the classification weights with respect to the element in the i-th row and j-th column of the embeddings during backpropagation of the model to the linear layer. and After multiplication, the ReLU function is used to filter out the positions that have a positive impact on classification.

8. The method according to claim 1 or 2, characterized in that, In S3, the contribution value of the amino acid sequence is... , The calculation formula is: （2） in, It is the contribution matrix of embeddings. It is the inverse of the attention weight matrix of the i-th transformer layer. This means that the mean of the contribution values ​​in the j-th column is used as the contribution of the j-th element of the amino acid.

9. The method according to claim 2, characterized in that, In S4, PDB2PPQR is used to convert the PDB file obtained by predicting the three-dimensional structure of the protein sequence using AlphaFold or ESMFold into a PQR file. The amino acid microenvironment information around the key amino acid sites is obtained from the PQR file and converted into point cloud data storage, with the central amino acid as the data tag.

10. The method according to claim 2, characterized in that, In S4, the amino acid microenvironment includes the atomic microenvironment, atomic charge, and atomic solvent accessible surface area. BioPython is used to sample atomic charge, Gaussian and Multiwfn are used to calculate atomic charge, and FreeSASA is used to calculate atomic solvent accessible surface area. The optimal amino acid prediction platform APCNet is constructed using the intermediate encoding layer and classification of the PointMLP framework, with the number of point cloud inputs ranging from 256 to 1536. L2 regularization is introduced to prevent overfitting.

Images