A risk prediction model of chikungunya virus mutation sites based on deep mutation scanning data

Abstract

The application discloses a Chikungunya Virus (CHIKV) mutation site risk prediction model, and belongs to the cross field of artificial intelligence and bioinformatics. The model takes an amino acid feature vector of a CHIKV protein sequence as input, cooperatively works through a local genome feature extraction module (LG module) and a multi-task prediction module (MT module), and simultaneously outputs two key biological indexes of a binding affinity change of a mutation site and an antibody escape probability. The application is specially trained and optimized according to CHIKV deep mutation scanning (DMS) experimental data, and the 5-fold cross-validation AUC reaches 0.869±0.029, the F1-Score reaches 0.801, and the Pearson correlation coefficient reaches 0.851 on the binding affinity prediction task, which is significantly better than a general virus mutation prediction framework without CHIKV special training. The application can be used for rapid screening of CHIKV high-risk mutation sites and biological threat early warning, and has important public health and national defense application values.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and bioinformatics, specifically relating to a risk prediction model for mutation sites of Chikungunya Virus (CHIKV). Background Technology

[0002] Chikungunya virus (CHIKV) is an RNA alphavirus transmitted by Aedes mosquitoes that causes chikungunya fever. The main symptoms are acute fever and severe joint pain; some patients may develop chronic arthritis, resulting in a high rate of disability. Since 2004, CHIKV has caused multiple large-scale outbreaks in Africa, Asia, Europe, and the Americas, infecting millions of people. It is listed as an important emerging infectious disease pathogen and also poses a potential biological threat.

[0003] The CHIKV genome is a single-stranded positive-sense RNA, approximately 11.8 kb in length, encoding four non-structural proteins (nsP1-nsP4) and five structural proteins (C, E3, E2, 6K, and E1). Among these, the E1 and E2 envelope proteins are key functional proteins for viral receptor binding and membrane fusion with host cells, and are also major targets for neutralizing antibodies. The E1-A226V mutation is a crucial adaptive mutation for CHIKV to Aedes albopictus mosquitoes, closely related to the 2005-2006 Indian Ocean outbreak, and serves as an important marker site for viral evolution monitoring.

[0004] Existing general-purpose viral mutation prediction frameworks (such as E2VD, Nie et al., Nature Machine Intelligence, 2024) are mainly trained and validated for SARS-CoV-2. Their pre-trained models do not include prior biological knowledge of CHIKV. To predict CHIKV mutation effects, it is necessary to collect CHIKV DMS data again and train from scratch, lacking out-of-the-box CHIKV prediction capabilities. This invention fills this gap, providing a ready-to-use prediction model trained specifically on CHIKV DMS data, with an AUC of 0.869±0.029 (5-fold cross-validation), an F1-Score of 0.801, and a Pearson correlation coefficient of 0.851. The main limitations of existing methods include: (1) Traditional bioinformatics methods rely on known structural information and cannot handle newly emerging variations; (2) General protein language models are not optimized for viral evolutionary characteristics and have limited prediction accuracy; (3) Existing methods can usually only predict a single biological indicator and cannot simultaneously assess the combined effects of mutations on binding affinity, expression level and immune escape; (4) DMS experiments are costly and time-consuming and cannot meet the needs of rapid emergency response. Summary of the Invention

[0005] The purpose of this invention is to provide a risk prediction model for chikungunya virus mutation sites, which solves the technical problems of insufficient prediction accuracy, limited functionality, and inability to quickly respond to emerging mutations in existing methods.

[0006] This invention provides a risk prediction model for chikungunya virus mutation sites, characterized in that the model comprises:

[0007] Feature extraction layer: Receives the CHIKV protein mutant sequence and encodes the amino acid sequence into a feature vector matrix of dimension 1024 with shape [N, L, 1024], where N is the number of samples and L is the sequence length;

[0008] The Local Genome Feature Enhancement Module (LG Module) consists of a three-layer one-dimensional convolutional network with a kernel size of 3 and a stride of 1. It adaptively aggregates the local sequence context at each amino acid position through a local normal distribution weighted context pooling mechanism, and outputs an enhanced position-specific feature matrix.

[0009] The multi-task prediction module (MT module) includes a shared feature representation layer and two independent prediction heads: the classification prediction head outputs the antibody escape probability of mutation sites, using Focal Loss to handle class imbalance; the regression prediction head outputs the binding affinity change value and the protein expression level change value, using weighted mean squared error loss.

[0010] The aforementioned feature extraction layer employs a BERT architecture protein language model pre-trained on a large-scale viral protein sequence corpus. The language model contains 24 Transformer hidden layers with a hidden layer dimension of 1024, 16 attention heads, and an intermediate layer dimension of 4096. The vocabulary includes 20 standard amino acids and 5 special markers: [PAD], [UNK], [CLS], [SEP], and [MASK].

[0011] The context pooling mechanism in the aforementioned local genomic feature enhancement module is implemented through the following steps:

[0012] a. Perform three-layer one-dimensional convolution on the input feature matrix, and output the number of channels as 2048, 2048 and 2 respectively to obtain the positional saliency score s and positional weight w;

[0013] b. Perform Softmax normalization on s, and construct a local normal distribution with each location as the center and s as the standard deviation parameter;

[0014] c. Multiply the local normal distribution by the weight w to perform weighted aggregation of neighborhood features and output the enhanced feature vector at each position.

[0015] The training data mentioned above are deep mutation scanning (DMS) experimental data of CHIKV proteins, including measured biological phenotypic data of single-site and multi-site mutations of CHIKV structural proteins E1, E2, and C non-structural proteins nsP1-nsP4. A 5-fold cross-validation strategy was used during training.

[0016] The model also includes an evolutionary trend perception function: by comprehensively scoring the predicted binding affinity, expression level and escape probability, it identifies rare beneficial mutations with high adaptive advantages and outputs a ranking list of high-risk mutation sites.

[0017] Furthermore, this invention also suggests the application of the above-mentioned prediction model in the screening of CHIKV vaccine targets.

[0018] Furthermore, this invention also suggests the application of the above-mentioned prediction model in the evaluation of antiviral drug targets.

[0019] Furthermore, this invention also suggests the application of the aforementioned prediction model in early warning of biological threats. Attached Figure Description

[0020] Figure 1 : Schematic diagram of the overall architecture of the prediction model of this invention

[0021] Figure 2 Schematic diagram of the Local Genome Feature Enhancement Module (LG module)

[0022] Figure 3 Schematic diagram of the multi-task prediction module (MT module)

[0023] Figure 4 CHIKV binding affinity prediction 5-fold cross-validation ROC curve

[0024] Figure 5 Visualization of CHIKV high-risk mutation site prediction results Detailed Implementation

[0025] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments, further clarifies the invention. Those skilled in the art should understand that the specific descriptions below are illustrative rather than restrictive, and should not be construed as limiting the scope of protection of the present invention.

[0026] All salinomycin used in the embodiments of this invention can be purchased commercially.

[0027] Example 1. Model Framework

[0028] The overall architecture of the CHIKV variant site risk prediction model provided by this invention is as follows: Figure 1 As shown, it is divided into three levels:

[0029] 1. Protein Language Model Feature Extraction Layer

[0030] A BERT-based protein language model, pre-trained on a large-scale viral protein sequence corpus, is used to transform amino acid sequences into high-dimensional semantic feature vectors. The language model contains 24 Transformer hidden layers with a hidden layer dimension of 1024, 16 attention heads, and an intermediate layer dimension of 4096. The vocabulary covers 20 standard amino acids and 5 special markers ([PAD], [UNK], [CLS], [SEP], [MASK]). The pre-training task is Masked Language Modeling (MLM), with a masking ratio of 15% and a maximum sequence length of 256 amino acids. For an input CHIKV mutant protein sequence, after encoding by the above language model, the output is a feature matrix of shape [N, 201, 1024].

[0031] 2. Local Genome Feature Enhancement Module (LG Module)

[0032] LG module structure as follows Figure 2 As shown, the core innovation lies in the local context pooling mechanism, which can adaptively capture the local sequence context information of each amino acid position.

[0033] The specific implementation steps are as follows:

[0034] Step 1: Perform three-layer one-dimensional convolution on the input feature matrix. The kernel size is 3, the stride is 1, the padding is 1, and the number of channels changes sequentially from [1024→2048→2048→2]. Each convolution is followed by LayerNorm normalization and LeakyReLU activation function.

[0035] Step 2: Interpret the 2-channel features output by the convolution as positional saliency score s (channel 0) and positional weight w (channel 1), respectively, and normalize them using Softmax.

[0036] Step 3: Using each position i in the sequence as the center and s[i] as the standard deviation parameter, construct a local normal distribution kernel function:

[0037] K_i(j) = (1 / √(2π)σ_i) × exp(-(ji)² / (2σ_i²)), where σ_i = r × L × s_i, L is the sequence length, and r is a hyperparameter (default 0.1).

[0038] Step 4: Multiply the local normal distribution kernel with the weight w element by element and normalize it. Then, perform weighted aggregation on the neighborhood features and output the enhanced feature vector at each position. The final output shape is still [N, L, 1024].

[0039] 3. Multi-task prediction module (MT module)

[0040] MT module structure as follows Figure 3 As shown, based on shared feature representations, classification and regression tasks are completed simultaneously using two independent prediction heads:

[0041] The classification prediction head uses the Focal Loss function to handle the imbalance between positive and negative samples:

[0042] L_cls = -α_t × (1-p_t)^γ × log(p_t), where α_t is the class weight and γ is the focusing parameter (default 2.0).

[0043] The regression prediction head uses a weighted mean squared error loss, assigning greater weight to high-value samples:

[0044] L_reg = Σ w_i × (y_i - _i)², where w_i ∝ |y_i|

[0045] The total loss function is:

[0046] L = L_cls + k × L_reg, where k is the regression loss weight hyperparameter.

[0047] 4. Specific training for CHIKV

[0048] The training data covers the experimental biological phenotypic data of single-site and multi-site mutations in CHIKV structural proteins (E1, E2, capsid protein C) and non-structural proteins (nsP1-nsP4), totaling 7713 samples (721 positive samples and 6992 negative samples). The training strategy employs 5-fold cross-validation, with each fold trained independently, and the final ensemble prediction results are used.

[0049]

Claims

1. A risk prediction model for chikungunya virus mutation sites, characterized in that, The model includes: Feature extraction layer: Receives the CHIKV protein mutant sequence and encodes the amino acid sequence into a feature vector matrix of dimension 1024 with shape [N, L, 1024], where N is the number of samples and L is the sequence length; The Local Genome Feature Enhancement Module (LG Module) consists of a three-layer one-dimensional convolutional network with a kernel size of 3 and a stride of 1. It adaptively aggregates the local sequence context at each amino acid position through a local normal distribution weighted context pooling mechanism, and outputs an enhanced position-specific feature matrix. The multi-task prediction module (MT module) includes a shared feature representation layer and two independent prediction heads: the classification prediction head outputs the antibody escape probability of mutation sites, using Focal Loss to handle class imbalance; the regression prediction head outputs the binding affinity change value and the protein expression level change value, using weighted mean squared error loss.

2. The prediction model according to claim 1, characterized in that, The feature extraction layer uses a BERT architecture protein language model pre-trained on a large-scale viral protein sequence corpus. The language model contains 24 Transformer hidden layers with a hidden layer dimension of 1024, 16 attention heads, and an intermediate layer dimension of 4096. The vocabulary contains 20 standard amino acids and 5 special markers: [PAD], [UNK], [CLS], [SEP], and [MASK].

3. The prediction model according to claim 1, characterized in that, The context pooling mechanism in the local genome feature enhancement module is implemented through the following steps: a. Perform three-layer one-dimensional convolution on the input feature matrix, and output the number of channels as 2048, 2048 and 2 respectively to obtain the positional saliency score s and positional weight w; b. Perform Softmax normalization on s, and construct a local normal distribution with each location as the center and s as the standard deviation parameter; c. Multiply the local normal distribution by the weight w to perform weighted aggregation of neighborhood features and output the enhanced feature vector at each position.

4. The prediction model according to claim 1, characterized in that, The training data is CHIKV protein deep mutation scanning (DMS) experimental data, which includes measured biological phenotypic data of single-site and multi-site mutations in CHIKV structural proteins E1, E2, and C non-structural proteins nsP1-nsP4. A 5-fold cross-validation strategy was used during training.

5. The prediction model according to claim 1, characterized in that, The model also includes an evolutionary trend perception function: by comprehensively scoring the predicted binding affinity, expression level and escape probability, it identifies rare beneficial mutations with high adaptive advantages and outputs a ranking list of high-risk mutation sites.

6. The application of the prediction model according to claim 1 in the screening of CHIKV vaccine targets.

7. The application of the prediction model according to claim 1 in the evaluation of antiviral drug targets.

8. The application of the prediction model according to claim 1 in early warning of biological threats.

Images