A protein language model and hyperbolic space embedding-based unaligned phylogenetic inference method and system

Abstract

The application relates to a construction method of a no-alignment phylogenetic inference model based on a protein language model and hyperbolic space embedding, and a no-alignment phylogenetic inference model is constructed based on the construction method. After global target sequences of a query sequence are input into the no-alignment phylogenetic inference model, cross-scale feature representations of fused local conservative signals of the global target sequences can be automatically extracted, and the cross-scale feature representations can be accurately mapped to a parameterized hyperbolic metric space, and phylogenetic analysis can be realized by combining the hyperbolic geodesic distance between the query sequence and the reference sequence. The application can realize direct inference of the global target sequence, does not require manual pre-interception of local functional domain sequences, and can break through the analysis limitation of high-variation sequences in phylogenetic analysis, can effectively extract deep evolutionary semantics from dark matter sequences with low homology, and realizes accurate phylogenetic positioning of distant and high-variation sequences.

Description

Technical Field

[0001] This application relates to the fields of molecular biology and artificial intelligence, and in particular to a method and system for inferring unpaired phylogenetic relationships based on protein language models and hyperbolic space embedding. Background Technology

[0002] In virology and biological evolution research, inferring phylogenetic relationships between sequences (i.e., phylogenetic tree construction) is a core task. The current "gold standard" for molecular phylogenetic inference usually relies on constructing multiple sequence alignments (MSA) and then using maximum likelihood (ML) or Bayesian methods to reconstruct the phylogenetic tree.

[0003] The existing "gold standard" usually follows the following process: (1) obtain multiple homologous sequences; (2) perform multiple sequence alignment using software such as MAFFT; (3) reconstruct the phylogenetic tree using the maximum likelihood method or Bayesian method based on the results of multiple sequence alignment (such as software such as IQ-TREE and RAxML); (4) for the insertion of new sequences (phylogenetic localization), it usually relies on tools such as pplacer or EPA-ng, which require that the query sequence be aligned to the reference skeleton before insertion.

[0004] Phylogenetic inference methods following the above process face the following insurmountable technical bottlenecks: (1) High dependence on sequence homology, making it difficult to handle highly variable distantly related sequences: Existing multiple sequence alignment methods have a hard premise, that is, there must be sufficient similarity between sequences for effective alignment. However, in many biological systems (especially high-frequency variable species such as RNA viruses), many newly discovered sequences are very different from known sequences (in the "twilight zone", and the sequence similarity is usually less than 30%), making it impossible to construct effective alignments. This forces existing tools to discard these "dark matter" sequences with high research value; (2) Computational complexity increases exponentially with the scale of data: With the explosive growth of sequencing data, the computational cost of constructing large-scale MSAs is extremely high and time-consuming, which seriously restricts the efficiency of high-throughput phylogenetic analysis; (3) Geometric distortion of Euclidean space limits the restoration of topological structure: Early attempts to use conventional deep learning to extract sequence features usually map the features to Euclidean space. However, phylogenetic trees are essentially hierarchical branching structures, and the number of nodes increases exponentially with depth, while the capacity of Euclidean space only increases polynomially. Forcibly embedding a tree structure into Euclidean space will cause serious geometric distortion, causing the inferred evolutionary distance to deviate from the actual evolutionary trajectory; (4) lack of semantic understanding of deep evolutionary constraints: existing alignment-free methods, such as K-mer frequency method or shallow learning model, usually have difficulty capturing complex, long-distance evolutionary constraints and functional patterns in biological sequences.

[0005] Therefore, there is an urgent need for a novel phylogenetic inference architecture that not only breaks free from the dependence on explicit sequence alignment to handle low-homology biomolecule sequences (especially sequences with extremely high genomic variation and very low homology, such as RNA viruses), but also captures deep evolutionary semantics and achieves accurate phylogenetic localization in a geometric space that conforms to the laws of natural evolution, while maintaining excellent topological reconstruction and computational scalability. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a method and system for inferring unpaired phylogenetic relationships based on protein language models and hyperbolic space embedding.

[0007] The primary objective of this invention is to provide a method for constructing a unique phylogenetic inference model based on a protein language model and hyperbolic space embedding.

[0008] The second objective of this invention is to provide a peerless phylogenetic inference model based on a protein language model and hyperbolic space embedding, constructed using the above-described method.

[0009] The third objective of this invention is to provide a method for inferring unpaired phylogenetic relationships based on protein language models and hyperbolic space embedding.

[0010] The fourth objective of this invention is to provide a peer-to-peer phylogenetic inference system based on a protein language model and hyperbolic space embedding.

[0011] The fifth objective of this invention is to provide the application of the above-mentioned unpaired phylogenetic inference model, the above-mentioned unpaired phylogenetic inference method, and / or the above-mentioned unpaired phylogenetic inference system in phylogenetic research.

[0012] To achieve the above objectives, the present invention is implemented through the following solution: This invention claims protection for a method for constructing a unique pairwise phylogenetic inference model based on a protein language model and hyperbolic space embedding, comprising the following steps: S1. Obtain a training set containing global target sequences, determine the taxonomic group label of each global target sequence in the training set, and extract the local key functional region sequences containing evolutionary conserved information from each global target sequence. For the training set, a reference phylogenetic tree representing the true evolutionary distance between sequences is constructed under each taxonomic group; S2. Using each global target sequence in the training set as input to the biological sequence pre-trained large language model, the model is adaptively trained through the masked language modeling task to obtain the Stage-1 model; The Stage-1 model outputs an initial high-dimensional embedding vector 1 for each global target sequence; S3. Using each global target sequence in the training set as input to the Stage-1 model obtained in step S2, and combining the class group label of each global target sequence in the training set in step S1, construct positive sample pairs with the same class group label and negative sample pairs with different class group labels in the embedding vector feature space. Optimize the Stage-1 model through a contrastive learning strategy to obtain a global feature representation containing coarse-grained classification structure information, and obtain the Stage-2 model. The Stage-2 model outputs a high-dimensional embedding vector 2 of the global target sequence; the high-dimensional embedding vector 2 maps the classification group information corresponding to the global target sequence; S4. Using the global target sequence and its corresponding local key functional region sequence in the training set as input to the Stage-2 model obtained in step S3, extract the global embedding vector of the global target sequence and the local embedding vector of the local key functional region sequence, and train the Stage-2 model to obtain the Stage-3 model. The Stage-3 model outputs a high-dimensional embedding vector 3 for the global target sequence; the high-dimensional embedding vector 3 maps the classification group information of the global target sequence and integrates the cross-scale comprehensive feature representation of the global embedding vector and the local embedding vector. During the training of the Stage-2 model, each global target sequence and its corresponding local key functional region sequence in the training set described in step S1 are used as the input of the Stage-2 model. The global target sequences are encoded using the Stage-2 model, and the global embedding vector Z corresponding to each global target sequence is extracted. poly，i Simultaneously, the corresponding local key functional region sequence is encoded using the Stage-2 model, and the corresponding local embedding vector Z is extracted. dom，i ; Then, by minimizing the global embedding vector Z poly，i With the corresponding local embedding vector Z dom，i The joint loss function between the two is used to train and optimize the Stage-2 model parameters; The joint loss function is a weighted sum of the cross-modal InfoNCE loss function, the orthogonal regularization penalty term, and the center separation penalty term; The calculation method for the center separation penalty term is shown in Formula I: Formula I: ; Where sim(·) represents the cosine similarity function; It is the mean vector of all global embedding vectors within the same training batch; It is the mean vector of all local embedding vectors within the same training batch; S5. Using each global target sequence in the training set as input to the Stage-2 model obtained in step S3 or the Stage-3 model obtained in step S4, extract the Euclidean space feature representation of the output. The Euclidean space feature representation is projected onto a learning curvature parameter through a geometric mapping function. k In the hyperbolic manifold space, obtain the hyperbolic coordinates of each global target sequence in that space; Based on the hyperbolic coordinates of each global target sequence in this space, the hyperbolic geodesic distance between every two global target sequences is calculated. At the same time, the evolutionary path distance between the corresponding global target sequences in the reference phylogenetic tree constructed in step S1 is extracted as a supervision signal for model training. During training and optimization, a loss function is constructed with the regression objective of minimizing the error between the hyperbolic geodesic distance and the true evolutionary distance. The backbone parameters of the network are frozen, and the parameters in the Stage-2 or Stage-3 model are updated using a Parametric Efficient Fine-Tuning (PEFT) strategy, along with the curvature parameters of the hyperbolic manifold space. k We obtained a unique phylogenetic inference model based on protein language model and hyperbolic space embedding; The model outputs a high-dimensional embedding vector 4 for each global target sequence in the training set, projects the high-dimensional embedding vector 4 of each global target sequence in the training set into hyperbolic space, and records the hyperbolic coordinates of each global target sequence in the training set in hyperbolic space.

[0013] Preferably, the global target sequence in step S1 includes a protein sequence, a deoxyribonucleic acid sequence, and / or a ribonucleic acid sequence.

[0014] More preferably, the global target sequence is a ribonucleic acid sequence.

[0015] More preferably, the ribonucleic acid sequence is data from 180 viral superclusters identified by LucaProt, as described in the prior art: "Using artificial intelligence to document the hidden RNAvirosphere" (https: / / doi.org / 10.1016 / j.cell.2024.09.027).

[0016] Preferably, the local key functional region sequence in step S1 is a marker gene and / or a housekeeping gene.

[0017] Preferably, the method for extracting local key functional region sequences containing evolutionarily conserved information in each global target sequence in step S1 is an extraction method based on bioinformatics homology or an AI method based on sequence semantics.

[0018] Preferably, the taxonomic group category label in step S1 is a biological taxonomic level category label.

[0019] More preferably, the biological classification level label is a label for kingdom, phylum, class, order, family, genus, species, or other taxonomic groups.

[0020] Preferably, the construction of the reference phylogenetic tree in step S1 specifically includes the following steps: S11. Extract local key functional region sequences from each global target sequence in the training set, and perform multi-sequence alignment to obtain the multi-sequence alignment results; S12. Based on the multiple sequence alignment results obtained in step S11, a reference phylogenetic tree is constructed using the maximum likelihood method, and the path distances between nodes in the tree are extracted to construct a reference evolutionary distance matrix.

[0021] More preferably, the multiple sequence alignment in step S11 is performed using MAFFT or MUSCLE.

[0022] More preferably, MAFFT is used for multiple sequence alignment; wherein the version of MAFFT is v7.525 and the parameters are set to --maxiterate 1000 --localpair.

[0023] More preferably, before constructing the reference phylogenetic tree in step S12, the multiple sequence alignment results need to be pruned using trimAl.

[0024] More preferably, the version of trimAl is v1.5, and the parameters are set to -gt 0.9 -cons 5.

[0025] More preferably, step S12 uses IQ-TREE, RAxML, FastTree, or PhyML to construct a reference phylogenetic tree.

[0026] More preferably, the maximum likelihood tree is constructed using IQ-TREE; the version number of IQ-TREE is v2.3.6, and the parameters are set to -m LG --alrt 1000 -B 1000.

[0027] Preferably, the biological sequence pre-trained large language model in step S2 includes a protein language model, a nucleic acid language model, or a unified biological large model that simultaneously supports nucleic acid and protein representation.

[0028] More preferably, the biological sequence language model is a protein language model.

[0029] More preferably, the protein language model is the ESM-2 model.

[0030] More preferably, the ESM-2 model is an ESM-2 model with 650 million (650M) parameters.

[0031] Preferably, in step S2, a masking language modeling task is used to mask 5-50% of the amino acids in the full-length polyprotein sequence.

[0032] More preferably, a masking language modeling task is used to mask 15% of the amino acids in the full-length polyprotein sequence.

[0033] Preferably, the optimization of the Stage-1 model through the contrastive learning strategy in step S3 is as follows: optimization is performed using the InfoNCE loss function, which is shown in Formula II; Formula II: ; Where N is the number of sequences within the same training batch; sim(·) is the cosine similarity; τ is the temperature parameter; (z 1,i , z 2，i () represents a positive sample pair consisting of the high-dimensional feature representations of two independent sequences sampled from the same taxonomic group, while the feature representations of other unpaired sequences in the global batch are considered as negative samples.

[0034] In step S3, during the training of the Stage-1 model obtained in step S2, global target sequences belonging to the same classification group (i.e., global target sequences with the same classification group category label) are regarded as positive samples, and the distance between positive samples in the feature space is reduced during the training process; global target sequences belonging to different classification groups (i.e., global target sequences with different classification group category labels) are regarded as negative samples, and the distance between negative samples in the feature space is increased.

[0035] Preferably, the calculation method of the cross-modal InfoNCE loss function in step S4 is as shown in Formula III; Formula III:

[0036] Where N is the number of sequences in the same training batch; sim(·) represents the cosine similarity function; τ is the temperature parameter; Z poly，i Z is the global embedding vector of the i-th global target sequence within this training batch; dom，i Z is the local embedding vector of the i-th global target sequence within this training batch; dom，j This is the local embedding vector of the j-th global target sequence within this training batch.

[0037] Preferably, the calculation method for the orthogonal regularization penalty term in step S4 is as shown in Formula IV; Formula IV: ; Among them, Z poly For all global embedding vectors Z within this training batch poly，i The global feature matrix composed of Z and subjected to L2 normalization; dom For all local embedding vectors Z within the same training batch dom，i The local feature matrix composed of and L2-normalized features; I is the identity matrix; ||·|| F This represents the Frobenius norm.

[0038] Preferably, the calculation method of the joint loss function in step S4 is as shown in formula V; Formula V: ; Where, λ ortho and λ sep This is a hyperparameter.

[0039] In step S4, when training and optimizing the Stage-2 model, the cosine similarity between the global embedding vector and the corresponding local embedding vector is optimized, and orthogonal regularization is added to maintain feature diversity. At the same time, the center separation penalty is combined to maintain the distribution characteristics between modes. The Stage-3 model obtained in step S4 establishes the semantic relationship between the global target sequence and its corresponding local key functional region sequence, so that the Stage-3 model can automatically locate the local key functional region sequence (key evolutionary signal) from the global target sequence.

[0040] Preferably, the hyperbolic manifold space mentioned in step S5 is a Lorentz manifold, a Poincaré sphere model, a Poincaré half-plane model, or a Klein model.

[0041] More preferably, the hyperbolic manifold space is a Lorentz manifold.

[0042] Preferably, when the hyperbolic manifold space in step S5 is a Lorentz manifold, the specific process of projecting the Euclidean space feature representation onto the hyperbolic manifold space with a learnable curvature parameter k through a geometric mapping function is as follows: The Euclidean space feature representation is mapped to a (d+1)-dimensional coordinate space through a learnable linear transformation, while preserving the spatial components. Calculate the time component The Lorentz hyperbolic feature representation is obtained. ; Where d is the dimension of the feature representation in Euclidean space; The two features in the Lorentz hyperbolic manifold space representing the geometric relationship between x and y are quantized by the Minkowski inner product: .

[0043] Preferably, in step S5, a parameter-efficient fine-tuning strategy is used to update the parameters in the Stage-2 or Stage-3 model, specifically as follows: The parameters of the network backbone attention layer of the Stage-2 or Stage-3 model are frozen, and the parameters of the attention mechanism module at a specific level in the Stage-2 or Stage-3 model are fine-tuned. The curvature parameter k of the hyperbolic manifold space is also updated.

[0044] Preferably, the efficient parameter fine-tuning strategy in step S5 includes: low-rank adaptive fine-tuning (LoRA), adapter fine-tuning, prefix fine-tuning, internal activation injection fine-tuning (IA3), or bias term fine-tuning (BitFit).

[0045] More preferably, the parameter efficient fine-tuning strategy includes low-rank adaptive fine-tuning.

[0046] This invention also claims protection for the unpaired phylogenetic inference model based on protein language model and hyperbolic space embedding constructed by any of the above-described construction methods.

[0047] This invention also claims protection for a method for inferring unpaired phylogenetic relationships based on protein language models and hyperbolic space embeddings, comprising the following steps: S111. Obtain ≥1 query sequence; the query sequence is derived from the taxonomic group contained in the global target sequence in the training set of any of the above construction methods; S112. Input the query sequence obtained in step S111 into the above-mentioned unpaired phylogenetic inference model. The unpaired phylogenetic inference model outputs the high-dimensional embedding vector 4 of the query sequence and projects it into hyperbolic space to obtain the hyperbolic coordinates of the query sequence. S113. Using each global target sequence or the key functional region sequence of each global target sequence in the training set of the above-described construction method as a reference sequence, based on the hyperbolic coordinates of the query sequence obtained in step S112 and the hyperbolic coordinates of each global target sequence in the training set shown in step S5 of the above-described construction method in hyperbolic space, calculate the hyperbolic geodesic distance between the query sequence and each reference sequence and the hyperbolic geodesic distance between each query sequence, and combine the maximum likelihood tree obtained in step S1 of the above-described construction method to obtain the evolutionary distance between each reference sequence; S114. Based on the hyperbolic geodesic distance between the query sequence and each reference sequence, the hyperbolic geodesic distance between each query sequence, and the evolutionary distance between each reference sequence obtained in step S113, a phylogenetic tree containing the query sequence is constructed by combining a distance-based phylogenetic tree construction algorithm.

[0048] Preferably, before inputting the query sequence into the above-mentioned unpaired phylogenetic inference model in step S112, the query sequence can be input into the Stage-2 model obtained in step S3 of the above construction method to obtain the high-dimensional embedding vector 1 of the query sequence, and the taxonomic group of the query sequence can be determined by combining the high-dimensional embedding vector 1 of the query sequence.

[0049] Preferably, the distance-based phylogenetic tree construction method in step S114 is any one of the following: minimum evolution method, adjacency method, BIONJ method, unweighted pairwise arithmetic mean method, and / or Fitch-Margoliash method.

[0050] This invention also claims protection for a peerless phylogenetic inference system based on protein language model and hyperbolic space embedding, including a data acquisition module, a processing module, a phylogenetic tree construction module and a result output module; The data acquisition module is used to acquire the query sequence; the query sequence is derived from the classification group to which the global target sequence belongs in the training set of the above construction method; The processing module is the aforementioned unpaired phylogenetic inference model, which takes the query sequence obtained by the data acquisition module as input to obtain the coordinate vector of the query sequence in hyperbolic space. The phylogenetic tree construction module takes the coordinate vector of the sequence to be analyzed in hyperbolic space obtained by the processing module as input, and processes it according to steps S113 and S114 in the above-mentioned unparalleled phylogenetic inference method to obtain a phylogenetic tree containing the query sequence. The result output module is used to output the phylogenetic tree containing the query sequence obtained by the evolutionary tree construction module.

[0051] This invention also claims protection for the application of the above-described unpaired phylogenetic inference model, the above-described unpaired phylogenetic inference method, and / or the above-described unpaired phylogenetic inference system in phylogenetic research.

[0052] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method for constructing a peer-to-peer phylogenetic inference model based on a protein language model and hyperbolic space embedding. A peer-to-peer phylogenetic inference model is constructed based on this method. After inputting the global target sequence of the query sequence into the peer-to-peer phylogenetic inference model, the cross-scale feature representation of the fusion local conserved signal of the global target sequence (including global sequence features and local key functional region sequence features) can be automatically extracted and accurately mapped to a parameterized hyperbolic metric space. Phylogenetic analysis can then be achieved by combining the hyperbolic geodesic distance between the query sequence and the reference sequence. This invention enables direct inference of global target sequences without the need for manual pre-extraction of local functional domain sequences. Furthermore, it overcomes the limitations of high-variability sequence analysis in phylogenetic analysis, effectively extracting deep evolutionary semantics from low-homology dark matter sequences, and achieving precise phylogenetic localization of distantly related, highly variable sequences. Moreover, leveraging a hyperbolic metric space that fits the exponential hierarchical branching structure of phylogenetic trees and a biological sequence language model, this invention possesses extremely high scalability for phylogenetic analysis of large-scale data. It not only provides crucial technical support for the diversity and evolutionary research of highly variable species (such as RNA viruses) but also offers a universal core computational foundation for analyzing the evolutionary relationships of other biological systems. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the training process of an unpaired phylogenetic inference model based on a protein language model and hyperbolic space embedding, as described in one embodiment. Detailed Implementation

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Unless otherwise specified, the experimental methods used in the following embodiments are conventional methods; the materials and reagents used, unless otherwise specified, are commercially available.

[0055] Example 1: Construction of a unique pair phylogenetic inference model based on protein language model and hyperbolic space embedding This embodiment details the training process of the peer-to-peer phylogenetic inference model. A schematic diagram of the training process is shown below. Figure 1 As shown, the dataset used in this embodiment is derived from publicly available global RNAvirosphere research resources (hereinafter referred to as RNAvirosphere data): Using artificial intelligence to document the hidden RNAvirosphere, https: / / doi.org / 10.1016 / j.cell.2024.09.027.

[0056] I. Training of Phylogenetic Inference Models The specific training process for the phylogenetic inference model is as follows: S1. Dataset Preparation: The datasets for Phases 1 through 3: Representative data of 180 ICTV-identified and newly defined viral supergroups (taxonomic group category labels) were obtained from the RNA Virus Circle Data (recorded in the RNA Virus Circle Data), containing 22,063 pairs of full-length polyprotein sequences and their corresponding RNA-dependent RNA polymerase (RdRP) domain sequences (local key functional region sequences), which were divided into training and validation sets in an 8:2 ratio.

[0057] Phase 4 Tree-Specific Metric Learning Dataset: In the RNA virus sphere data, superclusters with ≥100 sequences within a viral supergroup are designated as specific superclusters. Full-length polyprotein sequences and their corresponding RdRP domain sequences within specific superclusters are selected. Multiple sequence alignment of the RdRP domain sequences is performed within each supergroup using MAFFT (v7.525, --maxiterate 1000 --localpair). Trimming is performed using trimAl (v1.5, -gt 0.9 -cons 5). Finally, a maximum likelihood (ML) reference tree is constructed using IQ-TREE (multicore v2.3.6, -m LG --alrt 1000 -B 1000). The evolutionary path distance between nodes in the tree is extracted, and a distance matrix is ​​constructed. Stratified sampling is performed based on evolutionary distance at a 9:1 ratio to obtain training set 1 and validation set 1.

[0058] S2. Adaptive Training of Masked Language Modeling (MLM) (Phase 1): Using the ESM-2 protein language model with an initial parameter count of 650 million (650M) as the baseline model, and the full-length polyprotein sequence of the training set in step S1 as the input, 15% of the amino acid tokens were randomly masked through the masked language modeling task, and the standard MLM loss function was minimized; the learning rate was set to 1e-4, and the model was trained for 25 batches (epochs) to obtain the Stage-1 model. The Stage-1 model outputs the initial high-dimensional embedding vector 1 for each full-length polyprotein sequence; S3. Contrastive learning of super-sensor (Phase Two): Using the full-length polyprotein from the training set in step S1 as input to the Stage-1 model obtained in step S2, and combining the taxonomic group labels (viral supergroups) of the full-length polyprotein, positive sample pairs with the same taxonomic group label and negative sample pairs with different taxonomic group labels are constructed in the embedding vector feature space. Training is performed by minimizing the InfoNCE loss function to obtain a global feature representation containing coarse-grained classification structure information, thus obtaining the Stage-2 model; the temperature parameter during the training process... Set the learning rate to 0.2 and the learning rate to 1e-6, and train for 100 batches (epochs). The InfoNCE loss function is shown in Equation II; Formula II: ; Where N is the number of sequences in the same training batch; sim(·) is the cosine similarity; τ is the temperature parameter; (z 1，i , z 2，i (z) represents a positive sample pair; 1，i , z 2，j () represents a negative sample pair; The Stage-2 model takes a high-dimensional embedding vector 1 of the full-length polyprotein sequence as input; the high-dimensional embedding vector 1 maps the taxonomic group information corresponding to the full-length polyprotein sequence. S4. Multi-scale cross-modal joint optimization (Phase 3): Using the full-length polyprotein sequence and its corresponding local key functional region sequence (RdRP domain sequence) of the training set in step S1 as input to the Stage-2 model obtained in step S3, the global embedding vector of the full-length polyprotein and the local embedding vector of the local key functional region sequence are extracted, and the Stage-2 model is trained to obtain the Stage-3 model. The Stage-3 model outputs a high-dimensional embedding vector 3 for full-length polyprotein sequences. The high-dimensional embedding vector 3 maps the taxonomic group information of the full-length polyprotein sequence and integrates cross-scale comprehensive feature representations that combine global and local features. The specific process for training the Stage-2 model is as follows: the full-length polyprotein sequences and their corresponding local key functional region sequences (RdRP domain sequences) from the training set in step S1 are used as input to the Stage-2 model. After the full-length polyprotein sequences are encoded by the Stage-2 model, the global embedding vector Z corresponding to each full-length polyprotein sequence is extracted. poly，i Simultaneously, the RdRP domain sequence corresponding to the full-length polyprotein sequence is encoded using the Stage-2 model, and the corresponding local embedding vector Z is extracted. dom，i ; Then, by minimizing the global embedding vector Z shown... poly，iand the corresponding local embedding vector Z dom，i The joint loss function between the Stage-2 model and the Stage-3 model were used to train and optimize the Stage-2 model parameters to obtain the Stage-3 model. The temperature parameter τ was set to 0.07 and the learning rate was set to 5e-5 during the training and optimization process, and the training lasted for 7 epochs. The joint loss function is the cross-modal InfoNCE loss function ( Orthogonal regularization penalty term ( ) and the center separation penalty item ( The weighted summation of () is shown in formula V; Formula V: ; Where λ ortho =0.3; λ sep =2; The calculation method for the cross-modal InfoNCE loss function is shown in Equation III; Formula III: ; Where N is the number of sequences in the same training batch; sim(·) represents the cosine similarity function; τ is the temperature parameter; Z poly，i Z is the global embedding vector of the i-th full-length polyprotein sequence within this training batch; dom，i Z is the local embedding vector of the i-th full-length polyprotein sequence within this training batch; dom，j This is the local embedding vector of the j-th full-length polyprotein sequence within this training batch; The calculation method for the orthogonal regularization penalty term is shown in Formula IV; Formula IV: ; Among them, Z poly For all global embedding vectors Z within this training batch poly，i The global feature matrix composed of Z and subjected to L2 normalization; dom For all local embedding vectors Z within the same training batch dom，i The local feature matrix composed of and L2-normalized features; I is the identity matrix; ||·|| F Denotes the Frobenius norm; The calculation method for the center separation penalty term is shown in Formula I: Formula I: ; Where sim(·) represents the cosine similarity function; It is the mean vector of all global embedding vectors within the same training batch (epochs); It is the mean vector of all local embedding vectors within the same training batch.

[0059] S5. Hyperbolic Metric Space Learning (Stage Four): Using each full-length polyprotein sequence in training set 1 of step S1 as input to the Stage-3 model obtained in step S4, extract the Euclidean space feature representation h of each full-length polyprotein sequence output (dimension d, i.e., the high-dimensional embedding vector 3 output by the Stage-3 model, which is the Euclidean space feature representation by default). The Euclidean space feature representation h of each full-length polyprotein sequence is projected onto a Lorentz hyperbolic manifold space with a learnable curvature parameter k through a geometric mapping function, and the hyperbolic coordinates of each full-length polyprotein in the hyperbolic manifold space are obtained (after Minkowski inner product quantization). The Euclidean space feature representation h is mapped to the (d+1)-dimensional coordinate space v=Wh+b through a learnable linear transformation, while preserving the spatial components. Calculate the time component The Lorentz hyperbolic feature representation is obtained. ; Where d is the dimension of the Euclidean space feature representation; v is the coordinate vector in the (d+1)-dimensional coordinate space; W is the weight matrix of the learnable linear transformation mapping; h is the Euclidean space feature representation (feature vector); and b is the bias vector of the linear transformation mapping. The geometric relationship between two features representing x and y in a hyperbolic manifold space is quantized using the Minkowski inner product: ; Based on the hyperbolic coordinates of each full-length polyprotein in the hyperbolic manifold space, the hyperbolic geodesic distance between every two full-length polyprotein sequences is calculated. At the same time, the evolutionary path distance between corresponding full-length polyprotein sequences in the maximum likelihood reference tree of each full-length polyprotein sequence in step S1 training set 1 is extracted (the evolutionary path distance between each node in the maximum likelihood reference tree) as the training supervision signal of the model. The loss function is constructed with the regression objective of minimizing the mean squared error (MSE) between the hyperbolic geodesic distance between any two full-length polyprotein sequences and the evolutionary path distance between the corresponding full-length polyprotein sequences. The Stage-3 model obtained in step S4 is fine-tuned using a low-rank adaptive (LoRA, rank r=16, α=32) strategy combined with validation set 1, and the curvature parameter k of the Lorenz hyperbolic manifold space is updated to obtain a pairwise phylogenetic inference model based on protein language model and hyperbolic space embedding. During the fine-tuning (training) process, the AdamW optimizer was used with a base learning rate of 5e-5. The learning rate for the curvature parameter was set to 0.1 times the base learning rate. Early stopping was triggered when the loss on validation set 1 did not decrease for 3 consecutive epochs. The model outputs a high-dimensional embedding vector 4 for each global target sequence in the training set, and projects the high-dimensional embedding vector 4 of each full-length polyprotein sequence in the training set 1 onto the Lorentz hyperbolic manifold space, and records the hyperbolic coordinates of each full-length polyprotein sequence in the training set 1 in the Lorentz hyperbolic manifold space.

[0060] II. Validation of the Stage-2 model in the unparalleled phylogenetic inference model (classification performance validation) 1. Experimental Methods The full-length polyprotein sequences containing 615 RNA virus superclusters were used as test subject 1. The full-length polyprotein sequences in test subject 1 have been recognized and classified by the International Committee on Taxonomy of Viruses (ICTV), are derived from RNA virus sphere data (Using artificial intelligence to document the hidden RNA virosphere, https: / / doi.org / 10.1016 / j.cell.2024.09.027), and are independent and non-overlapping with the 22,063 full-length polyprotein sequences used in the training process of the unparalleled phylogenetic inference model (the 22,063 full-length polyprotein sequences are representative data in the RNA virus sphere). The 615 full-length polyprotein sequences are derived from 21 viral superclusters, and these 21 viral superclusters are included in the 180 viral superclusters identified and newly defined by ICTV during the training process of the unparalleled phylogenetic inference model.

[0061] Test subject 1 (615 full-length polyprotein sequences) was used as input to the Stage-2 model obtained in Stage 2, resulting in a high-dimensional embedding vector 1. The t-SNE algorithm was used to project the high-dimensional embedding vector 1 into a two-dimensional space, and K-means clustering algorithm was used to perform unsupervised clustering (the number of clusters was set to 21). The Adjusted Rand Index (ARI) was used to quantify the consistency between the unsupervised clustering and the true classification labels (21 viral superclusters).

[0062] Test object 1 was sequentially input into the ESM2 baseline model (ESM-650M) with 650M parameters and the ESM2 baseline model (ESM-3B) with 3B parameters. The feature vectors output by the models were then dimensionality-reduced and unsupervised clustering was performed. The Adjusted Rand Index (ARI) was used to quantify the consistency between the unsupervised clustering and the true classification labels (21 virus supergroups).

[0063] 2. Experimental Results As shown above, during the construction of the phylogenetic inference model, the Stage-2 model obtained in Stage 2 achieved an ARI value of 0.8881 when clustering test object 1; while the ESM2 baseline model with 650M parameters had an ARI value of 0.1408 when clustering test object 1, and the ESM2 baseline model with 3B parameters had an ARI value of 0.1439 when clustering test object 1, both of which are much smaller than the ARI value of the Stage-2 model.

[0064] This demonstrates that the Stage-2 model can accurately locate highly divergent viral sequences macroscopically without the need for multiple sequence alignment (MSA).

[0065] Example 2: A peerless phylogenetic inference method based on protein language model and hyperbolic space embedding A method for inferring unpaired phylogenetic relationships based on protein language models and hyperbolic space embeddings includes the following steps: S111. Obtain at least one query sequence; wherein the query sequence is a full-length polyprotein sequence of RNA; S112. Input the query sequence obtained in step S111 into the unpaired phylogenetic inference model obtained in Example 1. The unpaired phylogenetic inference model outputs the high-dimensional embedding vector 4 of the query sequence and projects it into the Lorentz hyperbolic manifold space to obtain the hyperbolic coordinates of the query sequence in the Lorentz hyperbolic manifold space. S113. Using each full-length polyprotein sequence in training set 1 of Example 1 as a reference sequence, based on the hyperbolic coordinates of the query sequence in the Lorenz hyperbolic manifold space obtained in step S112 and the hyperbolic coordinates of each reference sequence in the Lorenz hyperbolic manifold space (stored and recorded during the training of the unparalleled phylogenetic inference model shown in Example 1), calculate the hyperbolic geodesic distance between the query sequence and each reference sequence, as well as the hyperbolic geodesic distance between each query sequence, and combine the maximum likelihood reference tree of each full-length polyprotein sequence in training set 1 of Example 1 to obtain the evolutionary path distance between each reference sequence; S114. Based on the hyperbolic geodesic distance between the query sequence and each reference sequence, the hyperbolic geodesic distance between each query sequence, and the evolutionary path distance between each reference sequence obtained in step S113, a phylogenetic tree containing the query sequence is constructed by combining the adjacency-joining phylogenetic tree construction algorithm.

[0066] Example 3: Testing a peer-based phylogenetic inference method based on protein language models and hyperbolic space embedding. I. Experimental Methods Two supergroups, SG005 and SG009, with ≥100 sequences from the RNA virus virosphere data (Using artificial intelligence to document the hidden RNA virosphere, https: / / doi.org / 10.1016 / j.cell.2024.09.027), were used as test subjects, and a reference phylogenetic tree was constructed using the maximum likelihood method described in Example 1. SG005 and SG009 were independent of and did not overlap with the 22,063 full-length polyprotein sequences used in the training process of the unpaired phylogenetic inference model in Example 1.

[0067] The full-length polyprotein sequences and their corresponding RdRP domain sequences in SG005 and SG009 were aligned using MAFFT (v7.525, --maxiterate 1000 --localpair), and trimmed using trimAl (v1.5, -gt0.9 -cons 5). Finally, the maximum likelihood (ML) reference tree of the test object was constructed using IQ-TREE (multicore v2.3.6, -m LG --alrt 1000 -B 1000).

[0068] The test set sequences (query sequences) are uniformly sampled from the maximum likelihood (ML) reference tree of the test object based on the path distance between sequences in the reference tree. 50 sequences are extracted each time as query sequences, and the selected sequences cover a representative range of phylogenetic distances. The remaining sequences in the maximum likelihood (ML) reference tree of the test object are used as skeleton sequences. The phylogenetic tree of the skeleton sequences is constructed using the maximum likelihood method described in Example 1. The process is repeated 5 times independently to obtain 5 test samples (i.e., 5 sets of query sequences and skeleton sequences, SG005-1~SG005-5, SG009-1~SG009-5).

[0069] According to the unpaired phylogenetic inference method based on protein language model and hyperbolic space embedding shown in Example 2, the test set sequence is used as the query sequence and the skeleton sequence is used as the reference sequence. Phylogenetic inference is performed on the query sequence according to the unpaired phylogenetic inference method shown in Example 2 to construct a phylogenetic tree containing the query sequence. The Pearson correlation coefficient and the standardized Jaccard generalized Robinson-Foulds distance (JRF distance, used to quantify the divergence of topological structure) are calculated based on the maximum likelihood (ML) reference tree of the test object.

[0070] Simultaneously, using the test set sequence as the query sequence and the skeleton sequence as the reference sequence, two different phylogenetic mapping tools, pplacer (v1.1.alpha20) and EPA-ng (v0.3.8), were used to perform phylogenetic mapping on the query sequence. The Pearson correlation coefficient and JRF distance were calculated based on the maximum likelihood (ML) reference tree of the test object. Before performing phylogenetic mapping using the two phylogenetic mapping tools, the query sequence needed to undergo multiple sequence alignment. The sequences retained after multiple sequence alignment were considered as valid query sequences, and the number of valid query sequences was recorded.

[0071] Before performing phylogenetic localization, both pplacer and EPA-ng require the query sequence to be aligned with the skeleton sequence using multiple sequence alignment (MSA) software via UPP (v4.5.6). When performing phylogenetic localization, the replacement model parameters of EPA-ng are consistent with those used when constructing the reference tree and skeleton tree based on the maximum likelihood method, i.e., LG parameters are used; pplacer uses the default parameters.

[0072] II. Experimental Results The comparative experimental results for the five repeated tests in the SG005 dataset are shown in Table 1, and the comparative experimental results for the five repeated tests in the SG009 dataset are shown in Table 2.

[0073] Table 1 Comparative experimental results for the SG005 dataset

[0074] Table 2 Comparative experimental results for the SG009 dataset

[0075] Note: The Pearson correlation coefficients in Tables 1 and 2 are the Pearson correlation coefficients used for phylogenetic localization of common valid query sequences.

[0076] For query sequences with low homology, traditional control group tools (pplacer, EPA-ng) suffer from severe "sequence loss" due to multiple sequence alignment failures. In five replicate experiments of the SG005 supergroup (each planned to input 50 query sequences), traditional tools (EPA-ng or pplacer) could only successfully process about 40 to 42 sequences. In the SG009 supergroup test with higher divergence, the number of effective localizations of traditional tools was sharply reduced to 20 to 27 (i.e., nearly half of the sequences were discarded by the system because they could not be aligned).

[0077] In contrast, the unpaired phylogenetic inference method based on protein language model and hyperbolic space embedding shown in Embodiment 2 of the present invention successfully processes and locates all query sequences (50) due to its continuous feature representation space rather than explicit homologous discrete symbol matching, achieving a sequence retention rate of 100%. Furthermore, when performing fair comparisons on the intersection sequence set that traditional tools can successfully process, the phylogenetic tree inferred by the unpaired phylogenetic inference method shown in Embodiment 2 of the present invention is highly correlated with the evolutionary distance of the ML reference tree, with a Pearson correlation coefficient (Pearson's r) consistently around 0.95, demonstrating performance equal to or better than traditional maximum likelihood localization tools. In contrast, other comparison tools suffer from sequence loss issues and cannot compare the JRF distance index with the reference tree.

[0078] The method for inferring phylogenetic relationships shown in Example 2 can solve the problem of sequence loss during the alignment process of highly divergent sequences, achieving a 100% sequence retention rate.

[0079] The Pearson correlation coefficient and JRF distance results of the peer phylogenetic inference method shown in Example 2 for phylogenetic localization of the test object are shown in Table 3.

[0080] Table 3. Pearson correlation coefficient and JRF distance results for phylogenetic localization of test subjects.

[0081] The results show that when the range of query sequences is expanded to include all 50 query sequences successfully preserved by the unpaired phylogenetic method shown in Example 2 (i.e., including those extremely divergent sequences that traditional methods cannot handle), the unpaired phylogenetic method shown in Example 2 still exhibits excellent structural stability, with its topological divergence (JRF distance) consistently remaining at an extremely low level (approximately 0.1 to 0.2). This strongly demonstrates that the unpaired phylogenetic inference method shown in Example 2 not only greatly expands the applicable boundaries of phylogenetic analysis (solving the problem of locating unaligned sequences), but also that its inference accuracy is comparable to or even surpasses that of traditional gold standard methods that rely on high-quality alignments.

[0082] In summary, the results show that after performing phylogenetic inference on the test set sequences using the unpaired phylogenetic inference method shown in Example 2, the resulting phylogenetic tree is highly correlated with the maximum likelihood (ML) reference tree of the test object (the Pearson correlation coefficient r is significantly higher). In particular, for sequences with low homology, traditional control group tools (pplacer, EPA-ng) suffer from severe "sequence loss" due to multiple sequence alignment failures. However, this invention, based on feature representation rather than explicit homology sequence matching, successfully locates all query sequences without any failures.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description and ideas, and it is neither necessary nor possible to exhaustively describe all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a unique pair phylogenetic inference model based on protein language model and hyperbolic space embedding, characterized in that, Includes the following steps: S1. Obtain a training set containing global target sequences, determine the taxonomic group label of each global target sequence in the training set, and extract the local key functional region sequences containing evolutionary conserved information from each global target sequence. For the training set, a reference phylogenetic tree representing the true evolutionary distance between sequences is constructed under each taxonomic group; S2. Using each global target sequence in the training set as input to the biological sequence pre-trained large language model, the model is adaptively trained through the masked language modeling task to obtain the Stage-1 model; The Stage-1 model outputs an initial high-dimensional embedding vector 1 for each global target sequence; S3. Using each global target sequence in the training set as input to the Stage-1 model obtained in step S2, and combining the class group label of each global target sequence in the training set in step S1, construct positive sample pairs with the same class group label and negative sample pairs with different class group labels in the embedding vector feature space. Optimize the Stage-1 model through a contrastive learning strategy to obtain a global feature representation containing coarse-grained classification structure information, and obtain the Stage-2 model. The Stage-2 model outputs a high-dimensional embedding vector 2 of the global target sequence; the high-dimensional embedding vector 2 maps the classification group information corresponding to the global target sequence; S4. Using the global target sequence and its corresponding local key functional region sequence in the training set as input to the Stage-2 model obtained in step S3, extract the global embedding vector of the global target sequence and the local embedding vector of the local key functional region sequence, and train the Stage-2 model to obtain the Stage-3 model. The Stage-3 model outputs a high-dimensional embedding vector 3 for the global target sequence; the high-dimensional embedding vector 3 maps the classification group information of the global target sequence and integrates the cross-scale comprehensive feature representation of the global embedding vector and the local embedding vector. During the training of the Stage-2 model, each global target sequence and its corresponding local key functional region sequence in the training set described in step S1 are used as the input of the Stage-2 model. Wherein, the global target sequence is encoded by the Stage-2 model to obtain the global embedding vector Z corresponding to each global target sequence poly，i ; and the corresponding local key functional region sequence is encoded by the Stage-2 model to obtain the corresponding local embedding vector Z dom，i . Stage-2 model parameters are then trained and optimized by minimizing a joint loss function between the global embedding vector Z poly，i and the corresponding local embedding vector Z dom，i ​ The joint loss function is a weighted sum of the cross-modal InfoNCE loss function, the orthogonal regularization penalty term, and the center separation penalty term; The calculation method for the center separation penalty term is shown in Formula I: Formula I: ; Where sim(·) represents the cosine similarity function; It is the mean vector of all global embedding vectors within the same training batch; It is the mean vector of all local embedding vectors within the same training batch; S5. Using each global target sequence in the training set as input to the Stage-2 model obtained in step S3 or the Stage-3 model obtained in step S4, extract the Euclidean space feature representation of the output. The Euclidean space feature representation is projected onto a geometric mapping function with learnable curvature parameters. k In the hyperbolic manifold space, obtain the hyperbolic coordinates of each global target sequence in that space; Based on the hyperbolic coordinates of each global target sequence in this space, the hyperbolic geodesic distance between every two global target sequences is calculated. At the same time, the evolutionary path distance between the corresponding global target sequences in the reference phylogenetic tree constructed in step S1 is extracted as a supervision signal for model training. During training and optimization, a loss function is constructed with the regression objective of minimizing the error between the hyperbolic geodesic distance and the true evolutionary distance. The backbone parameters of the network are frozen, and a parameter-efficient fine-tuning strategy is used to update the parameters in the Stage-2 or Stage-3 model, as well as the curvature parameters of the hyperbolic manifold space. k We obtained a unique phylogenetic inference model based on protein language model and hyperbolic space embedding; The model outputs a high-dimensional embedding vector 4 for each global target sequence in the training set, projects the high-dimensional embedding vector 4 of each global target sequence in the training set into hyperbolic space, and records the hyperbolic coordinates of each global target sequence in the training set in hyperbolic space.

2. The construction method according to claim 1, characterized in that, The global target sequence mentioned in step S1 includes protein sequences, deoxyribonucleic acid sequences, and / or ribonucleic acid sequences.

3. The construction method according to claim 1, characterized in that, The construction of the reference phylogenetic tree in step S1 specifically includes the following steps: S11. Extract local key functional region sequences from each global target sequence in the training set, and perform multi-sequence alignment to obtain the multi-sequence alignment results; S12. Based on the multiple sequence alignment results obtained in step S11, a reference phylogenetic tree is constructed using the maximum likelihood method, and the path distances between nodes in the tree are extracted to construct a reference evolutionary distance matrix.

4. The construction method according to claim 1, characterized in that, The biological sequence pre-trained large language model mentioned in step S2 includes a protein language model, a nucleic acid language model, or a unified biological large model that simultaneously supports nucleic acid and protein representation.

5. The construction method according to claim 1, characterized in that, The Stage-1 model optimization strategy described in step S3 involves training using the InfoNCE loss function, which is calculated as shown in Formula II. Official II: ; Where N is the number of sequences within the same training batch; Cosine similarity; For temperature parameters; Positive sample pairs are represented by the high-dimensional feature representations of two independent sequences sampled from the same taxonomic group, while other unpaired sequence feature representations in the global batch are considered as negative samples.

6. The construction method according to claim 1, characterized in that, The calculation method for the cross-modal InfoNCE loss function in step S4 is shown in Formula III; Official III: Where N is the number of sequences in the same training batch; sim(·) represents the cosine similarity function; For temperature parameters; Z poly，i Z is the global embedding vector of the i-th global target sequence within this training batch; dom，i Z is the local embedding vector of the i-th global target sequence within this training batch; dom，j This is the local embedding vector of the j-th global target sequence within this training batch.

7. The construction method according to claim 1, characterized in that, The calculation method for the orthogonal regularization penalty term in step S4 is shown in Formula IV; Formula IV: ; Among them, Z poly For all global embedding vectors Z within this training batch poly，i The global feature matrix composed of Z and subjected to L2 normalization; dom For all local embedding vectors Z within the same training batch dom，i The local feature matrix composed of and L2-normalized features; I is the identity matrix; ||·|| F This represents the Frobenius norm.

8. The construction method according to claim 1, characterized in that, The calculation method of the joint loss function in step S4 is shown in formula V; Formula V: ; Where, λ ortho and λ sep This is a hyperparameter.

9. The unpaired phylogenetic inference model based on protein language model and hyperbolic space embedding constructed by any of the construction methods described in claims 1 to 8.

10. A method for inferring phylogenetic relationships based on protein language models and hyperbolic space embeddings, characterized in that, Includes the following steps: S111. Obtain ≥1 query sequence; the query sequence is derived from the classification group contained in the global target sequence in the training set of any of the construction methods described in claims 1 to 8; S112. Input the query sequence obtained in step S111 into the unpaired phylogenetic inference model of claim 9. The unpaired phylogenetic inference model outputs the high-dimensional embedding vector 4 of the query sequence and projects it into hyperbolic space to obtain the hyperbolic coordinates of the query sequence. S113. Using each global target sequence or the local key functional region sequence of each global target sequence in the training set of the construction method of claim 1 as a reference sequence, based on the hyperbolic coordinates of the query sequence obtained in step S112 and the hyperbolic coordinates of each global target sequence in the training set shown in step S5 of the construction method of claim 1 in hyperbolic space, calculate the hyperbolic geodesic distance between the query sequence and each reference sequence and the hyperbolic geodesic distance between each query sequence, and combine the maximum likelihood tree obtained in step S1 of the construction method of claim 1 to obtain the evolutionary distance between each reference sequence; S114. Based on the hyperbolic geodesic distance between the query sequence and each reference sequence, the hyperbolic geodesic distance between each query sequence, and the evolutionary distance between each reference sequence obtained in step S113, a phylogenetic tree containing the query sequence is constructed by combining a distance-based phylogenetic tree construction algorithm.

Images