A method for designing odor molecule-specific olfactory receptors based on multi-modal deep learning

Abstract

The application discloses a novel smell molecule-specific olfactory receptor design method based on a multi-modal deep learning method, adopts a multi-modal deep learning model, fuses the physicochemical, sequence and structural characteristics of smell molecules and the conserved structure and functional motif information of olfactory receptors, innovatively realizes the reverse technology of designing novel olfactory receptors for given smell molecules, has high design precision, breaks through the limitations of traditional dependence on natural receptors or random mutations, shortens the research and development cycle, reduces the research and development cost, provides a novel path for customized olfactory receptor development, and meets the accurate identification requirements of specific smell molecules.

Description

Technical Field

[0001] This invention relates to the field of bioelectronic nose odor-sensitive element design, and more particularly to a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning. Background Technology

[0002] Olfactory receptors, as an important class of G protein-coupled receptors (GPCRs), play a central role in odor perception by specifically recognizing odor molecules and initiating signal transduction. They have broad application prospects in fields such as fragrance screening, odor detection, and biosensing. Traditional olfactory receptor development primarily relies on screening natural olfactory receptor libraries or randomly mutagenesis of natural receptors to obtain receptors capable of binding to target odor molecules. However, natural olfactory receptor resources are limited, and most natural receptors have low specificity, making it difficult to meet the precise recognition requirements of specific odor molecules. Random mutagenesis lacks targeted design logic, failing to specifically optimize the binding site between the receptor and the target odor molecule, resulting in low development efficiency and long development cycles. Existing deep learning technologies mostly focus on predicting the response of known odor molecules to known olfactory receptors, without addressing reverse engineering techniques for designing entirely new olfactory receptors for given odor molecules. Therefore, developing a targeted olfactory receptor design method based on a precise computational model to achieve customized, highly specific, and highly active olfactory receptors for given odor molecules is crucial to solving these problems. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning. This method does not rely on natural receptors or random mutations. Through multimodal feature fusion and accurate model prediction, it can design olfactory receptors that can specifically bind to target odor molecules, thus achieving reverse design from odor molecules to functional olfactory receptors.

[0004] The objective of this invention is achieved through the following technical solution: a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning, comprising:

[0005] A multimodal feature dataset of odor molecules is constructed, and molecular physicochemical descriptors, molecular sequence data and molecular structure images of odor molecules are extracted to form a multimodal feature set of odor molecules;

[0006] Construct and train a multimodal deep learning model for olfactory receptor design:

[0007] Using the multimodal feature set of odor molecules as input, the multimodal features are fused to obtain multiscale molecular features, and the GPCR structure prior encoder is used to obtain the GPCR structure prior.

[0008] GPCR structural priors and multi-scale molecular features are input into AlphaFold2, and molecular feature injection and GPCR structural prior knowledge injection are performed in parallel.

[0009] The standard invariant point attention computation of AlphaFold2 is divided into two branches to predict the active and inactive states, and multiple conformations are generated using ensemble sampling. Rigid constraints are added to the invariant point attention computation to predict the receptor structure, and the receptor structure and constraints are obtained. The predicted receptor structure and constraints are encoded and decoded and amino acid prediction is performed to obtain the amino acid sequence of the olfactory receptor.

[0010] The trained model is used to generate receptors, and the generated receptors are validated and evaluated to output qualified olfactory receptors.

[0011] Furthermore, the molecular physicochemical descriptors include at least 10 of the following: molecular weight, three-dimensional molecular structure features, molecular polarity, lipophilicity parameters, number of hydrogen bond donors, number of hydrogen bond acceptors, topological surface area, number of aromatic rings, proportion of hydrophobic groups, and charge distribution; the molecular sequence data includes sequence features converted from molecular SMILES strings and more than 166 types of molecular fingerprint data, including MACCSFP, SubFP, and RDKit series fingerprints; the molecular structure image is a schematic diagram of the molecular planar structure, including atomic connection methods, chemical bond types, ring structures, and spatial distribution information of functional groups, obtained from the molecular SMILES format string using the RDKit tool.

[0012] Furthermore, the fusion of multimodal features includes: using a Transformer encoder with a hierarchical attention mechanism to perform cross-modal fusion of multimodal molecular features; adding a special token as a global representation for each modal feature, aligning the dimensions through linear projection, capturing the correlation information between features through intramodal attention and cross-modal attention, finally extracting a unified molecular feature vector from the global aggregated token, using hierarchical convolutional blocks of ConvNeXt to extract multi-scale features from the molecular feature vector, using parallel convolutional branches with convolutional kernels of different sizes to capture multi-scale information, extracting local and global patterns of molecular features, and performing multi-scale adaptive weight fusion to obtain multi-scale molecular features.

[0013] Furthermore, the GPCR structural prior includes Pair representation, single-residue features, and a global structural prior vector, and the acquisition process includes:

[0014] The input amino acids are topologically constrained and encoded to extract the amino acid sequence context relationship. The helical positions of G protein-coupled receptors are classified, and the amino acid residues and helical labeling matrix of each helical region are identified based on the classification results.

[0015] Based on the classification results, predict the helical space parameters and calculate the environmental information by combining the amino acid sequence context relationship;

[0016] The frequency of amino acid residue positions in homologous sequences in the MSA matrix is ​​calculated to generate position-specific scoring matrix features and conservation score codes;

[0017] Single-residue features are obtained by integrating amino acid sequence context, environmental information, specificity scoring matrix features, helical annotation matrix, and conservation score encoding.

[0018] Pair features are obtained by encoding the interaction between helices based on helical space parameters, and Pair representations are obtained by embedding the Pair features after calculating mutual information.

[0019] Global motif representations are generated based on databases; ligand binding pocket features are predicted using environmental information, amino acid sequence contextual relationships, and multi-scale molecular features.

[0020] Key motifs were labeled using G protein binding domain labelers to obtain pocket-ligand complementarity scores. The motifs were then state-encoded and the target rotation angle was calculated.

[0021] The global structural prior vector is obtained by integrating helical spatial parameters, pocket-ligand complementarity scores, pocket geometric centers, global motif representations, and GPCR structural statistical priors.

[0022] Furthermore, the molecular feature injection includes:

[0023] Multi-scale molecular features are fused to enhance the initial MSA embedding, generating position-specific biases for molecular conditions. The desired inter-residue distances are predicted using multi-scale molecular features and converted into radial basis function encodings. Molecular induced distances and multi-scale molecular feature biases are added to the Pair representation initialization to obtain a molecularly enhanced Pair representation. For the geometric center of the GPCR binding pocket, molecular consistency constraints are calculated based on multi-scale molecular features, the expected molecular distances on the three sides, and the pocket volume, and weighted and added to the Pair representation. A ligand-structure interaction module is inserted every 8 layers in Evoformer to extract pocket subgraphs from the Pair representation. A graph attention network is used to process pocket features, and then cross-attention is used to calculate the attention weights obtained from the interaction with multi-scale molecular features and feed them back to the Pair representation.

[0024] The GPCR structural prior knowledge injection includes:

[0025] Injecting GPCR structural priors into the input layer: fusing the MSA representation with the single-residue features generated by the GPCR structural prior encoder as the MSA representation initialization of Evoformer; integrating the Pair representation with the Pair features generated by the GPCR structural prior encoder as the Pair representation initialization of Evoformer; injecting GPCR structural priors into the intermediate layer of Evoformer for single-residue feature enhancement and Pair feature enhancement.

[0026] Furthermore, after molecular feature injection and GPCR structural prior knowledge injection, the update process of the Structure Module in AlphaFold2 includes molecular guidance and GPCR structural prior guidance. The two guidances are not sequential, and the residue coordinates are updated twice, including:

[0027] Structure Module Molecular Guidance: Molecular geometric bias is added to the original IPA attention to obtain the modified IPA attention; for pocket residues, multi-scale molecular features are used to predict the ideal position of the residue as a virtual force for molecular feature guidance to guide coordinate updates; in side chain prediction, when predicting the side chain rotation angle for each residue, multi-scale molecular features are used to correct the angle of the pocket residue; for aromatic residues, stereochemical constraints on the side chain direction are added to predict the optimal π-π stacking direction.

[0028] Use GPCR structural prior guidance in the Structure Module: add geometric bias to the IPA attention calculation; add virtual forces as constraints when updating coordinates in the IPA output coordinates.

[0029] Furthermore, the encoding / decoding and amino acid prediction of the predicted receptor structure and constraints includes:

[0030] Based on the three-dimensional structure obtained by AlphaFold2, a three-dimensional structural node encoder is constructed to extract comprehensive node features. A three-dimensional structural edge encoder is constructed, and a K-nearest neighbor graph is used to construct an interaction graph between residues. Edge features are encoded, and the distance and direction, relative orientation, sequence separation degree, and potential interaction type label of each edge are calculated to generate comprehensive edge features.

[0031] A graph neural network encoder is used to perform geometric encoding on the integrated node features and integrated edge features to obtain the structural encoding of each amino acid residue in the predicted receptor and the global structural encoding; constraint encoding is performed based on spatial constraints and interaction rules.

[0032] Conserved motif libraries, G protein binding interfaces, and ligand binding pocket residue constraints were used to obtain the comprehensive encoding of GPCR functional motifs.

[0033] A bidirectional LSTM decoder is used, with the bidirectional LSTM layer employing a three-layer stacked BiLSTM and multi-head self-attention enhancement. It combines structural coding, constraint coding, motif coding, and position coding features for decoding, and obtains amino acid prediction results through the prediction head.

[0034] Furthermore, the validation and evaluation of the generated receptor includes:

[0035] Molecular docking verification was performed using AutoDock software to dock the target odor molecule with the binding pocket of the designed receptor and calculate the binding energy and binding mode.

[0036] Molecular dynamics simulations were performed in aqueous solutions or membrane-like simulation environments for ≥100 ns to analyze the RMSD value of the receptor structure, conformational fluctuations of the binding pocket, and the persistence of molecular-receptor interactions.

[0037] Structural rationality assessment was conducted by analyzing the dihedral distribution of amino acid residues using the PROCHECK tool;

[0038] Compatibility assessment was conducted using the Verify3D tool to verify the compatibility of the structure and sequence.

[0039] According to another aspect of the specification, a device for designing odor molecule-specific olfactory receptors based on multimodal deep learning is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the aforementioned method for designing odor molecule-specific olfactory receptors based on multimodal deep learning.

[0040] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the aforementioned method for designing odor molecule-specific olfactory receptors based on multimodal deep learning.

[0041] The beneficial effects of this invention are that it adopts a multimodal deep learning model, which integrates the physicochemical, sequence, and structural characteristics of odor molecules with the conserved structural and functional motif information of olfactory receptors. It innovatively realizes the reverse technology of designing olfactory receptors for a given odor molecule, with high design accuracy. It breaks through the limitations of traditional reliance on natural receptors or random mutations, shortens the research and development cycle, reduces research and development costs, provides a new path for the development of customized olfactory receptors, and meets the needs of accurate identification of specific odor molecules. Attached Figure Description

[0042] Figure 1 The flowchart illustrates a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning, as provided in this invention.

[0043] Figure 2 This invention provides an overall framework diagram for a method of designing odor molecule-specific olfactory receptors based on multimodal deep learning.

[0044] Figure 3 The following are some physicochemical descriptors, molecular sequence features, and molecular structure images provided for embodiments of the present invention.

[0045] Figure 4 A typical structural diagram of the GPCR family olfactory receptors provided in an embodiment of the present invention.

[0046] Figure 5 A flowchart illustrating the model training process provided in this embodiment of the invention.

[0047] Figure 6 The diagram shows the docking results of the TAAR receptor and amphetamine molecules provided in the embodiments of the present invention.

[0048] Figure 7 This is a schematic diagram of a device for designing odor molecule-specific olfactory receptors based on multimodal deep learning, provided for an embodiment of the invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0050] like Figure 1 As shown, the present invention provides a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning, comprising the following steps:

[0051] Step 1: Multimodal Feature Extraction of Odor Molecules. For a given target odor molecule, starting from the SMILES string, three types of core features are comprehensively extracted;

[0052] Step Two: Design and Training of a Multimodal Deep Learning Model for Olfactory Receptors. A multimodal deep learning model integrating Transformer, ConvNeXt, and LSTM is constructed to establish a mapping relationship from odor molecule features to olfactory receptor structure and sequence. This model comprises three core modules: a feature fusion module, a structure prediction module, and a sequence generation module.

[0053] Step 3: Virtual verification and structural evaluation of the designed receptor. Through molecular docking, molecular dynamics simulation and structural rationality analysis, the binding potential, structural stability and signal transduction adaptability of the olfactory receptor to the target odor molecule are verified. After the evaluation is qualified, the final olfactory receptor design result is obtained.

[0054] Furthermore, step one is implemented through the following sub-steps:

[0055] (1.1) As Figure 3As shown, molecular physicochemical descriptors, such as basic physical properties (logP, TPSA, molecular weight, number of rotatable bonds), electronic properties (HOMO / LUMO band gap, dipole moment, polarizability), stereochemical features (moment of inertia ratio, radius of gyration), surface properties (van der Waals surface area, solvent-accessible surface area), and topological indices (Wiener index, Balaban index, Zagreb index), are obtained from molecular SMILES format strings using chemical information software tools such as PaDel, Mordred, and Dragon. In this embodiment, quantitative indicators including molecular weight, three-dimensional structural features, polarity, lipophilicity parameters, number of hydrogen bond donors / acceptors, topological surface area, number of aromatic rings, proportion of hydrophobic groups, and charge distribution are extracted as molecular physicochemical descriptor features. , where n is the feature dimension, see Figure 3 The modality-1 features are preprocessed using standardization.

[0056]

[0057] (1.2) As Figure 3 As shown, molecular fingerprint features are obtained from molecular SMILES format strings using chemical information software tools such as RDKit, PaDel, Mordred, and Dragon. One or more molecular fingerprint types can be used, such as 166-bit MACCS Keys, 881-bit PubChem Fingerprint, and 2048-bit RDKit Topological Fingerprint, with a molecular fingerprint dimension of m.

[0058]

[0059] Sequence encoding is performed on the SMILES identifier. Specifically, the SMILES string is segmented using an unsupervised trained BPE tokenizer. The vocabulary size is 5k, and the length of the encoded sequence is L, such as 128 or 256. Each token is encoded as a d-dimensional vector, and then sine-cosine positional encoding is performed. See [link to documentation]. Figure 3 Record the position information of the token in the original SMILES string, that is, for each position in the sequence Hidden layer dimension d, index i of each dimension in the hidden layer, and the i-th element of the position encoding vector. and Calculated using sine and cosine functions respectively:

[0060]

[0061]

[0062] in, represents the position index in the input sequence, ranging from (0-L); i represents the index in the hidden dimension (starting from 0), and d is the size of the hidden dimension of the model. Thus, we obtain... Figure 3 Molecular sequence characteristics.

[0063] (1.3) Use the RDKit tool to convert SMILES into a 300×300 pixel RGB image, clearly showing the spatial distribution of atoms, chemical bonds, ring structures, and functional groups. The image background is white or transparent, carbon is black, oxygen is red, nitrogen is blue, sulfur is yellow, chlorine is green, phosphorus is orange, hydrogen is light gray, single bonds are solid lines, double bonds are double parallel lines, triple bonds are triple parallel lines, and aromatic bonds are dashed or single lines. See details. Figure 3 The example molecular image is given in the image. Visual features are extracted using a ConvNeXt pre-trained model (ImageNet-22K) and finally reduced to 512 dimensions to obtain the desired result. .

[0064]

[0065] Furthermore, step two is implemented through the following sub-steps:

[0066] (2.1) For the physicochemical descriptor features in step (1.1) Add a special token of one dimension as the overall representation, denoted as [PHYS], and use linear projection to... Projecting [PHYS] onto the model dimension D yields Token_phys.

[0067] (2.2) For the molecular sequence features in step (1.2), the m-dimensional fingerprint features are... Block embedding involves adding a special token of one dimension as the overall representation, aggregating molecular sequence and fingerprint information, denoted as [SEQ], and using linear projection to project the molecular sequence features after block embedding + [SEQ] onto the model dimension D to obtain Token_seq.

[0068] (2.3) For the image features in step (1.3) Add a special token of one dimension as the overall representation, denoted as [IMG], and use linear projection to... Projecting [IMG] onto the model dimension D yields Token_img.

[0069] (2.4) After processing in steps (2.1), (2.2), and (2.3), the model input sequence Including features consistent across three modal dimensions, a [CLS] token is added for the final molecular representation:

[0070]

[0071] (2.5) See also Figure 2 The feature fusion module in the middle processes the model input sequence obtained in step (2.4). Transformer cross-modal fusion employs a hierarchical attention mechanism. The first layer is intra-modal attention. [CLS] focuses only on itself, and each modal token only focuses on its own modal information; the intermediate layer is for cross-modal attention. [CLS] focuses on all modal tokens, and each modal token focuses on other modalities; the last layer is global aggregation, where [CLS] aggregates all information.

[0072]

[0073]

[0074] in, For the attention mechanism calculation, , , The feature matrix is ​​obtained after transforming different modal features. , , This is a trainable parameter matrix.

[0075] (2.6) For the Transformer network in step (2.5), L is the number of network layers. The global representation is extracted from the [CLS] Token of the last layer L, and a 512-dimensional vector is output to fully integrate the multi-dimensional information of the molecule.

[0076]

[0077] in This is global information extracted from the last layer L of the Transformer encoder. For the process The 512-dimensional uniform molecular feature vector output after the normalization layer, as shown below. Figure 2 The feature fusion module is shown in the figure.

[0078] (2.7) Using contrastive learning loss As an auxiliary optimization objective of the feature fusion module, similar odor molecule samples (positive samples) are brought closer to each other in the feature space, while dissimilar samples are moved away from each other, thereby optimizing the alignment and learning of unified representations between different modalities.

[0079]

[0080]

[0081] in, , , They are the i-th, j-th, and k-th samples, respectively. This is the temperature coefficient.

[0082] (2.8) See also Figure 2 The multi-scale molecular feature part in the feature fusion module is based on the molecular features obtained in step (2.6) of the feature fusion module. Decoding. Using linear projection on 512-dimensional molecular features. Feature expansion and dimensionality reorganization are performed, followed by multi-scale feature extraction using hierarchical convolutional blocks of ConvNeXt. Parallel convolutional branches with kernels of different sizes capture multi-scale information, extracting local and global patterns of molecular features. Specifically, the coarse-grained branch (16×16 kernel size) captures global ligand-receptor matching information, such as the matching of overall hydrophobicity with the membrane environment, the compatibility of molecular size with pocket volume, and the complementarity of polarity distribution with the binding interface; the medium-grained branch (8×8 kernel size) captures functional domain-ligand interactions, such as the matching pattern between functional groups and key residues in the binding pocket, the spatial arrangement requirements of hydrogen bond donors and acceptors, and the π-π interactions between aromatic rings and aromatic residues; the fine-grained branch (4×4 kernel size) captures atomic-level interactions, including the interaction preferences of specific atom pairs, constraints on side-chain rotomers, and precise distance and angle requirements. Adaptive weighted fusion of the three scales of molecular features yields enhanced molecular features. .

[0083] (2.9) such as Figure 4 The typical structure of an olfactory receptor is shown. Known GPCR homologous receptor structures, such as rhodopsin and β2-adrenergic receptors, are used as input to the GPCR structural prior encoder. The input sequence length L = 320 (typical length of an olfactory receptor), and the amino acid sequence... .

[0084] (2.10) See also Figure 2 In the GPCR prior encoder part, topological constraint encoding is performed first. The amino acid sequence from step (2.9) is preprocessed. The amino acid sequence is then encoded to obtain the amino acid embedding. For each residue i, calculate a 5-dimensional physical feature vector. The fusion embeddings and linear projection yield a 512-dimensional embedding vector. .

[0085]

[0086] in, This is the Kyte-Doolittle value, which represents the hydrophilicity index of an amino acid; For charge quantity, For side chain volume, For Vihinen flexibility index, Indicates whether it belongs to the aromatic family.

[0087] (2.11) The amino acid embedding vector obtained in step (2.10) Use BiLSTM network layers to capture sequence context relationships .

[0088] (2.12) Contextual relationship of the amino acid sequence obtained in step (2.11) Linear projection and softmax processing are used for multi-task output to obtain the helical / loop classification result for each residue i. Helical numbering classification results of residue i in the helical region Boundary probability prediction results for each residue i Combined output spiral label matrix .in The classification results are of two types: if residue i is located in the helical region of the G protein-coupled receptor, it is denoted as... If it is located in a ring region, it is denoted as ; Residue i represents the position of residue i on the G protein-coupled receptor's transmembrane helix. The classification results are 7 types (G protein-coupled receptors have 7 transmembrane helices), denoted as follows: ; This indicates whether residue i is located at either end of the amino acid sequence. The classification results are into three types: sequence initiation amino acid, denoted as... The amino acids in the sequence are denoted as The amino acid at the end of the sequence is denoted as .

[0089] (2.13) According to step (2.12) The classification results identified each spiral That is, the amino acid residues corresponding to the seven transmembrane helical regions of the G protein-coupled receptor, based on the contextual relationships in step (2.11). For each spiral region Features are pooled using MeanPool and spiral space parameters are predicted. Including: membrane core range Inclination angle Azimuth Rotation angle Helix length .

[0090] (2.14) Construct a membrane environment coding network and identify 7 helical regions according to step (2.13). The obtained amino acid residues in the corresponding region The spiral label matrix and spiral space parameters Calculate the membrane depth for each residue. :

[0091]

[0092] in, Let z be the z-coordinate of the i-th residue. Let be the z-coordinate of the center point of the k-th helix, and i be the sequence position of the current residue. represents the center position of the residue index of the k-th helix, and 1.5 is the vertical projection step size between adjacent residues. Let be the tilt angle of the k-th spiral.

[0093] (2.15) Sine-cosine multi-frequency position encoding is performed on each residue in the amino acid sequence, and the membrane depth of each residue is obtained according to step (2.14). Discrete encoding of membrane regions This determines which region (extracellular, outer layer, core, inner layer, or intracellular) each residue belongs to, generating membrane environment embeddings. .

[0094]

[0095] (2.16) Construct a membrane environment feature generator. Calculate hydrophobicity matching features using a simplified geometric model and functions. and solvent exposure characteristics Finally, the multi-frequency location encoding, membrane environment embedding and hydrophobicity matching features, and solvent exposure features from step (2.15) are concatenated to generate a comprehensive environment vector. :

[0096]

[0097] in , The value is the Kyte-Doolittle value of the amino acid residues calculated in step (2.10), which represents the hydrophilicity index of the amino acid. For learnable weight vectors, It is the hyperbolic tangent activation function.

[0098]

[0099] in, The distance from residue i to the nearest helical axis, based on the membrane depth of each residue. calculate, This is the distance attenuation coefficient. This is the position offset parameter.

[0100] (2.17) Construct an inter-helical interaction encoder. Based on the GPCR helical structure identified in step (2.13), construct a learnable and enhanced 7×7 contact frequency matrix. :

[0101]

[0102] in, To statistically determine the prior contact frequency matrix, values ​​range from 0 to 1, and the matrix is ​​7×7. This matrix is ​​based on a large number of GPCR structures in a known database, which statistically determine the contact probability between each pair of helices.

[0103]

[0104] in, It is a 7×7 learnable parameter matrix, with initial values ​​of ; This indicates that Softmax normalization is performed in the row direction so that the sum of the elements in each row is 1.

[0105] (2.18) For residues Based on the contact frequency matrix obtained in step (2.17) Query The basic contact probability is obtained. The center-to-center distance matrix between the spirals is derived from the membrane core range in step (2.13). Calculated, according to Obtain distance geometric prior and to Perform radial basis function encoding, then combine with contact probability One-hot encoding of the helices to which residues i and j belong, membrane depth difference between residues i and j, and helical labeling matrix. Connect to generate Pair feature codes .

[0106] (2.19) Constructing a ligand binding pocket predictor. Based on the known pockets typically located in TM3 / 5 / 6 / 7, an attention mechanism is designed to score the pocket region for each residue i, yielding... For high-scoring regions ( The residues of ) are combined with the characteristics of upstream amino acid residues and the ligand molecule obtained in step (2.8). Enhancing pocket features :

[0107]

[0108] (2.20) Construct a G protein binding domain labeler. Detect key motifs, such as the DRY motif at the end of TM3 and the NPxxY motif in the middle of TM7, and label the residues in the G protein binding region to obtain... .

[0109] (2.21) Encode the active / inactive state. Define two learnable state embedding vectors. , For each intracellular residue i in step (2.15), calculate the state. Prediction using small networks (based on molecular features) (Assess the expected level of activation)

[0110]

[0111]

[0112] (2.22) Process the TM6 conformation constraint vector according to the activation state in step (2.21). Calculate the target rotation angle for the intracellular terminal residues (last 8 residues) of TM6. .

[0113]

[0114] in, The target rotation angle at the intracellular end of TM6; is the activation state coefficient of the i-th residue; Residue position weights; The maximum possible rotation angle (usually 15-30°); The distance from residue i to the intersection of TM6 and TM3; This is the distance attenuation coefficient.

[0115] (2.23) Multiple sequence alignment processing. Input MSA matrix (A multiple sequence alignment information used in the AlphaFold2 framework to represent the evolutionary relationships of amino acid sequences), where N is the number of homologous sequences and L is the aligned length. Amino acid statistics are obtained from the UniProt library. background frequency And calculate the amino acid residues in all homologous sequences. Frequency of occurrence at position i The location-specific score matrix (PSSM) is generated and converted into location-specific score matrix features. The location-specific scoring matrix is ​​characterized by:

[0116]

[0117] (2.24) In step (2.18), the pair feature encoding of residue pair (i, j) has already been obtained. The mutual information MI(i, j) of residue pairs (i, j) is calculated and embedded into the previously obtained pair representation to form a new pair representation. .

[0118] (2.25) Use the amino acid residues obtained in step (2.23) Frequency of occurrence at position i The Shannon entropy H(i) of each residue i is calculated, and the conservation score Conservation(i) is obtained, thus yielding the conservation score encoding. Highly conserved sites receive higher weights w in the loss function. i = 1 + 2 × Conservation(i), where the weight is 3 for fully conservative (Conservation=1) and 1 for fully variable.

[0119] (2.26) Obtain structural motif databases from existing GPCR homologous structures, such as DRY motif template (3 residues) main chain coordinates and NPxxY motif template (5 residues) main chain coordinates.

[0120] (2.27) Based on step (2.26), template matching and stacking are performed. For amino acid sequence detection, for example, if a DRY motif is detected at position p, main chain coordinate initialization and side chain rotatores are generated, and soft constraints are added to the loss:

[0121]

[0122] in, The functional motif structure constraint loss. For the set of functional motifs, To predict the 3D coordinates of atoms in the structure, These are the 3D coordinates of the corresponding atoms in the template structure. The weighting coefficients are at the atomic level. The weighting coefficients are larger for atoms that directly participate in the interaction, such as 1.0. The weighting coefficients are moderate for main chain atoms, such as 0.5. For non-critical side chain atoms, the weighting coefficients can be smaller, such as 0.2.

[0123] (2.28) For each detected motif m, the motif type, position encoding, template space parameters, average conservation of the motif region, and contextual information from BiLSTM are fused to generate a motif embedding vector. The global representation of functional motifs is obtained through average pooling. .

[0124] (2.29) Single residue feature integration ( , , , , )get See Figure 2 The GPCR structure prior encoder portion in the high-resolution pocket residues Add extra pocket enhancements .

[0125] (2.30) Pair feature integration (initial pair representation) The final Pair representation is obtained by combining outer product to capture single-residue interactions and distance-constrained encoding. See Figure 2 The GPCR structure prior encoder part in the text.

[0126] (2.31) The GPCR structure prior encoder finally outputs the global structure prior vector. See Figure 2 The GPCR structure prior encoder part in the code:

[0127]

[0128] in, The seven spiral space parameters obtained in step (2.13) have a dimension of 42; The pocket-ligand complementarity score obtained in step (2.20) has a dimension of 1; To establish the geometric center of the binding pocket, with a dimension of 3, the geometric core of the GPCR binding pocket is often composed of a triangular framework formed by conserved or functionally key residues on three helices. Following step (2.19), a ligand-binding pocket predictor is constructed to score the pocket region for each residue i, yielding... This allows us to obtain the pocket region residues with higher scores and calculate the geometric center coordinates of the binding pocket. This is an overall activation status indicator with a dimension of 1; The global representation of the functional motif obtained in step (2.28) has a dimension of 64; The statistical prior for GPCR structure has a dimension of 128, including the statistical prior contact frequency matrix of the contact probability between every two helices in step (2.17) based on the GPCR database statistics. And in step (2.23), amino acids are counted from the UniProt library. background frequency and the frequency of occurrence at position i This generates a location-specific score matrix (PSSM).

[0129] (2.32) In AlphaFold2, the Evoformer module performs molecular feature injection and GPCR structure prior knowledge injection in parallel. In this embodiment, the entire process of GPCR structure prior knowledge injection is described first, and then the process of molecular feature injection is described.

[0130] See Figure 2 The direction indicated by the red arrow in the structure prediction module is based on the GPCR single-residue features obtained in step (2.29). Based on the original AlphaFold2, the first step is to inject GPCR structural priors into the input layer, which involves combining the MSA representation with the single-residue features generated by the GPCR structural prior encoder. The fusion yields an improved MSA representation initialization for the Evoformer; see [link / reference]. Figure 2 The "MSA representation" in the core improvements of AlphaFold2:

[0131]

[0132] in, This is the original MSA representation; The features of the position-specific scoring matrix obtained in step (2.23) based on evolutionary information from multiple sequence alignment; These are single-residue features obtained based on the target sequence features. , These are learnable fusion weight coefficients.

[0133] Then the Pair representation is compared with the Pair features generated by the GPCR structure prior encoder in step (2.30). The improved Evoformer Pair representation initialization is obtained through integration; see [link / reference]. Figure 2 "Pair representation" in the AlphaFold2 core improvements:

[0134]

[0135] in, This represents the predicted inter-residue distance; Distance binning encoding maps continuous distances to discrete representations; For pairwise information based on sequence and structure; These are learnable fusion weight coefficients.

[0136] (2.33) The intermediate layer is injected with GPCR structural priors. Single residue feature enhancement and pair feature enhancement are performed in the k-th layer (k = 6, 12, 24, 36) of the Evoformer, and global information broadcasting is performed using Cross-Attention to obtain:

[0137]

[0138] (2.34) In the Structure Module of AlphaFold2, the residue coordinates are updated twice using GPCR structure prior guidance and molecular feature guidance. There is no restriction on the order of the two updates. For details of the molecular feature guidance update, see (2.39).

[0139] See Figure 2 The direction indicated by the green arrow in the structure prediction module uses GPCR structure prior guidance in the original AlphaFold2's StructureModule, and incorporates geometric bias (distance geometric prior in step (2.18)) into the IPA attention calculation. The contact probability matrix in step (2.17) Add virtual forces when outputting coordinates in IPA. (Based on distance geometric prior) As a constraint during coordinate updates, specifically, for each residue pair (i, j) to be constrained, their current actual distance and the target distance based on prior knowledge are first calculated. The difference between them is used to construct a potential energy function E related to the distance error, and finally the virtual force is calculated. .

[0140]

[0141]

[0142]

[0143] Where k is a force constant, the magnitude of which controls the strength of the constraint. The coordinates before the update in this step can be the coordinates after the update guided by molecular features. For learning rate, The gradient of the structural loss function with respect to atomic coordinates. These are weighting coefficients used to balance the influence of the two driving forces on the atomic coordinate update. To constrain the actual distance of the residues to the current position, This represents the gradient of the coordinate vector of the i-th residue.

[0144] (2.35) Injected molecular features in AlphaFold2 First, molecular conditionalization of the MSA representation is performed, followed by molecular feature enhancement of the initial MSA embedding to generate position-specific biases for molecular conditions. (Integration of multi-scale features) (Whether the location is in the pocket, membrane depth normalized), broadcast to all MSA sequences (with the same bias applied to each sequence), resulting in molecularly enhanced... See Figure 2 The "MSA representation" is a core improvement in AlphaFold2. Molecular characteristics tell the MSA which amino acids are more suitable for ligand binding. The MSA representation in the pocket region is guided by more molecular information, and different membrane depths have different amino acid preferences based on molecular polarity.

[0145]

[0146] in, This is the original MSA representation. For position encoding, These are learnable fusion weight coefficients.

[0147] (2.36) Add molecular gating to MSA row attention. In each layer l of the Evoformer (48 layers in total), add standard MSA row attention. Then, multi-scale molecular features were used. Gating Modulation, selective enhancement of pocket-related features is obtained The shallow layers (1-16) are injected every 4 layers to capture global ligand constraints; the middle layers (17-32) are injected every 2 layers to refine functional domain matching; and the deep layers (33-48) are injected every layer for precise atomic-level guidance.

[0148]

[0149]

[0150] in, The multi-scale molecular features obtained in step (2.8); For standard MSA line attention; To selectively enhance the attention of the MSA row after highlighting pocket-related features; Multi-scale molecular feature gating; These are learnable parameters; To determine whether amino acid residue i is related to the pocket based on the pocket scoring in step (2.19).

[0151] (2.37) Pair representation of ligand distance constraint. For pocket residue pairs (i, j), molecular characteristics are used to define the ligand distance constraint. Predict the expected inter-residue distance And convert it to radial basis function encoding. Then, molecular induced distance and multi-scale molecular feature bias are added during the pair representation initialization to obtain the molecularly enhanced pair representation. Finally, based on the original Evoformer's Triangle Multiplication update, for the pocket triangle region (the geometric core of the GPCR binding pocket is often composed of a triangular backbone formed by conserved or functionally key residues on three helices), according to multi-scale molecular characteristics... Expected molecular distances on three sides And the pocket volume calculation molecule consistency constraint Mol_triangle_constraint, and weighted addition of Pair representation. See also Figure 2 The "Pair representation" is an improvement in the AlphaFold2 core.

[0152] (2.38) Construct a dedicated ligand-structure interaction module, inserting one ligand-structure interaction module every 8 layers in Evoformer, specifically extracting pocket subgraphs from Pair representations. A graph attention network is used to process pocket features, and then cross-attention is used with molecular features. Interaction And weighted feedback is given back to Pair. See also Figure 2 The direction indicated by the "Cycle 3 times" arrow in the structure prediction module.

[0153] (2.39) Molecular guidance is performed in the original AlphaFold2 Structure Module. First, a molecular geometric bias is added to the original IPA attention to obtain the modified IPA attention. For pocket residues, use molecular characteristics. Predicting the ideal position of the residue (pocket coordinate system) as a virtual force guided by molecular characteristics. Guide coordinate updates:

[0154]

[0155] in, Let the updated residue i be represented; The attention weight represents the importance of residue j to residue i. For the change operations in AlphaFold2, The coordinates before the update can be the updated coordinates guided by GPCR structure priors. This is a rigid body transformation matrix that maps each residue j from local coordinates to global coordinates; Guided by molecular characteristics; Ligand guiding weights.

[0156] (2.40) Based on the improvement of the Structure Module pocket residue prediction in step (2.39), molecular guidance is finally added to the side chain prediction, predicting the side chain rotation angle for each residue. When using molecular enhancement features Correction angles for pocket residues and molecular characteristics guide the determination of which rotational isomer. It is more suitable for ligand binding. Especially for aromatic residues, it adds stereochemical constraints on the side chain direction and predicts the optimal π-π stacking direction.

[0157]

[0158] in, The weights are variable, the step size is updated, and the convergence speed is controlled.

[0159] (2.41) Improvement of AlphaFold2 Structure Module: Adding membrane environment sensing to the standard IPA. First, the single residue representation output from the Evoformer module is obtained through step (2.36), denoted as... By integrating normalized membrane depth, membrane region classification, desired hydrophobicity, and solvent accessibility estimation features, a membrane environment enhancement embedding is obtained. ,and The fusion yields a new single-residue representation:

[0160]

[0161] (2.42) Increase the side chain orientation constraint of the hydrophobic region within the membrane. The side chains of the hydrophobic residues within the membrane should face the membrane core, not the aqueous phase; therefore, a side chain orientation prediction head is added. Based on the results obtained in step (2.41) , Based on the type of residue i, predict the desired side chain orientation. Extract the current sidechain direction Calculate the total sidechain directional loss:

[0162]

[0163] in, For directional weighting coefficients, For dot product operation, This indicates the hydrophobic region inside the membrane.

[0164] (2.43) Increase solvent exposure constraints on extracellular hydrophilic regions. Extracellular / intracellular polar residues should be exposed to solvents, not buried; therefore, solvent accessibility prediction and constraints are performed. Based on step (2.41) , Based on the type of residue i, predict the expected solvent-accessible surface area. Calculate the current solvent-accessible surface area. Total side-chain direction solvent exposure loss:

[0165]

[0166] in, These are the weighting coefficients. This represents the hydrophilic region outside the membrane.

[0167] (2.44) A rigid constraint is introduced for the transmembrane helix. The TM helix should maintain a regular α-helix conformation to avoid excessive distortion; therefore, a helix geometric constraint is added. For each TM helix identified in the GPCR encoder, the soft constraint loss between the current dihedral angle and the ideal dihedral angle of the α-helix is ​​calculated. and ; Calculate the straightness constraint of the helical axis between the distance Cα to the helical axis and the ideal distance. ; Calculate the constraint loss between the current hydrogen bond distance and the ideal hydrogen bond distance. The total helical stiffness loss is obtained by weighting. :

[0168]

[0169] in, for Corner loss weighting coefficient for , for Corner loss weighting coefficient , This is the weighting coefficient for the helical shaft loss. This represents the total loss of the helical shaft. This is the hydrogen bond loss weighting coefficient. This represents the total hydrogen bond loss.

[0170] (2.45) Add multi-state prediction to simultaneously predict the receptor conformation of the active and inactive states. In the StructureModule, IPA is divided into two branches, which predict the active and inactive states based on motif conformation, TM6 conformational differences, G protein binding domain exposure, and molecular features, and use ensemble sampling to generate multiple conformations.

[0171] (2.46) Ligand binding pocket optimization module: Constructing a pocket-molecule matching scorer to encode the pocket's geometry and chemical properties. A pocket recognition and characterization network is constructed, the predicted preliminary structure is voxelized, and spatial features of the pocket are extracted using 3D ConvNeXt blocks to identify possible ligand binding sites, obtaining a complete pocket characterization vector. Molecular features are processed, and the spatial complementarity score (shape matching) and chemical complementarity score (hydrophobic / hydrophilic, hydrogen bond donor-receptor matching) between the molecule and the binding pocket are calculated. An iterative refinement network is constructed to optimize the local structure, mainly optimizing residues in the pocket region, and outputting the optimized structure, pocket information, matching quality, and docking results.

[0172] (2.47) Constraint optimization and loss function design: A multi-objective loss function is used, with the main loss terms including: FAPE (Frame Aligned Point Error) loss: measures the deviation between the predicted atomic coordinates and the actual structure. TM-score loss: evaluates the overall correctness of topological folding. 7-TM conservation loss in GPCR priors: helical tilt angle loss ensures that the angle between each helix and the membrane normal is within a reasonable range; helical spacing loss maintains the distance between adjacent helices at 6-12 Å. Conserved residue distance loss in GPCR priors: distance matrix constraints of known conserved sites (e.g., DRY motif, NPxxY motif). Ligand binding pocket loss: volume constraint loss ensures the pocket volume is between 200-500 Å; hydrophobic / hydrophilic distribution loss matches the polarity characteristics of the target molecule; accessibility loss ensures that the ligand can enter the pocket from the extracellular space. Signal transduction domain loss: G protein binding interface loss maintains the conformation of the intracellular ends of ICL3 and TM6; activation switch loss ensures the activated state characteristics of TM6 moving outward. Physicochemical rationality loss: van der Waals conflict penalty, hydrogen bond network integrity, Ramachandran angle distribution; total side chain orientation loss in supplementary step (2.42): increase side chain orientation constraints in hydrophobic regions within the membrane; total side chain orientation solvent exposure loss in supplementary step (2.43): increase solvent exposure constraints in hydrophilic regions outside the membrane; total helical rigidity loss in supplementary step (2.44): rigid constraints on transmembrane helices, TM helices should maintain a regular α-helical conformation to avoid excessive twisting, therefore helical geometric constraints are added.

[0173] (2.48) See also Figure 2In the output section, the predicted three-dimensional receptor structure (main chain coordinates, spatial constraints) and known constraints (GPCR topology, physicochemical rules, functional motifs) are obtained through the above steps.

[0174] (2.49) See also Figure 2 The sequence generation module performs 3D structural node feature encoding. First, local geometric features are extracted: dihedral angles of main chain residues are calculated, a local coordinate framework is established for each residue, side chain orientation vectors, solvent accessible surface area (SASA), burial density features, membrane environment geometric features (inherited from the GPCR structure prior encoder), local curvature, etc., to obtain comprehensive node features. .

[0175] (2.50) See also Figure 2 The sequence generation module performs three-dimensional structural edge feature encoding. It uses a K-nearest neighbor graph to construct residue interaction maps at three distance scales: local contacts. Each node connects to its 10 nearest neighbors; contact between secondary structures. Connecting nodes in the 10-20 Å range; long-range interactions Connect nodes within the 20-30 Å range. Perform edge feature encoding, calculating the distance and direction, relative orientation, sequence separation degree, and potential interaction type label for each edge to generate a comprehensive edge feature. .

[0176] (2.51) See also Figure 2 The sequence generation module uses a graph neural network encoder to perform node feature extraction. Sum of edge features The message passing, message aggregation, and node updates are processed through a 6-layer GNN to obtain the final structure encoding. The structure encoding L residues of the receptor .

[0177] (2.52) Constructing a spatial constraint and interaction rule encoder. This mainly includes hydrophobicity spatial distribution analysis, hydrogen bond network constraint encoder, salt bridge and electrostatic constraints, aromatic interactions, and finally obtaining comprehensive constraint features. .

[0178] (2.53) Construct GPCR functional motif localizers. Based on the conserved motif library, G protein binding interface and ligand binding pocket residue constraints, the comprehensive encoding of functional motifs is obtained. If residue i is at a certain motif position, the amino acid preference distribution is given; otherwise, a uniform distribution is given.

[0179] (2.54) See also Figure 2The sequence generation module constructs a bidirectional LSTM sequence decoder. The decoder input is the structure encoding of each position obtained in step (2.51). The constraint code obtained in step (2.21) The motif encoding obtained in step (2.53) The fusion features were analyzed, and standard sinusoidal position encoding was added to the bidirectional LSTM. Receptor amino acid sequence decoding was performed using a three-layer stacked BiLSTM and multi-head self-attention enhancement. In the final stage of decoding, constraints were injected a second time to further strengthen the constraints, yielding the final hidden layer representation of residue i. .

[0180] (2.55) Constructing the amino acid prediction head. First, the final hidden layer representation of residue i obtained in step (2.54) is constructed. Basic predictions are obtained Then, considering physicochemical constraints, hard constraints of functional motifs, special treatments for Pro and Gly, and disulfide bond constraints of Cys, the probability distribution of the corresponding amino acid residue i is weighted and adjusted, and the final probability distribution is obtained after Softmax normalization. See Figure 2 Output results section.

[0181] (2.56) Loss function design: Cross-entropy loss evaluates the error between the one-hot encoding of the real amino acid and the model prediction probability; constraint satisfaction loss; motif conservation loss ensures the motif of the prediction result is correct; sequence feasibility regularization avoids unreasonable dipeptides and repetitive sequences.

[0182] (2.57) See also Figure 5 Model Training Strategy: Model training utilizes a high-quality dataset containing "odor molecule-olfactory receptor binding pairs" and a joint loss function. The dataset includes multimodal features of odor molecules, 3D structures of olfactory receptors, amino acid sequences of olfactory receptors, binding activity data, and signal transduction efficiency data. Feature fusion and structure prediction modules are pre-trained on GPCR superfamily datasets, such as ChEMBL and GPCRdb, to transfer GPCR general knowledge. During the fine-tuning phase, end-to-end training is performed using olfactory receptor-specific datasets, starting with simple molecule-receptor pairs and gradually adding more complex ligands. Experiments are conducted on samples with high model uncertainty for validation, and new data is fed back into the training set for iterative model optimization.

[0183] Furthermore, step three is implemented through the following sub-steps:

[0184] (3.1) Molecular docking verification: Using AutoDock software, the target odor molecule was docked with the binding pocket of the designed receptor, and the binding energy and binding mode were calculated. The binding energy was required to be ≤-7.0 kcal / mol, and there must be at least two stable hydrogen bonds between the molecule and the receptor. Simultaneously, the hydrophobic group and the hydrophobic region of the receptor pocket must form an effective hydrophobic interaction. See details... Figure 6 Results of docking examples between the TAAR receptor and amphetamine molecules.

[0185] (3.2) Molecular dynamics simulation: Molecular dynamics simulations of ≥100 ns were performed in aqueous solution or membrane-like simulation environment to analyze the RMSD value of the receptor structure, conformational fluctuations of the binding pocket and the persistence of molecular-receptor interactions to ensure the stability of the binding mode.

[0186] (3.3) Structural rationality assessment: The dihedral distribution of amino acid residues was analyzed using the PROCHECK tool, requiring that the allowable region account for ≥90% of the Ramachandran diagram.

[0187] (3.4) Compatibility assessment: Verify the compatibility of the structure and sequence using the Verify3D tool, requiring a score ≥0.7 to ensure that the designed structure conforms to the natural folding rules of proteins.

[0188] Corresponding to the aforementioned embodiment of a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning, the present invention also provides an embodiment of a device for designing odor molecule-specific olfactory receptors based on multimodal deep learning.

[0189] See Figure 7 The present invention provides a device for designing odor molecule-specific olfactory receptors based on multimodal deep learning, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning as described in the above embodiment.

[0190] The embodiment of the odor molecule-specific olfactory receptor design device based on multimodal deep learning provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 7The diagram shown is a hardware structure diagram of any device with data processing capabilities, which includes a multimodal deep learning-based odor molecule-specific olfactory receptor design device provided by the present invention. (Except for...) Figure 7 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0191] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0192] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0193] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning as described in the above embodiments.

[0194] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0195] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for designing odor molecule-specific olfactory receptors based on multimodal deep learning.

[0196] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0197] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for designing odor molecule-specific olfactory receptors based on multimodal deep learning, characterized in that, include: A multimodal feature dataset of odor molecules is constructed, and molecular physicochemical descriptors, molecular sequence data and molecular structure images of odor molecules are extracted to form a multimodal feature set of odor molecules; Construct and train a multimodal deep learning model for olfactory receptor design: Using the multimodal feature set of odor molecules as input, the multimodal features are fused to obtain multiscale molecular features, and the GPCR structure prior encoder is used to obtain the GPCR structure prior. GPCR structural priors and multi-scale molecular features are input into AlphaFold2, and molecular feature injection and GPCR structural prior knowledge injection are performed in parallel. The molecular feature injection includes: Multi-scale molecular features are fused to enhance the initial MSA embedding, generating position-specific biases for molecular conditions. The desired inter-residue distances are predicted using multi-scale molecular features and converted into radial basis function encodings. Molecular induced distances and multi-scale molecular feature biases are added to the Pair representation initialization to obtain a molecularly enhanced Pair representation. For the geometric center of the GPCR binding pocket, molecular consistency constraints are calculated based on multi-scale molecular features, the expected molecular distances on the three sides, and the pocket volume, and weighted and added to the Pair representation. Ligand-structure interaction modules are inserted every 8 layers in Evoformer to extract pocket subgraphs from the Pair representation. A graph attention network is used to process pocket features, and then cross-attention is used to calculate the attention weights obtained from the interaction with multi-scale molecular features and feed them back to the Pair representation. The GPCR structural prior knowledge injection includes: Injecting GPCR structural priors into the input layer: fusing the MSA representation with the single-residue features generated by the GPCR structural prior encoder as the MSA representation initialization of Evoformer; integrating the Pair representation with the Pair features generated by the GPCR structural prior encoder as the Pair representation initialization of Evoformer; injecting GPCR structural priors into the intermediate layer of Evoformer for single-residue feature enhancement and Pair feature enhancement. The standard invariant point attention computation of AlphaFold2 is divided into two branches to predict the active and inactive states, and multiple conformations are generated using ensemble sampling. Rigid constraints are added to the invariant point attention computation to predict the receptor structure, and the receptor structure and constraints are obtained. The predicted receptor structure and constraints are encoded and decoded and amino acid prediction is performed to obtain the amino acid sequence of the olfactory receptor. The trained model is used to generate receptors, and the generated receptors are validated and evaluated to output qualified olfactory receptors.

2. The method according to claim 1, characterized in that, The molecular physicochemical descriptors include at least 10 of the following: molecular weight, three-dimensional molecular structural features, molecular polarity, lipophilic parameters, number of hydrogen bond donors, number of hydrogen bond acceptors, topological surface area, number of aromatic rings, proportion of hydrophobic groups, and charge distribution; the molecular sequence data includes sequence features converted from the molecular SMILES string and molecular fingerprint data; the molecular structure image is a schematic diagram of the molecular planar structure, including atomic connection methods, chemical bond types, ring structures, and spatial distribution information of functional groups, obtained from the molecular SMILES format string using the RDKit tool.

3. The method according to claim 1, characterized in that, The multimodal feature fusion process includes: using a Transformer encoder with a hierarchical attention mechanism to perform cross-modal fusion of multimodal molecular features; adding a special token to each modal feature as a global representation, aligning the dimensions through linear projection, capturing the correlation information between features through intramodal attention and cross-modal attention, finally extracting a unified molecular feature vector from the global aggregated token, using hierarchical convolutional blocks of ConvNeXt to extract multi-scale features from the molecular feature vector, using parallel convolutional branches with convolutional kernels of different sizes to capture multi-scale information, extracting local and global patterns of molecular features, and performing multi-scale adaptive weight fusion to obtain multi-scale molecular features.

4. The method according to claim 1, characterized in that, The GPCR structural prior includes Pair representation, single residue features, and a global structural prior vector, and the acquisition process includes: The input amino acids are topologically constrained and encoded to extract the amino acid sequence context relationship. The helical positions of G protein-coupled receptors are classified, and the amino acid residues and helical labeling matrix of each helical region are identified based on the classification results. Based on the classification results, predict the helical space parameters and calculate the environmental information by combining the amino acid sequence context relationship; The frequency of amino acid residue positions in homologous sequences in the MSA matrix is ​​calculated to generate position-specific scoring matrix features and conservation score codes; Single-residue features are obtained by integrating amino acid sequence context, environmental information, specificity scoring matrix features, helical annotation matrix, and conservation score encoding. Pair features are obtained by encoding the interaction between helices based on helical space parameters, and Pair representations are obtained by embedding the Pair features after calculating mutual information. Global motif representations are generated based on databases; ligand binding pocket features are predicted using environmental information, amino acid sequence contextual relationships, and multi-scale molecular features. Key motifs were labeled using G protein binding domain labelers to obtain pocket-ligand complementarity scores. The motifs were then state-encoded and the target rotation angle was calculated. The global structural prior vector is obtained by integrating helical spatial parameters, pocket-ligand complementarity scores, pocket geometric centers, global motif representations, and GPCR structural statistical priors.

5. The method according to claim 1, characterized in that, After molecular feature injection and GPCR structural prior knowledge injection, the update process of the Structure Module in AlphaFold2 includes molecular guidance and GPCR structural prior guidance. The two guidance processes are not sequential, and the residue coordinates are updated twice, including: Structure Module Molecular Guidance: Molecular geometric bias is added to the original IPA attention to obtain the modified IPA attention; for pocket residues, multi-scale molecular features are used to predict the ideal position of the residue as a virtual force for molecular feature guidance to guide coordinate updates; in side chain prediction, when predicting the side chain rotation angle for each residue, multi-scale molecular features are used to correct the angle of the pocket residue; for aromatic residues, stereochemical constraints on the side chain direction are added to predict the optimal π-π stacking direction. Use GPCR structural prior guidance in the Structure Module: add geometric bias to the IPA attention calculation; add virtual forces as constraints when updating coordinates in the IPA output coordinates.

6. The method according to claim 1, characterized in that, The encoding / decoding and amino acid prediction of the predicted receptor structure and constraints includes: Based on the three-dimensional structure obtained by AlphaFold2, a three-dimensional structural node encoder is constructed to extract comprehensive node features. A three-dimensional structural edge encoder is constructed, and a K-nearest neighbor graph is used to construct an interaction graph between residues. Edge features are encoded, and the distance and direction, relative orientation, sequence separation degree, and potential interaction type label of each edge are calculated to generate comprehensive edge features. A graph neural network encoder is used to perform geometric encoding on the integrated node features and integrated edge features to obtain the structural encoding of each amino acid residue in the predicted receptor and the global structural encoding; constraint encoding is performed based on spatial constraints and interaction rules. Conserved motif libraries, G protein binding interfaces, and ligand binding pocket residue constraints were used to obtain the comprehensive encoding of GPCR functional motifs. A bidirectional LSTM decoder is used, with the bidirectional LSTM layer employing a three-layer stacked BiLSTM and multi-head self-attention enhancement. It combines structural coding, constraint coding, motif coding, and position coding features for decoding, and obtains amino acid prediction results through the prediction head.

7. The method according to claim 1, characterized in that, The validation and evaluation of the generated receptor includes: Molecular docking verification was performed using AutoDock software to dock the target odor molecule with the binding pocket of the designed receptor and calculate the binding energy and binding mode. Molecular dynamics simulations were performed in aqueous solutions or membrane-like simulation environments for ≥100 ns to analyze the RMSD value of the receptor structure, conformational fluctuations of the binding pocket, and the persistence of molecular-receptor interactions. Structural rationality assessment was conducted by analyzing the dihedral distribution of amino acid residues using the PROCHECK tool; Compatibility assessment was conducted using the Verify3D tool to verify the compatibility of the structure and sequence.

8. A device for designing odor molecule-specific olfactory receptors based on multimodal deep learning, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning as described in any one of claims 1-7.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for designing odor molecule-specific olfactory receptors based on multimodal deep learning as described in any one of claims 1-7.

Images