A multimodal molecular representation prediction method
Through multimodal molecular representation learning methods, combining sequences, graphs and conformational graphs, the information loss and redundancy problems of single-modal molecular representation in existing technologies are solved, a more comprehensive capture of molecular features is achieved, and the accuracy and efficiency of drug discovery are improved.
Patent Information
- Application Number
- CN202311456933.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2025-10-03
- Estimated Expiration
- 2043-11-03
AI Technical Summary
In the drug discovery process, existing molecular representation learning methods only focus on a single modality, resulting in missing or redundant information and an inability to fully capture molecular features, affecting the accuracy and efficiency of drug discovery.
A multimodal approach based on sequences, graphs, and molecular conformations is adopted to construct a dataset, obtain molecular representations, introduce a global attention pooling layer and an inter-modal comparison mechanism, combine it with a multi-layer perceptron to predict molecular properties, and optimize the model to obtain the optimal molecular representation.
It improves the accuracy and consistency of molecular representation prediction, reduces information redundancy, enhances the ability to capture molecular features, and improves the efficiency and accuracy of drug discovery.
Smart Images
Figure CN117292764B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug discovery, and more particularly to a multimodal-based molecular representation prediction method. Background Art
[0002] Drug discovery is a challenging and costly endeavor using traditional experimental methods. Candidate compounds can only be used in drug design after undergoing comprehensive testing of their physical, biological, and chemical properties. Therefore, molecular property prediction is a critical step in the drug discovery and design process, involving the determination of the biochemical characteristics of unknown compounds. In recent years, with the rapid development of artificial intelligence (AI), its application in drug discovery and design has garnered widespread attention due to its potential to streamline the discovery cycle and effectively reduce costs. Within this field, molecular property prediction methods based on deep learning have demonstrated remarkable performance. Currently, there are many data-driven molecular representation learning methods, which can be categorized into three main categories: sequence-based, graph-based, and molecular conformation-based methods.
[0003] In sequence-based molecular representation learning methods, molecules can be represented using SMILES (similates) with ASCII strings. Sequence-based methods, primarily related to natural language processing, extract features from each string to capture broader global information. While they can provide a unique encoding for each molecule, their representational power is limited by the lack of molecular topology information. Alternatively, molecules can be represented as a graph, where nodes represent atoms and edges represent valence bonds. Molecular graphs provide molecular topology, such as atomic connectivity and the number and size of rings. Compared to sequence-based methods, graph-based molecular representation learning methods effectively exploit molecular topology. For example, graph neural network methods extract molecular representations by aggregating neighborhood information and have proven effective in tasks such as molecule generation and molecular property prediction. However, graph-based methods may encounter the problem of having similar topologies but different properties. Furthermore, molecular graphs do not include important information such as chirality and molecular conformation. Faced with these challenges, many existing methods have turned to conformation-based molecular representations. However, these conformation-based methods focus more on molecular information related to molecular conformation, which is generated by chemistry tools such as RDKit. Therefore, utilizing different data types can yield more comprehensive molecular representations. In addition, some methods obtain molecular representations by combining two modalities, such as GraSeq, which utilizes sequence and graph modalities, but ignores molecular conformational information, and GeomGCL, which uses molecular conformational modality as an enhanced view, but ignores the chirality information contained in the sequence modality.
[0004] Therefore, it is an urgent problem for those skilled in the art to propose a multimodal-based molecular representation prediction method to solve the difficulties existing in the prior art. Summary of the Invention
[0005] In view of this, the present invention provides a multimodal molecular representation prediction method, which uses molecular representations based on three modalities: sequence, graph, and molecular conformation graph to more comprehensively capture molecular features and information.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] A multimodal molecular representation prediction method comprises the following steps:
[0008] S1. Build a dataset of drug molecules and obtain the sequence representation of each molecule in the dataset;
[0009] S2, obtain a molecular graph representation of each molecule in the dataset;
[0010] S3, obtaining a molecular conformational representation of each molecule in the dataset;
[0011] S4, introduce the sequence representation, molecular graph representation, and molecular conformation graph representation of each molecule obtained in S1-S3 into a global attention pooling layer to obtain a molecular-level representation;
[0012] S5. For the molecular-level representation of different modalities obtained in S4, a similarity-based comparison mechanism is introduced between modalities to calculate the total comparison loss between different modalities;
[0013] S6. Design a fusion layer to obtain a joint representation of molecules based on the molecular-level representations of different modalities obtained in S4;
[0014] S7. For the molecular joint representation obtained in S6, a molecular attribute predictor is introduced, and the optimization model is trained by calculating the mixed supervised signal loss and the total contrast loss calculated in S5 to obtain the optimal model to obtain the final molecular representation.
[0015] Optionally, S1 specifically includes:
[0016] The input molecule is converted into a SMILES sequence, and the SMILES sequence is mapped into a feature vector using one-hot encoding. The one-hot encoding is then preprocessed using a Bi-LSTM unit to capture the SMILES contextual information representation. The sequence encoding representation is input into the transformer to learn the node-level representation of SMILES.
[0017] Optionally, S2 specifically includes:
[0018] The input molecules are converted into molecular graphs, and a multi-layer CMPNN is introduced to obtain atomic representations; the CMPNN module consists of two parts, the first part includes a message enhancer, and the second part uses a multi-layer perceptron.
[0019] Optionally, S3 specifically includes:
[0020] The input molecules are converted into molecular conformation graphs, GEMGNN is introduced, and the sum function in the original GIN is replaced by GINE to obtain the node representation of each input conformation graph.
[0021] Optionally, S5 specifically includes:
[0022] Input comes from sequence-, graph-, and conformation-based Figure 3 The numerator of the mode is H s 、H g and H h , the three molecular representations are mapped to the spatial representation Z for contrast mechanism through a common projection head s , Z g and Z h .
[0023] Optionally, S5 also includes: introducing NTXent contrast loss for each modality, and expressing the total contrast loss as:
[0024]
[0025] in, is the similarity calculation function, and T is the contrast loss ratio.
[0026] Optionally, S6 specifically includes: inputting and fusing data from sequence-based, graph-based, and conformation-based Figure 3 The numerator of the mode is H s 、H g and H h :
[0027] H fusion =W s ·H s +W g ·H g +W h ·H h +b
[0028] Among them, W s 、W g and W h is the learnable weight and b is the bias vector.
[0029] Optionally, S7 specifically includes:
[0030] Input fusion representation H fusion And introduce a multilayer perceptron as a molecular property predictor to obtain the predicted label y pred , the final joint loss function can be defined as:
[0031]
[0032] Where α is used to balance different losses, is the loss function between the predicted label and the ground truth.
[0033] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a multimodal molecular representation prediction method with the following beneficial effects:
[0034] (1) Existing technology solutions do not have a molecular representation learning model that simultaneously targets sequences, graphs, and conformational graphs, and can only focus on learning one modality of molecular representation. The present invention can utilize molecular representations of three modalities to more comprehensively capture molecular features and information, thereby improving the accuracy of molecular representation prediction.
[0035] (2) The present invention utilizes encoders based on sequences, graphs, and molecular conformations when acquiring molecular representations, and introduces a training mechanism based on contrastive learning and multimodal representations to maximize the similarity between representations of the same molecule and minimize the similarity between different molecular representations, thereby ensuring consistency between different modalities and reducing information redundancy. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0037] Figure 1 A flow chart of a multimodal molecular representation prediction method provided by the present invention;
[0038] Figure 2 A framework diagram of a multimodal molecular representation prediction method provided by the present invention;
[0039] Figure 3 Obtaining schematic diagrams for different modal representations of the present invention;
[0040] Figure 4 Schematic diagram of similarity learning for different modal representations of the present invention;
[0041] Figure 5 Schematic diagram of the present invention for the representation and model optimization of the fusion of different modalities. DETAILED DESCRIPTION
[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0043] Reference Figure 1 As shown, the present invention discloses a molecular representation prediction method based on multimodality, comprising the following steps:
[0044] S1. Build a dataset of drug molecules and obtain the sequence representation of each molecule in the dataset;
[0045] S2, obtain a molecular graph representation of each molecule in the dataset;
[0046] S3, obtaining a molecular conformational representation of each molecule in the dataset;
[0047] S4, introduce the sequence representation, molecular graph representation, and molecular conformation graph representation of each molecule obtained in S1-S3 into a global attention pooling layer to obtain a molecular-level representation;
[0048] S5. For the molecular-level representation of different modalities obtained in S4, a similarity-based comparison mechanism is introduced between modalities to calculate the total comparison loss between different modalities;
[0049] S6. For the molecular-level representations of different modalities obtained in S4, a fusion layer is designed to obtain the joint representation of molecules, such as Figure 2 As shown;
[0050] S7. For the molecular joint representation obtained in S6, a molecular attribute predictor is introduced, and the optimization model is trained by calculating the mixed supervised signal loss and the total contrast loss calculated in S5 to obtain the optimal model to obtain the final molecular representation.
[0051] Furthermore, S1 includes: the SMILES sequence of the input molecule S = {s i |i=1,…,s T}, where T represents the length of the SMILES sequence. The present invention uses one-hot encoding to map the SMILES sequence into a feature vector Use Bi-LSTM unit to preprocess the one-hot encoding as follows Figure 3 As shown, and captures SMILES context information:
[0052]
[0053] in and is a vector s i The hidden state of Can be spliced and The sequence encoding can be input into the transformer to learn the node-level representation of SMILES.
[0054] Specifically, S2 includes: input molecular graph G = (V, E), where V represents the set of atoms and E represents the set of valence bonds; the present invention introduces a multi-layer CMPNN to obtain atomic representations, and the CMPNN module includes two parts: AGGREGATE (·) and COMMUNICATE (·). The former part includes a message enhancer for generating the maximum pooling result of the sum of the edge hidden representations and calculating the element product thereof, and the latter part adopts a multi-layer perceptron, and the node hidden representation and edge hidden representation The k-th propagation process is as follows:
[0055]
[0056]
[0057]
[0058] in represents the message received by node v, W is the learnable weight matrix, and σ is the activation function; after L iterations, the following operations are performed to obtain the final message and node representation:
[0059]
[0060]
[0061] S3 includes: the conformational graph of the input molecule H = (V, E, A), where V represents a node, E represents an edge, and A represents a bond angle; the present invention introduces GEMGNN and replaces the AGGREGATE(·) and COMBINE(·) functions in the original GIN with GINE to obtain the node representation of each input conformational graph, and the node representation of the kth iteration is It can be expressed as follows:
[0062]
[0063]
[0064] Edge representation of the kth iteration It can be expressed as follows:
[0065]
[0066]
[0067] in and are the neighborhoods of nodes v and u respectively.
[0068] S4 includes: input node-level representation h based on sequence, graph and conformation s 、h g and h h , the present invention introduces a global attention pooling layer to obtain the molecular level representation H of the three modalities s 、H g and H h .
[0069] S5 includes: input from sequence-based, graph-based and conformation-based Figure 3 The numerator of the mode is H s 、H g and H h , the three molecular representations are mapped to the spatial representation Z for contrast mechanism through a common projection head s , Z g and Z h The contrast loss of NTXent is introduced for each modality as Figure 4 As shown, the total contrast loss is expressed as:
[0070]
[0071] in, is the similarity calculation function, and T is the contrast loss ratio.
[0072] S6 includes: input and fusion of sequence-based, graph-based and conformation-based Figure 3 The numerator of the mode is H s 、H g and H h like Figure 5 As shown:
[0073] H fusion =W s ·H s +W g ·H g +W h ·H h +b
[0074] Among them, W s 、W g and W h is the learnable weight and b is the bias vector.
[0075] S7 includes: input fusion representation H fusionAnd introduce a multilayer perceptron as a molecular property predictor to obtain the predicted label y pred ; For different categories of tasks, the loss function between the predicted label and the ground truth Differently, for classification tasks, this paper selects the BCELoss loss function, and for regression tasks, it selects the MSELoss loss function. The final joint loss function can be defined as:
[0076]
[0077] Among them, α is used to balance different losses and optimize the model parameters according to the loss calculation results.
[0078] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.
[0079] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A molecular representation prediction method based on multimodality, characterized in that: The following steps are involved: S1. Build a dataset of drug molecules and obtain the sequence representation of each molecule in the dataset; S2, obtain a molecular graph representation of each molecule in the dataset; S3, obtaining a molecular conformational representation of each molecule in the dataset; S4, introduce the sequence representation, molecular graph representation, and molecular conformation graph representation of each molecule obtained in S1-S3 into a global attention pooling layer to obtain a molecular-level representation; S5. For the molecular-level representation of different modalities obtained in S4, a similarity-based comparison mechanism is introduced between modalities to calculate the total comparison loss between different modalities; S6. Design a fusion layer to obtain a joint representation of molecules based on the molecular-level representations of different modalities obtained in S4; S7, for the molecular joint representation obtained in S6, introduce a molecular attribute predictor, and train the optimization model by calculating the mixed supervised signal loss and the total contrast loss calculated in S5 to obtain the optimal model to obtain the final molecular representation; S1 specifically includes: converting the input molecule into a SMILES sequence, mapping the SMILES sequence into a feature vector using one-hot encoding, and then preprocessing the one-hot encoding using a Bi-LSTM unit to capture the SMILES context information representation; inputting the sequence encoding representation into the transformer to learn the node-level representation of SMILES; S2 specifically includes: converting the input molecule into a molecular graph and introducing a multi-layer CMPNN to obtain atomic representation; the CMPNN module consists of two parts, the first part includes a message enhancer, and the second part uses a multi-layer perceptron; S3 specifically includes: converting the input molecules into molecular conformation graphs, introducing GEMGNN, and replacing the sum function in the original GIN with GINE to obtain the node representation of each input conformation graph; S5 specifically includes: input from the molecular representation H based on three modalities: sequence, graph and conformation graph s 、H g and H h , the three molecular representations are mapped to the spatial representation Z for contrast mechanism through a common projection head s , Z g and Z h ; S5 also includes: introducing the NTXent contrast loss for each modality, and expressing the total contrast loss as: in, is the similarity calculation function, and T is the contrast loss ratio.
2. A multimodal molecular representation prediction method according to claim 1, characterized in that: S6 specifically includes: inputting and fusing molecular representations H based on sequence, graph, and conformation graph s 、H g and H h : H fusion =W s ·H s +W g ·H g +W h ·H h +b Among them, W s 、W g and W h is the learnable weight and b is the bias vector.
3. A multimodal molecular representation prediction method according to claim 1, characterized in that: S7 specifically includes: input fusion representation H fusion And introduce a multilayer perceptron as a molecular property predictor to obtain the predicted label y pred , the final joint loss function is defined as: Among them, α is used to balance different losses, is the loss function between the predicted label and the ground truth.