Molecular property prediction method based on geometric feature encoding graph attention network

Abstract

The present invention proposes a molecular property prediction method based on a geometric feature coding graph attention network, comprising the following steps: Step 1, constructing a molecular graph G with atoms as nodes based on the two-dimensional and three-dimensional structural information of the molecule to be predicted a and a molecular graph G with chemically bonded atomic pairs as nodes b ; Step 2, by encoding the two-dimensional and three-dimensional features of the molecule, the node feature vectors of the two types of molecular graphs are generated; Step 3, the molecular graph node features obtained in step 2 are aggregated and updated through the graph neural network guided by the multi-head self-attention mechanism; Step 4, the molecular graph G a and G b The geometric information of the three-dimensional structure is encoded and fused; Step five, the molecular graph node features obtained in Step four are merged, and the molecular graph G is pooled using a fully connected network and graph pooling operation. a and G b The node-level feature vectors of the molecule are globally aggregated to generate graph-level features of the molecular graph and predict the properties of the target molecule.

Description

Technical Field

[0001] The present invention is directed to the field of artificial intelligence technology, specifically a molecular property prediction method based on a geometric feature coding graph attention network, which has good application prospects in drug design, chemical calculations, and material screening. Background Art

[0002] Scientific research in the fields of biology and chemistry usually requires processing large amounts of computational and experimental data, especially in research areas such as drug development and chemical synthesis. Faced with the high-throughput calculation and screening of large amounts of chemical molecular data, in order to minimize computational costs, researchers use machine learning and other methods to perform molecular characterization learning to achieve high-throughput molecular screening. For example, molecular descriptors are used to represent molecular features. Common descriptors include (1) molecular fingerprints: by splitting molecules into specific substructures and functional groups and using binary encoding to represent molecules, and (2) SMILES strings: by expressing molecular features such as elements and chemical bonds in the form of strings. Therefore, traditional machine learning algorithms require a lot of feature engineering to construct molecular descriptors, which is not conducive to high-throughput screening of large-scale data.

[0003] In recent years, with the rapid development of big data and artificial intelligence technologies, deep neural networks, as a mainstream artificial intelligence technology, have garnered widespread attention in areas such as assisted drug development and chemical catalytic reaction screening. However, related research has primarily focused on migrating deep learning methods from classic image processing and natural language processing to represent molecular data using one-dimensional sequences or two-dimensional molecular graph structures. However, the properties of actual drug molecules are highly complex, and the spatial structural characteristics of molecules are closely related to their physicochemical properties. Deep learning models for molecular property prediction should not only fully consider the sequence and graph structure information of molecules, but also encode the three-dimensional spatial structural characteristics of molecules to further improve their performance in molecular property prediction. Summary of the Invention

[0004] The present invention provides a molecular property prediction method based on a geometric feature encoding graph attention network. By constructing two types of molecular graphs, a multi-head attention mechanism is used to guide the aggregation and update of molecular graph node feature vectors and fuse the geometric features of the molecular three-dimensional structure, thereby improving the accuracy of molecular property prediction.

[0005] To achieve molecular property prediction, the present invention adopts the following technical solutions:

[0006] Step 1: Based on the two-dimensional and three-dimensional structural information of the molecule to be predicted, a molecular graph G with atoms as nodes is constructed. a and a molecular graph G with chemically bonded atomic pairs as nodes b , where the molecular graph Gb Each node consists of a chemical bond and two atoms at its endpoints, and the node information should also include the spatial position coordinates of the atoms.

[0007] Step 2: Generate node feature vectors of two types of molecular graphs by encoding the two-dimensional and three-dimensional features of the molecules. a The atomic number, atomic formal charge, atomic hybridization mode and the number of atomic chemical bonds of the node are encoded through the embedding layer to form the two-dimensional features of the node, and the distance information between the node atoms is encoded using the Gaussian kernel function to obtain the three-dimensional features of the node; the molecular graph G b The chemical bond type of the node, the atomic number at both ends, and the number of chemical bonds connecting the endpoint atoms are encoded through the embedding layer to form the two-dimensional features of the node. The Gaussian kernel function is used to convert the shortest distance D between the nodes into b The information is encoded to obtain the three-dimensional features of the node.

[0008] Step 3: Through the graph neural network guided by the multi-head self-attention mechanism, the molecular graph node features obtained in step 2 are aggregated and updated. a The shortest path information between node atoms and the spatial distance information between atoms are feature encoded as part of the graph attention parameters; the molecular graph G b The shortest path information and shortest distance D between the chemical bonds of the nodes b The information is feature encoded as part of the graph attention feature.

[0009] Step 4: the molecular graph G a and G b The geometric information of the three-dimensional structure is encoded, and the geometric features are fused with the molecular graph node features obtained in step 3 by using cross multiplication and residual connection. a The Euclidean distance between node atoms is used as a geometric feature encoding; the molecular graph G b The spatial angle between chemical bonds at nodes is encoded as a geometric feature.

[0010] Step 5: Merge the molecular graph node features obtained in step 4, and use the fully connected network and graph pooling operation to a and G b The node-level feature vectors of the molecule are globally aggregated to generate graph-level features of the molecular graph and predict the properties of the target molecule.

[0011] Among them, for step 2, further including the case where the molecular data lacks two-dimensional structure information such as atomic form charge, chemical bond, etc., according to the three-dimensional structure of the molecular data, the interatomic distance information is arranged in ascending order, and the minimum atomic pairs containing all atoms are selected, and these atomic pairs are used as the endpoints of the same type of chemical bonds, and then the molecular graph G can be constructed using step 1 b .

[0012] For step 3, the shortest path information between nodes of the two-dimensional molecular graph G is obtained by using the Floyd algorithm. a The shortest path information of a node is the atomic type information of the node on the path; the molecular graph G b The shortest path information of a node is the atomic type and chemical bond type information of the node on the path. The node features, the shortest path between nodes, and the distance information described in step 3 are added together through a learnable weight matrix to obtain the self-attention coefficient between the nodes in the molecular graph.

[0013] For step 4, further including the molecular graph G b The spatial angle between the chemical bonds of the nodes is the shortest distance D between the nodes b The corresponding atom pairs are used as common endpoints. a The inter-node atom pair distances and molecular graph G b The chemical bond angle information between nodes is obtained by using basis functions and learnable weight matrices for feature encoding to obtain high-dimensional geometric feature vectors.

[0014] Compared with the prior art, this application has the following beneficial effects:

[0015] i) Constructing two types of molecular graphs can obtain characteristic information of molecular graphs from the two aspects of atoms and chemical bonds, and more comprehensively encode the two-dimensional and three-dimensional structural information of molecular graphs.

[0016] ii) Using a multi-head self-attention mechanism to aggregate and update molecular graph node features, the two-dimensional topological information and three-dimensional spatial position information of the molecular graph can be used to guide node feature aggregation, thereby improving the effect of molecular graph node feature aggregation.

[0017] iii) Encoding the spatial geometric features of the molecular three-dimensional structure and further integrating the three-dimensional structural information of the molecule can improve the accuracy of molecular property prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A schematic flow chart of a molecular property prediction method based on a geometric feature encoding graph attention network provided in an embodiment of the present invention.

[0019] Figure 2A model structure diagram of the geometric feature encoding graph attention network provided in an embodiment of the present invention.

[0020] Figure 3 A schematic diagram of the method for generating customized nodes in a molecular graph provided by the present invention. DETAILED DESCRIPTION

[0021] The following describes exemplary embodiments of the present disclosure in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0022] The process of the molecular property prediction method based on the geometric feature coding graph attention network provided by the embodiment of the present invention is as follows: Figure 1 As shown in FIG, the molecular property prediction method based on the geometric feature encoding graph attention network includes the following steps:

[0023] S101 generates a molecular graph G with atoms as nodes based on the SMILES string of the molecule to be predicted a and a molecular graph G with chemically bonded atomic pairs as nodes b , where the molecular graph G a The edges are chemical bonds; the molecular graph G b The edges are atoms shared between the endpoints of the chemical bonds.

[0024] For molecular data in the field of molecular dynamics research, there are cases where only three-dimensional structural information is included but two-dimensional structural information such as chemical bonds and atomic features is missing. b In the process of customizing the molecular graph G according to the three-dimensional structure of the molecule b It is generally believed that the smaller the distance between atoms in a molecule, the more obvious the interaction and the higher the possibility of chemical bonds. Therefore, nodes can be constructed based on the information of the distance between atoms in the three-dimensional structure of the molecule. In most cases, the number of atoms in a molecule is close to the number of chemical bonds, such as Figure 3 As shown, here the atomic adjacency distance matrix of the molecule is arranged in ascending order, and the minimum number of atomic pairs containing all the atoms of the molecule is selected as the molecular graph G b Node.

[0025] S102 generates node features of two types of molecular graphs according to the two-dimensional and three-dimensional structural information of the molecules.

[0026] For molecular graph G aFor each node in the graph, the embedding layer encodes the node's atomic number, atomic formal charge, atomic hybridization mode, and the number of chemical bonds connecting atoms to form the node's two-dimensional features. The spatial distance between the node and other node atoms in the molecular three-dimensional structure is then encoded to form the node's three-dimensional features. The two are added together to obtain the molecular graph node features. The two-dimensional features of the node are expressed as: Indicates node atomic information, Deg i N is the number of chemical bonds connecting the nodes. a is the number of nodes. The three-dimensional characteristics of the nodes of the three-dimensional molecular structure are: W1, W2, W3 are learnable weight matrices, GELU() is the activation function, N i Representation node The complement of the molecular graph nodes, Represents the molecular graph G a The distance between nodes i and j.

[0027] For molecular graph G b For each node in the embedding layer, the chemical bond type, the atomic numbers at both ends of the chemical bond, and the number of chemical bonds with common endpoint atoms are feature-encoded to form the two-dimensional features of the node. The shortest distance between nodes in the three-dimensional structure of the molecule is then feature-encoded to form the three-dimensional features of the node. The two are added together to obtain the node features. The two-dimensional features of the node are: is the node chemical bond type, and Indicates the atomic information at both ends of the chemical bond. b The shortest distance D between nodes b The definition is as follows: Any node and It can be expressed as and B and A represent the chemical bond and endpoint atoms of the node respectively. The shortest distance between nodes is D b is the minimum spatial distance between the endpoint atoms of the node, and its expression is: R is the spatial position coordinate of the endpoint atom A of the node, and the shortest distance between nodes with a common endpoint atom is zero.

[0028] For the case where molecular data lacks two-dimensional structural information such as atomic form charge, the node features of the molecular graph only retain the atomic number information. a The node characteristics are Molecular Graph G b Node Features

[0029] The molecular graph node feature vector aggregation update based on the attention mechanism guidance described in S103 is as follows Figure 2 As shown in Figure 2, the two-dimensional and three-dimensional features of the molecular graph are combined with the multi-head self-attention mechanism of the Transformer model to aggregate and update the molecular graph node features. The model expression is: n =AttnG(LN(M n ))+M n , M n+1 =FNN(LN(X n ))+X n , n represents the number of layers in the network, M n is the molecular graph node feature of the network input, X n Represents the molecular graph node features aggregated by the attention mechanism, M n+1 is the molecular graph node feature output by the network, LN() represents the normalization function, and FFN() represents the feedforward neural network.

[0030] The molecular graph node feature aggregation model AttnG() based on the multi-head self-attention mechanism can be expressed as: H G Represents the geometric characteristics of the three-dimensional structure of the molecule. Attn() represents the fusion molecule Figure 2 The multi-head self-attention model for dimensional and 3D features is expressed as follows: Q=W Q M,K=W K M,V=W V M, M is the molecular graph node feature, W Q ,W K ,W V is the learnable weight matrix, H P is the encoding feature of the shortest path between nodes of the molecular two-dimensional structure, H D The shortest path information in a two-dimensional molecular structure reflects the local structural characteristics between molecular nodes, while the shortest distance information in a three-dimensional molecular structure reflects the degree of interaction between nodes. Therefore, combining these features with the self-attention mechanism can better guide the aggregation and update of molecular graph node features, improving model performance.

[0031] For molecular graph G a , get the node through Floyd algorithm and The node set with the shortest path between P N is the step length of the shortest path, and the node features and the number of nodes on the shortest path are encoded to form the shortest path feature: Representation node The features of , W1 and W2 are learnable weight matrices. and The spatial distance is R represents the spatial coordinates of the node. Since the distance information between nodes is discontinuous, in order to avoid overfitting during model training, the Gaussian kernel function is used to encode the distance information between atoms to form multidimensional features: μ, σ, and W are learnable parameters and weight matrices, and S represents the feature dimension.

[0032] For molecular graph G b ,node and The shortest path between nodes The chemical bond and atomic information contained in the node can be expressed as:

[0033]

[0034] Where A and B represent the molecular graph G b The atoms and chemical bonds contained in the nodes. Since there are common endpoint atoms between adjacent nodes in the shortest path, the nodes and The shortest path between can be expressed as:

[0035] The nodes The endpoint atom and nodes The endpoint atom Therefore, the shortest path feature encoding between overlapping nodes is expressed as: in and B pt Represents a node on the shortest path The shortest distance D between nodes b Gaussian kernel function is also used to encode and form multidimensional features

[0036] When molecular data lacks two-dimensional structural information such as chemical bonds, the molecular graph G is constructed by screening the distance between atoms. b , the feature encoding of the shortest path between molecular graph nodes only retains the shared endpoint atom information, and the expression is:

[0037] S104 molecular graph geometric feature encoding is to encode the spatial information of the molecular three-dimensional structure into geometric features H in the process of molecular graph attention-guided node feature aggregation. G Fusion with molecular graph node features further improves the performance of molecular graph attention network.

[0038] Molecular Graph G a The geometric information is the Euclidean distance D between node atoms a In order to avoid the overfitting phenomenon caused by the discontinuity of distance information, radial basis function is used to encode the distance information to form geometric features. The multidimensional features obtained by radial basis function are expressed as: Where S represents the feature dimension. The encoded geometric features are expressed as: Among them, γ, μ k is a learnable parameter, W1 φ and is the learnable weight matrix.

[0039] Molecular Graph G b The geometric information is the spatial angle E between the chemical bonds of the nodes b , with the shortest distance D between nodes b The corresponding endpoint atom pairs are used as common endpoints between the node chemical bonds to obtain the corresponding angle information. and The shortest distance between When the corresponding node chemical bond space angle is R represents the spatial coordinates of the endpoint atoms of the node. In order to avoid the overfitting problem that may be caused by the discontinuity of angle information, the angle information is encoded using basis functions to form multidimensional features. The multidimensional features obtained by the basis function are expressed as Where S represents the feature dimension, and the geometric features are encoded as:

[0040] Geometric characteristics of molecular three-dimensional structure H G The molecular graph node features aggregated by the multi-head attention mechanism are fused. The expression of this process is: X is the molecular graph node feature output by the multi-head self-attention model Attn(), represents the cross product operation, W G is the learnable weight matrix.

[0041] S105 targets molecular graph G a and G b The multi-dimensional feature vector of the node after aggregation by the graph attention network is firstly pooled to obtain the global graph-level features of the molecular graph: Pool() represents the graph pooling function, X N Aggregate the updated node features for the molecular graph, W1 and W2 are learnable weight matrices, and LeakyReLU() is the activation function. Then merge the graph-level features of the two molecular graphs: in Represents the molecular graph Ga and G b The graph-level feature vector, W G is a learnable weight matrix; finally, the weight matrix and activation function are used to reduce the dimensionality of the global features of the molecular graph to predict the properties of the target molecule.

[0042] The molecular property prediction method based on the geometric feature encoding graph attention network provided in the above embodiment encodes the two-dimensional and three-dimensional structural information of the molecule based on the multi-head self-attention mechanism, and further encodes the distance and angle information of the three-dimensional molecular structure, thereby improving the prediction effect of the chemical physics, quantum chemistry and other properties of the molecule.

[0043] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A molecular property prediction method based on a geometric feature encoding graph attention network, characterized by: The steps include: Step 1: Based on the two-dimensional and three-dimensional structural information of the molecule to be predicted, a molecular graph G with atoms as nodes is constructed. a and a molecular graph G with chemically bonded atomic pairs as nodes b ; Among them, the molecular graph G b Each node consists of a chemical bond and two atoms at its endpoints. The node information should also include the spatial position coordinates of the atoms. Step 2: Generate node feature vectors of two types of molecular graphs by encoding the two-dimensional and three-dimensional features of the molecules. a The atomic number, atomic formal charge, atomic hybridization mode and the number of chemical bonds connecting atoms of the node are encoded through the embedding layer to form the two-dimensional features of the node, and the distance information between the node atoms is encoded using the Gaussian kernel function to obtain the three-dimensional features of the node; the molecular graph G b The chemical bond type of the node, the atomic number at both ends, and the number of chemical bonds connecting the endpoint atoms are encoded through the embedding layer to form the two-dimensional features of the node. The Gaussian kernel function is used to convert the shortest distance D between the nodes into b The information is encoded to obtain the three-dimensional features of the node; Step 3: Through the graph neural network guided by the multi-head self-attention mechanism, the molecular graph node features obtained in step 2 are aggregated and updated. a The shortest path information between node atoms and the spatial distance information between atoms are feature-encoded as part of the graph attention parameters; the molecular graph G b The shortest path information and shortest distance D between the chemical bonds of the nodes b The information is feature encoded as part of the graph attention feature; Step 4: For the molecular graph G a and G b The geometric information of the three-dimensional structure is encoded, and the geometric features are fused with the molecular graph node features obtained in step 3 by using cross multiplication and residual connection. a The Euclidean distance between node atoms is used as a geometric feature encoding; the molecular graph G b The spatial angle between chemical bonds of nodes is encoded as a geometric feature; Step 5: Merge the molecular graph node features obtained in step 4, and use the fully connected network and graph pooling operation to a and G b The node-level feature vectors of the molecule are globally aggregated to generate graph-level features of the molecular graph and predict the properties of the target molecule. Among them, in step 4, the molecular graph geometric feature encoding is to encode the spatial information of the molecular three-dimensional structure to form the geometric feature H in the process of molecular graph attention-guided node feature aggregation. G Fusion with molecular graph node features to improve the performance of molecular graph attention networks; Molecular Graph G a The geometric information is the Euclidean distance D between node atoms a , radial basis function is used to encode the distance information to form geometric features; the multidimensional features obtained by radial basis function are expressed as: Where S represents the feature dimension; the encoded geometric features are expressed as: in, and is a learnable weight matrix; Molecular Graph G b The geometric information is the spatial angle E between the chemical bonds of the nodes b , with the shortest distance D between nodes b The corresponding endpoint atom pairs are used as common endpoints between the node chemical bonds to obtain the corresponding angle information; when the node and The shortest distance between When the corresponding node chemical bond space angle is R represents the spatial position coordinates of the endpoint atoms of the node. In order to avoid the overfitting problem caused by the discontinuity of angle information, the angle information is encoded using basis functions to form multidimensional features. The multidimensional features obtained by the basis functions are expressed as: Where S represents the feature dimension, and the geometric features are encoded as: Geometric characteristics of molecular three-dimensional structure H G The molecular graph node features aggregated by the multi-head attention mechanism are fused. The expression of this process is: X is the molecular graph node feature output by the multi-head self-attention model Attn(), represents the cross product operation, W G is the learnable weight matrix.

2. The molecular property prediction method based on the geometric feature encoding graph attention network according to claim 1, characterized in that: In step 2, if the molecular data lacks a two-dimensional structure including atomic formal charges and chemical bonds, the interatomic distance information is sorted in ascending order based on the three-dimensional structure of the molecular data, and the minimum atomic pairs containing all atoms are selected. These atomic pairs are used as the endpoints of the same type of chemical bonds, and then the molecular graph G is constructed using step 1. b .

3. The molecular property prediction method based on the geometric feature encoding graph attention network according to claim 1 is characterized in that: In step 2, for the molecular graph G a For each node in , the two-dimensional feature of the node is represented as: Indicates node atomic information, Deg i N is the number of chemical bonds connecting the nodes. a is the number of nodes; the three-dimensional characteristics of the nodes of the three-dimensional molecular structure are: W1, W2, W3 are learnable weight matrices, GELU() is the activation function, N i Representation node The complement of the molecular graph nodes, Represents the molecular graph G a The distance between nodes i and j.

4. The molecular property prediction method based on the geometric feature encoding graph attention network according to claim 1, characterized in that: In step 2, for the molecular graph G b For each node in , the two-dimensional features of the node are: is the node chemical bond type, and Indicates the atomic information at both ends of the chemical bond; Deg i The number of chemical bonds connecting the nodes; It is the three-dimensional feature of the node of the three-dimensional molecular structure; molecular graph G b The shortest distance D between nodes b The definition is as follows: Any node and Expressed as and B and A represent the chemical bond and endpoint atoms of the node respectively. The shortest distance between nodes is D b is the minimum spatial distance between the endpoint atoms of the node, and its expression is: R is the spatial position coordinate of the endpoint atom A of the node, and the shortest distance between nodes with a common endpoint atom is zero.

5. The molecular property prediction method based on geometric feature encoding graph attention network according to claim 1 or 2, characterized in that: In step 2, when the molecular data lacks atomic formal charges, the node features of the molecular graph only retain the atomic number information; the molecular graph G a The node characteristics are Molecular Graph G b Node Features 6. The molecular property prediction method based on geometric feature encoding graph attention network according to claim 1, characterized in that: In step 3, the Floyd algorithm is used to obtain the shortest path information between nodes of the two-dimensional molecular graph. a The shortest path information of a node is the atomic type information of the node on the path; the molecular graph G b The shortest path information of the node is the atomic type and chemical bond type information of the node on the path; the node features, the shortest path and the distance information between nodes in step three are added together through a learnable weight matrix to obtain the self-attention coefficient between the nodes of the molecular graph.

7. The molecular property prediction method based on geometric feature encoding graph attention network according to claim 1 or 6, characterized in that: In step 3, based on the attention mechanism-guided aggregation update of molecular graph node feature vectors, the two-dimensional and three-dimensional features of the molecular graph are combined with the multi-head self-attention mechanism of the Transformer model to aggregate and update the molecular graph node features; the model expression is: X n =AttnG(LN(M n ))+M n , M n+1 =FNN(LN(X n ))+X n , n represents the number of layers in the network, M n is the molecular graph node feature of the network input, X n Represents the molecular graph node features aggregated by the attention mechanism, M n+1 is the molecular graph node feature output by the network, LN() represents the normalization function, and FFN() represents the feedforward neural network; The molecular graph node feature aggregation model AttnG() based on the multi-head self-attention mechanism is expressed as: H G Represents the geometric features of the three-dimensional structure of the molecule; Attn() represents a multi-head self-attention model that integrates the two-dimensional and three-dimensional features of the molecular graph. The expression is as follows: Q=W Q M,K=W K M,V=W V M, M is the molecular graph node feature, W Q ,W K ,W V is the learnable weight matrix, H P is the encoding feature of the shortest path between nodes of the molecular two-dimensional structure, H D It is the encoding feature of the Euclidean distance between the nodes of the molecular structure; the shortest path information of the two-dimensional molecular structure reflects the local structural characteristics between the molecular nodes, and the shortest distance information of the three-dimensional molecular structure reflects the degree of interaction between the nodes.

8. The molecular property prediction method based on geometric feature encoding graph attention network according to claim 1, characterized in that: In step 4, the molecular graph G b The spatial angle between the chemical bonds of the nodes is the shortest distance D between the nodes b The corresponding atom pairs are used as common endpoints; for the molecular graph G a The inter-node atom pair distances and molecular graph G b The chemical bond angle information between nodes is obtained by using basis functions and learnable weight matrices for feature encoding to obtain high-dimensional geometric feature vectors.

9. The molecular property prediction method based on geometric feature encoding graph attention network according to claim 1, characterized in that: For molecular graph G a and G b The multi-dimensional feature vector of the node after aggregation by the graph attention network is firstly pooled to obtain the global graph-level features of the molecular graph: Pool() represents the graph pooling function, X N Aggregate the updated node features for the molecular graph, W1 and W2 are learnable weight matrices, and LeakyReLU() is the activation function. Then merge the graph-level features of the two molecular graphs: in Represents the molecular graph G a and G b The graph-level feature vector, W G is a learnable weight matrix; finally, the weight matrix and activation function are used to reduce the dimensionality of the global features of the molecular graph to predict the properties of the target molecule.

Images