System and method for predicting compound-protein interaction

Abstract

The invention discloses a system for predicting compound-protein interaction. The system is composed of a compound feature vector extraction model, a protein feature vector extraction model and an interaction prediction model. The compound feature vector extraction model comprises a segmentation module, a graph conversion module and a molecular feature extraction module; the protein feature vector extraction model comprises a preprocessing module, a semantic vector conversion module, a local feature extraction module and a protein sequence enhancement module; the compound feature vector extraction model processes an input compound SMILES character string to obtain a compound fusion feature vector; the protein feature vector extraction model is used for processing an amino acid sequence of input protein to obtain a protein enhanced feature vector; and the interaction prediction module carries out compound fusion feature vector and protein enhancement feature vector through a fully connected network. According to the invention, the interaction between the compound and the protein is accurately predicted by adopting space-time sequence intelligent processing of an artificial intelligence technology in the field of medicines.

Description

Technical Field

[0001] The present invention relates to the field of compound protein molecular feature extraction, and in particular to a system and method for predicting compound-protein interaction. Background Art

[0002] The study of compound-protein interactions (CPIs) is of vital importance in the field of drug discovery and biomedical research. However, experimental measurement of CPIs is time-consuming and expensive [1]. Despite the practicality of techniques such as high-throughput screening (HTS) [2] and virtual screening (VS), they still face limitations in accelerating the CPI prediction process due to cost and time constraints [3]. Given the remarkable success of artificial intelligence (AI) in natural language processing and computer vision, researchers have turned their attention to exploring the application of AI technologies to the challenges encountered in CPI research. AI-based tools are currently regarded by the pharmaceutical industry as a key "engine" for the discovery and development of innovative drugs [4]. By leveraging machine learning algorithms and deep neural networks, researchers can extract meaningful patterns from large amounts of data and achieve accurate prediction of CPIs. These AI-based methods have the potential to facilitate the identification of potential drug candidates and accelerate the drug discovery process by providing an efficient and cost-effective means of predicting compound-protein interactions [5].

[0003] In CPI research, compound representation methods have always been a research focus. Many methods use the powerful learning ability of neural networks to capture important information within compounds, providing new possibilities for CPI prediction and drug design. These methods use graph neural networks (GNNs) to iteratively update node representations to capture the interactions and information between atoms in the compound. This method is particularly effective in dealing with large molecules and complex compounds. However, most GNN-based compound modeling methods ignore high-order information such as pharmacophores and functional groups [5]. For example, existing methods typically only focus on information at a single scale, ignoring functional groups and pharmacophores that can provide molecular properties that are not fully utilized by current atom-based molecular graph representations. Pharmacophore modeling is currently a key area in drug design and discovery [6]. Its goal is to identify and describe structural patterns or functional groups associated with specific drug activities. Pharmacophores are key structural units in drug molecules that are responsible for interacting with protein targets, and they play a vital role in drug selectivity and activity. Through pharmacophore modeling, key pharmacophore features can be identified by analyzing known active compounds and their related protein structures, such as Figure 1 If a compound matches multiple features, it is considered to conform to the pharmacophore model. Pharmacophore modeling enables researchers to better understand and optimize the interactions between drug molecules[7][8]. Summary of the invention

[0004] In view of the technical difficulty that important information of compound molecules and protein molecules is seriously lost in the prediction process in the prior art, the present invention provides a compound-protein interaction prediction system and method based on multi-scale learning and pharmacophore information fusion; the present invention can encode the structural properties of compound molecules into SMILES strings, so that more information about the compound molecules can be extracted, and attention models are respectively added in the process of compound protein feature representation to obtain more accurate feature representation vectors, and a deep learning method is used to improve the accuracy of predicting compound protein affinity values.

[0005] In order to solve the technical problems existing in the prior art, the present invention adopts the following technical solutions:

[0006] A system for predicting compound-protein interactions, the system is composed of a compound feature vector extraction model, a protein feature vector extraction model and an interaction prediction model; the compound feature vector extraction model includes a segmentation module, a graph conversion module and a molecular feature extraction module; the graph conversion module is composed of a first multi-scale conversion layer and a second multi-scale conversion layer; the molecular feature extraction module is composed of a first gated recurrent unit and a second gated recurrent unit; each of the first gated recurrent unit and the second gated recurrent unit is composed of an update gate, a reset gate and an attention layer; the protein feature vector extraction model includes a preprocessing module, a semantic vector conversion module, a local feature extraction module and a protein sequence enhancement module; the local feature extraction module adopts a convolutional neural network, and the convolutional neural network is composed of a convolutional layer, an activation layer and a pooling layer; wherein:

[0007] The compound feature vector extraction model processes the input compound SMILES string to obtain a compound fusion feature vector;

[0008] The protein feature vector extraction model processes the input protein amino acid sequence to obtain a protein enhanced feature vector;

[0009] The interaction prediction module processes the compound fusion feature vector and the protein enhancement feature vector through a fully connected network to predict the interaction probability between the compound and the protein.

[0010] Furthermore, the compound feature vector extraction model processes the input compound SMILES string to obtain the compound feature vector; including:

[0011] The segmentation module converts the SMILES format of a compound into an atomic scale graph representation:

[0012] G a =(X a , E a )

[0013] in: The set of vertices representing atoms; n a The number of atoms, f atom is the dimension of atomic features; the feature x of each atom a ∈X a Encoded using one-hot vector connections based on their atomic symbol, degree, formal charge, radical electrons, hybridization, and aromaticity; E a is a bond in the molecular graph. Each bond e a ∈E a Join using one-hot vectors based on their key type;

[0014] The segmentation module converts the SMILES format of compounds into a pharmacophore scale map representation

[0015] G p =(X p , E p );

[0016] in: is the n in the numerator p The collection of pharmacophores, f pharm is the dimension of the pharmacophore node feature; each node feature x p ∈X P Use one-hot vector connection based on its pharmacophore features, BRICS features and MACCS fingerprints; where:

[0017] The pharmacophore characteristics are represented using a unique heat representation of the pharmacophore characteristics;

[0018] The BRICS feature consists of a BRICS key generated by RDKit; p is the set of BRICS keys; each BRICS key e p ∈E p Use one-hot vector connections based on their BRICS reaction type;

[0019] The MACCS fingerprint is a binary fingerprint generated based on molecular structure and functional group fragments;

[0020] The graph conversion module processes the atomic scale graph and the pharmacodynamic scale graph respectively to obtain the atomic scale graph feature representation and the pharmacodynamic scale graph feature representation;

[0021] The molecular feature extraction module reads the atomic scale image feature representation and the drug efficacy scale feature representation respectively to obtain a compound fusion feature vector.

[0022] Furthermore, the graph conversion module processes the atomic scale graph to obtain an atomic scale graph feature representation process, including:

[0023] Through the atomic scale graph Ga The node feature matrix is ​​obtained according to the following formula:

[0024]

[0025] Where: n a is the number of nodes, f atom is the dimension of the feature;

[0026] Pick Where: h is the index of the head in the multi-head attention mechanism, l is the index of the layer in the graph conversion module, is the weight matrix;

[0027] The multi-head attention update process of the node features is performed through the following formula:

[0028]

[0029] Where: W l is a learnable weight matrix, N(i) is the neighboring node of node i. Concat() is the head vector concatenation operation; multi-head calculation is performed according to the following formula,

[0030]

[0031] Among them: the number of heads in the multi-head attention mechanism is represented by H; the dimension of K is represented by d k ;

[0032] Similarly, the drug efficacy scale characteristics can be obtained from the atomic scale graph G p Obtained by the same processing method.

[0033] Furthermore, the protein feature vector extraction model processes the input protein amino acid sequence to obtain the protein feature vector; including:

[0034] The semantic vector conversion module converts the amino acid subsequence into a real value to obtain an amino acid subsequence embedding vector;

[0035] The local feature extraction module extracts local features of amino acid subsequence embedding vectors through a convolutional neural network to obtain a protein feature vector;

[0036] The protein sequence enhancement module establishes protein feature vector association through a self-attention mechanism to obtain a protein enhanced feature vector.

[0037] In order to solve the problems of the prior art, the present invention also adopts the following technical solutions:

[0038] A method for predicting compound-protein interactions, the method comprising constructing a system for predicting compound-protein interactions and training the system, comprising:

[0039] The system is optimized by minimizing the cross entropy loss between the predicted value and the true value; the initial learning rate is set to 0.0001, and the system weight is adjusted according to the error gradient using an optimizer; through repeated iterative training, the loss function value of the system is gradually reduced until the system converges on a given data set.

[0040] Furthermore, the given data set includes: a C.elegans data set, a Human data set, and a GPCR data set.

[0041] Beneficial Effects

[0042] 1. The present invention combines the multi-scale two-dimensional molecular graph structure information of compound molecules, especially conducts separate learning on the pharmacophore of the molecule, so as to extract more information about the compound molecule, uses the processing method in linguistics to obtain more information about the protein molecule from the amino acid sequence information, and uses deep learning methods to improve the accuracy of predicting the affinity value of the compound protein.

[0043] 2. In order to improve the performance of the model, the present invention uses a large number of existing compound-protein binding results for training. Through this process, the model obtains perfect parameters, enabling it to more accurately predict and evaluate potential drug candidates.

[0044] 3. The method of the present invention has been extensively experimented on multiple data sets, consistently demonstrating the superiority of multi-scale methods over previous methods that rely solely on atomic representation. In addition, the method of the present invention also has the ability to reveal molecular fragments that interact with proteins, which has important practical application value for drug design and screening.

[0045] 4. The present invention plays an important role in the intelligent processing of spatiotemporal sequences in the field of medicine using artificial intelligence technology. The present invention not only helps solve the problems of high cost, long cycle and high safety risk of new drug development, but also can screen out new drugs and therapeutic targets from old drugs that have been proven to be safe and compounds that have been abandoned. In gradually changing the traditional model of drug development, the present invention leads a new drug development strategy called drug repositioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a specific flow chart of the prediction model in the present invention.

[0047] Figure 2 These are the results of testing the present invention on three data sets.

[0048] Figure 3 It is the result of the present invention in visualizing the interaction area. DETAILED DESCRIPTION

[0049] The following is combined with Figure 1 -Attached Figure 3 The present invention is described in detail:

[0050] like Figure 1 As shown, the present invention discloses a system for predicting compound-protein interactions; the system is composed of a compound feature vector extraction model, a protein feature vector extraction model and an interaction prediction model; the compound feature vector extraction model includes a segmentation module, a graph conversion module and a molecular feature extraction module; the graph conversion module is composed of a first multi-scale conversion layer and a second multi-scale conversion layer; the molecular feature extraction module is composed of a bidirectional gated recurrent unit; each of the first gated recurrent unit and the second gated recurrent unit is composed of an update gate, a reset gate and an attention layer; the protein feature vector extraction model includes a preprocessing module, a semantic vector conversion module, a local feature extraction module and a protein sequence enhancement module; the local feature extraction module adopts a convolutional neural network, and the convolutional neural network is composed of a convolutional layer, an activation layer and a pooling layer; wherein:

[0051] The compound feature vector extraction model processes the input compound SMILES string to obtain a compound fusion feature vector; the one-dimensional SMILES sequence compound in the compound feature vector extraction model of the present invention constructs atomic-level and pharmacophore-level graphs according to its structure, and then inputs them into graph transformation modules (Graph Transformers) of different scales for learning, and finally reads them out using a bidirectional gated recurrent unit to obtain the final compound representation vector. The multi-scale Graph Transformer model is an innovative deep learning architecture, namely a graph transformation model, which is designed to process complex graph structure data and enhance the model's expressiveness and prediction accuracy by capturing information at different scales. The model combines the attention mechanism of the Transformer and the graph embedding capability of the graph neural network, and can learn the long-term dependencies between nodes and edges in the graph while maintaining sensitivity to local and global structures. In the multi-scale Graph Transformer model, information at different scales is explicitly modeled and integrated. The model extracts graph structure features of different scales through multi-layer graph convolution operations or graph attention mechanisms, which are then fed into the encoder part of the Transformer. In the encoder, the self-attention mechanism enables the model to focus on important nodes and edges in the graph and learn the complex interactions between them. In addition, the multi-scale Graph Transformer model also retains the topological information of the graph by introducing positional encoding or structural encoding. These encodings help the model distinguish different nodes and edges and understand their relative positions or roles in the graph.

[0052] The GRU gated recurrent unit is a variant of a recurrent neural network (RNN), which solves the gradient vanishing problem of traditional RNN by introducing an update gate and a reset gate. In the invention, GRU is used to read the final compound representation vector from the updated atomic scale and pharmacophore scale graphs. By considering the characteristics of adjacent nodes, GRU can capture molecular features more comprehensively, and by combining the update gate, reset gate and attention mechanism, the model effectively utilizes the relationship between atoms in the molecule, as well as the information of the pharmacophore, and enhances the modeling ability of molecular structure and properties. This GRU modeling method promotes a more accurate representation and understanding of the entire molecular feature. The bidirectional gated recurrent unit model includes a sequence processing model composed of two gated recurrent units (GRUs), one input is a forward input, and the other input is a reverse input, forming a bidirectional recurrent neural network.

[0053] The protein feature vector extraction model processes the amino acid sequence of the input protein to obtain a protein feature vector; the protein feature vector extraction model in the present invention uses a method to obtain an n-gram protein representation vector, and then processes it through a one-dimensional convolution method and a self-attention mechanism to obtain a final protein representation vector. The final output of the model is to predict whether the compound protein interacts.

[0054] The semantic vector conversion module uses Word2vec, a pre-trained word embedding technology, which is used in natural language processing to convert words into real-valued vector representations. In the invention, the Word2vec method is used to process protein sequences and convert amino acid subsequences into real-valued embedding vectors. These embedding vectors capture the semantic information of the amino acid subsequences and provide rich local context information for subsequent CNN feature extraction, thereby enhancing the model's understanding and representation capabilities of protein sequences.

[0055] The local feature extraction module is a convolutional neural network (CNN) composed of three structures: convolution layer, activation, and pooling. It extracts local features in the sequence through the convolution layer. In protein sequence representation learning, CNN uses one-dimensional convolution layers and GLU activation functions to extract and enhance features. Through residual connections, the output of each layer is fused with the input in a weighted manner to form a new amino acid subsequence embedding representation. This design of CNN allows the model to capture local contextual information in protein sequences, which helps to understand the structure and function of proteins.

[0056] The protein sequence enhancement module adopts a self-attention mechanism, which is a method for calculating the correlation between different positions in a sequence, which allows the model to establish connections and interactions between different positions in the sequence. In the present invention, the self-attention mechanism is used to enhance the feature representation ability of protein sequences. By calculating the correlation between each subsequence in the sequence, the model can capture long-range dependencies and establish associations between different positions. The introduction of this mechanism significantly improves the model's ability to learn protein sequence features, providing a richer and more robust feature representation for the final CPI prediction.

[0057] The interaction prediction module processes the compound fusion feature vector and the protein feature vector through a fully connected network to predict the interaction probability between the compound and the protein. The feature space output by the interaction prediction model CNN is used as the input of a fully connected neural network (FCN) to complete the mapping from the input feature vector to the affinity value.

[0058] The system also includes a metamodel training process, and a specific embodiment of the system training process is given below.

[0059] During the training phase of the system, we used RDKit, a powerful toolkit commonly used in cheminformatics and drug discovery research, to convert the SMILES string format of the compound into an atomic scale graph representation. a =(X a , E a ),in is the set of vertices representing atoms. Here, n a is the number of atoms, f atom is the dimension of the atomic features. The feature x of each atom a ∈X a Encoded using one-hot vector connections based on their atomic symbol, degree, formal charge, radical electrons, hybridization, and aromaticity. a is a bond in the molecular graph. Each bond e a ∈E a Connect using one-hot vectors based on their bond type, such as single, double, or aromatic.

[0060] The pharmacophore scale map is obtained by segmenting the atomic scale map into pharmacophores based on reaction information using the BRICS method in RDKit. The pharmacophore scale map is denoted as G p =(X p , E p ),in is the n in the numerator p The collection of pharmacophores, f pharmis the dimension of the pharmacophore node features. Each node feature x p ∈X P The molecules are linked using one-hot vectors based on their pharmacophore features, BRICS features and MACCS fingerprints. Pharmacophore features are represented using one-hot representations of pharmacophore features such as acidity, basicity, aromaticity, etc. MACCS fingerprints are binary fingerprints generated based on molecular structure and functional group fragments. They can be used to generate molecular fingerprints and calculate molecular similarity. BRICS features consist of BRICS bonds generated by RDKit. p is a set of BRICS keys. Each BRICS key p ∈E p The graphs are connected using one-hot vectors based on their BRICS reaction types. Atomic-scale and pharmacophore-scale graphs are fed into different Graph Transformer layers for feature learning. To effectively learn features from graphs, a multi-head attention mechanism is used in Graph Transformer. Multi-head attention can significantly enhance the representation power of the model by aggregating diverse and complex features. In the multi-head mechanism, the input is split into multiple heads for parallel computation, and their outputs are concatenated to construct the final representation for message passing and updating.

[0061] Taking the update of atomic-scale graph as an example, the input of Graph Transformer is graph G a The node feature matrix Where: n a is the number of nodes, f atom is the dimension of the feature. We take Where h is the index of the head in the multi-head attention mechanism, l is the index of the layer in the Graph Transformer, is the weight matrix. The node features are updated through multi-head attention.

[0062]

[0063] Where W l is a learnable weight matrix, N(i) is the neighboring node of node i. Concat() is the head vector concatenation operation. The multi-head calculation is as follows,

[0064]

[0065] The number of heads in the multi-head attention mechanism is represented by H. The dimension of K is represented by d k The updating process of the pharmacophore scale map is the same as that of the atomic scale map.

[0066] For protein representation learning, the word2vec method is used to process protein sequences to generate n-gram protein representation vectors. These vectors are then further processed by one-dimensional convolution (conv1d) and self-attention mechanism to obtain the final protein representation vector.

[0067] Next, the atomic-scale and pharmacophore-scale representations of the compound are fused through a multi-head attention mechanism and finally read out using a gated recurrent unit (GRU) to obtain the fused feature vector of the compound. Similarly, the representation vector of the protein is also obtained and prepared for the final interaction prediction module, such as Figure 3 shown.

[0068] The final prediction module concatenates the feature vectors of the compound and protein and processes them through a fully connected network (MLP) using the ReLU activation function to predict the interaction between the compound and the protein. During the training process, the parameters of the model are adjusted by minimizing the cross entropy loss between the predicted value and the true value. (Cross-entropy loss) is used for optimization. The initial learning rate is set to 0.0001, and the Adam optimizer is used to adjust the network weights according to the error gradient. Through repeated iterative training, the loss function value of the model gradually decreases until the network converges on the given data set.

[0069] In addition, in order to improve the generalization ability of the model, MS-CPI conducted extensive experiments on multiple datasets, including C.elegans dataset, Human dataset and GPCR dataset. Finally, through case studies, the performance of the MS-CPI model in virtual screening was also demonstrated. Through visualization tools, the model can also help researchers understand the interaction between compounds and proteins.

[0070] The model described is specifically designed to process two selected variables as input, which contain key information from different data sources: HUMAN, GPCR, BindingDB, and compound SMILES and protein amino acid sequences in C.elegans. These data provide the model with rich chemical and biological information, which helps to gain a deeper understanding of the interaction between compounds and proteins, such as Figure 2 shown.

[0071] In order to improve the performance of the model, the present invention uses a large number of existing compound-protein binding results for training. Through this process, the model obtains perfect parameters, enabling it to more accurately predict and evaluate potential drug candidates.

[0072] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The technicians in the field can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. The technicians in the field can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

Claims

1. A system for predicting compound-protein interactions, characterized in that: The system is composed of a compound feature vector extraction model, a protein feature vector extraction model and an interactive prediction model; the compound feature vector extraction model includes a segmentation module, a graph conversion module and a molecular feature extraction module; The graph conversion module is composed of a first multi-scale conversion layer and a second multi-scale conversion layer; the molecular feature extraction module is composed of a first gated recurrent unit and a second gated recurrent unit; each of the first gated recurrent unit and the second gated recurrent unit is composed of an update gate, a reset gate and an attention layer; the protein feature vector extraction model includes a preprocessing module, a semantic vector conversion module, a local feature extraction module and a protein sequence enhancement module; the local feature extraction module adopts a convolutional neural network, and the convolutional neural network consists of a convolutional layer, an activation layer and a pooling layer; wherein: The compound feature vector extraction model processes the input compound SMILES string to obtain a compound fusion feature vector; The protein feature vector extraction model processes the input protein amino acid sequence to obtain a protein enhanced feature vector; The interaction prediction module processes the compound fusion feature vector and the protein enhancement feature vector through a fully connected network to predict the interaction probability between the compound and the protein.

2. A system for predicting compound-protein interactions according to claim 1, characterized in that: The process of extracting compound feature vector model to process the input compound SMILES string to obtain the compound feature vector includes: The segmentation module converts the SMILES format of a compound into an atomic scale graph representation: G a =(X a .E a ) in: The set of vertices representing atoms; n a The number of atoms, f atom is the dimension of atomic features; the feature x of each atom a ∈X a Encoded using one-hot vector connections based on their atomic symbol, degree, formal charge, radical electrons, hybridization, and aromaticity; E a is a bond in the molecular graph. Each bond e a ∈E a Join using one-hot vectors based on their key type; The segmentation module converts the SMILES format of compounds into a pharmacophore scale map representation G p =(X p ,E p ); in: is the n in the numerator p The collection of pharmacophores, f pharm is the dimension of the pharmacophore node feature; each node feature x p ∈X P Use one-hot vector connection based on its pharmacophore features, BRICS features and MACCS fingerprints; where: The pharmacophore characteristics are represented using a unique heat representation of the pharmacophore characteristics; The BRICS feature consists of a BRICS key generated by RDKit; p is the set of BRICS keys; each BRICS key e p ∈E p Use one-hot vector connections based on their BRICS reaction type; The MACCS fingerprint is a binary fingerprint generated based on molecular structure and functional group fragments; The graph conversion module processes the atomic scale graph and the pharmacodynamic scale graph respectively to obtain the atomic scale graph feature representation and the pharmacodynamic scale graph feature representation; The molecular feature extraction module reads the atomic scale image feature representation and the drug efficacy scale feature representation respectively to obtain a compound fusion feature vector.

3. A system for predicting compound-protein interactions according to claim 2, characterized in that: The graph conversion module processes the atomic scale graph to obtain the atomic scale graph feature representation process, including: Through the atomic scale graph G a The node feature matrix is ​​obtained according to the following formula: Where: n a is the number of nodes, f atom is the dimension of the feature; Pick Where: h is the index of the head in the multi-head attention mechanism, l is the index of the layer in the graph conversion module, is the weight matrix; The multi-head attention update process of the node features is performed through the following formula: Where: W l is a learnable weight matrix, N(i) is the neighboring node of node i. Concat() is the head vector concatenation operation; multi-head calculation is performed according to the following formula, Among them: the number of heads in the multi-head attention mechanism is represented by H; the dimension of K is represented by d k ; Similarly, the drug efficacy scale characteristics can be obtained from the atomic scale graph G p Obtained by the same processing method.

4. A system for predicting compound-protein interactions according to claim 1, characterized in that: The protein feature vector extraction model processes the amino acid sequence of the input protein to obtain the protein feature vector; comprising: The semantic vector conversion module converts the amino acid subsequence into a real value to obtain an amino acid subsequence embedding vector; The local feature extraction module extracts local features of amino acid subsequence embedding vectors through a convolutional neural network to obtain a protein feature vector; The protein sequence enhancement module establishes protein feature vector association through a self-attention mechanism to obtain a protein enhanced feature vector.

5. A method for predicting compound-protein interactions, characterized in that: The method adopts any one of claims 1 to 4 to construct a system for predicting compound-protein interactions; the method further comprises the following steps to train the system, including: Optimizing the system by minimizing the cross entropy loss between predicted values ​​and true values; By setting the initial learning rate to 0.0001, the optimizer is used to adjust the system weight according to the error gradient; Through repeated iterative training, the loss function value of the system gradually decreases until the system reaches convergence on a given data set.

6. A method for predicting compound-protein interaction according to claim 5, characterized in that: The given data sets include: C.elegans data set, Human data set and GPCR data set.

Images