Drug target prediction method based on cross-modal attention and uncertainty evaluation
By employing a cross-modal attention and uncertainty assessment approach, we have addressed the challenges of reliability and complex interaction modeling in drug target prediction. This approach enables accurate prediction and risk identification of drug target interactions, thereby enhancing the credibility and safety of drug screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-15
AI Technical Summary
Existing drug target prediction methods lack quantitative assessment of the reliability of prediction results, make it difficult to distinguish between high-confidence predictions and high-risk predictions, and have limited modeling capabilities in the process of fusing drug and target features, making it difficult to fully establish the complex nonlinear interaction relationship between drugs and targets.
We employ a cross-modal attention and uncertainty assessment approach. By fusing graph structure features and sequence features, we utilize a bidirectional cross-attention mechanism to align and fuse drug and target features. We also introduce an uncertainty quantification mechanism to output predicted labels for drug-target interactions. Finally, we train the model using Monte Carlo Dropout and adaptive optimization algorithms.
It achieves accurate prediction of drug target interactions, can identify high-confidence samples and reduce the risk of model overconfidence, improves the usability and safety of prediction results, and provides robust decision support for drug development.
Smart Images
Figure CN122050487A_ABST
Abstract
Description
Technical Field
[0001] This invention provides a drug target prediction method based on cross-modal attention and uncertainty assessment, belonging to the field of drug target prediction technology. Background Technology
[0002] Drug development is typically characterized by long cycles and low success rates. Identifying the interaction between drug molecules and target proteins is a crucial step in the drug development process. Whether a drug can effectively bind to a specific target protein directly determines its therapeutic effect and potential side effects. Therefore, accurately predicting drug-target interactions before experimental verification is of great significance for improving the efficiency of candidate drug screening and reducing development costs.
[0003] Traditional drug-target interaction identification methods mainly rely on high-throughput screening or biological experiments. These methods usually require a large investment of human and material resources and are difficult to cover large-scale drug and target combinations. With the rapid accumulation of biological data, computational methods based on machine learning and deep learning have gradually become important means of drug target prediction. By modeling drug structural information and target protein characteristics, it is possible to quickly predict potential interactions.
[0004] However, most existing prediction methods still have the following shortcomings: (1) Most methods only output a single interaction prediction probability and lack quantitative evaluation of the reliability of the prediction results. In real application scenarios, it is difficult to distinguish between high confidence prediction and high risk prediction; (2) Some methods have limited modeling capabilities in the process of fusing drug features and target features, and it is difficult to fully establish the complex nonlinear interaction relationship between drugs and targets. Summary of the Invention
[0005] To address the technical problems existing in the background art, the present invention provides a drug target prediction method based on cross-modal attention and uncertainty assessment, comprising the following prediction steps:
[0006] Step S1: Collect data related to drug molecules, target proteins, and drug-target interactions, including drug molecule structure information, target protein amino acid sequence information, and corresponding drug-target interaction annotation data;
[0007] Step S2: Extract the graphical structural features of drug molecules;
[0008] Step S3: Extract sequence features of drug molecules;
[0009] Step S4: After fusing the graph structure features and sequence features of the drug molecule, the final code of the drug molecule is obtained;
[0010] Step S5: Extract amino acid sequence features of the target protein and construct a protein sequence feature representation;
[0011] Step S6: Input the drug molecule features obtained in step S4 and the protein sequence features obtained in step S5 into the cross-modal attention module respectively, and use the bidirectional cross-attention mechanism to align and fuse the drug features and protein features to obtain the joint feature representation of drug-target.
[0012] Step S7: Introduce an uncertainty quantification mechanism to output drug target interaction prediction labels;
[0013] Step S8: Input the predicted label data into the model for forward propagation and define the loss function required for training;
[0014] Step S9: Update the model parameters using the backpropagation algorithm;
[0015] Step S10: Train the model using an adaptive optimization algorithm, and predict drug targets based on the trained model.
[0016] The specific method for collecting drug molecule, target protein, and drug-target interaction related data in step S1 is as follows:
[0017] Step S11: Collect relevant data such as drug, target sequence and drug target protein interaction tags from public datasets BindingDB, Biosnap, and Human;
[0018] Step S12: Divide the collected data to ensure that the training set, validation set and test set are distributed in a consistent manner to avoid data leakage.
[0019] The specific method for step S2 is as follows:
[0020] Step S21: In the dataset loading section, use a function to convert the SMILES string into a molecular graph, define the atomic and bond features of the drug, and define the corresponding features for nodes and edges;
[0021] Step S22: Encode the above molecular graph using a graph neural network, aggregate neighborhood node information, and generate a graph-level embedding representation of the drug molecule.
[0022] The specific method for step S3 is as follows:
[0023] Step S31: Standardize and segment the SMILES string representation of the drug, and map it into a continuous vector representation through the embedding layer to construct the embedding matrix of the SMILES sequence;
[0024] Step S32: Input the SMILES sequence embedding matrix into a one-dimensional convolutional neural network, perform convolution operations on the sequence using multi-scale convolution kernels, and aggregate the convolution features through global pooling operations to generate a fixed-dimensional drug sequence feature representation containing local and global information.
[0025] The specific method for step S4 is as follows:
[0026] Step S41: Obtain the graph structure feature representation and sequence feature representation from steps S2 and S3, respectively;
[0027] Step S42: Concatenate the two feature vectors and dynamically calculate the weights through an attention mechanism to balance the contributions of graph features and sequence features.
[0028] The specific method for extracting the amino acid sequence features of the target protein in step S5 is as follows:
[0029] Step S51: Use One-Hot encoding to convert the protein sequence from text to numerical form, representing each amino acid as a fixed-length vector;
[0030] Step S52: Construct convolutional layers to capture inherent features in the sequence, extract complex pattern features, and output the embedded representation of the protein sequence after pooling and fully connected layers.
[0031] The specific method for step S6 is as follows:
[0032] Step S61: Feature representation preparation: The drug multimodal unified feature vector extracted in step S4 and the target protein sequence feature vector extracted in step S5 are mapped to the same feature space dimension through independent linear projection layers to prepare for subsequent bidirectional attention calculation.
[0033] Step S62: Feature Interaction: By combining protein features and drug features through a bidirectional cross-attention mechanism, the deep interaction between them is captured.
[0034] The uncertainty quantification mechanism described in step S7 is implemented by Monte Carlo Dropout. Specifically, the mean and standard deviation of the predicted probability are calculated through multiple Dropout samplings. The mean is used as a drug-target interaction prediction label, and the standard deviation is used as an uncertainty measure to identify high-confidence predictions and uncertain samples.
[0035] The global interaction features obtained from step S6 are max-pooled and then mapped to the classification probability space in a fully connected layer. Combined with the uncertainty threshold, predictions are directly output for high-confidence samples, while uncertain samples are manually reviewed.
[0036] The specific method for step S8 is as follows:
[0037] The data is input into the model for forward propagation, and a loss function is constructed by combining binary cross-entropy and KL divergence and introducing an uncertainty weighting mechanism.
[0038] The specific method for step S9 is as follows:
[0039] The loss is backpropagated back to the model, the gradient of the loss function with respect to the model parameters is calculated, and the chain rule is used to propagate the loss layer by layer, storing the gradient information for optimization.
[0040] The advantages of this invention compared to existing technologies are as follows: This invention proposes a drug target prediction method based on cross-modal attention and uncertainty assessment. The cross-modal attention mechanism can guide the model to dynamically focus on key information that contributes more to the prediction results during feature interaction, achieving deep fusion of two types of heterogeneous features, thereby effectively capturing the potential complex relationships between drug molecules and target proteins. On this basis, an uncertainty assessment strategy is introduced in the prediction stage. By performing multiple random inferences on the model output results, the stability and reliability of the prediction results are quantified. This invention can not only output the prediction results of the interaction between drugs and targets, but also identify high-uncertainty prediction samples, thereby reducing the potential risks caused by model overconfidence, improving the usability and safety of prediction results in real drug screening scenarios, and providing more robust and reliable decision support for the drug development process. Attached Figure Description
[0041] The present invention will be further described below with reference to the accompanying drawings:
[0042] Figure 1 This is a flowchart of the steps of the drug target prediction method based on cross-modal attention and uncertainty assessment of the present invention;
[0043] Figure 2 This is a framework diagram of the drug target prediction method based on cross-modal attention and uncertainty assessment of the present invention. Detailed Implementation
[0044] like Figure 1 and Figure 2As shown, this invention provides a drug target prediction method based on cross-modal attention and uncertainty assessment. This method effectively predicts drug-target interactions by separately modeling drug molecule features and target protein features, and jointly learning their interaction relationships. Specifically, the feature representation of the drug molecule can be constructed based on its structural information. By encoding the structural description information of the drug molecule, a feature representation reflecting the molecular composition and connectivity is obtained. The feature representation of the target protein can be obtained from its amino acid sequence information, used to characterize the protein's compositional sequence and spatial conformational properties. Furthermore, the drug features and target features are input into a bidirectional cross-attention fusion module, where information interaction between the two is achieved through attention weight calculation, thereby fully exploring the potential complex interaction relationships between the drug and the target. The fused joint features are further input into a prediction module, outputting the predicted interaction results between the drug and the target. Combined with an uncertainty assessment mechanism, this provides a reliable decision-making basis for subsequent drug screening and development processes.
[0045] To more clearly illustrate the technical problems, technical solutions, and beneficial effects involved in this invention, exemplary embodiments of this application will now be described in conjunction with the accompanying drawings, and the invention will be further elaborated in detail. It should be noted that the specific embodiments described herein are merely illustrative of the invention and not intended to limit its scope. The technical solutions of this invention will be explained in detail below with reference to the embodiments and accompanying drawings, but their scope of protection is not limited thereto.
[0046] The drug target prediction method based on cross-modal attention and uncertainty assessment provided by this invention can predict drug target interactions, and specifically includes the following prediction steps:
[0047] Step S1: Collect data on drugs, targets, and interactions;
[0048] The data used in this invention was collected from public datasets BindingDB, Biosnap, and Human. The data information index specifically includes drug SMILES strings, target amino acid sequences, and interaction tags, with a data volume of approximately 30,000 to 50,000 records.
[0049] Step S2: Process the SMILES string of the drug, define the characterization methods of atoms and bonds, and construct graph structure features;
[0050] In an embodiment of the invention, the collected drug target data is first divided into a training set, a validation set, and a test set in a 7:1:2 ratio. During processing, the data is filtered using code; if any data is missing, that drug target pair is discarded to ensure the program runs correctly and eliminate any impact on the model's learning. The drug's SMILES string representation is then converted into a molecular graph structure. ,in, Represents the set of atomic nodes in a molecule. This represents the chemical bond connections between atoms. A corresponding feature vector is defined for each atomic node and chemical bond to characterize the basic chemical properties of the atom and the type of bond connection. Since the number of atoms varies among different drug molecules, virtual nodes and self-connecting edges are introduced during the molecular graph construction process to improve the model's adaptability to molecular structures of different sizes. This allows the model to uniformly handle molecular graph structures of different sizes during learning. Simultaneously, to distinguish between real atomic nodes and filled virtual nodes, corresponding identification information is set in the node features, thereby reducing noise interference introduced by the filling operation to some extent.
[0051] After constructing the molecular graph, a graph convolutional network is used to encode the features of the drug molecule graph. First, node features are mapped to a predefined feature dimension space. Then, a message-passing mechanism is used to aggregate information from the neighboring nodes of each node, achieving iterative updates of node features. The graph convolution process can be represented as:
[0052] ;
[0053] in, = + This represents the adjacency matrix after adding self-connections. This is the adjacency matrix corresponding to the molecular graph. It is the identity matrix. For degree matrix, Indicates the first Trainable weight parameters of layered graph convolutional networks, For the first Layer-level node feature representation. By stacking multiple layers of graph convolutional networks, layer-by-layer abstraction and expression of drug molecule structural information are achieved, ultimately generating a graph-level embedding representation of the drug molecule. , used for subsequent predictions.
[0054] Step S3: Encode the SMILES sequence of the drug molecule and extract the sequence feature representation of the drug, specifically including:
[0055] Step S31: Map each character in the SMILES sequence to its corresponding index number, and map the discrete characters to continuous vector representations through an embedding layer, thereby converting the original string sequence into a numerical matrix representation. This process can be represented as:
[0056] ;
[0057] in, This represents a sequence of SMILES characters of length L. This indicates an embedded mapping function. This is the embedded sequence feature matrix. For the embedded vector dimension.
[0058] Step S32: Feature extraction is performed on the embedded SMILES sequence features using a convolutional neural network. By applying multiple different convolutional kernels along the sequence dimension, local subsequence patterns are modeled to capture the implicit chemical structure information in the SMILES sequence. The calculation process for the k-th convolutional kernel is as follows:
[0059] ;
[0060] in, Indicates the kernel width. and These represent the trainable weights and bias parameters of the k-th convolutional kernel, respectively. After extracting features in parallel using multiple convolutional kernels, local feature maps for different channels are obtained. To obtain a fixed-length sequence representation, pooling is performed on the convolutional output features along the sequence dimension, preferably using max pooling. The calculation process can be represented as follows:
[0061] ;
[0062] Finally, the pooling results of each convolutional channel are concatenated to obtain the sequence-level feature representation of the drug molecule:
[0063] ;
[0064] in, Indicates the number of convolution kernels. This represents the SMILES sequence feature vector of the drug molecule. These sequence features are used for subsequent fusion with drug molecule map features and participate in the prediction process of drug-target interactions.
[0065] Step S4: Combine the features obtained in steps S2 and S3 and By dynamically weighting and fusing attention mechanisms, the final representation of the drug can be obtained. Introducing weighting factors The expression is:
[0066] ;
[0067] Step S5: Feature extraction of the target protein amino acid sequence. A convolutional neural network is used to learn the numerically encoded amino acid sequence. After multiple layers of convolution and pooling operations, the feature encoding of the target in the sequence is completed. Specifically, this includes:
[0068] Step S51: The protein sequence is numerically processed using One-Hot encoding, mapping each amino acid to a fixed-length binary vector, where only the element corresponding to the amino acid position is 1, and the rest are 0; in this way, a protein sequence of length 𝐿 can be represented as a matrix:
[0069] ;
[0070] in, Indicates the first One-Hot vector representation of amino acids, This indicates the dimension of the amino acid dictionary.
[0071] Step S52: A one-dimensional convolutional neural network is used to model protein sequence features to capture local patterns and their contextual dependencies in the amino acid sequence. By sliding the convolution kernel along the sequence dimension, features of adjacent amino acid fragments are aggregated. The convolution calculation process can be represented as follows:
[0072] ;
[0073] in, Indicates the size of the convolution kernel window. and They represent the first The weight and bias parameters corresponding to each convolution kernel are used to perform pooling operations on the sequence dimension of the convolution output. The pooled feature vectors of each channel are concatenated and then feature-mapped through a fully connected layer to obtain the final embedded representation P of the target protein.
[0074] Step S6: A cross-modal bidirectional cross-attention module is employed to achieve deep, bidirectional interaction between drug and target features. In existing methods, feature fusion between drugs and targets often uses self-attention mechanisms. While computationally less complex, this typically only involves interaction at a single feature layer, failing to fully utilize the bidirectional semantic associations between different modalities. This limits the model's ability to capture complex interaction patterns. The cross-modal bidirectional cross-attention mechanism designed in this invention aims to solve this problem. Its core is to enable bidirectional, intensive cross-modal information exchange between drug and target features. Specific steps include:
[0075] Step S61: Map the unified drug feature vector D obtained in step S4 and the final protein representation P obtained in step S5 to the same feature dimension d through two independent linear projection layers to generate the basic representation of query, key, and value projections: This ensures that features from different modalities reside in the same semantic space.
[0076] Step S62: Bidirectional cross-attention interaction calculation, including attention calculation in two directions:
[0077] Based on the characteristics of the projected drug As a query (Q), the projected target features As keys (K) and values (V), attention weights are calculated to inject relevant target information into the drug representation.
[0078] ;
[0079] in, , , Let be a learnable projection matrix, and let the output be denoted as... .
[0080] Target features after projection As a query (Q), the projected drug characteristics As keys (K) and values (V), attention weights are calculated to inject relevant drug information into the target representation:
[0081] ;
[0082] in, , , Let the output be another set of learnable projection matrices. The enhanced features from the two directions are concatenated and then fused and nonlinearly transformed through a feedforward network to generate the final cross-modal interactive representation. .
[0083] Step S7: Drug target interaction prediction and uncertainty determination, using Monte Carlo Dropout model to quantify the uncertainty predicted.
[0084] Specifically, the global interaction features of the drug target obtained in step S6 are input into the prediction module. First, a fixed-length feature representation is obtained through max pooling and then input into a fully connected network containing Dropout layers. All Dropout layers are kept active, and T independent forward propagation calculations are performed on the same drug target pair to be predicted.
[0085] Since each forward propagation produces a slightly different prediction output due to the randomness of Dropout, a set of prediction probabilities is obtained. Subsequently, the mean of the T predictions is calculated as the final interaction probability estimate. And calculate its standard deviation. As a measure of the uncertainty of this prediction;
[0086] Finally, an uncertainty threshold is set. ,when Below At that time, the prediction results were considered to have a high degree of confidence and could be directly adopted. Make classification decisions; when Higher than If the sample is not found to be "uncertain," manual review or further experimental verification is recommended.
[0087] Step S8: Input the data into the model for forward propagation. Based on this, construct a composite loss function consisting of a binary cross-entropy loss term and a KL divergence constraint term, and introduce a weighting mechanism based on prediction uncertainty to balance the model's learning process for high-confidence samples and uncertain samples; the expression of the loss function is:
[0088] ;
[0089] in, This represents the predicted probability output by the model. Indicates the true label, This is the binary cross-entropy loss term. The KL divergence term is used to constrain the distribution of latent features. and These are hyperparameters used to adjust the weights of each loss term. For uncertainty indicators obtained in step S7 The constructed weighting function is used to guide the model to reduce the risk of overfitting to high-uncertainty samples during training, thereby improving the model's generalization ability and prediction stability.
[0090] Step S9: After constructing the loss function described in step S8, the backpropagation algorithm is used to calculate the gradient value of the loss function with respect to each trainable parameter layer by layer according to the chain rule. The gradient information is stored in the corresponding model parameters for subsequent parameter update process.
[0091] Step S10: Update the model parameters using the Adam optimizer, achieving efficient optimization through an adaptive learning rate mechanism. The specific steps are as follows:
[0092] Step S101: Calculate the exponentially weighted moving average of the gradient. Preserve historical gradient directions to accelerate convergence:
[0093] ;
[0094] in, For the current gradient, This is the attenuation rate, which defaults to 0.9.
[0095] Step S102: Calculate the exponentially weighted moving average of the squared gradient. This reflects the magnitude of the parameter gradient change.
[0096] ;
[0097] in, For the current gradient, This is the attenuation rate, which defaults to 0.999.
[0098] Step S103: Correct the moment estimation bias at the initial time to ensure stability in the early stages of training.
[0099] ;
[0100] Step S104: Update each parameter by combining the corrected moment estimate:
[0101] ;
[0102] The Adam optimizer not only dynamically adjusts the learning rate of each parameter, but also ensures that the model converges efficiently and stably in complex drug target prediction tasks through hierarchical configuration, bias correction, and optional scheduling strategies.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A drug target prediction method based on cross-modal attention and uncertainty assessment, characterized in that: The prediction steps include the following: Step S1: Collect data related to drug molecules, target proteins, and drug-target interactions, including drug molecule structure information, target protein amino acid sequence information, and corresponding drug-target interaction annotation data; Step S2: Extract the graphical structural features of drug molecules; Step S3: Extract sequence features of drug molecules; Step S4: After fusing the graph structure features and sequence features of the drug molecule, the final code of the drug molecule is obtained; Step S5: Extract amino acid sequence features of the target protein and construct a protein sequence feature representation; Step S6: Input the drug molecule features obtained in step S4 and the protein sequence features obtained in step S5 into the cross-modal attention module respectively, and use the bidirectional cross-attention mechanism to align and fuse the drug features and protein features to obtain the joint feature representation of drug-target. Step S7: Introduce an uncertainty quantification mechanism to output drug target interaction prediction labels; Step S8: Input the predicted label data into the model for forward propagation and define the loss function required for training; Step S9: Update the model parameters using the backpropagation algorithm; Step S10: Train the model using an adaptive optimization algorithm, and predict drug targets based on the trained model.
2. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 1, characterized in that: The specific method for collecting drug molecule, target protein, and drug-target interaction related data in step S1 is as follows: Step S11: Collect relevant data such as drug, target sequence and drug target protein interaction tags from public datasets BindingDB, Biosnap, and Human; Step S12: Divide the collected data to ensure that the training set, validation set and test set are distributed in a consistent manner to avoid data leakage.
3. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 2, characterized in that: The specific method for step S2 is as follows: Step S21: In the dataset loading section, use a function to convert the SMILES string into a molecular graph, define the atomic and bond features of the drug, and define the corresponding features for nodes and edges; Step S22: Encode the above molecular graph using a graph neural network, aggregate neighborhood node information, and generate a graph-level embedding representation of the drug molecule.
4. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 3, characterized in that: The specific method for step S3 is as follows: Step S31: Standardize and segment the SMILES string representation of the drug, and map it into a continuous vector representation through the embedding layer to construct the embedding matrix of the SMILES sequence; Step S32: Input the SMILES sequence embedding matrix into a one-dimensional convolutional neural network, perform convolution operations on the sequence using multi-scale convolution kernels, and aggregate the convolution features through global pooling operations to generate a fixed-dimensional drug sequence feature representation containing local and global information.
5. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 4, characterized in that: The specific method for step S4 is as follows: Step S41: Obtain the graph structure feature representation and sequence feature representation from steps S2 and S3, respectively; Step S42: Concatenate the two feature vectors and dynamically calculate the weights through an attention mechanism to balance the contributions of graph features and sequence features.
6. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 5, characterized in that: The specific method for extracting the amino acid sequence features of the target protein in step S5 is as follows: Step S51: Use One-Hot encoding to convert the protein sequence from text to numerical form, representing each amino acid as a fixed-length vector; Step S52: Construct convolutional layers to capture inherent features in the sequence, extract complex pattern features, and output the embedded representation of the protein sequence after pooling and fully connected layers.
7. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 6, characterized in that: The specific method for step S6 is as follows: Step S61: Feature representation preparation: The drug multimodal unified feature vector extracted in step S4 and the target protein sequence feature vector extracted in step S5 are mapped to the same feature space dimension through independent linear projection layers to prepare for subsequent bidirectional attention calculation. Step S62 Feature Interaction: By combining protein and drug features through a bidirectional cross-attention mechanism, deep interactions between them are captured.
8. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 7, characterized in that: The uncertainty quantification mechanism described in step S7 is implemented by Monte Carlo Dropout. Specifically, the mean and standard deviation of the predicted probability are calculated through multiple Dropout samplings. The mean is used as a drug-target interaction prediction label, and the standard deviation is used as an uncertainty measure to identify high-confidence predictions and uncertain samples. The global interaction features obtained from step S6 are max-pooled and then mapped to the classification probability space in a fully connected layer. Combined with the uncertainty threshold, predictions are directly output for high-confidence samples, while uncertain samples are manually reviewed.
9. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 8, characterized in that: The specific method for step S8 is as follows: The data is input into the model for forward propagation, and a loss function is constructed by combining binary cross-entropy and KL divergence and introducing an uncertainty weighting mechanism.
10. The drug target prediction method based on cross-modal attention and uncertainty assessment according to claim 9, characterized in that: The specific method for step S9 is as follows: The loss is backpropagated back to the model, the gradient of the loss function with respect to the model parameters is calculated, and the chain rule is used to propagate the loss layer by layer, storing the gradient information for optimization.