Drug molecule multi-attribute optimization method and system based on multi-distribution mapping

Abstract

The invention provides a drug molecule multi-attribute optimization method based on multi-distribution mapping, and belongs to the field of molecule optimization. According to the number of attributes, a drug molecule data set is decomposed into non-overlapping sub-data sets, and the two sub-data sets are selected as a source distribution molecule set and a target distribution set respectively; sample input is obtained after the pilot molecules and the distribution labels are fused; the samples are input into a distribution mapper to generate new molecules, the new molecules are input into a pre-trained multi-distribution classifier, the output of the multi-distribution classifier and the distribution labels are subjected to mapping loss, the new molecules and the distribution labels are fused and then input into the distribution mapper to generate reconstructed molecules, and all mapping paths are traversed; fusing a to-be-optimized drug molecule and the target distribution tag, and inputting the fused drug molecule and the target distribution tag into the trained distribution mapper to generate an optimized molecule; the invention further provides an optimization system. The problem that an existing method is insufficient in generalization ability in a multi-attribute optimization task with sparse data is solved.

Description

Technical Field

[0001] This invention relates to the fields of molecular optimization, artificial intelligence and bioinformatics, and in particular to a method and system for optimizing multiple attributes of drug molecules based on multi-distribution mapping. Background Technology

[0002] Drug discovery is a lengthy and costly process, typically requiring over 10 years and billions of dollars. Discovering lead compounds with novel structures, good biological activity, and druggability is fundamental to innovative drug development. Molecular optimization of lead compounds is a crucial step in drug discovery, aiming to improve their biological activity, reduce toxicity, enhance chemical stability, and improve pharmacokinetic properties through a series of chemical and drug design strategies, while maintaining or increasing molecular characteristics similar to the original lead compound to ensure drug efficacy and safety. Traditional drug molecule design techniques, such as virtual screening and combinatorial chemistry-based methods, are insufficient to meet these requirements due to limited chemical space, heavy reliance on expert chemical knowledge, and poor synthetic feasibility.

[0003] In recent years, molecular optimization algorithms have made significant progress due to the successful use of deep generative models. These methods can be broadly categorized into two types: matched molecular pair mapping and molecular distribution mapping. Among them, molecular distribution mapping, because it does not require a large number of matched molecular pairs as training data, has the advantage of being more realistic and has greater research potential than matched molecular pair mapping. In the past, methods based on molecular distribution mapping have achieved good results in single-attribute optimization tasks. However, since most of these methods only focus on the mapping process from the source distribution to the target distribution, and multi-attribute optimization tasks generally lack a distribution set composed of samples with good attributes, these methods still have certain limitations.

[0004] In the prior art, Chinese invention patent application CN117497090A, entitled "A Drug Molecule Optimization Method Guided by Multi-Task Learning Attribute Predictor", adopts a network architecture of encoder-molecular converter network-decoder-multi-attribute predictor. The datasets for training the autoencoder-molecular converter network-decoder and the multi-attribute predictor are M and M, respectively. VAE (composed of smile sequences), M pro(Composed of smiles sequences and molecular attribute values), the problems with this method are: (1) The discrete smiles molecules are mapped to the latent space by the autoencoder and then decoded back to the optimized new smiles molecules. The autoencoder decoder is pre-trained and is not related to the mapping, which will cause information loss; (2) The molecular optimization task is summarized as mapping + regression prediction. After the molecular is mapped, it is input into the multi-attribute predictor and the loss is used to guide the molecular converter network optimization with the set target attribute. This results in the optimized molecular attribute values ​​being only near the set target attribute values, making it difficult to generate molecules with better attributes, which affects the diversity of generated molecules. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to solve the problem of insufficient generalization ability of existing molecular distribution mapping methods in multi-attribute optimization tasks with sparse data.

[0006] This invention solves the above-mentioned technical problems through the following technical solution: a multi-attribute optimization method for drug molecules based on multi-distribution mapping, comprising: Based on the number of attributes of the optimization task Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... At least one of the property values ​​of the drug molecule does not meet the success threshold; From subdataset , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... Any two in the set of source distribution molecules Middle lead molecule and target distribution set The sample input is obtained after the distribution labels are fused; The sample is input into the distribution mapper to generate a new molecule. The new molecule is then input into a pre-trained multi-distribution classifier. The output of the multi-distribution classifier is compared with the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input into the distribution mapper again to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated. The distribution mapper is trained based on the total contrast loss, mapping loss and cyclic reconstruction loss. All mapping paths are traversed to obtain the trained distribution mapper. The drug molecule to be optimized is fused with the target distribution label and then input into the trained distribution mapper to generate the optimized molecule.

[0007] This invention introduces an auxiliary intermediate domain between the source and target domains. The intermediate domains act as transitional bridges in the chemical space. The system learns a complete mapping network between the source, target, and all intermediate domains. By traversing all mapping paths, the model can form a robust and clear understanding of the attribute space. The intermediate domains provide architectural support for rare mappings between the source and target domains. All mapping paths train a distribution mapper with shared parameters. Regardless of the complexity of the domain or task, while maintaining a constant model size, it can learn a unified chemical space representation, thereby solving the problem of insufficient generalization ability of existing molecular distribution mapping methods in data-sparse multi-attribute optimization tasks.

[0008] Preferred source distribution molecular set Middle lead molecule and target distribution set The process of obtaining sample input after fusing the distribution labels includes: calculating the source distribution molecule set. Middle lead molecule Molecular fingerprints Leader molecule After word segmentation, it is compared with molecular fingerprint. The fusion yields sample data Sample data With the target distribution set The sample input is obtained by concatenating the distribution labels. .

[0009] Preferably, the pre-trained multi-distribution classifier includes an encoder, a linear mapping layer, and a log_softmax function connected in sequence. The new molecule is input into the pre-trained multi-distribution classifier, which performs word segmentation on the new molecule. The characters after word segmentation are numbered according to the vocabulary to obtain the encoding tag corresponding to each character in the new molecule. The position of each bit in the encoding tag is recorded to obtain the position encoding of the encoding tag. The encoding tag and the position encoding are added together and each bit is expanded to obtain the initial latent feature vector of the new molecule. The initial latent feature vector is input into the encoder to output the context-aware vector of the new molecule. The context-aware vector is input into the linear mapping layer, and the mapped output is calculated by the log_softmax function to obtain the distribution probability of the new molecule in each subset.

[0010] This invention first performs a linear mapping after the encoder in the Transformer, and then obtains the probability distribution of the new molecule with respect to each distribution through the log_softmax function, thereby enabling the prediction of the classification of multiple distributions.

[0011] Preferably, the training process of a pre-trained multi-distribution classifier includes: Based on sub-datasets , , ... , Training data is constructed using the distribution labels of each subset; drug molecules in the subsets are segmented into words, and the characters after segmentation are numbered according to the vocabulary to obtain the encoding label corresponding to each character in the drug molecule. The position of each bit in the encoding label is recorded to obtain the position encoding of the encoding label; the encoding label and the position encoding are added together and then expanded for each bit to obtain the initial latent feature vector of the drug molecule. The initial latent feature vector of the drug molecule is passed through the encoder, the Linear mapping layer and the log_softmax function in sequence to obtain the distribution probability prediction value of the drug molecule in each subset. When the loss value between the distribution probability prediction value and the distribution label is minimized, the pre-trained multi-distribution classifier is obtained.

[0012] A pre-trained multi-distribution classifier determines which distribution the mapped molecule belongs to, and the loss is calculated between the mapped distribution and the target distribution to help guide the distribution mapper to optimize towards the correct distribution.

[0013] Preferably, the distribution mapper includes an encoder and a decoder, which take the sample input... Input distribution mapper, sample input Enter the encoder and output sample input. Potential mapping representation Calculate the contrast loss for a single mapping. Input the sample The latent mapping represents the input decoder and the output new molecule. Calculate new molecules Molecular fingerprints New molecules After word segmentation, it is compared with molecular fingerprint. After fusion, the molecular set is distributed with the source. After the distributed labels are fused, they are input into the encoder again, and the output latent mapping representation is obtained. Calculate the contrast loss of the quadratic mapping. Represent the potential mapping Input decoder, output reconstructed molecule .

[0014] This invention abandons the traditional encoding-mapping-decoding process and completes the mapping process through encoder-decoder, which reduces information loss caused by encoding-decoding process compared with previous methods and further improves the performance of the model.

[0015] Preferred total contrast loss for:

[0016] Contrast loss of a single mapping Contrast loss with quadratic mapping The calculation method is the same:

[0017] , These are the intra-class aggregation loss and the inter-class separation loss, respectively.

[0018]

[0019] in, The number of distributions in an optimization task. For the first All samples from a distribution, , The first , All samples from a distribution, , These represent the vectors obtained after encoder and dimensionality reduction for the same or different samples, respectively. It represents a quantity that infinitely approaches 0 but is not zero.

[0020] Contrast loss improves domain separation by enhancing the latent representation in the Transformer encoder.

[0021] Preferably, mapping loss for:

[0022] in, Indicates the leader molecule With the target distribution set Distribution labels Fusion input distribution mapper The newly generated molecules, This represents a pre-trained multi-distribution classifier.

[0023] Mapping loss It is used to measure the difference between the distribution to which a molecule belongs and the target distribution after a single distribution mapping.

[0024] Preferably, the cyclic reconstruction loss for:

[0025] in, As a lead molecule, To reconstruct molecules, Source distribution molecular set The number of samples in the sample, , .

[0026] Cyclic reconstruction loss is used to measure the ability of a molecular sample in a distribution to be restored back to the original molecule after two distribution mappings, in order to maintain a certain degree of similarity between the molecules before and after optimization.

[0027] Preferably, the process of fusing the drug molecule to be optimized with the target distribution label and inputting it into a trained distribution mapper to generate the optimized molecule includes: Calculate the drug molecule to be optimized Molecular fingerprints Drug molecules to be optimized After word segmentation, it is compared with molecular fingerprint. Fusion to obtain sample data Sample data The actual input is obtained by concatenating the target distribution label with the actual input. This actual input is then fed into the encoder of the trained distribution mapper, which outputs the drug molecule to be optimized. Potential mapping representation Randomly generate a mapping representation with the potential Gaussian noise of the same shape Represent the potential mapping With Gaussian noise The concatenated molecules are then input into the decoder of the trained distribution mapper to obtain the optimized molecules.

[0028] This invention also provides a drug molecule multi-attribute optimization system based on multi-distribution mapping, comprising: The dataset decomposition module is used to decompose data based on the number of attributes in the optimization task. Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... At least one of the property values ​​of the drug molecule does not meet the success threshold; The feature fusion module is used to fuse features from a subset of datasets. , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... Any two in the set of source distribution molecules Middle lead molecule and target distribution set The sample input is obtained after the distribution labels are fused; The multi-distribution mapping training module is used to input samples into a distribution mapper to generate new molecules. These new molecules are then input into a pre-trained multi-distribution classifier, whose output is compared to the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input into the distribution mapper again to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated. The distribution mapper is trained based on the total contrast loss, mapping loss and cyclic reconstruction loss. All mapping paths are traversed to obtain the trained distribution mapper. The inference module is used to fuse the drug molecule to be optimized with the target distribution label and input it into the trained distribution mapper to generate the optimized molecule.

[0029] The advantages provided by this invention are as follows: Compared with existing methods, this invention uses the encoder-decoder as a mapper to directly perform mapping, which solves the problem of information loss caused in the self-encoding and decoding process; by expanding the optimization target attribute value from a single value (target attribute value) to a target range (greater than or less than the threshold range), this invention has a larger optimization range and is more likely to produce better attributes and more diverse molecules. Attached Figure Description

[0030] Figure 1 The flowchart shows the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the training of a pre-trained multi-distribution classifier in the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 3 This is a flowchart of the distribution mapper training process in the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 4 This is a schematic diagram of all mapping paths trained in the multi-attribute optimization method for drug molecules based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 5 This is a schematic diagram of the principle of training a distribution mapper based on total contrast loss, mapping loss and cyclic reconstruction loss in the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 6 The source distribution molecule set in the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention With the target distribution set Schematic diagram of bidirectional mapping; Figure 7 This is a flowchart illustrating the generation of optimized molecules in the drug molecule multi-attribute optimization method based on multi-distribution mapping provided in Embodiment 1 of the present invention. Figure 8 This is a schematic diagram of a drug molecule multi-attribute optimization system based on multi-distribution mapping provided in Embodiment 2 of the present invention; Figure 9 This is a schematic diagram illustrating the working principle of the feature fusion module in the drug molecule multi-attribute optimization system based on multi-distribution mapping provided in Embodiment 2 of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0032] Example 1 like Figure 1 and Figure 3 As shown, this embodiment provides a method for optimizing multiple attributes of drug molecules based on multi-distribution mapping, including the following steps: Step 1: Obtain the drug molecule dataset ,in, Indicates the first One drug molecule, n The total number of drug molecules, based on the number of attributes of the optimization task. Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... In this invention, at least one attribute value of a drug molecule fails to meet the success threshold. The subset of data where none of the drug molecule attribute values ​​meet the success threshold is used as the source domain, and the subset of data where all drug molecule attribute values ​​meet the success threshold is used as the target domain. , ... These are the intermediate fields between the source domain and the target domain. There are a total of [number missing] intermediate fields. indivual.

[0033] This invention will extract drug molecule datasets The drug molecules sampled in the middle are denoted as For those with The optimization task for each attribute makes Indicates the first drug molecules The Each attribute value This represents the success threshold for each attribute. Non-overlapping subsets , , ... , By satisfying The success threshold for each attribute is defined using a binary decision. For example... When, that is, for a two-attribute The optimization task aims to optimize these two attributes to a level higher or lower than the thresholds set for these attributes. Based on these two thresholds, the drug molecule dataset can be... M Decomposed into four non-overlapping subsets , , , Where a, b, c, and d are subsets of the dataset. A , , , C The number of samples in the dataset, where a+b+c+d=n. Taking molecules with attribute values ​​higher than the success threshold as optimized molecules as an example, the subset dataset... A Both attribute ranges of the samples are below the set threshold range, i.e. , Sub-dataset The sample contains one attribute range below the set threshold range and one attribute range above the set threshold range, i.e. , Sub-dataset The sample contains one attribute range that is above the set threshold range and one attribute range that is below the set threshold range, i.e. , Sub-dataset C The samples in the dataset have both attribute ranges higher than the set threshold range, i.e. , .

[0034] For a two-attribute Optimize tasks by setting thresholds. Drug molecule datasets can be used Decomposed into four subsets A , , , C Four sub-datasets A , , , C This represents four different data distributions. This invention abstractly represents each data distribution using a one-bit label, with each subset of data corresponding to a unique distribution label. For example, for a subset of data...A Its distribution label is 0; sub-dataset The distribution label is 1; the subset of datasets The distribution label is 2; the subset of datasets C The distribution label is 3.

[0035] Current molecular distribution mapping methods mainly focus on mapping directly from the source domain to the target domain. Since direct mapping from the source domain to the target domain mainly focuses on learning direct mapping relationships, isolated source and target distributions (e.g., from suboptimal regions to ideal regions) cannot capture the rich and structured interdependencies in the entire spectrum of attribute-defined distributions. This simplistic perspective ignores the guiding and stabilizing role of intermediate auxiliary distributions in the final transformation process. Therefore, in multi-attribute optimization scenarios with sparse data, existing methods often exhibit insufficient generalization ability and unreliable mapping results.

[0036] Unlike previous methods that only model isolated distribution pairs, this invention introduces an auxiliary intermediate domain between the source and target domains. The intermediate domain acts as a transitional bridge in the chemical space, and the system learns the complete mapping network between the source domain, target domain, and all intermediate domains. This enables the mapper to more effectively capture complex and multidimensional attribute landscapes and build a richer learning framework.

[0037] Step 2, from the subset dataset , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... "Any two" means that this invention does not consider the mapping paths of any two intermediate distributions. See also Figure 9 , Figure 9 This is a feature fusion module used to integrate source distribution molecules. Middle lead molecule and target distribution set The distribution labels are fused to obtain the sample input. The specific process of feature fusion is as follows: calculate the source distribution molecular set. Middle lead molecule Molecular fingerprints Leader molecule After word segmentation, it is compared with molecular fingerprint. The fusion yields sample data Sample data With the target distribution set The sample input is obtained by concatenating the distribution labels. .

[0038] Using token segmentation from a drug molecule dataset Extracting vocabulary ,in, Indicates the first One word, Indicates the maximum length of the vocabulary; leading molecule After word segmentation, the segmented characters are obtained, which are the SMILES sequence features of the leader molecule. The segmented characters are then numbered according to the vocabulary to obtain the encoding tag corresponding to each character in the segmented characters. ,in, This indicates the first character in the segmented text. l The encoding tags corresponding to each character convert the SMILES sequence features of the leader molecule into a numerical sequence, which is then input into the distribution mapper for training.

[0039] Step 3: Input the sample into the distribution mapper to generate a new molecule. Input the new molecule into a pre-trained multi-distribution classifier. The output of the multi-distribution classifier is compared with the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input again into the distribution mapper to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated, based on the total contrast loss. Mapping loss and cyclic reconstruction loss Train the distribution mapper and use gradient descent to train the distribution mapper. Perform training, traverse all mapping paths, and calculate the total loss. The model parameters are updated until the total loss converges or the maximum number of iterations is reached. Training then stops, and the trained distribution mapper is considered the optimal distribution mapper, used to generate new target drug molecules. The total loss of this invention... The loss for all mapping paths is the total contrastive loss for each mapping path. Mapping loss and cyclic reconstruction loss The weighted sum.

[0040] The pre-trained multi-distribution classifier consists of an encoder, a linear mapping layer, and a log_softmax function connected in sequence. New molecules are input into the pre-trained multi-distribution classifier, which then processes the new molecules... The process involves word segmentation, padding or truncating the character sequence according to a preset maximum length L to obtain a certain number of characters. Specifically, a maximum length is set for each character sequence; if the length exceeds this range, subsequent characters need to be truncated. If the sequence length is insufficient, the special symbol [PAD] is used for padding to reach the maximum length.

[0041] The characters after word segmentation are numbered according to the vocabulary to obtain the encoding tag corresponding to each character in the new molecule. ,in, Indicates new molecules The Middle l Each character corresponds to an encoding tag. The position of each bit in the encoding tag is recorded to obtain the position encoding of the encoding tag. ; Encode the tag and position encoding After adding them together, we expand each molecule to obtain new molecules. Initial latent feature vectors ,in, express The i-th position Features in each dimension The total number of feature dimensions represents the initial latent feature vector. Input the encoder of the Transformer, output the context-aware vector of the new molecule. ,in, express The Features in one dimension; This represents the total number of feature dimensions; context-aware vectors Input a Linear mapping layer and obtain the mapped output. The mapped output is then processed by the log_softmax function to obtain the probability distribution of the new molecule in each subset of the dataset. .

[0042]

[0043]

[0044]

[0045] in, This represents the distribution quantity.

[0046] This invention first performs a linear mapping after the encoder in the Transformer, and then obtains the probability distribution of the new molecule with respect to each distribution through the log_softmax function, thereby enabling the prediction of the classification of multiple distributions.

[0047] See Figure 2 The training process of a pre-trained multi-distribution classifier includes: Based on sub-datasets , , ... , Training data is constructed using the distribution labels of each subset. Drug molecules in the subsets are segmented into words, and the characters after segmentation are numbered according to the vocabulary to obtain the encoding label corresponding to each character in the drug molecule. The position of each bit in the encoding label is recorded to obtain the positional encoding of the encoding label. The encoding label and positional encoding are added together, and each bit is expanded to obtain the initial latent feature vector of the drug molecule. The initial latent feature vector of the drug molecule is then passed through an encoder, a Linear mapping layer, and a log_softmax function to obtain the predicted distribution probability value of the drug molecule in each subset. A multi-distribution classifier is trained, and the loss value between the predicted distribution probability value and the distribution label is calculated. To update the encoder parameters when the predicted probability distribution is compared with the loss value of the distributed label. When the minimum value is reached, a pre-trained multi-distribution classifier is obtained.

[0048] See Figure 5 The distribution mapper includes the Transformer's encoder and decoder, taking the sample input... Input distribution mapper, sample input Enter the encoder and output sample input. Potential mapping representation Calculate the contrast loss for a single mapping. Input the sample The latent mapping represents the input decoder and the output new molecule. Calculate new molecules Molecular fingerprints New molecules After word segmentation, it is compared with molecular fingerprint. After fusion, the molecular set is distributed with the source. After the distributed labels are fused, they are input into the encoder again, and the output latent mapping representation is obtained. Calculate the contrast loss of the quadratic mapping. Represent the potential mapping Input decoder, output reconstructed molecule .

[0049] The distribution mapper of this invention is based on the Transformer model and introduces a multi-distribution classifier to guide the optimization process of the distribution mapper. Combining the distribution mapper and the multi-distribution classifier results in better performance compared to previous deep generative models. This invention builds upon the traditional method of mapping from one distribution to another by introducing a multi-distribution classifier as additional guidance to help the multi-distribution mapper find the optimal model parameters.

[0050] Furthermore, this invention abandons the traditional encoding-mapping-decoding process and completes the mapping process through encoder-decoder, which reduces information loss caused during the encoding-decoding process compared to previous methods and further improves the performance of the model.

[0051] See Figure 6 In training the distribution mapper, this invention traverses all mapping paths, i.e., the set of source distribution molecules. Target distribution set Select all subsets that meet the conditions, including all mapping paths for each subset. Each with a subset , ... , The bidirectional mapping path, and the subset dataset Each with a subset , ... A bidirectional mapping path. This embodiment uses four subsets of datasets. A , , , C For an example, please refer to the section on all paths. Figure 4 There are ten mapping paths in total, namely: when the subset of data A Source distribution molecular set Sub-dataset For the target distribution set At that time, subset dataset A To sub-dataset The first mapping path; when the subset A Source distribution molecular set Sub-dataset For the target distribution set At that time, subset dataset A To sub-dataset The second mapping path; when the subset A Source distribution molecular set Sub-dataset C For the target distribution set At that time, subset dataset ATo sub-dataset C The third mapping path; when the subset of data Source distribution molecular set Sub-dataset A For the target distribution set At that time, subset dataset To sub-dataset A The fourth mapping path; when the subset of data Source distribution molecular set Sub-dataset C For the target distribution set At that time, subset dataset To sub-dataset C The fifth mapping path; when the subset of data Source distribution molecular set Sub-dataset A For the target distribution set At that time, subset dataset To sub-dataset A The sixth mapping path; when the subset of data Source distribution molecular set Sub-dataset C For the target distribution set At that time, subset dataset To sub-dataset C The seventh mapping path; when the subset C Source distribution molecular set Sub-dataset For the target distribution set At that time, subset dataset C To sub-dataset The eighth mapping path; when the subset C Source distribution molecular set Sub-dataset For the target distribution set At that time, subset dataset C To sub-dataset The ninth mapping path; when the subset of data C Source distribution molecular set Sub-dataset For the target distribution set At that time, subset dataset C To sub-dataset The tenth mapping path.

[0052] This invention introduces an auxiliary distribution , and related mapping paths , , , This allows the model to not only focus on the mapping relationship between the source distribution and the target distribution, but also to take into account the distribution of other molecules in the chemical space. The model has a greater expansion and understanding of spatial distribution than previous methods.

[0053] Total comparative loss for:

[0054] Contrast loss of a single mapping Contrast loss with quadratic mapping The calculation method is the same:

[0055] , These are the intra-class aggregation loss and the inter-class separation loss, respectively.

[0056]

[0057] in, The number of distributions in an optimization task. For the first All samples from a distribution, , The first , All samples from a distribution, , These represent the vectors obtained after encoder and dimensionality reduction for the same or different samples, respectively. It represents a quantity that infinitely approaches 0 but is not zero.

[0058] Contrast loss improves domain separation by enhancing the latent representations in the Transformer encoder, using a contrast loss for a single mapping. For example, the contrast loss of a single mapping Including intra-class aggregation loss Inter-class separation loss By employing a contrastive learning strategy of clustering intra-class samples and dispersing inter-class samples, molecules achieve clustering effects across different domains through their representations mapped from the Encoder in the Transformer to the latent space. Specifically, for a batch E of samples sampled in a single run, containing the same number of samples from all distributions... Indicates the first All samples from a distribution, , These represent the vectors obtained after passing through the encoder and dimensionality reduction of the same or different samples, respectively. Representing vectors , Cosine similarity between them.

[0059]

[0060] To achieve intra-class sample clustering, we want molecules with the same distribution to be close together in the latent space. The closer two vectors from the same distribution are, the better; that is, the smaller the cosine similarity between the vectors, the better. Cosine similarity is minimized when two vectors are parallel. To achieve inter-class sample clustering, we want molecules with different distributions to be far apart in the latent space. The greater the difference between two vectors from different distributions, the better; that is, the greater the cosine similarity between the vectors, the better. Cosine similarity is maximized when two vectors are perpendicular.

[0061] For intra-class aggregation loss, our goal is to make samples in the same domain as close as possible. , As similar as possible. Limit. hour, , making For inter-class distance loss, our goal is to keep samples from different domains as far apart as possible. , Try to be as different as possible. or When the magnitude of the molecular embedding vectors from two different domains is the same (or opposite in direction), it indicates that the two vectors are very close in the latent space. , Approaching 0 maximizes the penalty; hour, , This minimizes the penalty. Experiments revealed that due to the existence of other losses, It will not cluster all sample vectors into a single, identical vector.

[0062] Mapping loss Used to measure the difference between the distribution to which a molecule belongs and the target distribution after a single distribution mapping. Mapping loss. for:

[0063] in Represents the calculation of the leading molecule The difference between the probability distribution of the molecule and the target distribution after a single distribution mapping. Indicates the leader molecule With the target distribution set Distribution labels Fusion input distribution mapper New molecules generated , Source distribution molecular set The number of samples in the sample, , . This represents a pre-trained multi-distribution classifier.

[0064] This invention, while training the distribution mapper, also trains a pre-trained multi-distribution classifier, and uses the multi-distribution classifier... Determine which distribution the mapped molecule belongs to, and calculate the loss between the mapped distribution and the target distribution to help guide the distribution mapper. Optimize towards the correct distribution direction.

[0065] Cyclic reconstruction loss measures the ability of a molecular sample in a distribution to be reconstructed back into the original molecule after two distribution mappings, in order to maintain a certain degree of similarity between the molecules before and after optimization. for:

[0066] in, This indicates the difference between the reconstructed molecule obtained after the lead molecule undergoes two distribution mappings and the original molecule. As a lead molecule, To reconstruct molecules, , Source distribution molecular set Distribution labels, For the target distribution set Distribution labels, Source distribution molecular set The number of samples in the sample, , .

[0067] Step 4: After fusing the drug molecule to be optimized with the target distribution label, input it into the trained distribution mapper to generate the optimized molecule. See also Figure 7 The specific process is as follows: Calculate the drug molecule to be optimized. Molecular fingerprints Drug molecules to be optimized After word segmentation, it is compared with molecular fingerprint. Fusion to obtain sample data Sample data The actual input is obtained by concatenating the target distribution label with the actual input. This actual input is then fed into the encoder of the trained distribution mapper, which outputs the drug molecule to be optimized. Potential mapping representation Randomly generate a mapping representation with the potential Gaussian noise of the same shape Represent the potential mapping With Gaussian noise The concatenated molecules are then input into the decoder of the trained distribution mapper to obtain the optimized molecules.

[0068] This invention enables the model to form a robust and clear understanding of the attribute space by traversing all mapping paths. The intermediate domain provides architectural support for rare mappings between the source and target domains. All mapping paths train a distributed mapper with shared parameters, allowing the model to learn a unified chemical space representation while maintaining a constant model size, regardless of the complexity of the domain or task. Furthermore, this invention combines the sequence modeling capabilities of the Transformer with adversarial domain transformation, and employs contrastive learning to enforce discriminative alignment of features between domains.

[0069] Example 2 See Figure 8 This embodiment provides a drug molecule multi-attribute optimization system based on multi-distribution mapping, including: The dataset decomposition module is used to decompose data based on the number of attributes in the optimization task. Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... At least one of the property values ​​of the drug molecule does not meet the success threshold; The feature fusion module is used to fuse features from a subset of datasets. , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... Any two in the set of source distribution molecules Middle lead molecule and target distribution set The sample input is obtained after the distribution labels are fused.

[0070] See Figure 9 Source distribution molecular set Middle lead molecule and target distribution set The process of obtaining sample input after fusing the distribution labels includes: calculating the source distribution molecule set. Middle lead molecule Molecular fingerprints Leader molecule After word segmentation, it is compared with molecular fingerprint. The fusion yields sample data Sample data With the target distribution set The sample input is obtained by concatenating the distribution labels. .

[0071] The multi-distribution mapping training module is used to input samples into a distribution mapper to generate new molecules. These new molecules are then input into a pre-trained multi-distribution classifier, whose output is compared to the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input into the distribution mapper again to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated. The distribution mapper is trained based on the total contrast loss, mapping loss and cyclic reconstruction loss. All mapping paths are traversed to obtain the trained distribution mapper.

[0072] The pre-trained multi-distribution classifier consists of an encoder, a linear mapping layer, and a log_softmax function connected in sequence. The new molecule is input into the pre-trained multi-distribution classifier, which performs word segmentation. The segmented characters are numbered according to the vocabulary to obtain the encoding tag corresponding to each character in the new molecule. The position of each bit in the encoding tag is recorded to obtain the positional encoding of the encoding tag. The encoding tag and positional encoding are added together, and each bit is expanded to obtain the initial latent feature vector of the new molecule. The initial latent feature vector is input into the encoder, which outputs the context-aware vector of the new molecule. The context-aware vector is input into the linear mapping layer, and the mapped output is calculated by the log_softmax function to obtain the distribution probability of the new molecule in each subset.

[0073] The training process of a pre-trained multi-distribution classifier includes: Based on sub-datasets , , ... , Training data is constructed using the distribution labels of each subset; drug molecules in the subsets are segmented into words, and the characters after segmentation are numbered according to the vocabulary to obtain the encoding label corresponding to each character in the drug molecule. The position of each bit in the encoding label is recorded to obtain the position encoding of the encoding label; the encoding label and the position encoding are added together and then expanded for each bit to obtain the initial latent feature vector of the drug molecule. The initial latent feature vector of the drug molecule is passed through the encoder, the Linear mapping layer and the log_softmax function in sequence to obtain the distribution probability prediction value of the drug molecule in each subset. When the loss value between the distribution probability prediction value and the distribution label is minimized, the pre-trained multi-distribution classifier is obtained.

[0074] The distribution mapper includes an encoder and a decoder, which take the sample input... Input distribution mapper, sample input Enter the encoder and output sample input. Potential mapping representation Calculate the contrast loss for a single mapping. Input the sample The latent mapping represents the input decoder and the output new molecule. Calculate new molecules Molecular fingerprints New molecules After word segmentation, it is compared with molecular fingerprint. After fusion, the molecular set is distributed with the source. After the distributed labels are fused, they are input into the encoder again, and the output latent mapping representation is obtained. Calculate the contrast loss of the quadratic mapping. Represent the potential mapping Input decoder, output reconstructed molecule .

[0075] Total comparative loss for:

[0076] Contrast loss of a single mapping Contrast loss with quadratic mapping The calculation method is the same:

[0077] , These are the intra-class aggregation loss and the inter-class separation loss, respectively.

[0078]

[0079] in, The number of distributions in an optimization task. For the first All samples from a distribution, , The first , All samples from a distribution, , These represent the vectors obtained after encoder and dimensionality reduction for the same or different samples, respectively. It represents a quantity that infinitely approaches 0 but is not zero.

[0080] Mapping loss for:

[0081] in, Indicates the leader molecule With the target distribution set Distribution labels Fusion input distribution mapper The newly generated molecules, This represents a pre-trained multi-distribution classifier.

[0082] Cyclic Reconstruction Loss for:

[0083] in, As a lead molecule, To reconstruct molecules, Source distribution molecular set The number of samples in the sample, , .

[0084] The inference module is used to fuse the drug molecule to be optimized with the target distribution label and input it into the trained distribution mapper to generate the optimized molecule.

[0085] The process of fusing the drug molecule to be optimized with the target distribution label and inputting it into a trained distribution mapper to generate the optimized molecule includes: Calculate the drug molecule to be optimized Molecular fingerprints Drug molecules to be optimized After word segmentation, it is compared with molecular fingerprint. Fusion to obtain sample data Sample data The actual input is obtained by concatenating the target distribution label with the actual input. This actual input is then fed into the encoder of the trained distribution mapper, which outputs the drug molecule to be optimized. Potential mapping representation Randomly generate a mapping representation with the potential Gaussian noise of the same shape Represent the potential mapping With Gaussian noise The concatenated molecules are then input into the decoder of the trained distribution mapper to obtain the optimized molecules.

[0086] Comparative experiment This embodiment designs three multi-attribute optimization tasks to optimize specific molecular properties while maintaining minimal structural similarity to the original molecule. The three multi-attribute optimization tasks are as follows: Task 1: Optimize the similarity between QED (≥0.7) and DRD2 (≥0.3) to ≥0.3.

[0087] Task 2: Optimize the similarity between QED (≥0.7) and SA (≥0.7) to ≥0.3.

[0088] Task 3: Optimize the similarity of GSK3B (≥0.4), QED (≥0.6), and SA (≥0.6) to ≥0.2.

[0089] Among these, QED refers to drug-likeness, an indicator that measures whether a compound is more closely related to the basic structure of a synthetic drug, ranging from 0 to 1, with higher values ​​indicating better conformity. SA is synthetic accessibility, used to measure the ease with which a compound can be chemically synthesized, ranging from 0 to 1, with higher values ​​indicating easier synthesis. DRD2 is the dopamine D2 receptor, generally used to measure the binding activity of a molecule to the DRD2 receptor, ranging from 0 to 1, with higher values ​​being better. GSK3B is glycogen synthase kinase 3β, and the inhibitory activity of a molecule on GSK3B is related to this, ranging from 0 to 1, with higher values ​​being better.

[0090] Molecular domains were partitioned based on the satisfaction of attribute thresholds. In the two-attribute task, four domains were defined: source domain A (not satisfying any threshold), target domain C (satisfying both thresholds), and auxiliary domains B1 and B2 (satisfying only one threshold). For the three-attribute task, eight domains were defined, including several intermediate domains. All molecular samples were selected from the ZINC250K database, with a balanced domain distribution: Task 1 contained 4,281 training molecules and 1,070 test molecules per domain; Tasks 2 and 3 contained 4,000 training molecules and 1,000 test molecules per domain.

[0091] Performance evaluation employed three metrics: the average attribute value of the optimized effective molecules, the mean cosine similarity between the optimized molecules and the original molecules, and the success rate (the percentage of optimized effective molecules that meet all target attribute thresholds and minimum similarity requirements). Following the IPCA protocol, 20 optimization candidate molecules were generated for each test set. A repeated random sampling procedure was used for 10 independent runs, each time randomly selecting a candidate molecule from the source molecules and calculating the metrics for the resulting set. Source molecules that did not yield effective candidate molecules from any model during this run were excluded from evaluation by all models.

[0092] Baseline methods include MIMOA, UGMMT, IPCA, and MOMP. MIMOA employs a Markov chain Monte Carlo (MCMC) sampling strategy, combining two pre-trained graph neural networks (mGNN for substructure type prediction and bGNN for molecular topology prediction). UGMMT consists of a Molecular Embedding Translation Network (METN) and an Embedding-Embedding Translation Network (EETN). METN is based on a Variational Autoencoder (VAE), while EETN uses a bottleneck-structured converter. IPCA integrates a multi-recurrent architecture, including multiple attribute-specific encoders and converters, as well as a shared central converter and decoder, built on GRU and a convolutional bottleneck network. MOMP fuses METN, EETN, and an extended EETN, introducing a molecular attention mechanism and patent fingerprint constraints, and employs a multi-domain transformation architecture (A→B→C). MIMOA performs well on the complete dataset (A, B, C). , The GNN and latent space model retrained on C) are UGMMT source training set as follows: The target domain is C; the source training set of IPCA is domain A, and the target is... MOMP maps the domains to the desired three-domain structure (A, B, C, D, E, F ... The evaluation results of the QED and DRD2 optimization tasks are shown in Table 1, the evaluation results of the QED and SA optimization tasks are shown in Table 2, and the evaluation results of the GSK3B, QED, and SA optimization tasks are shown in Table 3.

[0093] Table 1 Evaluation results of QED and DRD2 optimization tasks

[0094] As shown in Table 1, our invention achieved the highest success rate (SR) of 0.194, which is almost 4.5 times that of the second-best baseline method, IPCA (0.042). While maintaining a competitive QED score (0.795 ± 0.003), our method significantly improved the DRD2 score to 0.471 ± 0.015, far exceeding the best baseline method (approximately 50% higher than IPCA and UGMMT).

[0095] Table 2 Evaluation results of QED and SA optimization tasks

[0096] As shown in Table 2, this invention achieved the highest SA score of 0.835 and a success rate of 0.121. Notably, this invention is the only method that effectively balances high SA with high similarity (0.242), while mimosa failed to do so. QED (0.440) was used to maintain similarity, while IPCA failed to achieve any success (SR of 0.000).

[0097] Table 3 Evaluation results of GSK3B, QED and SA optimization tasks

[0098] As can be seen from Table 3, the present invention exhibits robust performance, achieving the best QED (0.669) and GSK3B (0.108) scores, as well as the highest success rate (0.048). Most baseline methods, such as mimosa and IPCA, fail completely in this task with a success rate of only 0.000.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing multiple attributes of drug molecules based on multi-distribution mapping, characterized in that: include: Based on the number of attributes of the optimization task Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... At least one of the property values ​​of the drug molecule does not meet the success threshold; From subdataset , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... Any two in the set of source distribution molecules Middle lead molecule and target distribution set The sample input is obtained after the distribution labels are fused; The sample is input into the distribution mapper to generate a new molecule. The new molecule is then input into a pre-trained multi-distribution classifier. The output of the multi-distribution classifier is compared with the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input into the distribution mapper again to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated. The distribution mapper is trained based on the total contrast loss, mapping loss and cyclic reconstruction loss. All mapping paths are traversed to obtain the trained distribution mapper. The drug molecule to be optimized is fused with the target distribution label and then input into the trained distribution mapper to generate the optimized molecule.

2. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: Source distribution molecular set Middle lead molecule and target distribution set The process of obtaining sample input after fusing the distribution labels includes: calculating the source distribution molecule set. Middle lead molecule Molecular fingerprints Leader molecule After word segmentation, it is compared with molecular fingerprint. The fusion yields sample data Sample data With the target distribution set The sample input is obtained by concatenating the distribution labels. .

3. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: The pre-trained multi-distribution classifier consists of an encoder, a linear mapping layer, and a log_softmax function connected in sequence. The new molecule is input into the pre-trained multi-distribution classifier, which performs word segmentation. The characters after segmentation are numbered according to the vocabulary to obtain the encoding tag corresponding to each character in the new molecule. The position of each bit in the encoding tag is recorded to obtain the positional encoding of the encoding tag. The encoding tag and positional encoding are added together, and each bit is expanded to obtain the initial latent feature vector of the new molecule. The initial latent feature vector is input into the encoder, which outputs the context-aware vector of the new molecule. The context-aware vector is input into the linear mapping layer, and the mapped output is calculated by the log_softmax function to obtain the distribution probability of the new molecule in each subset.

4. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: The training process of a pre-trained multi-distribution classifier includes: Based on sub-datasets , , ... , Training data is constructed using the distribution labels of each subset; drug molecules in the subsets are segmented into words, and the characters after segmentation are numbered according to the vocabulary to obtain the encoding label corresponding to each character in the drug molecule. The position of each bit in the encoding label is recorded to obtain the position encoding of the encoding label; the encoding label and the position encoding are added together and then expanded for each bit to obtain the initial latent feature vector of the drug molecule. The initial latent feature vector of the drug molecule is passed through the encoder, the Linear mapping layer and the log_softmax function in sequence to obtain the distribution probability prediction value of the drug molecule in each subset. When the loss value between the distribution probability prediction value and the distribution label is minimized, the pre-trained multi-distribution classifier is obtained.

5. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: The distribution mapper includes an encoder and a decoder, which take the sample input... Input distribution mapper, sample input Enter the encoder and output sample input. Potential mapping representation Calculate the contrast loss for a single mapping. Input the sample The latent mapping represents the input decoder and the output new molecule. Calculate new molecules Molecular fingerprints New molecules After word segmentation, it is compared with molecular fingerprint. After fusion, the molecular set is distributed with the source. After the distributed labels are fused, they are input into the encoder again, and the output latent mapping representation is obtained. Calculate the contrast loss of the quadratic mapping. Represent the potential mapping Input decoder, output reconstructed molecule .

6. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: Total comparative loss for: Contrast loss of a single mapping Contrast loss with quadratic mapping The calculation method is the same: , These are the intra-class aggregation loss and the inter-class separation loss, respectively. in, The number of distributions in an optimization task. For the first All samples from a distribution, , The first , All samples from a distribution, , These represent the vectors obtained after encoder and dimensionality reduction for the same or different samples, respectively. It represents a quantity that infinitely approaches 0 but is not zero.

7. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: Mapping loss for: in, Indicates the leader molecule With the target distribution set Distribution labels Fusion input distribution mapper The newly generated molecules, This represents a pre-trained multi-distribution classifier.

8. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: Cyclic Reconstruction Loss for: in, As a lead molecule, To reconstruct molecules, Source distribution molecular set The number of samples in the sample, , .

9. The method for optimizing multiple attributes of drug molecules based on multi-distribution mapping according to claim 1, characterized in that: The process of fusing the drug molecule to be optimized with the target distribution label and inputting it into a trained distribution mapper to generate the optimized molecule includes: Calculate the drug molecule to be optimized Molecular fingerprints Drug molecules to be optimized After word segmentation, it is compared with molecular fingerprint. Fusion to obtain sample data Sample data The actual input is obtained by concatenating the target distribution label with the actual input. This actual input is then fed into the encoder of the trained distribution mapper, which outputs the drug molecule to be optimized. Potential mapping representation Randomly generate a mapping representation with the potential Gaussian noise of the same shape Represent the potential mapping With Gaussian noise The concatenated molecules are then input into the decoder of the trained distribution mapper to obtain the optimized molecules.

10. A drug molecule multi-attribute optimization system based on multi-distribution mapping, characterized in that: include: The dataset decomposition module is used to decompose data based on the number of attributes in the optimization task. Decompose the drug molecule dataset into Non-overlapping subsets , , ... , And set distribution labels for each subset, where the subsets The attribute values ​​of the drug molecules in the subset did not meet the success threshold. The attribute values ​​of the drug molecules in the subset all meet the success threshold. , ... At least one of the property values ​​of the drug molecule does not meet the success threshold; The feature fusion module is used to fuse features from a subset of datasets. , , ... , Two subsets of the dataset were selected as the source distribution molecule sets. Target distribution set , , Non-overlapping and not simultaneously , ... Any two in the set of source distribution molecules Middle lead molecule and target distribution set The sample input is obtained after the distribution labels are fused; The multi-distribution mapping training module is used to input samples into a distribution mapper to generate new molecules. These new molecules are then input into a pre-trained multi-distribution classifier, whose output is compared to the target distribution set. The distribution labels are used as the mapping loss, and the new molecule is mapped to the source distribution molecule set. After the distribution labels are fused, they are input into the distribution mapper again to generate reconstructed molecules. The cyclic reconstruction loss between the reconstructed molecules and the leader molecules is calculated. The distribution mapper is trained based on the total contrast loss, mapping loss and cyclic reconstruction loss. All mapping paths are traversed to obtain the trained distribution mapper. The inference module is used to fuse the drug molecule to be optimized with the target distribution label and input it into the trained distribution mapper to generate the optimized molecule.

Images