A deep learning-based tumor drug sensitivity prediction method and system

Abstract

The application discloses a kind of based on deep learning's tumor drug sensitivity prediction method and system, it is related to artificial intelligence medical technology field, including, using node edge joint graph neural network constructs the molecular graph representation between the atomic node of candidate drug, chemical bond edge and molecular descriptor, and by drug substructure mask reconstruction restores the atom type, bond type and functional group category being masked, generates drug substructure sequence and drug overall graph representation;Pathogenic gene's multiomics representation is arranged as gene sequence, and the abnormal conduction relationship of gene in pre-training learning signal path is recovered using transformer encoder network combined with path mask, generates gene path activity sequence and path overall representation.The application improves the consistency of half-inhibitory concentration prediction results in multi-drug, multiple sample scenarios by correcting drug sensitivity order and sample response order.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence in the field of medical technology, and in particular to a method and system for predicting tumor drug sensitivity based on deep learning. Background Technology

[0002] Tumor drug sensitivity prediction is a crucial computational task in precision medicine and anti-tumor drug screening. Common methods typically characterize the state of tumor samples using multi-omics data such as gene expression, mutations, methylation, and copy number variations, and characterize drug properties using drug molecular structure, molecular descriptors, and drug target information. With the development of deep learning technology, graph neural networks are often used to process drug molecular graph structures, transformer-encoding networks are often used to extract long-range dependencies in gene sequences or multi-omics sequences, and attention mechanisms are often used to enhance the fusion expression between drug features and sample features, thereby achieving the prediction of half-maximal inhibitory concentration, drug response score, or drug sensitivity status.

[0003] Existing deep learning-based methods for predicting tumor drug sensitivity often splice or fuse the overall representation of the drug molecule with the overall multi-omics representation of the tumor sample, and then output the sensitivity prediction value through a regression network. Although these methods can utilize drug structure information and tumor omics information, they usually lack consistent mechanistic expression between drug substructures, drug target genes, signaling pathway abnormalities, and sensitivity results. It is difficult to explain which pathway abnormality a specific drug substructure acts on, and it is also difficult to maintain consistency between multi-drug rankings within the same sample and multi-sample response rankings of the same drug. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for predicting tumor drug sensitivity to solve the problem of inconsistent matching between drug substructures and abnormal tumor pathways.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a deep learning-based method for predicting tumor drug sensitivity, comprising,

[0008] A node-edge joint graph neural network is used to construct a molecular graph representation of candidate drugs between atomic nodes, chemical bond edges and molecular descriptors. The masked atom types, bond types and functional group categories are recovered by drug substructure mask reconstruction to generate drug substructure sequences and overall drug graph representation.

[0009] The multi-omics representation of pathogenic genes is arranged into gene sequences, and the abnormal gene transmission relationship within the pre-trained learning signal pathway is recovered by using a transformer coding network combined with pathway masking, generating gene pathway activity sequences and overall pathway representations;

[0010] Bidirectional cross-attention matching is performed on drug substructure sequences and gene pathway activity sequences, and drug target pathway embedding and target pathway contrast learning are combined to generate drug pathway interaction representations.

[0011] The drug pathway interaction representation, drug overall graph representation, and pathway overall representation are fed into the sensitivity regression network, and the sensitivity order between drugs and the response order between samples are corrected through sensitivity ranking consistency learning to generate the half-maximum inhibitory concentration prediction value and sensitivity ranking results.

[0012] Based on the predicted half-maximal inhibitory concentration (MCI), sensitivity ranking results, and drug pathway interaction representations, the contributing drug substructures, target genes, and signaling pathways are identified, and tumor drug sensitivity prediction results are generated.

[0013] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the construction of the molecular graph representation specifically includes:

[0014] The normalized molecular structure of candidate drugs is analyzed to form atomic index sequences, chemical bond connection sequences and drug adjacency matrices. Atomic properties, chemical bond properties and molecular physicochemical properties are extracted respectively to generate node feature matrices, edge feature matrices and molecular descriptor vectors.

[0015] The node-edge joint graph neural network defines the range of adjacent atoms based on the drug adjacency matrix, maps the node feature matrix to the initial atom latent vector, maps the edge feature matrix to the chemical bond latent vector, and uses the chemical bond latent vector to modulate the message passing between adjacent initial atom latent vectors to obtain the updated atom latent vector.

[0016] The molecular descriptor vector is extended to the atomic positions corresponding to each updated atom latent vector, and then concatenated with the updated atom latent vector to generate a molecular graph representation.

[0017] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the generation of drug substructure sequences and overall drug graph representation specifically includes,

[0018] Based on the molecular diagram representation, ring structures, functional groups, and molecular skeleton fragments are identified, a set of substructure fragments is generated, and atomic masking fragments, chemical bond masking fragments, and functional group masking fragments are determined based on the set of substructure fragments to generate a drug masking molecular diagram.

[0019] The node-edge joint graph neural network uses the unmasked atomic environment, chemical bond connections, and molecular descriptor vectors in the drug-masked molecular graph to recover the atom type, bond type, and functional group category corresponding to the masked segment, and generates a reconstructed molecular graph representation.

[0020] The molecular graph representation is reconstructed by segmenting the drug substructure sequence according to the positional order of the ring structure, functional groups, and molecular backbone fragments. The fragment vectors in the drug substructure sequence are then compressed according to the molecular backbone connection order and combined with the molecular descriptor vector to generate the overall drug graph representation.

[0021] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, wherein: the arrangement of the multi-omics representation of the pathogenic gene into a gene sequence specifically includes,

[0022] Gene sequences are generated by determining the linking positions of pathogenic genes in protein-protein interaction networks and their positions in signaling pathways.

[0023] The expression level, mutation status, methylation level, and copy number variation status of each pathogenic gene are mapped to the same gene position in a fixed gene sequence, and positional encoding is performed to generate multi-omics gene sequences.

[0024] Transformer coding networks capture long-distance gene dependencies in multi-omics gene sequences and generate gene sequence representations.

[0025] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the generation of gene pathway activity sequences and overall pathway representation specifically includes,

[0026] Gene sequences are divided into pathway gene segments according to the membership of the signaling pathway, and some gene positions within the same signaling pathway are masked to generate pathway masked gene sequences.

[0027] The transformer coding network uses unmasked genes, co-pathway genes, and protein-protein interaction neighboring genes in the pathway-masked gene sequence to restore the pathway state of the masked gene and generate a restored gene sequence representation.

[0028] Extract member gene representations of the same signaling pathway from the recovered gene sequence representation, and incorporate the position encoding of the member gene into the corresponding member gene representation to generate a set of pathway member representations;

[0029] Intra-path attention is calculated on the pathway member representation set. The pathway member representation set is compressed according to the contribution response of each member gene to its respective signaling pathway to generate a pathway activity vector. The pathway activity vectors are then arranged into a gene pathway activity sequence according to the signaling pathway connection relationship.

[0030] The attention response between pathways is calculated using the pathway activity vectors in the gene pathway activity sequence. Based on the attention response between pathways, the activity vectors of each pathway are compressed to generate an overall representation of the pathway.

[0031] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the generation of the drug pathway interaction representation specifically includes:

[0032] The drug target gene is mapped to the pathway location in the gene pathway activity sequence to generate the drug target pathway embedding.

[0033] The drug substructure sequence carries the drug target pathway embedded in the query gene pathway activity sequence to generate a pathway response representation corresponding to the drug substructure.

[0034] Gene pathway activity sequences carrying drug target pathways are embedded into reverse lookup drug substructure sequences to generate drug action representations corresponding to pathway abnormalities.

[0035] Cross-consistency verification is performed on pathway response representations and drug action representations to generate substructure pathway matching representations;

[0036] A target-pathway comparison sample set is generated by constructing positive interaction samples between the pathways where the actual drug targets are located and the corresponding drug substructures, and negative interaction samples between non-target pathways and disordered drug substructures.

[0037] Target pathway contrastive learning clusters the sub-structural pathway matching representations corresponding to positive interaction samples and separates the sub-structural pathway matching representations corresponding to negative interaction samples, generating drug pathway interaction representations with target pathway consistency.

[0038] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the generation of the half-maximal inhibitory concentration (WMC) prediction value specifically includes:

[0039] Sensitivity regression network converts drug pathway interaction representation, drug global graph representation, and pathway global representation into target-pathway interaction components, drug structure components, and pathway state components, respectively.

[0040] The residual fusion layer compresses the target pathway interaction component, drug structure component, and pathway state component to generate a sensitivity prediction vector. The sensitivity prediction vector is then mapped by a regression layer to the half-maximum inhibitory concentration (IC50) prediction value of the tumor sample corresponding to the candidate drug.

[0041] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, wherein: the sensitivity ranking consistency learning includes,

[0042] For the same tumor sample, candidate drugs are compared in pairs based on known half-maximal inhibitory concentration (IC50) values. The drug with the lower IC50 value is marked as the pre-sensitive drug, and a drug sensitivity order sample is generated.

[0043] For tumor samples under the same candidate drug, the known half-maximal inhibitory concentration values ​​are compared in pairs. The side with the lower value is marked as the preceding response sample, and the sample response sequence sample is generated.

[0044] The sensitivity regression network uses drug sensitivity order samples and sample response order samples to correct the drug prediction ranking and sample prediction ranking, respectively;

[0045] The half-inhibitory concentration predictions are rearranged according to the corrected predictions, and the candidate drugs and tumor samples are labeled with their ranking positions to generate sensitivity ranking results.

[0046] As a preferred embodiment of the deep learning-based tumor drug sensitivity prediction method of the present invention, the generation of the tumor drug sensitivity prediction result includes:

[0047] Extract the cross-attention response between drug substructures and gene pathways from the drug pathway interaction representation to generate a substructure pathway contribution map;

[0048] The contributing drug substructures, target genes, and signaling pathways are determined by the substructure pathway contribution map, and consistency is judged with the half-maximal inhibitory concentration prediction value and sensitivity ranking results to generate sensitivity judgment information;

[0049] The half-maximum inhibitory concentration (MCI) prediction, ranking, and contribution localization results in the sensitivity determination information are mapped into structured prediction entries. The sensitivity status of candidate drugs is then determined by the structured prediction entries, generating tumor drug sensitivity prediction results.

[0050] Secondly, this invention provides a deep learning-based tumor drug sensitivity prediction system, comprising:

[0051] The drug graph encoding module uses a node-edge joint graph neural network to construct a molecular graph representation of candidate drugs between atomic nodes, chemical bond edges, and molecular descriptors. It also recovers the masked atom types, bond types, and functional group categories through drug substructure mask reconstruction, generating drug substructure sequences and overall drug graph representations.

[0052] The pathway coding module encodes the multi-omics representation of pathogenic genes into gene sequences, and uses a transformer coding network combined with pathway masking to recover the abnormal gene transmission relationships within the pre-trained learning signal pathway, generating gene pathway activity sequences and overall pathway representations.

[0053] The interactive learning module performs bidirectional cross-attention matching on drug substructure sequences and gene pathway activity sequences, and combines drug target pathway embedding and target pathway contrast learning to generate interactive representations of drug pathways.

[0054] The sensitivity prediction module feeds the drug pathway interaction representation, the overall drug graph representation, and the overall pathway representation into the sensitivity regression network, and corrects the sensitivity order between drugs and the response order between samples through sensitivity ranking consistency learning, generating the half-maximum inhibitory concentration prediction value and sensitivity ranking results.

[0055] The results generation module generates tumor drug sensitivity prediction results based on the half-maximal inhibitory concentration prediction, sensitivity ranking results, and drug pathway interaction representation to locate contributing drug substructures, target genes, and signaling pathways.

[0056] The beneficial effects of this invention are as follows: Atomic nodes, chemical bond edges, and molecular descriptors of candidate drugs are expressed through a node-edge joint graph neural network. Drug substructure masks are used to reconstruct the local structural relationships in the learning loop structure, functional groups, and molecular backbone fragments, ensuring that drug-side features are no longer limited to the overall molecular vector level. Simultaneously, the multi-omics representation of pathogenic genes is arranged into fixed gene sequences, and pathway masks are used to recover the abnormal gene transmission relationships within the pre-trained signaling pathways, enabling sample-side features to reflect pathway activity status. Bidirectional cross-attention matching of drug substructure sequences and gene pathway activity sequences, combined with drug target pathway embedding and target pathway contrastive learning, forms a drug pathway interaction representation with target pathway consistency, establishing a predictable correspondence between drug structure, target genes, and pathway abnormalities. Sensitivity ranking consistency learning corrects the sensitivity order between drugs and the response order between samples, improving the consistency of half-maximal inhibitory concentration (ICD) prediction results in multi-drug, multi-sample scenarios. Substructure pathway contribution maps locate contributing drug substructures, target genes, and signaling pathways, enhancing the interpretability and clinical reference value of the prediction results. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart of a deep learning-based method for predicting tumor drug sensitivity.

[0059] Figure 2 A flowchart for encoding drug substructure diagrams.

[0060] Figure 3 This is a flowchart encoding gene pathway activity.

[0061] Figure 4 A flowchart for sensitivity prediction and result generation. Detailed Implementation

[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0063] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0064] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0065] Reference Figures 1-4 This is one embodiment of the present invention, which provides a deep learning-based method for predicting tumor drug sensitivity, comprising the following steps:

[0066] S1. A node-edge joint graph neural network is used to construct a molecular graph representation of the candidate drug between atomic nodes, chemical bond edges and molecular descriptors. The masked atom types, bond types and functional group categories are recovered by drug substructure masking reconstruction to generate drug substructure sequences and overall drug graph representation.

[0067] S1.1. The drug sensitivity prediction task consists of the tumor sample to be analyzed, the drug identifier to be evaluated, and the target of half-maximal inhibitory concentration prediction. The drug identifier to be evaluated includes at least one of the following: drug name, CAS number, and PubChem CID. Candidate drugs are determined based on the drug identifier to be evaluated in the tumor drug sensitivity prediction task, and the corresponding SMILES molecular structure string is determined based on the candidate drug. Salt fragment removal, maximum organic fragment retention, atomic valence state verification, aromaticity annotation, stereoconfiguration retention, and normalized SMILES conversion are performed on the SMILES molecular structure string to obtain the normalized molecular structure of the candidate drug. The normalized molecular structure is a molecular diagram structure that can uniquely represent the atomic composition, chemical bond connection relationship, aromaticity state, and stereoconfiguration state of the candidate drug.

[0068] The normalized molecular structure of candidate drugs is analyzed, and each atom is assigned an atom index according to the order of atomic appearance in the normalized molecular structure. Bonding atom index pairs are recorded according to the chemical bond connections in the normalized molecular structure, resulting in an atom index sequence and a chemical bond connection sequence. The bonding state between any two atoms is marked according to the chemical bond connection sequence to form a drug adjacency matrix. At the same time, atom type, aromaticity, formal charge, hybridization mode, hydrogen bond donor state, hydrogen bond acceptor state, and chiral state are extracted from the atom index sequence to form a node feature matrix. Bond type, conjugation state, intra-loop state, and stereoconfiguration are extracted from the chemical bond connection sequence to form an edge feature matrix. Molecular weight, polar surface area, lipid-water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, and number of rotatable bonds are numerically encoded to form a molecular descriptor vector.

[0069] S1.2. The node-edge joint graph neural network determines the range of neighboring atoms for each atom based on the drug adjacency matrix. It linearly maps the node feature matrix to obtain the initial atom latent vector and the edge feature matrix to obtain the chemical bond latent vector. For each atom, the node-edge joint graph neural network concatenates the adjacent initial atom latent vectors with the corresponding chemical bond latent vectors to form an adjacent message vector. It then uses the chemical bond latent vector to perform gating modulation on the adjacent message vectors. After compressing all adjacent message vectors of the same atom, it performs residual connection with the corresponding initial atom latent vector to obtain the updated atom latent vector. The node-edge joint graph neural network copies the molecular descriptor vector to the position corresponding to each atom index and concatenates it with the corresponding updated atom latent vector to generate a molecular graph representation.

[0070] Based on molecular diagram representation, ring structures, functional groups, and molecular skeletal fragments in candidate drugs are identified. Ring structures are determined according to the closed-loop relationship of chemical bonds, i.e., a closed set of atoms formed by the first-to-last atomic paths in the normalized molecular structure. Functional groups are determined according to the SMARTS substructure matching rule, which uses SMARTS patterns to describe atom type, bond type, aromaticity, charge state, and adjacency relationship, and searches for sets of atoms and chemical bonds that match the corresponding patterns in the molecular diagram structure of the candidate drug. Molecular skeletal fragments are determined according to the Bemis-Murcko skeleton rule, which retains the ring structures and connecting chains of the ring structures in the candidate drug, and removes side chain atoms that do not belong to the ring structures and inter-ring connecting chains, thus obtaining the set of atoms and chemical bonds representing the core skeleton of the candidate drug.

[0071] When a candidate drug has ring structures, functional groups, and molecular skeleton fragments, the ring structures, functional groups, and molecular skeleton fragments are collectively considered as a set of substructure fragments. When a candidate drug does not have ring structures or the Bemis-Murcko skeleton rule results in an empty skeleton, the heavy atom connecting chains in the normalized molecular structure are considered as molecular skeleton fragments. When a candidate drug does not have matching functional groups, the heavy atoms not classified into ring structures and molecular skeleton fragments are considered as single-atom fragments, generating a set of substructure fragments.

[0072] S1.3. Based on the set of substructure fragments, determine the atomic masking fragments, chemical bond masking fragments, and functional group masking fragments. Replace the atomic properties of the masked atoms with mask markers, replace the chemical bond properties of the masked chemical bonds with mask markers, and replace the atomic and chemical bond properties within the masked functional groups with mask markers. At the same time, preserve the connection relationships in the drug adjacency matrix to generate a drug masking molecular graph.

[0073] The node-edge joint graph neural network receives the drug-masked molecular graph, the unmasked atomic environment, chemical bond connections, and molecular descriptor vectors. It recalculates the updated atomic latent vectors at the masked positions according to the message passing method in S1.2. The node-edge joint graph neural network predicts the atom type of the masked atoms, the bond type of the masked chemical bonds, and the functional group category of the masked functional groups through a classification layer. The prediction results are then cross-entropy trained with the corresponding real atom types, real bond types, and real functional group categories in the normalized molecular structure to complete the pre-training of drug substructure mask reconstruction and generate a reconstructed molecular graph representation.

[0074] S1.4. The molecular skeleton connection order is determined by the core skeleton atom set and chemical bond set obtained according to the Bemis-Murcko skeleton rule. The molecular skeleton connection order is the order of segments formed by the skeleton atom with the smallest atomic index in the core skeleton as the starting position and passing through adjacent skeleton atoms along the chemical bond connection relationship of the core skeleton. When there are branches in the core skeleton, the branch segments are arranged in ascending order according to the starting atomic index of the branch. When the core skeleton contains a ring structure, the ring segments are formed in the order of the smallest atomic index in the ring structure as the ring starting point and along the chemical bond connection direction.

[0075] The molecular graph representation is reconstructed by segmenting the molecular structure according to the positional order of ring structures, functional groups, and molecular skeleton fragments in the atomic index sequence. The updated atomic latent vectors within each substructure fragment are extracted, and the updated atomic latent vectors within the same substructure fragment are compressed at the fragment level to obtain fragment vectors. The fragment vectors are arranged according to the molecular skeleton connection order to generate the drug substructure sequence. The fragment vectors in the drug substructure sequence are compressed at the sequence level according to the molecular skeleton connection order and combined with the molecular descriptor vector to generate the overall drug graph representation.

[0076] S2. The multi-omics representation of the pathogenic gene is arranged into a gene sequence, and the abnormal gene transmission relationship within the pre-trained learning signal pathway is recovered by using a transformer coding network combined with a pathway mask, generating gene pathway activity sequences and overall pathway representation.

[0077] S2.1. A tumor drug sensitivity prediction task is used to identify the tumor samples to be analyzed. The tumor drug sensitivity prediction task is used in S1 to identify candidate drugs and in S2 to identify the tumor samples to be analyzed. Both belong to the same task object: drug-side data and sample-side data. Based on the tumor drug sensitivity prediction task, the tumor samples to be analyzed and the set of pathogenic genes are determined. Each pathogenic gene in the set of pathogenic genes is matched with the protein-protein interaction network and the member relationship of the signaling pathway to obtain the protein-protein interaction connection position and the member position of the signaling pathway of the pathogenic gene. The pathogenic genes are grouped according to the member position of the signaling pathway, and within the same signaling pathway, the pathogenic genes are arranged according to the connection distance in the protein-protein interaction network. Pathogenic genes with the same connection distance are arranged according to the gene arrangement order in the member relationship of the signaling pathway to generate a fixed gene sequence.

[0078] For each pathogenic gene in a fixed gene sequence, the expression level, mutation status, methylation level, and copy number variation status of the corresponding pathogenic gene in the tumor sample to be analyzed are extracted. The expression level is represented by a normalized expression value, the mutation status is represented by the presence of mutations, the methylation level is represented by the methylation site level, and the copy number variation status is represented by the copy number increase or decrease status. The expression level, mutation status, methylation level, and copy number variation status are mapped to the same gene position in the fixed gene sequence to generate a multi-omics gene position representation.

[0079] S2.2. Add positional encoding to each gene position in the fixed gene sequence so that the multi-omics positional representation of the gene simultaneously includes the sequence information of the pathogenic gene in the fixed gene sequence, the protein-protein interaction connection position information, and the signaling pathway member position information; arrange the positionally encoded multi-omics positional representation of the gene according to the fixed gene sequence to generate a multi-omics gene sequence.

[0080] The Transformer Encoding Network receives multi-omics gene sequences and uses Transformer scaled dot product attention to calculate the dependencies between different pathogenic genes in a fixed gene sequence. The expression is as follows:

[0081] ;

[0082] in, This is a query matrix obtained by query mapping from multiple omics gene sequences. The key matrix is ​​obtained by key mapping from multiple omics gene sequences. This is a value matrix obtained by value mapping from multiple omics gene sequences. This represents the transpose of the key matrix. The dimension of the key vector. This is used to normalize inner product similarity into a dependency coefficient between different pathogenic genes. This represents the gene dependency representation obtained by weighting based on the dependency coefficient.

[0083] Multi-head self-attention concatenates the results of multiple scaled dot product attention and obtains a gene-dependent representation of a fixed gene sequence through linear mapping. A transformer coding network converts the multi-omics information of each pathogenic gene into a gene coding vector using a feedforward neural network. The transformer coding network outputs the gene coding vectors corresponding to each pathogenic gene according to the fixed gene sequence, generating a gene sequence representation. The gene sequence representation is divided into pathway gene segments according to signaling pathway membership. Each pathway gene segment contains the representation of member genes within the same signaling pathway. Within each pathway gene segment, the multi-omics position representations of some member genes are masked. The masked content includes expression level, mutation state, methylation level, and copy number variation state, generating a pathway-masked gene sequence. The masked member genes still retain their positional codes and signaling pathway member positions in the fixed gene sequence, ensuring that the transformer coding network can identify the position of the masked member genes within the signaling pathway.

[0084] S2.3. The transformer coding network receives the pathway-masked gene sequence and uses the gene coding vectors of unmasked genes, genes in the same pathway, and protein-interacting neighboring genes to recover the pathway state of the masked gene. The pathway state of the masked gene includes the expression level, mutation state, methylation state, copy number variation state, and signaling pathway member position of the masked gene. The transformer coding network compares the recovery result with the real pathway state before masking to complete the pathway masking recovery pre-training and generate the recovered gene sequence representation.

[0085] Extract member gene representations corresponding to the same signaling pathway from the recovered gene sequence representation, and encode the position of the member gene in the fixed gene sequence and incorporate it into the corresponding member gene representation to generate a pathway member representation set. The pathway member representation set retains the multi-omics abnormal states and gene positional relationships of each member gene in the same signaling pathway, providing gene-level data for attention computation within the pathway.

[0086] S2.4. Calculate intra-path attention for the set of path member representations, expressed as:

[0087] ;

[0088] ;

[0089] in, Indicates the signal path index. Indicates member gene index, Indicates the first The number of member genes within each signaling pathway The path member represents the first element in the set. The first signal path Individual member genes represent, Indicates the first The mean vector of all member genes within a signaling pathway. , , These represent the query mapping matrix, key mapping matrix, and value mapping matrix, respectively. This represents the transpose of the query vector. Represents the dimension of the key vector. Indicates the first The member gene pair The contribution response of each signal path Indicates the first The pathway activity vector of each signaling pathway represents the abnormal activity state of the same signaling pathway in multiple omics in the tumor sample to be analyzed.

[0090] The activity vectors of each pathway are arranged according to the connection relationships between different signaling pathways to generate gene pathway activity sequences. The connection relationships between signaling pathways are determined by cross-pathway gene connections in the shared member gene and protein interaction network. The activity vectors of each pathway in the gene pathway activity sequence are further subjected to inter-pathway attention calculation to obtain the inter-pathway attention response of each signaling pathway to the overall pathway state of the tumor sample to be analyzed. The activity vectors of each pathway are synthesized according to the inter-pathway attention response to generate the overall pathway representation.

[0091] S3. Perform bidirectional cross-attention matching on drug substructure sequences and gene pathway activity sequences, and combine drug target pathway embedding and target pathway contrast learning to generate drug pathway interaction representations.

[0092] S3.1. Based on the drug identifiers to be evaluated in the tumor drug sensitivity prediction task, determine the drug target genes of candidate drugs, match the drug target genes with the signaling pathway members in the gene pathway activity sequence, and determine the signaling pathway position where the drug target genes are located; mark the signaling pathway positions of the drug target genes that are matched as target pathway positions, and mark the signaling pathway positions of the drug target genes that are not matched as non-target pathway positions, and encode the target pathway positions and non-target pathway positions into the corresponding pathway activity vectors to generate drug target pathway embeddings; when the candidate drug does not have a confirmed drug target gene, mark all signaling pathway positions in the gene pathway activity sequence as non-target pathway positions to generate drug target pathway embeddings.

[0093] The drug substructure sequences are arranged into a drug substructure sequence matrix. Gene pathway activity sequences and drug target pathway embeddings are spliced ​​at the same position to generate a target embedding pathway sequence matrix. The drug substructure sequence matrix represents each substructure fragment of the candidate drug, and the target embedding pathway sequence matrix represents the activity status of each signaling pathway and the location of the drug target pathway in the tumor sample to be analyzed. Both are used as matching objects for the bidirectional cross-attention network.

[0094] S3.2. A bidirectional cross-attention network enables the drug substructure sequence matrix to query the target embedding pathway sequence matrix, obtaining the pathway response representation corresponding to the drug substructure. The calculation follows the Transformer scaled dot product attention model, limiting the query object to the drug substructure sequence matrix and the queried object to the target embedding pathway sequence matrix. The expression is as follows:

[0095] ;

[0096] in, Represents the drug substructure sequence matrix. This represents the target embedding pathway sequence matrix. This represents the transpose of the target embedding pathway sequence matrix. This represents the common vector dimension after mapping the drug substructure sequence matrix and the target embedding pathway sequence matrix. Represents the normalization function. This represents the pathway response corresponding to the drug substructure.

[0097] A bidirectional cross-attention network enables the target embedding pathway sequence matrix to perform a reverse lookup of the drug substructure sequence matrix, obtaining the drug action representation corresponding to pathway abnormalities. The computation also uses Transformer scaled dot product attention, limiting the query object to the target embedding pathway sequence matrix and the query object to the drug substructure sequence matrix. The expression is as follows:

[0098] ;

[0099] in, This represents the transpose of the drug substructure sequence matrix. This indicates the drug effect corresponding to pathway abnormalities.

[0100] S3.3. Map the pathway response representation corresponding to the drug substructure and the drug action representation corresponding to the pathway abnormality to the same vector dimension, and perform cross-consistency verification according to the drug substructure sequence position and the gene pathway activity sequence position; the cross-consistency verification compares whether the signal pathway position pointed to by the drug substructure corresponds to the drug substructure position pointed to in the opposite direction by the signal pathway, and retains the substructure pathway pairing consistent with the drug target pathway embedding to generate a substructure pathway matching representation.

[0101] Using the target pathway location of the drug target pathway embedding marker as the actual drug target pathway, positive interaction samples are formed by combining the actual drug target pathway with the corresponding drug substructure that passes the cross-consistency check; negative interaction samples are formed by combining non-target pathways with drug substructures that fail the cross-consistency check; and negative interaction samples are formed by combining the shuffled drug substructures with the actual drug target pathway. A target pathway comparison sample set is then generated from the positive interaction samples, the first negative interaction samples, and the second negative interaction samples.

[0102] S3.4. Target-pathway contrastive learning uses vector similarity of substructure path matching representations to distinguish between target-consistent and target-inconsistent relationships. Positive interaction samples are labeled as target-consistent categories, while the first and second types of negative interaction samples are labeled as target-inconsistent categories. The bidirectional cross-attention network increases the category probability corresponding to positive interaction samples and decreases the category probability corresponding to the first and second types of negative interaction samples through cross-entropy training, causing the substructure path matching representations corresponding to positive interaction samples to cluster and the substructure path matching representations corresponding to negative interaction samples to separate.

[0103] The substructure pathway matching representations that have completed target pathway comparison learning are compressed according to the positional order of drug substructure sequences and gene pathway activity sequences, and the substructure pathway matching representations corresponding to positive interaction samples are retained to generate drug pathway interaction representations with target pathway consistency.

[0104] S4. The drug pathway interaction representation, drug overall graph representation, and pathway overall representation are fed into the sensitivity regression network, and the sensitivity order between drugs and the response order between samples are corrected through sensitivity ranking consistency learning to generate the half-maximum inhibitory concentration prediction value and sensitivity ranking results.

[0105] S4.1. The sensitivity regression network receives drug pathway interaction representation, drug overall graph representation, and pathway overall representation, and converts the drug pathway interaction representation into target-pathway interaction components, the drug overall graph representation into drug structure components, and the pathway overall representation into pathway state components, respectively. The target-pathway interaction components are used to represent the target consistency relationship between the candidate drug substructure and the signaling pathway, the drug structure components are used to represent the overall molecular structure state of the candidate drug, and the pathway state components are used to represent the overall pathway abnormality state of the tumor sample to be analyzed.

[0106] The residual fusion layer receives target-pathway interaction components, drug structure components, and pathway state components. It uses the target-pathway interaction components as the primary prediction information and the drug structure and pathway state components as supplementary prediction information. Within the residual fusion layer, it performs same-dimensional mapping, channel splicing, and residual compression to generate a sensitivity prediction vector. The sensitivity prediction vector is then mapped by a regression layer to the predicted half-maximum inhibitory concentration (WMC) value of the candidate drug for the tumor sample to be analyzed. The predicted WMC value is consistent with the WMC label commonly used in drug sensitivity testing; the lower the value, the higher the inhibitory sensitivity of the candidate drug to the tumor sample to be analyzed.

[0107] S4.2. The predicted half-inhibition concentration (WIC) is trained using supervised regression. The supervised regression uses mean squared error (MSE) to measure the numerical difference between the predicted WIC and the known WIC. The expression is:

[0108] ;

[0109] in, This represents the regression error of the half-maximal inhibition concentration. This indicates the number of candidate drug and tumor sample combinations used in the training. Indicates the first A combination of candidate drugs and tumor samples, Indicates the first Half-maximal inhibitory concentration (MCC) predictions for combinations of candidate drugs with tumor samples. Indicates the first The known half-maximal inhibitory concentration (MCC) values ​​for each candidate drug combined with a tumor sample.

[0110] For the same tumor sample, candidate drugs are compared in pairs based on known half-maximal inhibitory concentrations (WMCs). The drug with the lower WMC is designated as the preceding sensitive drug, generating a drug sensitivity order sample. For the same candidate drug, tumor samples are compared in pairs based on known WMCs. The drug with the lower WMC is designated as the preceding response sample, generating a sample response order sample. The drug sensitivity order sample reflects the sensitivity difference of the same tumor sample to different candidate drugs, while the sample response order sample reflects the response difference of the same candidate drug to different tumor samples.

[0111] S4.3. Sensitivity Ranking Consistency Learning utilizes drug sensitivity order samples and sample response order samples to correct the predicted ranking of the sensitivity regression network output. For two candidate drugs under the same tumor sample, if the known half-maximum inhibitory concentration (WMC) values ​​satisfy the condition that the preceding sensitive drug is lower than the following drug, then the sensitivity regression network constrains the predicted WMC value of the preceding sensitive drug to be lower than the predicted WMC value of the following drug. For two tumor samples under the same candidate drug, if the known WMC values ​​satisfy the condition that the preceding response sample is lower than the following sample, then the sensitivity regression network constrains the predicted WMC value of the preceding response sample to be lower than the predicted WMC value of the following sample.

[0112] Sensitivity ranking consistency learning uses pairwise ranking loss to correct the predicted permutation, constraining the predicted half-inhibition concentration of preceding objects to be lower than that of subsequent objects. The expression is:

[0113] ;

[0114] in, This indicates the sensitivity ranking error. This represents the number of pairwise comparisons in the drug sensitivity order samples and the sample response order samples. Indicates the first A pairwise comparison, Indicates the first The half-maximum inhibition concentration prediction of the preceding object in a pairwise comparison Indicates the first In a pairwise comparison, the half-maximal inhibition concentration prediction value of the subsequent object is... Used to generate sorting error when the predicted value of the preceding object is not lower than the predicted value of the following object.

[0115] S4.4. The sensitivity regression network updates the network parameters together based on the half-inhibition concentration regression error and the sensitivity ranking error. At the same time, it retains the representation relationship obtained from drug substructure mask reconstruction pre-training, pathway mask recovery pre-training, and target pathway comparison learning, so that the half-inhibition concentration prediction value conforms to the numerical supervision results, the sensitivity order among drugs, and the response order among samples.

[0116] Candidate drugs for the same tumor sample are ranked according to the corrected prediction ranking, and tumor samples for the same candidate drug are ranked. The ranked half-maximum inhibitory concentration (MCIC) prediction values ​​are arranged from low to high. Candidate drugs with low values ​​are marked as more sensitive, and tumor samples with low values ​​are marked as more responsive, thus generating sensitivity ranking results.

[0117] S5. Based on the predicted half-maximal inhibitory concentration (IC50), sensitivity ranking results, and drug pathway interaction representations, the contributing drug substructures, target genes, and signaling pathways are located to generate tumor drug sensitivity prediction results.

[0118] S5.1. Extract the cross-attention response between the drug substructure sequence position and the gene pathway active sequence position from the drug pathway interaction representation, and arrange the cross-attention responses according to the drug substructure sequence position, the gene pathway active sequence position, and the target pathway position in the drug target pathway embedding to generate a substructure pathway contribution map; the substructure pathway contribution map is used to represent the response relationship between each drug substructure fragment and each signaling pathway, and the position coordinates in the substructure pathway contribution map correspond to the fragment position in the drug substructure sequence and the pathway position in the gene pathway active sequence.

[0119] The contributing drug substructures and signaling pathways are determined based on the order of cross-attention responses in the substructure-pathway contribution map, and the target genes are determined based on the location of the drug target pathway embedded in the signaling pathway. The contributing drug substructures, target genes, and signaling pathways are arranged in descending order of cross-attention responses to generate contribution localization results. The contribution localization results are used to indicate the source of the drug substructure, the source of the target gene, and the source of the signaling pathway corresponding to the half-maximum inhibitory concentration (WMC) prediction.

[0120] S5.2. The consistency between the contribution localization results and the half-maximal inhibitory concentration (WMC) prediction values ​​and the sensitivity ranking results is determined. The WMC prediction values ​​are arranged from low to high and compared with the ranking position in the sensitivity ranking results. The signaling pathways of the target genes in the contribution localization results are compared with the signaling pathway responses in the substructure pathway contribution map. When the WMC prediction values ​​are consistent with the ranking position in the sensitivity ranking results, and the signaling pathways of the target genes in the contribution localization results have corresponding cross-attention responses, sensitivity determination information containing numerical consistency status and pathway consistency status is generated.

[0121] The half-maximal inhibitory concentration (WMC) prediction, ranking, contributing drug substructure, target gene, and signaling pathway in the sensitivity determination information are mapped into structured prediction entries. The structured prediction entries include the tumor sample to be analyzed, candidate drugs, WMC prediction, sensitivity ranking results, contributing drug substructure, target gene, signaling pathway, numerical consistency status, and pathway consistency status.

[0122] A drug sensitivity ranking is established based on the predicted half-maximal inhibitory concentration (MCC) of all candidate drugs for the same tumor sample under analysis, from low to high. The ranking in the structured prediction entries is then compared with the ranking of candidate drugs in the drug sensitivity ranking. When both numerical and pathway consistency are met, and the candidate drug is ranked first in the drug sensitivity ranking, its sensitivity status is determined as the preferred sensitivity status. When both numerical and pathway consistency are met, and the candidate drug is in the first half of the drug sensitivity ranking but not at the top, its sensitivity status is determined as sensitive. When both numerical and pathway consistency are met, and the candidate drug is in the second half of the drug sensitivity ranking, its sensitivity status is determined as low sensitivity. When either numerical or pathway consistency is not met, the candidate drug's sensitivity status is determined as pending verification. The sensitivity status of the candidate drugs is then incorporated into the structured prediction entries to generate the tumor drug sensitivity prediction results.

[0123] This embodiment also provides a deep learning-based tumor drug sensitivity prediction system, including: a drug graph encoding module, which uses a node-edge joint graph neural network to construct a molecular graph representation between atomic nodes, chemical bond edges and molecular descriptors of candidate drugs, and recovers the masked atom types, bond types and functional group categories through drug substructure mask reconstruction to generate drug substructure sequences and overall drug graph representation;

[0124] The pathway coding module encodes the multi-omics representation of pathogenic genes into gene sequences, and uses a transformer coding network combined with pathway masking to recover the abnormal gene transmission relationships within the pre-trained learning signal pathway, generating gene pathway activity sequences and overall pathway representations.

[0125] The interactive learning module performs bidirectional cross-attention matching on drug substructure sequences and gene pathway activity sequences, and combines drug target pathway embedding and target pathway contrast learning to generate interactive representations of drug pathways.

[0126] The sensitivity prediction module feeds the drug pathway interaction representation, the overall drug graph representation, and the overall pathway representation into the sensitivity regression network, and corrects the sensitivity order between drugs and the response order between samples through sensitivity ranking consistency learning, generating the half-maximum inhibitory concentration prediction value and sensitivity ranking results.

[0127] The results generation module generates tumor drug sensitivity prediction results based on the half-maximal inhibitory concentration prediction, sensitivity ranking results, and drug pathway interaction representation to locate contributing drug substructures, target genes, and signaling pathways.

[0128] In summary, this invention expresses the atomic nodes, chemical bond edges, and molecular descriptors of candidate drugs through a node-edge joint graph neural network. It also utilizes drug substructure masks to reconstruct the local structural relationships within the learning loop structure, functional groups, and molecular backbone fragments, ensuring that drug-side features are no longer confined to the overall molecular vector level. Simultaneously, it arranges the multi-omics representation of pathogenic genes into fixed gene sequences and recovers the abnormal gene transmission relationships within the pre-trained signaling pathways using pathway masks, enabling sample-side features to reflect pathway activity status. By matching drug substructure sequences and gene pathway activity sequences through bidirectional cross-attention, and combining drug target pathway embedding and target pathway contrastive learning, it forms a drug pathway interaction representation with target pathway consistency, establishing a predictable correspondence between drug structure, target genes, and pathway abnormalities. Through sensitivity ranking consistency learning, it corrects the sensitivity order between drugs and the response order between samples, improving the consistency of the half-maximal inhibitory concentration (ICD) prediction results in multi-drug, multi-sample scenarios. Finally, by locating contributing drug substructures, target genes, and signaling pathways through substructure pathway contribution maps, it enhances the interpretability and clinical reference value of the prediction results.

[0129] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for predicting tumor drug sensitivity, characterized in that: include, A node-edge joint graph neural network is used to construct a molecular graph representation of candidate drugs between atomic nodes, chemical bond edges and molecular descriptors. The masked atom types, bond types and functional group categories are recovered by drug substructure mask reconstruction to generate drug substructure sequences and overall drug graph representations. The multi-omics representation of pathogenic genes is arranged into gene sequences, and the abnormal gene transmission relationship within the pre-trained learning signal pathway is recovered by using a transformer coding network combined with pathway masking, generating gene pathway activity sequences and overall pathway representations; Bidirectional cross-attention matching is performed on drug substructure sequences and gene pathway activity sequences, and drug target pathway embedding and target pathway contrast learning are combined to generate drug pathway interaction representations. The drug pathway interaction representation, drug overall graph representation, and pathway overall representation are fed into the sensitivity regression network, and the sensitivity order between drugs and the response order between samples are corrected through sensitivity ranking consistency learning to generate the half-maximum inhibitory concentration prediction value and sensitivity ranking results. Based on the predicted half-maximal inhibitory concentration (MCI), sensitivity ranking results, and drug pathway interaction representations, the contributing drug substructures, target genes, and signaling pathways are identified, and tumor drug sensitivity prediction results are generated.

2. The tumor drug sensitivity prediction method based on deep learning as described in claim 1, characterized in that: The construction of the molecular diagram representation specifically includes, The normalized molecular structure of candidate drugs is analyzed to form atomic index sequences, chemical bond connection sequences and drug adjacency matrices. Atomic properties, chemical bond properties and molecular physicochemical properties are extracted respectively to generate node feature matrices, edge feature matrices and molecular descriptor vectors. The node-edge joint graph neural network defines the range of adjacent atoms based on the drug adjacency matrix, maps the node feature matrix to the initial atom latent vector, maps the edge feature matrix to the chemical bond latent vector, and uses the chemical bond latent vector to modulate the message passing between adjacent initial atom latent vectors to obtain the updated atom latent vector. The molecular descriptor vector is extended to the atomic positions corresponding to each updated atom latent vector, and then concatenated with the updated atom latent vector to generate a molecular graph representation.

3. The tumor drug sensitivity prediction method based on deep learning as described in claim 2, characterized in that: The generation of drug substructure sequences and overall drug diagram representations specifically includes, Based on the molecular diagram representation, ring structures, functional groups, and molecular skeleton fragments are identified, a set of substructure fragments is generated, and atomic masking fragments, chemical bond masking fragments, and functional group masking fragments are determined according to the set of substructure fragments to generate a drug masking molecular diagram. The node-edge joint graph neural network uses the unmasked atomic environment, chemical bond connections, and molecular descriptor vectors in the drug-masked molecular graph to recover the atom type, bond type, and functional group category corresponding to the masked segment, and generates a reconstructed molecular graph representation. The molecular graph representation is reconstructed by segmenting the drug substructure sequence according to the positional order of the ring structure, functional groups, and molecular backbone fragments. The fragment vectors in the drug substructure sequence are then compressed according to the molecular backbone connection order and combined with the molecular descriptor vector to generate the overall drug graph representation.

4. The tumor drug sensitivity prediction method based on deep learning as described in claim 3, characterized in that: The arrangement of the multi-omics representation of the pathogenic gene into a gene sequence specifically includes, Gene sequences are generated by determining the linking positions of pathogenic genes in protein-protein interaction networks and their positions in signaling pathways. The expression level, mutation status, methylation level, and copy number variation status of each pathogenic gene are mapped to the same gene position in a fixed gene sequence, and positional encoding is performed to generate multi-omics gene sequences. Transformer coding networks capture long-distance gene dependencies in multi-omics gene sequences and generate gene sequence representations.

5. The deep learning-based tumor drug sensitivity prediction method as described in claim 4, characterized in that: The gene pathway activity sequence and overall pathway representation specifically include, Gene sequences are divided into pathway gene segments according to the membership of the signaling pathway, and some gene positions within the same signaling pathway are masked to generate pathway masked gene sequences. The transformer coding network uses unmasked genes, co-pathway genes, and protein-protein interaction neighboring genes in the pathway-masked gene sequence to restore the pathway state of the masked gene and generate a restored gene sequence representation. Extract member gene representations of the same signaling pathway from the recovered gene sequence representation, and incorporate the position encoding of the member gene into the corresponding member gene representation to generate a set of pathway member representations; Intra-path attention is calculated on the pathway member representation set. The pathway member representation set is compressed according to the contribution response of each member gene to its respective signaling pathway to generate a pathway activity vector. The pathway activity vectors are then arranged into a gene pathway activity sequence according to the signaling pathway connection relationship. The attention response between pathways is calculated using the pathway activity vectors in the gene pathway activity sequence. Based on the attention response between pathways, the pathway activity vectors are compressed to generate an overall representation of the pathway.

6. The tumor drug sensitivity prediction method based on deep learning as described in claim 5, characterized in that: The generation of the drug pathway interaction representation specifically includes, The drug target gene is mapped to the pathway location in the gene pathway activity sequence to generate the drug target pathway embedding. The drug substructure sequence carries the drug target pathway embedded in the query gene pathway activity sequence to generate a pathway response representation corresponding to the drug substructure. Gene pathway activity sequences carrying drug target pathways are embedded into reverse lookup drug substructure sequences to generate drug action representations corresponding to pathway abnormalities. Cross-consistency verification is performed on pathway response representations and drug action representations to generate substructure pathway matching representations; A target-pathway comparison sample set is generated by constructing positive interaction samples between the pathways where the actual drug targets are located and the corresponding drug substructures, and negative interaction samples between non-target pathways and disordered drug substructures. Target pathway contrastive learning clusters the sub-structural pathway matching representations corresponding to positive interaction samples and separates the sub-structural pathway matching representations corresponding to negative interaction samples, generating drug pathway interaction representations with target pathway consistency.

7. The tumor drug sensitivity prediction method based on deep learning as described in claim 6, characterized in that: The generation of the half-maximal inhibition concentration prediction value specifically includes, Sensitivity regression network converts drug pathway interaction representation, drug global graph representation, and pathway global representation into target-pathway interaction components, drug structure components, and pathway state components, respectively. The residual fusion layer compresses the target pathway interaction component, drug structure component, and pathway state component to generate a sensitivity prediction vector. The sensitivity prediction vector is then mapped by a regression layer to the half-maximum inhibitory concentration (IC50) prediction value of the tumor sample corresponding to the candidate drug.

8. The tumor drug sensitivity prediction method based on deep learning as described in claim 7, characterized in that: The sensitivity ranking consistency learning includes, For the same tumor sample, candidate drugs are compared in pairs based on known half-maximal inhibitory concentration (IC50) values. The drug with the lower IC50 value is marked as the pre-sensitive drug, and a drug sensitivity order sample is generated. For tumor samples under the same candidate drug, the known half-maximal inhibitory concentration (IC50) values ​​are compared in pairs. The side with the lower value is marked as the preceding response sample, and the sample response sequence sample is generated. The sensitivity regression network uses drug sensitivity order samples and sample response order samples to correct the drug prediction ranking and sample prediction ranking, respectively; The half-inhibitory concentration predictions are rearranged according to the corrected predictions, and the candidate drugs and tumor samples are labeled with their ranking positions to generate sensitivity ranking results.

9. The tumor drug sensitivity prediction method based on deep learning as described in claim 8, characterized in that: The generation of the tumor drug sensitivity prediction results includes, Extract the cross-attention response between drug substructures and gene pathways from the drug pathway interaction representation to generate a substructure pathway contribution map. The contributing drug substructures, target genes, and signaling pathways are determined by the substructure pathway contribution map, and consistency is judged with the half-maximal inhibitory concentration prediction value and sensitivity ranking results to generate sensitivity judgment information; The half-maximum inhibitory concentration (MCI) prediction, ranking, and contribution localization results in the sensitivity determination information are mapped into structured prediction entries. The sensitivity status of candidate drugs is then determined by the structured prediction entries, generating tumor drug sensitivity prediction results.

10. A deep learning-based tumor drug sensitivity prediction system, based on the deep learning-based tumor drug sensitivity prediction method according to any one of claims 1 to 9, characterized in that: include, The drug graph encoding module uses a node-edge joint graph neural network to construct a molecular graph representation of candidate drugs between atomic nodes, chemical bond edges, and molecular descriptors. It also recovers the masked atom types, bond types, and functional group categories through drug substructure mask reconstruction, generating drug substructure sequences and overall drug graph representations. The pathway coding module encodes the multi-omics representation of pathogenic genes into gene sequences, and uses a transformer coding network combined with pathway masking to recover the abnormal gene transmission relationships within the pre-trained learning signal pathway, generating gene pathway activity sequences and overall pathway representations. The interactive learning module performs bidirectional cross-attention matching on drug substructure sequences and gene pathway activity sequences, and combines drug target pathway embedding and target pathway contrast learning to generate interactive representations of drug pathways. The sensitivity prediction module feeds the drug pathway interaction representation, the overall drug graph representation, and the overall pathway representation into the sensitivity regression network, and corrects the sensitivity order between drugs and the response order between samples through sensitivity ranking consistency learning, generating the half-maximum inhibitory concentration prediction value and sensitivity ranking results. The results generation module generates tumor drug sensitivity prediction results based on the half-maximal inhibitory concentration prediction, sensitivity ranking results, and drug pathway interaction representation to locate contributing drug substructures, target genes, and signaling pathways.

Images