Virtual screening method and device based on quantum manifold dense search

Abstract

This invention discloses a virtual screening method and apparatus based on quantum manifold dense retrieval, relating to computer-aided drug design. The method includes: acquiring target protein binding pocket maps and ligand molecule maps; extracting quantum manifold embedding vectors through a small molecule encoding module and extracting high-dimensional features through a protein pocket encoding module; fusing and interacting through a physically-aware cross-modal, attitude-free alignment module to obtain a protein-ligand cross-modal joint feature representation; based on this joint feature, performing an approximate nearest neighbor search in a pre-constructed quantum manifold embedding vector library through a quantum manifold dense retrieval module to obtain a preliminary list of hit molecules; and then outputting the final virtual screening result through molecular clustering and deduplication, druggability filtering, and affinity prediction. This invention combines free energy regression reconstruction into a unified physical latent space for cross-modal dense vector retrieval, achieving ultra-high throughput and high-precision screening of hundreds of millions of molecular libraries, breaking through existing technological bottlenecks.

Description

Technical Field

[0001] This invention relates to the field of drug-aided design technology, and more specifically to a virtual screening method and apparatus based on quantum manifold dense retrieval. Background Technology

[0002] New drug development is the core track for innovation and development in the global biopharmaceutical industry. However, its inherent characteristics of high investment, long cycle, and high risk remain the core industry bottleneck restricting the efficient output of innovative drugs. Data shows that the average development cycle of an innovative drug from target discovery to successful market launch exceeds 10 years, with a cumulative development cost of over US$2 billion. As the source of new drug development, the efficiency of preclinical lead compound discovery directly determines the pace of the entire new drug development pipeline, development costs, and ultimate success or failure.

[0003] Structure-Based Virtual Screening (SBVS), a core technology in modern computer-aided drug design (CADD), is currently the most mainstream and essential technique for discovering promising compounds for innovative drugs. This technology uses computational methods to simulate the intermolecular interactions between target proteins and small molecule compounds, enabling the rapid enrichment of compounds with potential target-binding activity from massive molecular libraries. This can reduce the development cost of traditional in vitro high-throughput screening by more than 90% and shorten the development cycle by more than 60%.

[0004] In the post-genomic era, a large number of novel and difficult-to-drug targets closely related to the occurrence and development of diseases are being continuously discovered and validated. Virtual screening technology provides key technical support for the development of drugs for these targets from scratch and has become an indispensable core link in the new drug development pipeline of global pharmaceutical companies, biotechnology companies and research institutions.

[0005] With the rapid development of synthetic chemistry and compound library construction, commercial and open-source virtual molecular libraries have reached a scale of hundreds of millions. For example, the ZINC15 library contains over 2 billion synthetic drug-like molecules, and the Enamine library covers over 1 billion physically accessible compounds. Traditional in vitro screening technologies are completely unable to cover such a vast chemical space, making high-throughput virtual screening the most feasible large-scale prospective compound discovery solution. Its core industrial value lies not only in its ability to complete the full evaluation of hundreds of millions of molecular libraries within hours to days, increasing throughput by more than three orders of magnitude compared to in vitro high-throughput screening, thus completely solving the industry pain point that traditional methods cannot cover such a large chemical space; but also in its ability to significantly reduce the financial and time costs of compound synthesis and biological testing from the source. It enables the discovery of active molecules from scratch for novel targets without known active ligands and difficult-to-drug targets. At the same time, it incorporates drug-likeness, ADMET, toxicity, and other drug-likeness constraints during the screening process, managing R&D risks upfront and reducing the probability of failure in subsequent drug chemistry optimization and clinical development.

[0006] The current mainstream virtual screening technologies are mainly divided into two categories: traditional physical docking methods and data-driven deep learning methods. Although the two have their own technical advantages in different application scenarios, neither can simultaneously meet the core industrial needs of innovative drug development for ultra-high throughput, high enrichment accuracy, strong generalization and physical interpretability, and there are underlying technical bottlenecks that are difficult to overcome.

[0007] Traditional physical docking methods, such as AutoDock Vina, Glide, and Surflex, simulate ligand-target binding modes and calculate binding free energies based on molecular force fields and empirical scoring functions. These methods have three inherent limitations: First, the computational cost is prohibitively high. A single docking of a small molecule with a target can take several seconds to tens of seconds, and the screening cycle for a full library of hundreds of millions of molecules can take weeks or even months, making them completely unsuitable for the industrial demands of ultra-high-throughput screening. Second, the generalization ability of empirical scoring functions is insufficient. For complex scenarios such as different target types, metalloprotein systems, and flexible binding pockets, the prediction accuracy fluctuates wildly, and there is a generally high false positive rate. Third, the early enrichment ability is weak, and it is impossible to effectively enrich highly active compounds in the top 1% of molecules in the screening results, which greatly increases the workload and R&D costs of subsequent experimental verification.

[0008] Deep learning-based virtual screening methods, such as DrugCLIP, RTMScore, and Gnina, have significantly improved screening speed through a data-driven approach, to some extent compensating for the efficiency shortcomings of traditional methods. However, they still have not overcome the underlying logical bottlenecks of virtual screening technology and have a series of core defects that cannot be ignored:

[0009] First, the models generally suffer from attitude dependence. Most methods rely on predefined docking attitudes or crystal conformations for modeling, making it impossible to achieve pose-free and accurate interactive modeling. When dealing with flexible conformational systems, the prediction accuracy will decrease significantly.

[0010] Secondly, there is a core shortcoming of missing microscopic physical information. Existing models rely only on 2D molecular topological fingerprints or pure 3D geometric structures to extract features, and cannot encode core quantum chemical information such as microscopic stereoelectronic effects, molecular potential energy surface topology, charge transfer, and local dipole moments that determine the real binding affinity. As a result, the model is essentially a statistical memory of the molecular topological framework, rather than a real modeling of the target-ligand physical binding mechanism.

[0011] Furthermore, the models exhibit extremely poor robustness in challenging scenarios. In demanding environments where the structures of active and decoy molecules are highly similar (such as the LIT-PCBA dataset), the Top 1% enrichment factor (EF1%) of most methods decreases significantly, failing to meet the requirements of real-world drug development. Simultaneously, the models suffer from severely insufficient zero-shot generalization ability; for novel targets without known active ligands, their prediction performance declines significantly, making them unsuitable for the crucial novel target screening scenarios in innovative drug development.

[0012] Finally, the curse of dimensionality problem under the classical computing framework remains unsolved. Similarity matching of high-dimensional molecular features still relies on classical computing systems. When faced with a library of hundreds of millions of molecules, it is impossible to simultaneously balance screening throughput and matching accuracy, which severely limits its large-scale industrial application in innovative drug development.

[0013] Therefore, there is an urgent need for a high-throughput, rapid, and accurate method to achieve large-scale virtual screening, fundamentally breaking through the efficiency and accuracy bottlenecks of existing technologies, and providing new technical support for the efficient discovery of innovative drug lead compounds. Summary of the Invention

[0014] The technical problem to be solved by the present invention is to provide a virtual screening method, device, computer equipment and storage medium based on quantum manifold dense retrieval, so as to realize the ultra-fast, accurate and highly interpretable virtual screening of molecular libraries of hundreds of millions, and break through the core bottleneck of existing virtual screening technology that cannot simultaneously take into account ultra-high throughput, high enrichment accuracy, strong generalization and physical interpretability.

[0015] In a first aspect, the present invention provides a virtual screening method based on quantum manifold dense retrieval, the method comprising:

[0016] Obtain the protein binding pocket diagram representation of the target protein to be screened, and the molecular diagram representation of the ligand molecule to be screened;

[0017] The protein binding pocket map and the molecular map are input into a virtual screening model, which includes: a small molecule encoding module, a protein pocket encoding module, a physical sensing cross-modal attitude-free alignment module, a quantum manifold dense retrieval module, a molecular clustering module, and a drug-likeness and affinity assessment output module.

[0018] The molecular graph is subjected to feature extraction by the small molecule encoding module to obtain the quantum manifold embedding vector representation corresponding to the molecular graph; the protein binding pocket graph is subjected to feature extraction by the protein pocket encoding module to obtain the high-dimensional feature representation of the protein binding pocket graph.

[0019] Based on the quantum manifold embedding vector representation of the molecular graph and the high-dimensional feature representation of the protein binding pocket graph, feature fusion and interactive computation are performed through the physical-aware cross-modal non-pose alignment module to obtain the protein-ligand cross-modal joint feature representation;

[0020] Based on the protein-ligand cross-modal joint feature representation, an approximate nearest neighbor search is performed in the pre-constructed quantum manifold embedding vector library through the quantum manifold dense search module to obtain a list of initial screening hit molecules with high binding tendency;

[0021] The molecular clustering module performs structural clustering and deduplication on the list of molecules that were initially screened, selects the central molecules of each category, and obtains a set of candidate molecules with excellent structural diversity.

[0022] The druggability and affinity assessment output module is used to filter the candidate molecule set for druggability and predict binding affinity, thereby obtaining the final virtual screening results and binding affinity ranking.

[0023] Secondly, the present invention provides a virtual screening device based on quantum manifold dense retrieval, comprising:

[0024] The data preprocessing module is used to obtain the protein binding pocket map corresponding to the target protein to be screened and the molecular map corresponding to the ligand molecule to be screened, and input the protein binding pocket map and the molecular map into the virtual screening model. The virtual screening model includes: a data preprocessing module, a node feature acquisition module, a cross-modal alignment module, a quantum manifold dense retrieval module, a molecular clustering module, and a result output module.

[0025] The node feature acquisition module is used to call the ligand quantum coding branch in the dual-track parallel encoder to extract features from the molecular graph, obtain the node feature representation and quantum manifold embedding vector of the molecular graph, and call the protein geometry coding branch in the dual-track parallel encoder to extract features from the protein binding pocket graph, obtain the protein node feature representation of the protein binding pocket graph.

[0026] The cross-modal alignment module is used to call the physical perception cross-modal poseless alignment module to complete poseless cross-modal feature alignment based on the node feature representations of the protein binding pocket diagram and the molecular diagram, and generate a protein-ligand cross-modal joint feature representation.

[0027] The quantum manifold dense retrieval module is used to call the quantum manifold dense retrieval module, complete the offline quantum manifold embedding vector library construction of a large-scale molecular library based on the pre-trained converged LiTENCLIP model, and perform ultra-fast approximate nearest neighbor retrieval in the vector library using target pocket features as query vectors to generate a list of initially screened hit molecules.

[0028] The molecular clustering module is used to call the molecular clustering module to perform structural clustering and deduplication on the molecules hit in the initial screening, and to screen out a set of candidate compounds with high structural diversity and good drug potential.

[0029] The results output module is used to call the results output module to quickly predict and filter the drug-likeness and toxicity risks of the clustered candidate compounds, sort the remaining molecules in descending order according to binding affinity, output the structural information of the Top-N hit molecules, the predicted binding affinity and drug-likeness parameters, and obtain the final virtual screening results.

[0030] Thirdly, the present invention also provides a computer device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, the computer program being used to perform the aforementioned virtual screening method steps based on quantum manifold dense retrieval.

[0031] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the aforementioned virtual screening method based on quantum manifold dense retrieval.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] This invention constructs a novel paradigm of quantum manifold dense retrieval, reconstructing the traditional virtual screening problem of combined free energy regression into a cross-modal dense vector retrieval task in a unified physical latent space. This fundamentally breaks through the underlying bottleneck of traditional virtual screening technology and achieves ultra-high throughput and high precision virtual screening of molecular libraries with hundreds of millions of molecules.

[0034] This invention extracts quantum manifold embeddings rich in microscopic three-dimensional electronic information through ligand quantum encoding branches. Combined with a physically-aware cross-modal attitude-free alignment module, it can accurately model the physical binding mechanism between targets and ligands without predefined docking attitudes, completely solving the core defects of existing methods such as lack of microscopic physical information and attitude dependence. In industry standard benchmark tests, the LiTENCLIP model of this invention achieved an EF1% of 44.06 on the DUD-E dataset, which is 38.2% and 62.6% higher than the state-of-the-art DrugCLIP and RTMScore, respectively. On the more challenging LIT-PCBA dataset, it achieved an EF1% of 6.77, which is 22.9% and 130.3% higher than DrugCLIP and RTMScore, respectively. At the same time, it maintained a stable high enrichment ability on 102 independent targets, demonstrating strong generalization and robustness.

[0035] This invention, through offline pre-computation and an approximate nearest neighbor retrieval architecture, can complete the full screening of a library of hundreds of millions of molecules within one to two days, which is more than three orders of magnitude faster than traditional molecular docking methods. At the same time, it maintains clear physical interpretability and can accurately capture subtle physical state differences under the activity cliff scenario. It provides a new computational solution that is efficient, accurate and highly interpretable for the discovery of innovative drug lead compounds, which significantly reduces the cost of new drug development and shortens the development cycle.

[0036] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0037] Figure 1 A flowchart illustrating the virtual screening method based on quantum manifold dense retrieval provided in this application embodiment;

[0038] Figure 2 A schematic diagram illustrating a virtual screening process executed by LiTENCLIP, provided for an embodiment of this application;

[0039] Figure 3 A schematic diagram illustrating the similarity of LiTENCLIP protein-ligand encoding in an embodiment of this application;

[0040] Figure 4 A schematic diagram illustrating the virtual filtering capabilities of LiTENCLIP on a classic virtual filtering dataset, provided as an embodiment of this application;

[0041] Figure 5 A schematic diagram illustrating the effect of LiTENCLIP in an active cliff scenario, provided as an embodiment of this application;

[0042] Figure 6A schematic diagram of the structure of a virtual screening device based on quantum manifold dense retrieval provided in an embodiment of this application;

[0043] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the purpose, description and advantages of this application clearer, the technical solutions in the examples of this application will be described in detail below with reference to the accompanying drawings.

[0045] It should be understood that the embodiments described in the examples of this application are only some examples of the technical solution, and not all embodiments. The terminology used in the embodiments and claims of this application is only used to describe specific implementation methods and is not intended to limit the scope of protection of this application. Unless the context clearly indicates otherwise, the singular forms "a," "the," and "the" in the embodiments and claims of this application also include their plural forms.

[0046] It should be noted that relational terms such as "first" and "second" in this document are used only to distinguish different objects or operational steps and do not indicate any actual relationship or sequence between them. Furthermore, the terms "comprising," "including," and any synonymous variations indicate a non-exclusive inclusion relationship. Therefore, an article, method, terminal, or process that comprises several elements may include, in addition to the listed elements, other elements not expressly listed or well-known in the art.

[0047] Unless otherwise specified, the phrase "comprising a..." does not exclude the inclusion of other identical or similar elements. Or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article, or terminal that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes said element.

[0048] Referring to Figure 1, this embodiment of the application illustrates a flowchart of a poseless ultra-high-throughput virtual drug screening method based on quantum manifold dense retrieval. As shown in Figure 1, the execution of this method may include the following steps:

[0049] Step 101: Obtain the protein binding pocket map corresponding to the target protein to be screened, and the molecular map corresponding to the ligand molecule to be screened.

[0050] In this embodiment, a molecular graph corresponding to the ligand molecule to be screened can be obtained. Specifically, the ligand to be screened can be characterized as a three-dimensional molecular graph. This molecular graph uses atoms as nodes and covalent bonds and non-bonded interactions between atoms as edges to construct an atomic-level subgraph. The node features and edge features of the atomic-level subgraph are shown in Table 1 below. The construction and characterization of the molecular graph in Table 1 can be completed using open-source Python libraries such as RDKit.

[0051] Table 1

[0052]

[0053] The target protein can be represented as a three-dimensional protein-binding pocket diagram structure, composed of residue-level subgraphs and atomic-level subgraphs. The residue-level subgraph uses amino acid residues as nodes, with node coordinates derived from the Cα atom coordinates of each residue. Using the K-nearest neighbor algorithm, each node is connected to its 30 nearest neighbors in Euclidean space, thus constructing an overall residue-level topological graph. The atomic-level subgraph uses atoms within the binding pocket as nodes, establishing edge connections based on the chemical covalent bonds or virtual non-bonded interactions within the residues to reflect local chemical structural features. The specific graph structure definitions are shown in Table 2.

[0054] Table 2

[0055] In Table 2, `bool2value` represents converting a Boolean variable to a numerical form, where `True` represents 1 and `False` represents 0. In practical applications, the construction and characterization of protein binding pocket maps can be accomplished using open-source Python libraries such as RDKit, MDAnalysis, and Uni-Mol. After obtaining the molecular map corresponding to the ligand molecule to be screened and the protein binding pocket map corresponding to the target protein, step 102 is executed.

[0056] Step 102: In this example, the virtual screening model can consist of the following modules: a small molecule encoding module, a protein pocket encoding module, a cross-modal alignment module, a molecular clustering module, and a drug-likeness and affinity assessment output module. After obtaining the molecular map corresponding to the ligand molecule to be screened and the protein binding pocket map corresponding to the target protein, both can be input into the corresponding encoding modules, and step 103 can be executed.

[0057] Step 103: This step extracts and encodes the feature representations of the protein pocket and ligand molecular map by calling the corresponding encoding modules. See [link to step 103]. Figure 2In this embodiment, feature extraction of the ligand molecular map is accomplished through a ligand quantum coding branch, the core of which includes a tensor quadrilateral attention module and a scalar-vector fusion module.

[0058] Furthermore, Tensorized Quadrangle Attention (TQA) is a highly efficient molecular technique with underlying physical support, proposed in this application. Figure 2 The face-corner interaction capture mechanism implicitly models the geometric relationships of four bodies during the edge feature update process, maintaining high feature expressiveness while keeping linear computational complexity. The calculation formula for the TQA mechanism is as follows:

[0059]

[0060] in, and Embed the direction vectors of nodes i and j. Let i be the normalized direction vector from node i to node j. Features of the edge at layer l and The weight matrix and bias vector are learnable; in the formula The term captures the relationship of torsion, which is similar to the physical meaning of dihedral angles, and is realized in a differentiable and vectorized form. This value modulates the nonlinear transformation result of the edge state of the previous layer through the Hadamard product (element-wise multiplication) to complete the update of edge features.

[0061] The above formula can implicitly model the torsional coupling between bonded atoms by encoding local quadrilateral interactions through edges, without explicitly enumerating four-body terms. Since each step of message passing contains information about node orientation and edge vectors, the effective receptive field of the model is significantly expanded. Under the condition that the cutoff radius of edge construction is 5 Å, the effective receptive field of single-layer tensor attention can reach about 10 Å, enabling the network to efficiently capture medium- and long-range structural correlations.

[0062] Furthermore, to incorporate node-level information and enrich feature representations, this module introduces a learnable attention mechanism. The formula for calculating the attention score is:

[0063]

[0064] in, and Let α represent the scalar features of atoms i and j, and let α∈RC be the trainable parameter vector for modulating attention weights.

[0065] Subsequently, attention is applied using a continuous truncation function based on inter-atomic distance. The gating process is performed using the following formula:

[0066]

[0067] This hybrid design achieves adaptive feature weighting that simultaneously considers both learned chemical relevance and geometric locality, and then completes the calculation of updated node features and vector features using the following formula:

[0068]

[0069] in, Represents a scalar message passing item. Represents a vector message passing item.

[0070] Finally, the scalar and vector features of neighboring nodes are aggregated through scattering-based summation and residual concatenation, as shown in the following formula:

[0071]

[0072] in, Scalar characteristics representing nodes at level l. The vector feature representing the node at layer l.

[0073] This structure constructs a fully vectorized, geometrically aware message-passing mechanism that can model complex conformational dependencies in a scalable and differentiable manner. By capturing implicit four-body interactions through local vector operations, the tensor quadrilateral attention enhances the model's ability to characterize torsional potential surfaces and anisotropic molecular features while maintaining computational efficiency for large-scale systems.

[0074] Through the six-layer stacking of the tensor quadrilateral attention module and the scalar-vector fusion module, the ligand quantum encoding branch completes the multipole moment decomposition and tensor contraction operation of the molecular electronic structure, extracts microscopic physical information including molecular potential energy surface topology, charge shielding effect, local dipole moment, and stereoelectronic effect, and finally encodes and generates a fixed-dimensional quantum manifold embedding (QME) vector.

[0075] Feature extraction of the protein binding pocket map is accomplished through a protein geometry coding branch, which uses a pre-trained Uni-Mol isomorphic graph neural network to extract features in order to capture the local chemical environment and spatial geometry of the binding pocket.

[0076] Furthermore, the protein geometry encoding branch employs an 8-layer pre-trained Uni-Mol isomorphic graph neural network as a 3D geometric field encoder. The inputs include the residue sequence of the target binding pocket, the 3D coordinates of the Cα atom, the side chain orientation vector, and the chemical properties of the residues. Through a multi-layer isomorphic graph neural network, spatial distribution features, local chemical environment features, and potential interaction region features at the pocket residue level are extracted. This ultimately generates a 1024-dimensional high-dimensional feature representation of the protein binding pocket, with the same dimension as the QME vector, and outputs it to the physically-aware cross-modal pose-free alignment module. After completing the multi-level feature encoding and fusion of the protein binding pocket map and the ligand molecule map, the model obtains a comprehensive representation characterizing the spatial structure information of the target protein and the chemical topological features of the ligand, and then proceeds to step 104.

[0077] Step 104: This step uses a physically-aware cross-modal attitude-free alignment module to achieve cross-modal precise alignment of ligand quantum manifold embedding and protein binding pocket features. Implicit modeling of target-ligand binding interactions can be completed without explicitly defining the ligand docking attitude.

[0078] After obtaining the encoded protein binding pocket graph node feature representation and ligand molecule graph node feature representation, the residue-level feature representation of the protein binding pocket and the atomic-level feature representation of the ligand molecule can be jointly input into the physical-aware cross-modal attitude-free alignment module.

[0079] Furthermore, this module first maps ligand atomic-level features to query vectors and protein residue-level features to bond and value vectors using a linear layer, unifying the dimensions of the three types of vectors. Then, fine-grained cross-attention calculation is performed on the query vectors and bond-value pairs, enabling the model to autonomously learn the spatial correspondence between ligand atoms and protein residues. This process does not require pre-defining the docking posture and binding position of the ligands. During attention calculation, the module adaptively adjusts attention weights to autonomously capture key non-covalent interactions at the binding interface, including the directionality of hydrogen bond networks, the electrostatic complementarity of hydrophobic cavities, the higher-order coupling of overlapping π-π electron clouds, and the geometric constraints of metal coordination, achieving a physically-aware model of the target-ligand binding mechanism. Finally, a spatial pooling strategy is used to globally and locally aggregate the features after cross-attention interaction, generating a protein-ligand cross-modal joint feature representation. Simultaneously, the interaction attention weight map of the binding interface is output for subsequent model training and physical mechanism analysis. After completing cross-modal feature alignment and joint feature representation generation, step 105 is executed.

[0080] Step 105: After cross-modal feature alignment and joint feature representation generation, this step completes the construction of an offline quantum manifold embedding vector library for a large-scale molecular library. Using the target pocket feature as the query vector, a fast approximate nearest neighbor search is performed in the vector library to quickly match the set of small molecules with the highest binding tendency to the target, generating a list of initially screened hit molecules. Then, step 106 is executed.

[0081] Step 106: After completing the quantum manifold dense search, this step uses the molecular clustering module to cluster the initially screened set of small molecules with high binding tendency, and screen out candidate molecules with high structural diversity and strong representativeness.

[0082] Specifically, based on the similarity of small molecule QME vectors, clustering algorithms (such as K-means and hierarchical clustering) are used to classify the initially screened small molecules, grouping those with similar structures and binding mechanisms into one class. After clustering, the most representative central molecule (i.e., the class center molecule) is selected from each class, eliminating redundant and repetitive small molecules with excessively high structural similarity, and retaining the candidate molecule set with excellent structural diversity and outstanding binding potential. This effectively reduces the computational load of subsequent evaluations while ensuring the diversity and representativeness of the candidate molecules. After completing clustering and determining the central molecules of each class, step 107 is executed.

[0083] Step 107: This step uses the druggability filtering and result output module to complete the druggability assessment, filtering, and result ranking of the molecules that were initially screened, and outputs the final virtual screening results and the predicted binding affinity value.

[0084] Furthermore, after obtaining the initial list of target molecules, the drug-likeness (Lipinski's five rules) of the target molecules is rapidly predicted simultaneously, filtering out molecules that do not meet the drug-likeness requirements; finally, the remaining molecules are sorted in descending order according to the cosine similarity score, and the SMILES sequences, similarity scores, predicted binding affinity, drug-likeness parameters and 3D conformations of the Top-N target molecules are output, completing the entire ultra-high throughput virtual screening process.

[0085] The advantages of the virtual screening model based on quantum manifold dense retrieval invented in this application are described below in conjunction with several test results:

[0086] I. Visualization of Latent Space Representation and Validation of Model Discriminative Ability

[0087] To systematically verify the ability of the LiTENCLIP model in this application to identify target-ligand binding compatibility, the latent space representation learned by the model was first visualized and analyzed. The results are as follows: Figure 3 As shown.

[0088] To systematically validate the model's feature learning performance, eight representative targets were selected from the DUD-E dataset, covering various core protein categories such as VEGFR-2 kinases, ADRB1 GPCRs, proteases, and nuclear receptors. After projecting the LiTEN-Base encoded protein pocket embeddings and small molecule embeddings onto the same shared latent space, a significant distribution pattern emerged: all active small molecules tightly clustered around their corresponding target protein embeddings, with active molecules from different targets forming independent clusters. This result directly demonstrates that the model in this application possesses excellent molecular discrimination capabilities, spontaneously bringing physically compatible target-active ligand pairs closer together in the latent space and accurately encoding the binding compatibility characteristics between targets and ligands.

[0089] To further verify the robustness of this recognition capability under massive chemical noise, the distribution of the eight selected protein pockets and their corresponding active molecules was visualized against the background of the entire DUD-E dataset, with all inactive decoy molecules marked in gray. The visualization results show that the active molecules of each target form clear and independent clusters within the vast gray background of the inactive decoys, while the inactive decoy molecules are evenly distributed in the background area. This result conclusively demonstrates that even in highly complex and noisy data environments, LiTENCLIP can easily and accurately identify and enrich the active molecules corresponding to specific proteins, exhibiting strong anti-interference capabilities and target-ligand recognition accuracy.

[0090] II. Validation of Virtual Screening Performance and Target Generalization on Standard Benchmark Datasets

[0091] After completing the latent space representation validation, the macroscopic screening performance of the LiTENCLIP model was comprehensively evaluated on industry-standard benchmark datasets. Simultaneously, target-level generalization validation was conducted, and the results are as follows: Figure 4 As shown in Tables 3 and 4.

[0092] 1. Performance testing of the DUD-E benchmark dataset

[0093] The DUD-E dataset is the most universal standard test set in the field of structure-based virtual screening. It covers 102 different types of drug targets, including active molecules and structure-matched decoy molecules, and can comprehensively measure the overall screening performance and early enrichment ability of the model.

[0094] This embodiment compares the LiTENCLIP model of this application with 11 mainstream traditional docking methods and deep learning models, including Glide-SP, Surflex, DrugCLIP, RTMScore, and DrugHash. The three core evaluation indicators, AUROC, BEDROC80.5, and EF1%, are calculated uniformly, and the results are shown in Table 3.

[0095] Table 3: Performance Comparison of Various Models on the DUD-E Dataset

[0096]

[0097] Quantitative results show that LiTENCLIP achieved industry-leading results on the DUD-E dataset with AUROC 0.90, BEDROC 80.5-0.68, and EF1% 44.06, surpassing all comparable models across the three core evaluation metrics. Specifically, the core metric EF1%, representing early enrichment capability, improved by 18.5% compared to the closest performing DrugHash, 38.2% compared to DrugCLIP, and 172% compared to the traditional docking software Glide-SP. The AUROC metric, representing overall discriminative ability, improved by 7.1% compared to the second-best DrugHash. The BEDROC 80.5 metric, representing early enrichment weight, improved by 19.3% compared to the second-best DrugHash. These results fully demonstrate that dynamically aligning the quantum manifold of the ligand with the chemical environment of the protein pocket achieves significantly stronger discriminative ability compared to traditional empirical topological matching, achieving early enrichment results far exceeding existing methods and significantly reducing the workload and R&D costs of subsequent experimental validation.

[0098] 2. Performance testing on the challenging LIT-PCBA benchmark dataset.

[0099] To eliminate the possibility of the model exploiting human bias in the DUD-E dataset and to rigorously verify the effectiveness of the physical binding logic of the retrieval mechanism in real unbiased scenarios, further evaluation was conducted on the highly challenging LIT-PCBA dataset. The active molecules and decoy molecules in this dataset have extremely high structural similarity, which rigorously verifies the model's ability to model real physical structures, rather than relying on statistical memory of topological frameworks. The performance comparison of each model is shown in Table 4.

[0100] Table 4: Performance Comparison of Various Models on the LIT-PCBA Dataset

[0101]

[0102] Test results show that under stringent conditions where the activity and decoy molecule structures are highly similar, the performance of most comparative models experienced a precipitous drop. However, the LiTENCLIP model maintained its leading screening performance, achieving world-class results with AUROC 0.61, BEDROC 80.5 0.08, and EF1% 6.77. Specifically, the EF1% was 22.9% higher than DrugCLIP and 130.3% higher than the RTMS score, while the traditional docking method Glide-SP had an EF1% of only 3.41, and the deep learning method had an RTMS score of only 2.94, both showing performance degradation exceeding 50%. These results objectively confirm that the quantum chemical information injected into the model successfully captured the differences in three-dimensional electronic effects that determine the actual binding affinity, rather than relying solely on statistical memory of the topological framework. Therefore, it possesses extremely strong resistance to distribution shifts and scenario robustness, perfectly adapting to the complex screening scenarios in real drug development.

[0103] 3. Target-level generalization validation

[0104] This embodiment further delves into the micro-target level from the macro-benchmark, comprehensively validating the model's generalization ability. For the 102 independent targets in the DUD-E dataset, the EF1% and AUROC metrics of each model on a single target were calculated. The results show that LiTENCLIP achieves an EF1% exceeding 30 on 92% of the targets and an EF1% exceeding 40 on over 80% of the targets. Simultaneously, all targets maintain an AUROC performance above 0.82, with no targets exhibiting significant performance weaknesses, demonstrating highly smooth and stable enrichment robustness.

[0105] For 15 challenging targets in the LIT-PCBA dataset, the EF1% performance of the model on each target was decomposed using an ensemble divergent sunburst plot. The results show that LiTENCLIP still achieved a significant early enrichment effect with EF1% greater than 6 on highly challenging druggable targets such as PPARG, TP53, ADRB2, and FEN1. The EF1% on all 15 targets exceeded 4, demonstrating highly stable performance. These results fully demonstrate that the physically driven quantum manifold latent space constructed in this application has broad applicability in different types of protein folding, binding pocket microenvironments, and ligand binding modes. For novel targets without known active ligands, it can still maintain excellent zero-sample screening capability, perfectly meeting the core R&D needs of novel and druggable targets in innovative drug development.

[0106] III. Microscopic resolution test of active cliff

[0107] This application further extends the model's resolution to the finest molecular level, validating the microscopic physical feature resolution capability of the LiTENCLIP model through an active cliff case study. The results are as follows: Figure 5 As shown.

[0108] The activity cliff refers to a scenario where two small molecules have extremely similar topological structures but differ in activity by 2-3 orders of magnitude. This represents the ultimate test of the microscopic resolution capability of virtual screening models and is a core weakness of traditional virtual screening models. This example selects two classic drug development targets, GSK-3β and AR, to address two pairs of structurally highly similar molecules with a difference in binding free energy (…). Molecular pairs with scores of -1.83 and -2.58 were scored using LiTENCLIP and the mainstream deep learning model DrugCLIP, respectively.

[0109] The results showed that in the molecular pairs targeting GSK-3β, the two molecules differed only in the substitution of a single chlorine atom on the benzene ring. DrugCLIP scored the active molecule 4.05 and the inactive analogue 2.83, showing only weak discrimination and failing to accurately reflect the activity difference of more than 100 times. In contrast, LiTENCLIP scored the active molecule 0.26 and the inactive analogue 0.44, with the scoring trend perfectly matching the actual activity trend and perfectly reproducing the activity difference of the molecular pairs. In the molecular pair tests targeting AR, the two molecules differed only in the substitution of a single methyl group on the side chain. DrugCLIP scored the two molecules by only 0.10, almost unable to distinguish between the active molecule and the inactive analogue. In contrast, LiTENCLIP scored by 0.23, accurately reflecting the activity difference between the two molecules.

[0110] These results demonstrate that traditional deep learning models can only capture differences in molecular topology and cannot identify subtle changes in local stereoelectronic effects. Therefore, they have extremely weak resolution capabilities in activity cliff scenarios, easily leading to the problem of "missing highly reactive molecules and mistakenly selecting inactive analogs." In contrast, the LiTENCLIP model in this application implicitly encodes microscopic physical information such as molecular potential energy surfaces, local dipole moments, and charge distribution through quantum manifold embedding. This enables it to accurately capture local physical state changes caused by single-atom substitution, achieving ultra-high precision resolution of activity cliffs and effectively solving the core pain points of traditional screening methods.

[0111] IV. Validation of Virtual Screening Applications for Popular Drug Targets

[0112] To explore the broad applicability of LiTENCLIP in real-world drug discovery scenarios, a large-scale virtual screening application was conducted using LiTENCLIP to target multiple popular clinical drug targets, covering three core drug target categories: kinase targets, epigenetic targets, and GPCR targets.

[0113] During the screening process, based on commercial molecular libraries, target structure preprocessing, binding pocket feature extraction, full-library quantum manifold dense search, drug-likeness filtering, and structure clustering were performed for the aforementioned targets. Finally, a Top 100 high-potential candidate compounds were output for each target. The screening results show that LiTENCLIP exhibits stable screening performance for targets with different mechanisms and structural types. For each target, it can enrich a large number of highly active potential compounds with matching pharmacophores and conforming to the spatial and chemical environment characteristics of the target binding pocket. Furthermore, all screening procedures can be completed within one day, fully demonstrating the broad applicability and ultra-high throughput advantages of the proposed method on different drug targets.

[0114] This virtual screening application validation covers the three core target categories in innovative drug development: kinases, epigenetics, and GPCRs. It confirms that the LiTENCLIP model in this application can be adapted to different types of drug development scenarios, providing an efficient and universal technical solution for the discovery of lead compounds for various disease-related targets, and has extremely high industrial application value.

[0115] Reference Figure 6 This illustration shows a schematic diagram of a posture-free, ultra-high-throughput virtual drug screening device based on quantum manifold dense retrieval, provided in an embodiment of this application. Figure 6 As shown, the ultra-high-throughput virtual drug screening device 600 based on quantum manifold dense retrieval may include the following modules:

[0116] The data preprocessing module 610 is used to obtain the protein binding pocket map corresponding to the target protein to be screened and the molecular map corresponding to the ligand molecule to be screened, and input the protein binding pocket map and the ligand molecule map into the virtual screening model.

[0117] The node feature acquisition module 620 is used to call the ligand quantum coding branch in the dual-track parallel encoder to extract features from the ligand molecular map, obtain the node feature representation and quantum manifold embedding vector of the ligand molecular map, and call the protein geometry coding branch in the dual-track parallel encoder to extract features from the protein binding pocket map, obtain the protein node feature representation of the protein binding pocket map.

[0118] The cross-modal alignment module 630 is used to call the physical perception cross-modal poseless alignment module to complete poseless cross-modal feature alignment based on the node feature representation of the protein binding pocket map and the node feature representation of the ligand molecule map, and generate a protein-ligand cross-modal joint feature representation.

[0119] The quantum manifold dense retrieval module 640 is used to construct an offline quantum manifold embedding vector library for a large-scale molecular library based on the pre-trained convergent LiTENCLIP model. It then performs an extremely fast approximate nearest neighbor retrieval in the vector library using target pocket features as query vectors to generate a list of initially screened hit molecules.

[0120] The molecular clustering module 650 is used to perform structural clustering and deduplication on the molecules that have been initially screened, and to screen out a set of candidate compounds with high structural diversity and good drug potential, so as to avoid the interference of duplicate structures in subsequent evaluation.

[0121] The results output module 660 is used to quickly predict and filter the drug-likeness and toxicity risks of clustered candidate compounds, sort the remaining molecules in descending order of binding affinity, output the structural information of the top-N hit molecules, predict the binding affinity and drug-likeness parameters, and obtain the final virtual screening results.

[0122] The node feature acquisition module includes a ligand quantum coding unit, a protein geometric coding unit, and a feature dimension alignment unit.

[0123] Furthermore, the ligand quantum encoding unit is used to perform dimension mapping processing on the atomic node features and edge features within the ligand molecule graph to obtain dimensionally unified atomic node features and edge features. The dimensionally unified atomic node features and edge features are then processed by the tensor quadrilateral attention module and the scalar-vector fusion module to obtain the quantum manifold embedding vector of the ligand molecule.

[0124] Furthermore, the protein geometry encoding unit is used to perform dimension mapping processing on the residue node features and edge features within the protein binding pocket graph to obtain residue node features and edge features with uniform dimensions. Then, the uniform residue node features, edge features, and protein sequence information encoding features are processed by an isovariant graph neural network to obtain a high-dimensional feature representation of the protein binding pocket.

[0125] Furthermore, the feature dimension alignment unit is used to perform dimension unification and standardization processing on the ligand quantum manifold embedding vector and the high-dimensional feature representation of the protein binding pocket, and outputs it to the cross-modal alignment module.

[0126] The cross-modal alignment module includes a feature linear mapping unit, a physical awareness cross-attention unit, and a feature aggregation output unit.

[0127] Furthermore, the feature linear mapping unit is used to map ligand atomic-level features to query vectors and protein residue-level features to bond vectors and value vectors, unifying the dimensions of the three types of vectors.

[0128] Furthermore, the physical perception cross-attention unit is used to perform fine-grained cross-attention calculations on the query vector and key-value pairs, autonomously capturing key non-covalent interactions of the interface and completing pose-free cross-modal feature interaction.

[0129] Furthermore, the feature aggregation output unit is used to perform global and local aggregation on the features after cross-attention interaction through a spatial pooling strategy, generating a protein-ligand cross-modal joint feature representation, and simultaneously outputting the interaction attention weight map of the binding interface.

[0130] The quantum manifold dense retrieval module includes an offline vector library construction unit, a query vector generation unit, and an approximate nearest neighbor retrieval unit.

[0131] Furthermore, the offline vector library construction unit is used to perform offline pre-computation on a large-scale virtual molecular library through a pre-trained and converged model, generating a quantum manifold embedding vector database of hundreds of millions of units, and establishing a retrieval index.

[0132] Furthermore, the query vector generation unit is used to extract the high-dimensional feature representation of the binding pocket of the target protein to be screened, as the retrieval query vector.

[0133] Furthermore, the approximate nearest neighbor retrieval unit is used to perform ultra-fast approximate nearest neighbor retrieval in a pre-built vector database using target pocket features as the query vector, to complete the matching and sorting of all molecules in the database and generate a list of molecules that have been initially screened and hit.

[0134] The result output module includes a drug-likeness filtering unit and a result sorting and output unit.

[0135] Furthermore, the drug-likeness filtering unit is used to quickly predict the drug-likeness and toxicity risks of molecules that have been initially screened, and to filter out molecules that do not meet the drug-likeness requirements.

[0136] Furthermore, the result sorting and output unit is used to sort the filtered molecules in descending order of binding tendency and output the structural information, predicted binding affinity and druggability parameters of the Top-N hit molecules.

[0137] The postureless ultra-high-throughput drug virtual screening device based on quantum manifold dense retrieval provided in this application embodiment acquires the protein-binding pocket map corresponding to the target protein to be screened and the molecular map corresponding to the ligand molecule to be screened, and inputs both into the virtual screening model. A dual-track parallel encoder completes the extraction of microscopic physical features and the generation of quantum manifold embedding. A physical perception cross-modal postureless alignment module completes the postureless interaction modeling of the target-ligand. A quantum manifold dense retrieval module completes the ultra-high-throughput screening of a large-scale molecular library. Finally, a druggability filtering and result output module outputs the final screening results. This application embodiment can effectively improve the early enrichment capability, screening efficiency, and generalization of ultra-high-throughput virtual screening, while maintaining clear physical interpretability, providing efficient and accurate technical support for the discovery of innovative drug lead compounds.

[0138] Figure 7 Figure 7 illustrates a schematic diagram of the structural composition of the electronic device 700 provided in an embodiment of the present invention. As shown in Figure 7, the electronic device 700 is mainly equipped with a central processing unit (CPU) 701. This processing unit can execute corresponding operations and control logic according to program instructions pre-stored in a read-only memory (ROM) 702, and can also read instructions loaded from the storage unit 708 to the random access memory (RAM) 703 to complete various data processing and process scheduling operations. The RAM 703 is also used to temporarily store various intermediate data, program variables, and system resources required during the operation of the electronic device 700, providing temporary storage space for the operation process. The CPU 701, ROM 702, and RAM 703 communicate and transmit signals through a system bus 704 to ensure the coordinated operation of each hardware unit. In addition, the system bus 704 is also connected to an input / output (I / O) interface 705 to realize signal transfer and data communication between internal processing units and external devices.

[0139] In the overall architecture of the electronic device 700, various peripherals and functional expansion modules are connected to the system bus through the I / O interface 705, thereby achieving interconnection with the core processing unit. The input unit 706 may include a keyboard, mouse, touch panel, or voice acquisition device, used to receive user commands and external input information; the output unit 707 may include a display screen, audio playback components, etc., used to present processing results and operating status; the storage unit 708 may use magnetic storage media, optical storage media, or solid-state storage devices for long-term storage of program code, model parameters, and filtering result data; the communication unit 709 may integrate a network adapter, wireless communication module, or modem, supporting the electronic device 700 to transmit data and exchange information with external terminals, cloud servers, or database platforms via the Internet or other dedicated communication networks.

[0140] The virtual screening method flow and processing steps described in the foregoing embodiments of this specification can all be executed by the central processing unit 701 calling the corresponding computer program. The computer program can be permanently stored in a computer-readable storage medium, such as in the storage unit 708. In some implementation scenarios, the relevant program can be directly loaded and run via ROM 702, or it can be transmitted over a network, deployed, installed, or updated via communication unit 709. When the program instructions are loaded into RAM 703 and scheduled for execution by CPU 701, the electronic device 700 can be driven to complete all or part of the processing steps in the aforementioned drug virtual screening method based on quantum manifold dense retrieval, achieving automated, high-throughput target-ligand screening operations.

[0141] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the aforementioned posture-free ultra-high-throughput virtual drug screening method based on quantum manifold dense retrieval.

[0142] In the relevant embodiments of this application, the computer-readable storage medium used to store program instructions can take many forms. Its carrier is not limited to USB flash drive, mobile storage device, read-only memory (ROM), magnetic recording medium or optical medium, etc. Any medium capable of storing computer executable code is applicable.

[0143] It should be noted that those skilled in the art will understand that the functional structures and algorithm steps described herein can be implemented through hardware circuits, software instructions, or a combination of hardware and software. To avoid unnecessary repetition, this specification mainly describes these components and operations in an abstract manner according to their functions. As for whether the actual implementation is undertaken by hardware or executed by software, this can be determined by those skilled in the art based on specific application requirements, system constraints, and design trade-offs. This choice will not affect the applicability and effectiveness of the technical solution of this application.

[0144] Furthermore, the apparatus and method proposed in this application are not limited to the division method used in the specification. The foregoing embodiments are merely illustrative of the structure and are used to explain the technical principles, and are not intended to limit the actual structure. In specific implementations, the functional units in the apparatus can be recombined, adjusted, or simplified according to the system architecture. For example, some units can be physically independent or integrated into the same processing module; multiple units can also be further split or merged to improve system performance or simplify deployment.

[0145] Similarly, the order of steps described in this application is not necessarily fixed and can be adjusted, combined, or omitted according to different needs of business processes, execution efficiency, or hardware logic. In software implementation scenarios, if certain functional units exist in the form of program modules and can be distributed independently as a product, these modules can be encapsulated and stored in a computer-readable medium. When the program is loaded into the processor, the processor can execute instructions to complete all or part of the processing steps described in this application.

[0146] In summary, although several embodiments are provided in the specification for reference, these embodiments are not intended to limit the scope of protection of this application. Those skilled in the art will recognize that various equivalent modifications and alternatives can be made to the described solutions without departing from the technical concept of this application. All such modifications should be considered to be included within the scope of protection claimed in this application, and the final scope of protection is determined by the submitted claims.

[0147] Although this application has described the technical solutions in conjunction with preferred embodiments, those skilled in the art, upon understanding the basic concept of the invention, may still make various modifications, substitutions, or extensions to these embodiments. Therefore, the appended claims are intended to cover all equivalent adjustments and improvements made based on the core ideas of this application, and are not limited to the embodiments illustrated in the specification.

[0148] This application relates to a posture-free, ultra-high-throughput virtual drug screening method and corresponding apparatus, electronic equipment, and computer-readable storage medium based on quantum manifold dense retrieval, which has been described in detail through specific examples. The above embodiments are intended to aid in understanding the principles and working mechanisms of this technical solution, and are not intended to limit this application. For those skilled in the art, different variations or extensions can be made to the specific operating procedures, system structures, or application scenarios without departing from the basic concept of this application. Therefore, the content of this specification should not be considered as a limitation on the scope of protection of this application; the actual scope of protection should be determined by the claims.

Claims

1. A virtual screening method based on quantum manifold dense retrieval, characterized in that... The method includes the following steps: Obtain the protein binding pocket diagram representation of the target protein to be screened, and the molecular diagram representation of the ligand molecule to be screened; The protein binding pocket map and the molecular map are input into a virtual screening model, which includes: a small molecule encoding module, a protein pocket encoding module, a physical sensing cross-modal attitude-free alignment module, a quantum manifold dense retrieval module, a molecular clustering module, and a drug-likeness and affinity assessment output module. The molecular graph is subjected to feature extraction by the small molecule encoding module to obtain the quantum manifold embedding vector representation corresponding to the molecular graph; the protein binding pocket graph is subjected to feature extraction by the protein pocket encoding module to obtain the high-dimensional feature representation of the protein binding pocket graph. Based on the quantum manifold embedding vector representation of the molecular graph and the high-dimensional feature representation of the protein binding pocket graph, feature fusion and interactive computation are performed through the physical-aware cross-modal non-pose alignment module to obtain the protein-ligand cross-modal joint feature representation; Based on the protein-ligand cross-modal joint feature representation, an approximate nearest neighbor search is performed in the pre-constructed quantum manifold embedding vector library through the quantum manifold dense search module to obtain a list of initial screening hit molecules with high binding tendency; The molecular clustering module performs structural clustering and deduplication on the list of molecules that were initially screened, selects the central molecules of each category, and obtains a set of candidate molecules with excellent structural diversity. The druggability and affinity assessment output module is used to filter the candidate molecule set for druggability and predict binding affinity, thereby obtaining the final virtual screening results and binding affinity ranking.

2. The method according to claim 1, characterized in that, The feature extraction of the molecular map through the small molecule encoding module includes: The ligand quantum coding branch in the small molecule coding module is invoked, and the molecular graph is subjected to multi-level feature extraction through the tensor quadrilateral attention module and the scalar-vector fusion module to obtain the quantum manifold embedding vector representation; The tensor quantization quadrilateral attention module captures the torsional coupling between atomic nodes in the molecular graph by implicitly modeling the geometric relationship of the four bodies during the edge feature update process.

3. The method according to claim 2, characterized in that, The tensor-quantized quadrilateral attention module updates edge features according to the following formula: in, and Embed the direction vectors of nodes i and j. Let i be the normalized direction vector from node i to node j. Features of the edges in the l-th layer and These are the learnable weight matrix and bias vector.

4. The method according to claim 1 or 2, characterized in that, Feature extraction of the protein binding pocket map is performed using the protein pocket encoding module, including: The protein geometry coding branch in the protein pocket coding module is invoked, and a pre-trained Uni-Mol isomorphic graph neural network is used to extract features from the protein binding pocket map to obtain a high-dimensional feature representation of the protein binding pocket map. The inputs to the protein geometric coding branch include: the residue sequence of the target binding pocket, the three-dimensional coordinates of the Cα atom, the side chain direction vector, and the chemical properties of the residues.

5. The method according to claim 1, characterized in that, Feature fusion and interactive computation are performed through the physical perception cross-modal pose-free alignment module, including: The quantum manifold embedding vector representation of the molecular graph is mapped to a query vector through a linear layer, and the high-dimensional feature representation of the protein binding pocket graph is mapped to a bond vector and a value vector, and the dimensions of the query vector, bond vector and value vector are unified. Fine-grained cross-attention calculations are performed on the query vector, the bond vector, and the value vector, enabling the model to autonomously learn the spatial correspondence between ligand atoms and protein residues and capture non-covalent interactions at the binding interface. The protein-ligand cross-modal joint feature representation is generated by globally and locally aggregating the features after cross-attention interaction using a spatial pooling strategy.

6. The method according to claim 1, characterized in that, The quantum manifold dense retrieval module performs approximate nearest neighbor retrieval in a pre-built quantum manifold embedding vector library, including: Based on the pre-trained and converged LiTENCLIP model, offline pre-computation is performed on a large-scale virtual molecular library to construct the quantum manifold embedding vector library and establish a retrieval index; Using the protein-ligand cross-modal joint feature representation as the query vector, a fast approximate nearest neighbor search is performed in the quantum manifold embedding vector library to obtain a list of initial screening hit molecules with the highest binding tendency to the target.

7. The method according to claim 1, characterized in that, The molecular clustering module performs structural clustering and deduplication on the list of molecules that were initially screened, including: Based on the similarity of the quantum manifold embedding vectors of each molecule in the initial screening molecule list, the K-means clustering algorithm or hierarchical clustering algorithm is used to classify the initial screening molecule list, and molecules with similar structures and similar binding mechanisms are grouped into one category. The most representative central molecules from each category are selected, and redundant, repetitive, or structurally similar molecules are removed to obtain a set of candidate molecules with excellent structural diversity.

8. A virtual screening device based on quantum manifold dense retrieval, characterized in that, include: The data preprocessing module is used to obtain the protein binding pocket map corresponding to the target protein to be screened and the molecular map corresponding to the ligand molecule to be screened, and input the protein binding pocket map and the molecular map into the virtual screening model. The virtual screening model includes: a data preprocessing module, a node feature acquisition module, a cross-modal alignment module, a quantum manifold dense retrieval module, a molecular clustering module, and a result output module. The node feature acquisition module is used to call the ligand quantum coding branch in the dual-track parallel encoder to extract features from the molecular graph, obtain the node feature representation and quantum manifold embedding vector of the molecular graph, and call the protein geometry coding branch in the dual-track parallel encoder to extract features from the protein binding pocket graph, obtain the protein node feature representation of the protein binding pocket graph. The cross-modal alignment module is used to call the physical perception cross-modal poseless alignment module to complete poseless cross-modal feature alignment based on the node feature representations of the protein binding pocket diagram and the molecular diagram, and generate a protein-ligand cross-modal joint feature representation. The quantum manifold dense retrieval module is used to call the quantum manifold dense retrieval module, complete the offline quantum manifold embedding vector library construction of a large-scale molecular library based on the pre-trained converged LiTENCLIP model, and perform ultra-fast approximate nearest neighbor retrieval in the vector library using target pocket features as query vectors to generate a list of initially screened hit molecules. The molecular clustering module is used to call the molecular clustering module to perform structural clustering and deduplication on the molecules hit in the initial screening, and to screen out a set of candidate compounds with high structural diversity and good drug potential. The results output module is used to call the results output module to quickly predict and filter the drug-likeness and toxicity risks of the clustered candidate compounds, sort the remaining molecules in descending order according to binding affinity, output the structural information of the Top-N hit molecules, the predicted binding affinity and drug-likeness parameters, and obtain the final virtual screening results.

9. A computer device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and executable on the processor, the computer program being used to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-7.

Images