Multi-mechanism molecular glue inducibility prediction method

Abstract

The invention discloses a multi-mechanism molecular gel inducibility prediction method, relates to the technical field of computer-aided drug design, and discloses a multi-mechanism molecular gel inducibility prediction method which combines multi-modal features such as a one-dimensional sequence, a two-dimensional graph structure and a three-dimensional structure of a coding small molecule and two types of proteins to predict the inducibility of the multi-mechanism molecular gel. A three-element cross-entity feature interaction mechanism is introduced, unified prediction of the inducibility of the ternary complex under different E3, substrates and chemical space backgrounds is realized, prediction requirements under different E3, different substrates and diversified chemical space backgrounds can be met, the influence of the conformational collaboration of the ternary complex on the inducibility is comprehensively captured, and the method has the advantages of being high in adaptability and high in practicability. And the prediction efficiency and applicability of the molecular glue are improved.

Description

Technical Field

[0001] This application relates to the field of computer-aided drug design technology, and in particular to a method for predicting the inducibility of multi-mechanism molecular colloids. Background Technology

[0002] Molecular glues are a class of small molecule compounds that can reshape intracellular interaction networks by inducing or stabilizing protein proximity. They hold significant application value in targeting "undruggable" targets such as transcription factors, disordered proteins, and flexible regulatory elements in ubiquitination systems. Unlike traditional small molecule inhibitors that rely on deep pockets, molecular glues promote inducible binding at smooth or superficial protein-protein interfaces through a "proximity-driven" mechanism, enabling various regulatory modes such as target protein degradation, enhanced native protein-protein interactions, or driving homologous association. Currently, the discovery of molecular glues mainly relies on experimental screening methods, including activity-based methods based on functional readout and interaction-based methods based on interaction readout. Activity-based methods capture the overall effects of molecular glues by detecting signals such as cell activity, transcriptional response, or changes in protein abundance. These methods are suitable for exploratory studies with unknown mechanisms, but are susceptible to interference from non-specific interactions. Interaction-based methods, on the other hand, utilize techniques such as TR-FRET, NanoBRET, surface plasmon resonance, or chemical proteomics to directly monitor the induction proximity and formation of small molecule-protein-protein ternary complexes. While beneficial for mechanism elucidation, these methods are costly and have limited throughput. Because the inducibility of molecular glues is highly dependent on the synergistic conformation and interfacial stability of the ternary complexes, a single experimental system cannot cover a broad compound and protein interaction space within acceptable time and cost, thus limiting the large-scale discovery and systematic study of molecular glues.

[0003] With the development of artificial intelligence technology, computational methods have been gradually introduced into the field of molecular glue research. Existing work has attempted to predict E3 ligase-substrate-small molecule ternary complexes from a ternary structure perspective using isomorphic graph networks to assist in interface localization and conformational sampling. Frameworks integrating molecular dynamics, interface energy assessment, and deep learning have also been developed to narrow the search space for inducible protein pairs. Furthermore, proteochemometric models and generative design models targeting specific E3 pathways have shown promise in local systems. However, these computational methods often rely on specific E3-substrate combinations or limited protein sets for modeling, and their performance is often only effective within local protein spaces and specific chemical backbones, failing to adapt to prediction needs with different E3s, different substrates, and diverse chemical spatial backgrounds. Currently, there is a lack of a universal prediction model that can jointly encode multimodal features such as one-dimensional sequences, two-dimensional graph structures, and three-dimensional structures of small molecules and two types of proteins within a unified framework, and directly discriminate the inducibility of "compound-protein-protein" ternary systems. This deficiency makes it difficult for existing methods to fully capture the influence of conformational synergy of ternary complexes on inducibility, which severely restricts the efficient prediction and application expansion of molecular gels in cross-mechanism and cross-protein contexts.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a multi-mechanism method for predicting the inducibility of molecular adhesives, aiming to improve the efficiency and applicability of molecular adhesive prediction.

[0006] To achieve the above objectives, this application proposes a multi-mechanism molecular adhesive inducibility prediction method, the method comprising: Obtain a molecular glue ternary complex dataset, which includes small molecule compound data, first protein data, second protein data, and corresponding inducible tag data; Multimodal feature extraction processing is performed on the small molecule compound data to obtain small molecule multimodal feature data; multimodal feature extraction processing is performed on the first protein data to obtain first protein multimodal feature data; multimodal feature extraction processing is performed on the second protein data to obtain second protein multimodal feature data; The small molecule multimodal feature data is subjected to intra-molecule feature fusion processing to obtain small molecule fusion feature data; the first protein multimodal feature data is subjected to intra-protein feature fusion processing to obtain first protein fusion feature data; the second protein multimodal feature data is subjected to intra-protein feature fusion processing to obtain second protein fusion feature data. The small molecule fusion feature data, the first protein fusion feature data, and the second protein fusion feature data are subjected to ternary cross-entity feature interaction processing to obtain small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data. Based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data, feature construction and classification are performed to obtain the predictive probability data of the inducibility of the ternary complex. Based on the inducible label data and the inducible prediction probability data of the ternary complex, a total loss function data is constructed, and the prediction model is trained using the total loss function data to obtain a trained prediction model. The data of the small molecule compound to be tested, the data of the first protein to be tested, and the data of the second protein to be tested are input into the trained prediction model for prediction processing to obtain the inducibility prediction results of the ternary system to be tested.

[0007] In one embodiment, the step of obtaining a dataset of molecular glue ternary complexes includes: Experimentally validated ternary complex data were obtained from pre-defined publicly available literature as a positive sample dataset; Perform the following operations on the positive sample dataset: replace the first protein data while keeping the small molecule compound data unchanged, replace the second protein data while keeping the small molecule compound data unchanged, or replace both the first protein data and the second protein data simultaneously to obtain a negative sample dataset. The positive sample dataset and the negative sample dataset are combined in a preset ratio to obtain a class-balanced molecular glue ternary complex dataset.

[0008] In one embodiment, the step of performing multimodal feature extraction processing on the small molecule compound data to obtain small molecule multimodal feature data includes: The small molecule compound data is subjected to one-dimensional sequence encoding to obtain small molecule one-dimensional sequence feature data; The small molecule compound data is subjected to two-dimensional graph structure encoding processing to obtain small molecule two-dimensional graph feature data; The small molecule compound data is subjected to three-dimensional structure encoding processing to obtain small molecule three-dimensional structure feature data; The small molecule multimodal feature data includes the small molecule one-dimensional sequence feature data, the small molecule two-dimensional graph feature data, and the small molecule three-dimensional structure feature data.

[0009] In one embodiment, the step of performing intra-molecule feature fusion processing on the small molecule multimodal feature data to obtain small molecule fused feature data includes: The small molecule multimodal feature data is input into the feature remodeling unit for remodeling processing to obtain small molecule remodeling feature data; Assign trainable weight data to different modal features in the small molecule reforming feature data; The weighted small molecule reforming feature data are summed and fused to obtain the initial fusion feature data of the small molecules. The initial fusion feature data of the small molecules are subjected to feature optimization processing to obtain the fusion feature data of the small molecules.

[0010] In one embodiment, the step of performing ternary cross-entity feature interaction processing on the small molecule fusion feature data to obtain small molecule interaction feature data includes: Using the small molecule fusion feature data as query data and the first protein fusion feature data and the second protein fusion feature data as reference data, interactive computation processing is performed to obtain small molecule intermediate representation data. The small molecule intermediate representation data and the small molecule fusion feature data are combined to obtain small molecule combined feature data. The small molecule combination feature data are subjected to nonlinear transformation processing to obtain the small molecule interaction feature data.

[0011] In one embodiment, the interactive computing process includes: The query data is projected onto the query subspace to obtain the projected query data. The reference data is projected onto a reference subspace to obtain projected reference data; Calculate the similarity between the projection query data and the projection reference data to obtain the original association data; The original associated data is normalized to obtain associated weight data; The reference data is weighted and combined based on the associated weight data to obtain the intermediate representation data of the small molecule.

[0012] In one embodiment, the step of performing feature construction and classification processing based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data to obtain the predictive probability data of the inducibility of the ternary complex includes: Calculate the similarity between the small molecule interaction feature data and the first protein interaction feature data to obtain the first similarity feature data; Calculate the similarity between the small molecule interaction feature data and the second protein interaction feature data to obtain the second similarity feature data; The small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data are spliced ​​together to obtain ternary combination feature data. The first similarity feature data, the second similarity feature data, and the ternary combination feature data are input into the classification unit for processing to obtain the predictability probability data of the ternary complex.

[0013] In one embodiment, the step of constructing total loss function data based on the inducibility label data and the ternary complex inducibility prediction probability data includes: Prediction error data is calculated based on the inducibility tag data and the inducibility prediction probability data of the ternary complex; Calculate structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data; The prediction error data and the structure regularization term data are weighted and combined to obtain the total loss function data.

[0014] In one embodiment, the step of calculating structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data includes: The small molecule interaction feature data of the same ternary complex and the first protein interaction feature data are used as positive correlation pairs; The entity interaction feature data of different ternary complexes are used as negative correlation pair data; Calculate the correlation strength data of the positively correlated pairs; The regularization value data is calculated based on the correlation strength data to obtain the structure regularization term data.

[0015] In one embodiment, the method further includes: The small molecule compound data is subjected to sequence random masking to obtain enhanced small molecule compound data; Structural perturbation processing is performed on the first protein data and the second protein data to obtain enhanced first protein data and enhanced second protein data; The enhanced small molecule compound data is used to replace the original small molecule compound data, the enhanced first protein data is used to replace the original first protein data, and the enhanced second protein data is used to replace the original second protein data.

[0016] The multi-mechanism molecular gel inducibility prediction method proposed in this application achieves unified prediction of the inducibility of ternary complexes under different E3, substrate, and chemical spatial backgrounds by jointly encoding multimodal features such as one-dimensional sequences, two-dimensional graph structures, and three-dimensional structures of small molecules and two types of proteins, and by introducing a ternary cross-entity feature interaction mechanism. It can adapt to the prediction needs of different E3, different substrates, and diverse chemical spatial backgrounds, comprehensively capture the influence of conformational synergy of ternary complexes on inducibility, and improve the efficiency and applicability of molecular gel prediction. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an embodiment of the multi-mechanism molecular adhesive inducibility prediction method of this application; Figure 2 This is a schematic diagram of the overall structure provided in an embodiment of the multi-mechanism molecular adhesive inducibility prediction method of this application; Figure 3 This is the ROC curve on an external independent validation set provided by an embodiment of the multi-mechanism molecular adhesive inducibility prediction method of this application.

[0020] Explanation of icon numbers: 10. Memory; 20. Processor.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] With the development of artificial intelligence technology, computational methods have been gradually introduced into the field of molecular glue research. Existing work has attempted to predict E3 ligase-substrate-small molecule ternary complexes from a ternary structure perspective using isomorphic graph networks to assist in interface localization and conformational sampling. Frameworks integrating molecular dynamics, interface energy assessment, and deep learning have also been developed to narrow the search space for inducible protein pairs. Furthermore, proteochemometric models and generative design models targeting specific E3 pathways have shown promise in local systems. However, these computational methods often rely on specific E3-substrate combinations or limited protein sets for modeling, and their performance is often only effective within local protein spaces and specific chemical backbones, failing to adapt to prediction needs with different E3s, different substrates, and diverse chemical spatial backgrounds. Currently, there is a lack of a universal prediction model that can jointly encode multimodal features such as one-dimensional sequences, two-dimensional graph structures, and three-dimensional structures of small molecules and two types of proteins within a unified framework, and directly discriminate the inducibility of "compound-protein-protein" ternary systems. This deficiency makes it difficult for existing methods to fully capture the influence of conformational synergy of ternary complexes on inducibility, which severely restricts the efficient prediction and application expansion of molecular gels in cross-mechanism and cross-protein contexts.

[0025] Based on this, embodiments of this application provide a method for predicting the inducibility of multi-mechanism molecular adhesives, see [link to relevant documentation]. Figure 2 The diagram shown is an overall structural schematic of an embodiment of the multi-mechanism molecular adhesive inducibility prediction method of this application. Wherein, Figure 2 A illustrates the multimodal input types for constructing the ternary system in this application: for the small molecule side, one-dimensional sequence modes (e.g., normalized SMILES sequences), two-dimensional graph modes (e.g., atomic-bond molecular graphs), and three-dimensional structure modes (e.g., molecular three-dimensional conformation / structure representation) are constructed respectively; for the protein side, one-dimensional sequence modes (e.g., standard amino acid sequences), two-dimensional graph modes (e.g., residue contact / interaction graphs based on three-dimensional structures), and three-dimensional structure modes (e.g., protein three-dimensional structure representation) are constructed respectively. Through the above methods, the ternary system of "small molecule—first protein—second protein" forms a multimodal data channel that can be processed in parallel at the input end.

[0026] Figure 2B illustrates the feature encoding methods for different modalities: In the one-dimensional sequence modality, small molecule sequences can be encoded by a pre-trained molecular sequence encoding model (e.g., Molformer) to obtain molecular sequence embeddings; protein sequences can be represented using one-hot encoding and input into a convolutional neural network text model (e.g., TextCNN) to obtain protein sequence embeddings. In the two-dimensional graph modality, small molecule molecular graphs and protein graphs can be message-passed and aggregated by a graph neural network model (e.g., a graph network based on GIN convolution) to obtain graph structure embeddings. In the three-dimensional structure modality, small molecule three-dimensional structures can be encoded by a three-dimensional structure pre-trained model (e.g., UniMol2) to obtain three-dimensional embeddings; protein three-dimensional structures can be encoded by a protein structure pre-trained model (e.g., ESM2) to obtain three-dimensional embeddings. The embeddings of the above modalities can be further aligned to a unified dimension through linear mapping for subsequent fusion and interaction.

[0027] Figure 2 C illustrates the overall prediction framework of this application: First, one-dimensional / two-dimensional / three-dimensional modal features are extracted from the small molecule, the first protein, and the second protein, respectively, and aligned to a unified dimension; then, in the intra-entity fusion stage, the multi-modal features of the same entity are weighted and fused to obtain the small molecule embedding, the first protein embedding, and the second protein embedding; in the cross-entity / cross-modal interactive fusion stage, through attention or equivalent information interaction mechanisms, the three entity representations of the ternary system can be explicitly modeled with collaborative dependencies; finally, the interactive representations are constructed as hybrid features and input into the discriminant module to output the inducibility prediction results of the ternary system under the target mechanism.

[0028] The following combination Figure 1 The flowchart shown further illustrates the specific steps of the multi-mechanism molecular adhesive inducibility prediction method. The method includes steps S100 to S700, wherein: Step S100: Obtain the molecular glue ternary complex dataset, which includes small molecule compound data, first protein data, second protein data and corresponding inducible tag data. Step S200: Perform multimodal feature extraction processing on the small molecule compound data to obtain small molecule multimodal feature data; perform multimodal feature extraction processing on the first protein data to obtain first protein multimodal feature data; perform multimodal feature extraction processing on the second protein data to obtain second protein multimodal feature data. Step S300: Perform intra-molecule feature fusion processing on the small molecule multimodal feature data to obtain small molecule fusion feature data; perform intra-molecule feature fusion processing on the first protein multimodal feature data to obtain first protein fusion feature data; perform intra-molecule feature fusion processing on the second protein multimodal feature data to obtain second protein fusion feature data. Step S400: Perform ternary cross-entity feature interaction processing on the small molecule fusion feature data, the first protein fusion feature data, and the second protein fusion feature data to obtain small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data. Step S500: Based on the small molecule interaction feature data, the first protein interaction feature data and the second protein interaction feature data, feature construction and classification processing are performed to obtain the predictive probability data of the inducibility of the ternary complex. Step S600: Construct total loss function data based on the inducible label data and the inducible prediction probability data of the ternary complex, and use the total loss function data to train the prediction model to obtain the trained prediction model. Step S700: Input the data of the small molecule compound to be tested, the data of the first protein to be tested, and the data of the second protein to be tested into the trained prediction model for prediction processing to obtain the inducibility prediction results data of the ternary system to be tested.

[0029] In this embodiment, the molecular gel ternary complex dataset is a collection containing small molecule compounds, a first protein, a second protein, and their corresponding inducible tags. It primarily serves as the foundational data for model training and prediction. Specifically, this dataset can be constructed by selecting known ternary complex experimental data from publicly available databases, such as integrating protein structure information from the PDB database and small molecule activity data from the ChEMBL database. Furthermore, to ensure dataset diversity, additional ternary complex data can be supplemented through artificial synthesis or virtual generation. Multimodal feature extraction processing refers to the process of extracting multiple types of features from the small molecule compound data, the first protein data, and the second protein data. Specifically, for small molecule compound data, feature extraction can be achieved through methods such as chemical fingerprinting, molecular descriptor calculation, or molecular field analysis; for the first protein data and the second protein data, features can be obtained through amino acid sequence analysis, secondary structure prediction, or contact map generation. The choice of these feature extraction methods depends primarily on the specific format of the data and the requirements of subsequent processing.

[0030] In this embodiment, the intra-entity feature fusion processing of small molecules is a process of integrating multimodal feature data of small molecules to form a unified representation. For example, principal component analysis (PCA) or linear discriminant analysis (LDA) can be used to reduce the dimensionality and fuse multimodal features; alternatively, an attention-based approach can be used to achieve weighted fusion by assigning weights to different modal features. Similarly, the intra-entity feature fusion processing of the first and second proteins can also employ the above methods or other parallel processing techniques to ensure that the fused features fully reflect the information of the original data. Ternary cross-entity feature interaction processing refers to the process of establishing associations between small molecules, the first protein, and the second protein. For example, cosine similarity or Euclidean distance can be used to measure the feature correlation between different entities; alternatively, graph neural networks (GNNs) or attention mechanisms can be used to explicitly model higher-order dependencies in the ternary system. The main purpose of this processing method is to capture the influence of conformational synergy of ternary complexes on inducibility.

[0031] In this embodiment, feature construction and classification processing can be understood as the process of generating discriminative features for classification based on interactive feature data. For example, multiple feature vectors can be combined into a single feature vector through concatenation; alternatively, kernel function mapping can be used to transform the original features into a high-dimensional space to enhance classification capabilities. Subsequently, classifiers such as Support Vector Machines (SVM), Random Forests, or Deep Neural Networks can be used to achieve the final classification task. The process of constructing the total loss function data can be understood as the process of combining prediction error and regularization terms to optimize model performance. For example, the prediction error can be calculated using the mean squared error (MSE) or cross-entropy loss function; simultaneously, structural constraints can be introduced through L1 or L2 regularization to prevent model overfitting.

[0032] In this embodiment, this application constructs a unified deep learning framework to jointly process multimodal information from small molecules and two types of proteins, thereby achieving general prediction of the inducibility of ternary complexes. Compared to the limitations of existing technologies that only model single proteins or local pathways, this application ensures the comprehensiveness of input information through multimodal feature extraction and intra-entity fusion processing, explicitly models the dynamics of the ternary system through cross-entity feature interaction processing, and achieves end-to-end prediction through feature construction and classification processing, thus solving the problem of insufficient model generalization ability in existing technologies.

[0033] In this embodiment, the working principle of this application embodiment is as follows: First, a molecular glue ternary complex dataset is acquired, including small molecule compound data, first protein data, second protein data, and corresponding inducible label data, to provide basic data support for subsequent model training and prediction. Further, multimodal feature extraction processing is performed on the small molecule compound data, first protein data, and second protein data respectively to comprehensively capture the chemical and biological characteristics of each entity. Specifically, the small molecule compound data is processed through one-dimensional sequence encoding, two-dimensional graph structure encoding, and three-dimensional structure encoding to generate small molecule multimodal feature data; similarly, the first protein data and second protein data are also processed through multimodal feature extraction to generate corresponding first protein multimodal feature data and second protein multimodal feature data.

[0034] Next, the multimodal feature data of small molecules undergoes intra-entity feature fusion processing to integrate information from different modalities within the molecule, forming robust fused feature data of small molecules. Similarly, the multimodal feature data of the first and second proteins also undergo intra-entity feature fusion processing to generate fused feature data of the first and second proteins, respectively. Thus, the multimodal information of each entity is adaptively integrated, eliminating the limitations of a single modality and providing high-quality feature input for subsequent cross-entity interactions.

[0035] Then, the small molecule fusion feature data, the first protein fusion feature data, and the second protein fusion feature data are subjected to ternary cross-entity feature interaction processing to explicitly model the high-order dependencies between the small molecule and the two types of proteins, thereby capturing the key influence of ternary complex conformational synergy on inducibility. In this process, the small molecule fusion feature data is used as query data, and the first protein fusion feature data and the second protein fusion feature data are used as reference data. Interactive computation processing is used to generate small molecule intermediate representation data, which is then combined with the small molecule fusion feature data and subjected to nonlinear transformation to finally output the small molecule interaction feature data; similarly, the first protein interaction feature data and the second protein interaction feature data are generated in the same way.

[0036] Based on the aforementioned interaction feature data, further feature construction and classification processing are performed. Specifically, by calculating the similarity between the small molecule interaction feature data and the first protein interaction feature data, and the similarity between the small molecule interaction feature data and the second protein interaction feature data, first similarity feature data and second similarity feature data are generated. Simultaneously, the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data are concatenated to generate ternary combination feature data. These feature data are then input into the classification unit for processing, directly outputting the predictive probability data of the ternary complex inducibility, achieving accurate discrimination of the ternary system.

[0037] Furthermore, a total loss function is constructed based on inducible label data and ternary complex inducible prediction probability data. This loss function is then used to train the prediction model, optimizing the model parameters to improve prediction accuracy. One possible implementation involves using a stochastic gradient descent-type optimization algorithm (e.g., the Adam optimizer) to train the model. Specifically, the batch size can be set to 16, the total number of training epochs to 50, and the initial learning rate to 1×10⁻⁴. If the validation set performance does not significantly improve after several consecutive training epochs, an early stopping strategy can be used to terminate training, for example, setting the early stopping patience value to 15 epochs. Additionally, the weight coefficient for the contrastive learning loss in the total loss function can be set to 0.1. All of the above hyperparameters can be adjusted according to different data scales and application scenarios, and do not limit the scope of protection of this application.

[0038] Finally, the data of the small molecule compound to be tested, the data of the first protein to be tested, and the data of the second protein to be tested are input into the trained prediction model to quickly generate the inducibility prediction results of the ternary system to be tested, thereby supporting efficient screening. That is, after the model training is completed, the user inputs any combination of "small molecule-first protein-second protein" to be tested, and constructs a multimodal input in the same way as in the training stage. The input is then passed through the multimodal feature extraction module, the intra-entity multimodal fusion module, the ternary cross-modal feature fusion module, and the discriminant head module to obtain the inducibility prediction probability of the ternary system under a given mechanism background. In terms of application, the method of this application can be used to conduct virtual screening of a large number of candidate small molecules and protein combinations, and prioritize the selection of ternary combinations with higher inducibility probabilities for experimental verification; it can also screen candidate protein 1 / protein 2 combinations one by one for a given molecular glue compound, predict the target proteins that it may recruit, and provide clues for mechanism research and off-target risk assessment; it can also pre-screen candidate small molecules and generate mechanism hypotheses under different mechanism backgrounds such as degradable molecular glue, non-degradable molecular glue, and induced self-association. The overall technical solution, through the coordinated operation of steps such as multimodal feature extraction, intra-entity feature fusion, ternary cross-entity feature interaction, and feature construction and classification processing, jointly encodes ternary entity information under a unified framework, solving the problems of insufficient model generalization ability and inability to adapt to diverse chemical and protein spaces in existing technologies.

[0039] In one feasible implementation, the steps of obtaining a molecular glue ternary complex dataset include: obtaining experimentally verified ternary complex data from a pre-defined public literature as a positive sample dataset; performing the following operations on the positive sample dataset: replacing the first protein data while keeping the small molecule compound data unchanged, replacing the second protein data while keeping the small molecule compound data unchanged, or simultaneously replacing the first protein data and the second protein data to obtain a negative sample dataset; and combining the positive sample dataset and the negative sample dataset in a pre-defined ratio to obtain a class-balanced molecular glue ternary complex dataset.

[0040] In this embodiment, the pre-defined public literature refers to scientific literature containing experimentally verified research on molecular glue ternary complexes. It can be journal articles, patent documents, or public database records, with the aim of ensuring the authenticity and reliability of positive sample data. In one possible implementation, this application first uses the "small molecule-first protein-second protein" ternary relationship as the basic recording unit, and manually selects experimentally verified positive samples of molecular glue ternary complexes by searching molecular glue-related literature and public databases. Each positive sample includes at least: (1) structural information of the small molecule compound, such as the normalized SMILES representation, two-dimensional structural formula, and necessary stereochemical information; (2) identity information of the first and second proteins, such as protein name, Uniprot number, gene name, or other standardized identifiers; (3) experimental evidence supporting the small molecule's ability to induce or stabilize the interaction between the two proteins under a specific mechanism, including but not limited to target protein degradation, enhanced natural interaction, or driving homologous association. The negative sample dataset refers to a set of ternary complex data that does not have inducing ability, generated by specific rules. Its generation method closely matches the molecular glue action mechanism, aiming to simulate non-inducible scenarios caused by protein mismatch. To construct negative samples and maintain class balance, in one possible way, negative samples are generated with each molecular glue compound as the center, while keeping the small molecule unchanged: (1) Replace only the first protein: replace the first protein with a first protein different from the original positive sample, keeping the small molecule and the second protein unchanged; (2) Replace only the original second protein: replace the original second protein with a different second protein, keeping the small molecule and the first protein unchanged; (3) Replace both the first and second proteins simultaneously: replace the original protein combination with both the first and second proteins.

[0041] The three strategies described above can be used for random sampling in roughly equal proportions, ensuring that the number of negative samples roughly matches the number of corresponding positive samples. All samples are deduplicated, and combinations proven by literature to still exhibit molecular glue induction are removed, resulting in the final ternary complex sample set. Subsequently, the dataset is divided into training and testing sets in an 8:2 ratio. When evaluating cold-start performance, three cold-start partitioning datasets are constructed based on this dataset: "new molecules only," "new proteins only," and "neither molecules nor proteins found." In this embodiment, the scheme first extracts experimentally validated ternary complex data from pre-defined publicly available literature as a positive sample dataset. This data typically includes structural information of small molecule compounds, identification information of the first and second proteins, and experimental evidence supporting the inducible effect. Subsequently, a negative sample dataset is generated by systematically replacing the positive sample dataset. This involves three strategies: replacing only the first protein, replacing only the second protein, or replacing both simultaneously. This ensures the negative samples have biological significance and avoids invalid interference that might result from random replacements. Based on this, the positive and negative sample datasets are combined in a pre-defined ratio to form a class-balanced dataset, eliminating model training bias caused by the scarcity of positive samples in the original data. This design not only improves the robustness of the prediction model in discriminating the inducibility of ternary complexes but also provides a high-quality data foundation for subsequent multimodal feature fusion and cross-entity interaction. It effectively addresses the core challenge of unbalanced training sample distribution in molecular glue prediction, significantly improving the prediction model's ability to balance inducible and non-induced modes during the learning process, thereby enhancing its performance in recognizing real molecular glue-induced events.

[0042] In one feasible implementation, the step of performing multimodal feature extraction processing on the small molecule compound data to obtain small molecule multimodal feature data includes: performing one-dimensional sequence encoding processing on the small molecule compound data to obtain small molecule one-dimensional sequence feature data; performing two-dimensional graph structure encoding processing on the small molecule compound data to obtain small molecule two-dimensional graph feature data; and performing three-dimensional structure encoding processing on the small molecule compound data to obtain small molecule three-dimensional structure feature data; wherein, the small molecule multimodal feature data includes the small molecule one-dimensional sequence feature data, the small molecule two-dimensional graph feature data, and the small molecule three-dimensional structure feature data.

[0043] In this embodiment, one-dimensional sequence encoding refers to the process of feature extraction based on the chemical sequence information of small molecules (such as the SMILES string), aiming to capture the linear arrangement of atoms and basic chemical bond characteristics. In practical applications, normalized one-dimensional sequence inputs can be generated through standardization processes (such as salt stripping, valence normalization, and aromaticity normalization), thereby providing a highly interpretable and computationally inexpensive sequence-level representation for subsequent modeling. Two-dimensional graph structure encoding can be understood as the process of transforming small molecules into a topological network of atomic nodes and chemical bond edges, aiming to accurately model the local chemical environment and bonding relationships of molecules. Specifically, graph neural networks (such as isomorphic networks GIN) can be used to aggregate neighbor information through message passing mechanisms, thereby compensating for the loop structures and branching topological defects that one-dimensional sequences cannot express. Three-dimensional structure encoding refers to the process of obtaining the spatial coordinate information of small molecules in the biological environment through molecular force fields or conformational sampling, aiming to reflect the three-dimensional shape and electronic distribution characteristics of molecules. Typically, three-dimensional structure pre-trained models (such as Uni-Mol2) can be used to encode the three-dimensional conformation of molecules, thereby supplementing the geometric constraint information missing in low-dimensional features.

[0044] In this embodiment, the above technical solution ensures the complementarity and integrity of different modal features by extracting multimodal features from small molecule compound data, including one-dimensional sequences, two-dimensional graph structures, and three-dimensional structures. First, the one-dimensional sequence feature extraction module encodes normalized SMILES sequences based on a small molecule language model with a Transformer architecture, generating molecular-level global sequence feature vectors. Second, the two-dimensional graph feature extraction module uses a graph isomorphic network (GIN) to encode the small molecule atom-bond molecular graph, aggregating neighbor information through message passing and nonlinear mapping, and summarizing node representations during the graph readout stage to obtain graph structure features of a unified dimension. Finally, the three-dimensional structure feature extraction module uses a pre-trained model to encode the low-energy conformations of small molecules, generating molecular-level three-dimensional embedding vectors. Based on this, each modal feature can be unified to a preset latent space dimension through a linear projection layer, and then processed by layer normalization and vector unitization to reduce the scale differences between modalities. The above technical solution systematically integrates full-spectrum structural information from sequence to space, providing rich and hierarchical inputs for subsequent intra-entity feature fusion, thereby enhancing the ability to characterize the dynamic behavior of small molecules in ternary complexes and supporting more accurate inducibility prediction.

[0045] In addition, the first protein data and the second protein data are subjected to multimodal feature extraction processing to obtain the first protein multimodal feature data. Constructing protein multimodal input (applicable to the first protein data and the second protein data) (1) One-dimensional sequence mode: the standard amino acid single letter sequence after removing the signal peptide and tag sequence is used as input; (2) Two-dimensional graph mode: based on the known three-dimensional structure or predicted three-dimensional structure of the protein, the residue-residue distance or contact relationship is calculated, the residue is regarded as a graph node, and when the distance or energy meets the preset threshold, an edge is established to obtain the residue contact graph; (3) Three-dimensional structure mode: the three-dimensional structure of the protein is input into the protein three-dimensional structure pre-training model, and after obtaining the residue-level three-dimensional embedding, the protein-level three-dimensional structure features are obtained through pooling operation. Thus, each ternary sample is represented at the input end as: the one-dimensional sequence, two-dimensional molecular graph and three-dimensional structure of the small molecule, and the one-dimensional sequence, two-dimensional residue contact graph and three-dimensional structure of the first protein and the second protein respectively.

[0046] In one possible implementation, the protein multimodal feature extraction module includes a one-dimensional sequence encoding submodule, a two-dimensional graph encoding submodule, and a three-dimensional structure encoding submodule, which respectively process the above three types of modal inputs: (1) One-dimensional sequence feature extraction: For small molecules, a pre-trained small molecule language model based on Transformer is used to encode the normalized SMILES sequence to obtain a molecular-level global sequence feature vector; For proteins, a text convolutional neural network (TextCNN) structure is used, with the one-hot embedding representation of the amino acid sequence as input, and various convolutional kernel sizes are used to extract local n-gram functional fragments, which are then mapped to a unified dimension through pooling and fully connected layers to obtain protein-level one-dimensional sequence features.

[0047] (2) Two-dimensional graph feature extraction: The molecular graph of small molecule atoms-bonds and the contact graph of protein residues are encoded by Graph Isomorphism Network (GIN). GIN aggregates neighbor information through message passing and nonlinear mapping, and performs global summarization of node representations in the graph readout stage to obtain graph structure features of a unified dimension.

[0048] (3) Three-dimensional structural feature extraction: For small molecules, the three-dimensional conformation of the molecules is encoded using the three-dimensional structure pre-training model Uni-Mol2 to obtain molecular-level three-dimensional embedding vectors; for proteins, the three-dimensional structure of the proteins is encoded using the structure pre-training model ESM2 to obtain residue-level three-dimensional embeddings first, and then protein-level three-dimensional structural features are obtained through pooling. To facilitate the fusion between different modalities, the above one-dimensional sequence features, two-dimensional graph features and three-dimensional structural features can be unified to the preset latent space dimension through linear projection layers, and then processed by layer normalization and vector unitization to reduce the scale difference between modalities.

[0049] In one feasible implementation, the step of performing intra-molecule feature fusion processing on the small molecule multimodal feature data to obtain small molecule fused feature data includes: inputting the small molecule multimodal feature data into a feature remodeling unit for remodeling processing to obtain small molecule remodeled feature data; assigning trainable weight data to different modal features in the small molecule remodeled feature data; performing summation and fusion processing on the weighted small molecule remodeled feature data to obtain small molecule initial fused feature data; and performing feature optimization processing on the small molecule initial fused feature data to obtain the small molecule fused feature data.

[0050] In this embodiment, the feature remodeling unit is a functional module used to standardize heterogeneous features. It can be implemented using multilayer perceptrons, normalization layers, or self-attention mechanisms. Its purpose is to eliminate representational conflicts caused by differences in the source and scale of different modal features, enabling subsequent operations on each modality within a unified space. The trainable weight data is a set of parameters that dynamically adjusts the contribution of each modality. The weights can be normalized using a softmax function or optimized using a gradient descent algorithm to adapt to specific task requirements. Its purpose is to automatically quantify the importance of each modality based on the specific characteristics of the small molecule, thereby avoiding the limitations of fixed weights. In practical applications, feature optimization can be understood as a nonlinear transformation or dimensionality reduction operation, which can be implemented using activation functions, fully connected layers, or principal component analysis. Its purpose is to further refine the expressive power of the fused features, reduce redundant information, and enhance feature discriminability.

[0051] In this embodiment, the above-mentioned technical solution significantly improves the coordination and discriminativeness of feature representation by optimizing the fusion process of multimodal features of small molecules in stages. First, the feature reshaping unit aligns and standardizes the differences in the original output dimensions and distributions of different modal features. This step ensures the comparability of heterogeneous features such as sequence encoding, graph structure, and three-dimensional conformation in a unified space, providing a basis for subsequent weighting. Second, by introducing a learnable dynamic weighting mechanism, the model can automatically quantify the importance of each modality based on the specific chemical characteristics of the current small molecule. For example, it prioritizes strengthening three-dimensional structural features in flexible molecules and emphasizes sequence features in rigid molecules, thereby overcoming the defect that fixed weights cannot adapt to diverse molecular scenarios. Next, by summing and fusing the weighted features, this scheme effectively suppresses noise interference from low-contribution modalities while retaining the synergistic effect of high-value features. The generated initial fused feature data of small molecules is more in line with the requirements of induced discrimination of ternary complexes. It is understandable that performing intra-protein feature fusion processing on the first protein multimodal feature data to obtain first protein fusion feature data, and performing intra-protein feature fusion processing on the second protein multimodal feature data to obtain second protein fusion feature data, can also be performed using a similar procedure. Specifically, for the first protein multimodal feature data, its one-dimensional sequence, two-dimensional residue contact map, and three-dimensional structural features are first standardized using a feature remodeling unit to eliminate representational conflicts between different modalities. Subsequently, trainable weight data is assigned to each modal feature. These weights can be dynamically adjusted according to the specific structural characteristics of the protein. For example, the importance of two-dimensional map features is enhanced in proteins rich in secondary structure, while the role of three-dimensional structural features is highlighted in proteins with complex three-dimensional conformations. Through weighted summation fusion processing, initial fusion feature data of the first protein is generated, and its expressive power is further refined through feature optimization processing to finally obtain the first protein fusion feature data. Similarly, the same process is used to process the second protein multimodal feature data to ensure that its fusion feature data can accurately reflect the dynamic behavior of the protein in the ternary complex. This series of operations not only improved the robustness of protein feature representation but also provided high-quality input for subsequent cross-entity interaction modeling, thereby enhancing the model's accuracy in predicting the inductive properties of molecular glue. Finally, feature optimization further refined the feature representation by eliminating redundant information through nonlinear transformations or dimensionality reduction operations, enhancing the compactness and discriminative power of the features and providing high-quality input for subsequent ternary cross-entity interactions.

[0052] In one feasible implementation, the step of performing ternary cross-entity feature interaction processing on the small molecule fusion feature data to obtain small molecule interaction feature data includes: using the small molecule fusion feature data as query data and the first protein fusion feature data and the second protein fusion feature data as reference data, performing interactive calculation processing to obtain small molecule intermediate representation data; performing feature combination processing on the small molecule intermediate representation data and the small molecule fusion feature data to obtain small molecule combined feature data; and performing nonlinear transformation processing on the small molecule combined feature data to obtain the small molecule interaction feature data.

[0053] In this embodiment, query data refers to data that acts as the active party in the feature interaction process. This can be implemented using small molecule fusion feature data, aiming to simulate the biological mechanism of small molecules as the inducing core. Reference data refers to data that acts as the passive party in the feature interaction process. This can be implemented using first protein fusion feature data and second protein fusion feature data, aiming to provide key regional information of protein interfaces to support dynamic dependency modeling. Interactive computation processing can be understood as a feature extraction method based on attention mechanisms or similarity calculations. It can be implemented through multi-head cross-attention mechanisms, dot product similarity calculations, etc., aiming to capture key interface information in the conformational synergy of ternary complexes. Feature combination processing refers to the technical means of integrating features from different sources. It can be implemented using weighted summation, splicing fusion, etc., aiming to preserve the structural integrity of the original features and enhance semantic expressive power. Nonlinear transformation processing can be understood as a method to improve feature expressive power through activation functions or complex mappings. It can be implemented through structures such as multilayer perceptrons and convolutional neural networks, aiming to mine complex nonlinear relationships between features.

[0054] In this embodiment, the scheme performs interactive computation processing using small molecule fusion feature data as query data and two types of protein fusion feature data as reference data. This allows small molecules to actively focus on key regions of the protein interface, thereby accurately extracting dynamic correlation information related to conformational synergy. Based on this, the small molecule intermediate representation data and small molecule fusion feature data are combined, preserving the structural integrity of the original features while incorporating contextual enhancement information generated during the interaction process. This avoids semantic loss due to single feature transmission and significantly improves the robustness of feature expression. Simultaneously, by performing nonlinear transformation processing on the small molecule combined feature data, the complex nonlinear relationships between features are further explored, strengthening the high-order characterization ability of the ternary complex interface stability. The above process closely aligns with the biological essence of molecular glue interactions, effectively solving the problem of insufficient modeling of high-order dependencies between small molecules and two types of proteins, improving the reliability of ternary system predictions, achieving refined modeling of the dynamic dependencies between small molecules as the inducing core and two types of proteins, overcoming the problem of overly general feature interaction processes, and providing a more discriminative feature basis for subsequent inducibility discrimination.

[0055] Furthermore, for both the first and second protein fusion feature data, ternary cross-entity feature interaction processing can also be performed to obtain their respective interaction feature data. Specifically, using the first protein fusion feature data as query data and the small molecule fusion feature data and the second protein fusion feature data as reference data, interactive computation processing is performed to obtain the intermediate representation data of the first protein. This step simulates the interaction mechanism between the first protein and the small molecule and the second protein in the ternary complex, aiming to capture its key interface regions and dynamic dependencies. Subsequently, the intermediate representation data of the first protein and the first protein fusion feature data are combined using feature combination processing. Through weighted summation or splicing fusion, the original feature structure is preserved while incorporating interactive context information to generate the first protein combined feature data. This process enhances the semantic expressive power of the features and avoids information loss. Next, the first protein combined feature data undergoes nonlinear transformation processing. Activation functions or complex mapping structures are used to mine the nonlinear relationships between features, improving the high-order representation ability of the stability of the ternary complex interface. Similarly, using the second protein fusion feature data as query data and the small molecule fusion feature data and the first protein fusion feature data as reference data, similar interactive calculations, feature combinations, and nonlinear transformations are performed to obtain the second protein interactive feature data. This series of operations not only achieves refined modeling of the dynamic behavior of the first and second proteins in the ternary complex, but also provides comprehensive and discriminative feature inputs for subsequent inducibility discrimination, further improving the accuracy and reliability of molecular glue inducibility prediction.

[0056] In one feasible implementation, the interactive computation process includes: projecting the query data to a query subspace to obtain projected query data; projecting the reference data to a reference subspace to obtain projected reference data; calculating the similarity data between the projected query data and the projected reference data to obtain original association data; normalizing the original association data to obtain association weight data; and performing weighted combination processing on the reference data based on the association weight data to obtain the small molecule intermediate representation data.

[0057] In this embodiment, query data refers to the target data that requires feature matching or association analysis, which can be implemented using small molecule fusion feature data. Reference data refers to the data source used to provide comparative or supplementary information, which can be implemented using first protein fusion feature data and second protein fusion feature data. Projected query data is a representation of the original query data transformed into a specific subspace through linear transformation or nonlinear mapping, aiming to eliminate redundancy and noise interference between feature dimensions. Projected reference data is the result of spatial transformation of reference data in a similar manner, aiming to establish a unified comparison benchmark. Original association data is a quantitative indicator measuring the similarity between the projected query data and the reference data, which can be calculated using methods such as cosine similarity or Euclidean distance. Association weight data is the result of standardizing the original association data, aiming to balance the importance of different feature dimensions.

[0058] In this embodiment, the scheme first normalizes the feature space by projecting the query data and reference data into their respective subspaces. This approach effectively addresses the inconsistency in the original feature dimensions, providing a clear foundation for subsequent similarity calculations. Subsequently, by calculating the similarity between the projected query and reference data, original association data reflecting the strength of their association is obtained. This process not only simplifies computational complexity but also delves deeper into the nonlinear higher-order dependencies between the query and reference. Based on this, normalization of the original association data ensures the stability and generalization ability of the association weights, avoiding the influence of extreme values ​​on the calculation results. Finally, the reference data is weighted and combined based on the association weight data to generate small-molecule intermediate representation data that accurately reflects the interaction relationships between the ternary entities.

[0059] The above technical solution effectively captures high-order dependencies between ternary entities. Based on the acquisition of small molecule fusion feature data, first protein fusion feature data, and second protein fusion feature data, a systematic approach involving projection, similarity calculation, normalization, and weighted combination significantly improves the accuracy of feature interactions. This processing method is particularly suitable for predicting the inducibility of "compound-protein-protein" ternary systems, providing high-quality input data for subsequent feature combination and prediction.

[0060] In one feasible implementation, the step of performing feature construction and classification processing based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data to obtain the predictive probability data of the inducibility of the ternary complex includes: calculating the similarity between the small molecule interaction feature data and the first protein interaction feature data to obtain first similarity feature data; calculating the similarity between the small molecule interaction feature data and the second protein interaction feature data to obtain second similarity feature data; concatenating the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data to obtain ternary combination feature data; and inputting the first similarity feature data, the second similarity feature data, and the ternary combination feature data into a classification unit for processing to obtain the predictive probability data of the inducibility of the ternary complex.

[0061] In this embodiment, similarity refers to a technical indicator that reflects the strength of association by quantifying the proximity between two feature vectors. It can be implemented using methods such as cosine similarity, Euclidean distance, or Manhattan distance, aiming to explicitly model the interaction between small molecules and two types of proteins in terms of interface stability. Ternary combination feature data refers to a high-dimensional feature vector formed by concatenating the interaction features of multiple entities in a fixed order. It can be generated through direct concatenation, weighted fusion, or attention mechanisms, aiming to retain the complete interaction information of all entities to avoid the loss of key details due to feature compression. A classification unit is a module used to perform nonlinear transformations on the input features and output predicted probabilities. It can be implemented using structures such as multilayer perceptrons, support vector machines, or deep neural networks, aiming to integrate local similarity features with global combination features to improve prediction performance.

[0062] In this embodiment, the scheme calculates the similarity between small molecule interaction feature data and first protein interaction feature data to accurately quantify the correlation strength between the two in terms of interface stability, thereby highlighting the key role of conformational synergy in the ternary complex. Similarly, calculating the similarity between small molecule interaction feature data and second protein interaction feature data further captures the specific interactions between the small molecule and the second protein, ensuring that the interface relationships of the two types of proteins are independently and explicitly modeled. The small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data are concatenated, preserving the complete interaction information of all entities rather than simply fusing them, providing a comprehensive structured input for subsequent classification. Finally, the first similarity feature data, second similarity feature data, and ternary combination feature data are input into the classification unit for processing, enabling the classifier to simultaneously utilize interface stability criteria and overall structural information, thus taking into account both the microscopic mechanism and macroscopic manifestation of molecular gel induction.

[0063] The above technical solution effectively addresses the problem of insufficient conformational synergy modeling in the prediction of ternary complex inducibility, significantly improving the model's ability to capture protein proximity induction mechanisms and its predictive robustness. Furthermore, this solution, combined with the aforementioned steps of acquiring molecular glue ternary complex datasets, multimodal feature extraction, intra-entity feature fusion, and cross-entity feature interaction, forms a complete technical chain from data preprocessing to final prediction, further enhancing the model's generalization ability under different E3, different substrates, and diverse chemical spatial backgrounds.

[0064] In one feasible implementation, the step of constructing total loss function data based on the inducible tag data and the ternary complex inducible prediction probability data includes: calculating prediction error data based on the inducible tag data and the ternary complex inducible prediction probability data; calculating structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data; and performing weighted combination processing on the prediction error data and the structure regularization term data to obtain the total loss function data.

[0065] In this embodiment, prediction error data refers to the numerical value obtained by measuring the deviation between the predicted probability data of the ternary complex inducibility output by the quantization model and the experimentally verified true inducibility label data. Its purpose is to directly reflect the model's ability to fit known samples during training. In practical applications, prediction error data can be implemented in various ways. For example, a binary cross-entropy loss function can be used to measure the difference between the predicted and true values, or other forms of classification loss functions can be introduced to adapt to different task requirements.

[0066] In this embodiment, the structure regularization term data can be understood as a regularization value used to constrain the model learning process, aiming to guide the model to generate feature representations that conform to biophysical characteristics. Specifically, the calculation of the structure regularization term data depends on the higher-order dependencies between small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data. These dependencies can reflect the conformational coordination and interface stability of the ternary complex. In practical implementation, the structure regularization term data can be constructed using contrastive learning methods (such as InfoNCE loss) or similarity-based regularization strategies. Weighted combination processing is a technique that integrates prediction error data and structure regularization term data into a unified optimization objective. Its core lies in balancing the relationship between prediction accuracy and structural consistency by dynamically adjusting the weight coefficients of the two parts, thereby ensuring that the model is both faithful to the experimentally verified label information and able to capture the intrinsic structural regularity of the ternary complex during training.

[0067] In this embodiment, the above-mentioned technical solution solves the problem of insufficient generalization ability in model training by fusing prediction error data and structure regularization term data. First, prediction error data is calculated based on inducibility label data and ternary complex inducibility prediction probability data, providing a basic direction for model optimization. Second, by introducing structure regularization term data and utilizing the higher-order dependencies between small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data, regularization constraints on conformational synergy and interface stability of ternary complexes are constructed. This design not only avoids excessive divergence in the feature space but also significantly improves the model's prediction reliability on unknown samples.

[0068] Building upon this foundation, the optimization capability of the total loss function is further enhanced by weighted combination of prediction error data and structure regularization term data. This mechanism enables the model to maintain prediction accuracy while taking into account the structural correlation constraints of ternary complexes, thereby effectively mitigating overfitting and improving the model's generalization performance. Particularly in the scenario of predicting the inducibility of multi-mechanism molecular gels, this technical approach can better adapt to the complex modeling needs of different E3 ligases, different substrates, and diverse chemical spatial backgrounds.

[0069] In one feasible implementation, the step of calculating structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data includes: taking the small molecule interaction feature data and the first protein interaction feature data of the same ternary complex as positive correlation pairs; taking the entity interaction feature data of different ternary complexes as negative correlation pairs; calculating the association strength data of the positive correlation pairs; calculating the regularization value data based on the association strength data, and obtaining the structure regularization term data.

[0070] In this embodiment, positively correlated pairs refer to the interaction feature data of small molecules and the first protein within the same ternary complex that exhibit high correlation. This can be achieved by calculating cosine similarity or Euclidean distance. The purpose of this design is to leverage the intrinsic synergistic relationship between molecules and target proteins during the molecular glue induction process, ensuring that the model focuses on the high correlation of entity features within the same complex. Negatively correlated pairs can be understood as entity interaction feature data from different ternary complexes. Their purpose is to enhance the model's ability to distinguish unrelated entities by introducing cross-sample differential features as interference terms. This design effectively prevents excessive convergence of the feature space, thus avoiding the loss of discriminative information. Specifically, correlation strength data refers to the numerical value obtained by quantitatively evaluating the similarity of features between entities, which can be achieved using dot product operations or kernel function-based methods. The introduction of this feature provides a differentiable optimization basis for regularization, enabling the model to dynamically adjust feature embeddings to reflect real biological correlations.

[0071] In this embodiment, the technical solution optimizes the calculation of the structure regularization term by constructing a contrastive learning mechanism, thereby improving the discriminative power of feature representation and the model's generalization ability. In specific implementation, firstly, the small molecule interaction feature data and the first protein interaction feature data of the same ternary complex are used as positively correlated pairs. This fully utilizes the inherent synergistic relationship between molecules and target proteins during molecular glue induction, ensuring that the model can accurately capture the high correlation of entity features within the same complex. Simultaneously, by using entity interaction feature data of different ternary complexes as negatively correlated pairs, cross-sample differential features are introduced as interference terms, effectively strengthening the model's ability to distinguish unrelated entities.

[0072] Building upon this foundation, by calculating the association strength data of positively correlated pairs and quantitatively evaluating the feature similarity between entities, a differentiable optimization basis for regularization is provided. Finally, regularization values ​​are calculated based on the association strength data to obtain structural regularization terms. Through a contrastive loss mechanism guided by association strength, positive sample pairs are explicitly brought closer together and negative sample pairs are pushed further apart during training, thereby optimizing the topological structure of the feature space, improving the robustness of cross-sample representation, and enhancing the prediction accuracy of molecular glue inducibility. This technical solution not only solves the problem of the lack of a specific mechanism for calculating structural regularization terms but also effectively constructs a contrastive learning framework for positive and negative sample pairs, significantly improving the sufficiency of entity association representation in the feature embedding space, and thus enhancing the model's generalization ability to discriminate the inducibility of ternary complexes.

[0073] In one feasible implementation, the method further includes: performing sequence random masking on the small molecule compound data to obtain enhanced small molecule compound data; performing structural perturbation on the first protein data and the second protein data to obtain enhanced first protein data and enhanced second protein data; replacing the original small molecule compound data with the enhanced small molecule compound data, replacing the original first protein data with the enhanced first protein data, and replacing the original second protein data with the enhanced second protein data.

[0074] In this embodiment, sequence random masking refers to the operation of randomly selecting some chemical group information in the one-dimensional sequence features of small molecule compounds for masking or replacement. This can be achieved by setting a specific masking probability, such as using a random masking ratio of 5% to 20%, thereby simulating local deletions or modifications in real molecular structures. The purpose of this operation is to enable the model to learn more robust contextual dependencies during the feature extraction stage, thereby improving its adaptability to the diversity of small molecule skeletons. Structural perturbation processing can be understood as introducing controllable atomic coordinate shifts based on the three-dimensional structural features of proteins. Specifically, methods such as Gaussian noise perturbation and local rotation and translation can be used. The purpose is to simulate the natural conformational fluctuations and subtle structural differences of proteins in vivo, enabling the model to effectively capture the sensitive influence of interface stability on inducibility.

[0075] In this embodiment, the above-mentioned technical solution significantly alleviates the model generalization bottleneck caused by the scarcity of training data through data augmentation strategies. First, in the processing of small molecule compound data, a diverse range of small molecule representations is generated by performing a random masking operation on one-dimensional sequence features. This augmentation not only expands the distribution range of training samples but also encourages the model to focus on global contextual information during learning, rather than over-relying on local features. Second, in the processing of the first and second protein data, a protein representation with minor conformational changes is generated by introducing structural perturbation operations. This augmentation helps the model better understand the impact of conformational synergy of ternary complexes on inducibility during cross-entity feature interactions. Finally, the enhanced data is used to replace the original data for model training, enabling the total loss function to cover a wider range of potential structural combinations during the optimization process. This avoids overfitting of the model to specific patterns of limited experimental samples, thereby improving the extrapolation performance for unseen ternary systems while maintaining prediction accuracy. This not only solves the problem of insufficient model generalization ability caused by the limited scale of experimental validation data, but also achieves efficient discrimination of the inducibility of "compound-protein-protein" ternary systems within a unified framework, significantly improving the stability and adaptability of the prediction model in practical application scenarios.

[0076] In the embodiments of this application, the multi-mechanism molecular glue inducibility prediction method jointly encodes multimodal features such as one-dimensional sequences, two-dimensional graph structures, and three-dimensional structures of small molecules and two types of proteins, and introduces a ternary cross-entity feature interaction mechanism. This enables unified prediction of the inducibility of ternary complexes under different E3, substrate, and chemical spatial backgrounds. It can adapt to the prediction needs of different E3, different substrates, and diverse chemical spatial backgrounds, comprehensively capture the influence of ternary complex conformational synergy on inducibility, and improve the efficiency and applicability of molecular glue prediction.

[0077] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the multi-mechanism molecular adhesive inducibility prediction method of this application. Any simple modifications based on this technical concept are within the scope of protection of this application. It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0078] To further verify the effectiveness of the multi-mechanism molecular glue inducibility prediction method proposed in this application, this embodiment conducted a systematic experimental evaluation of the model after the above-mentioned basic model and training process were constructed. This included five-fold cross-validation performance evaluation, ablation experiments of key modules, and cold-start experiments under conditions of no physical entity or no combination. All experiments were conducted based on the aforementioned dataset and training strategy, aiming to verify the generalization ability, stability, and technical advantages of the proposed method in different application scenarios.

[0079] In one feasible implementation, to fully evaluate the classification performance of the model under the full configuration, five-fold cross-validation is performed on the constructed dataset to obtain Table 1.

[0080] Table 1. Statistical results of five-fold cross-validation performance. As shown in Table 1, the model in this application shows stability in multiple indicators, with excellent performance in AUC, F1, and ACC, indicating that the multi-mechanism molecular glue inducibility prediction method has strong predictive ability and generalization under conventional model evaluation conditions.

[0081] In one feasible implementation, to verify the independent contributions of each key module of this application (including multimodal feature input, dynamic modality weighting, ternary cross-entity feature interaction, self-supervised strategy, gated attention mechanism, hybrid feature construction strategy, etc.), a systematic ablation experiment was conducted to obtain Table 2.

[0082] Table 2 Comparison of ablation test performance results As shown in Table 2, the complete model of this application achieves the best performance across all metrics. The performance degradation is most pronounced after removing feature enhancements, indicating that surface multimodal enhancement plays a crucial role in the model's generalization ability. Removing the dynamic weight mechanism also significantly reduces performance, demonstrating that adaptive modeling of different modal contributions is essential for improving overall performance. Ablation analysis of modal features reveals that removing protein 3D features has the greatest impact, indicating that 3D structures play a core role in characterizing interface synergy. Simultaneously, the disappearance of the hybrid feature module leads to another significant performance degradation, further demonstrating that the innovative design of cross-modal feature construction in this application is a key driver for achieving high-performance prediction. Overall, the ablation experiments fully validate that each module of this application makes an irreplaceable contribution to the final performance, and the complete system best reflects the synergistic advantages of the ternary complex inducibility model.

[0083] In one feasible implementation, to verify the model's generalization ability under conditions of no compound, no protein, or completely new combination, three cold start validation sets were constructed, including three settings: new compound only (UM), new protein only (UP), and no compound and protein seen (AU), as shown in Table 3. Dataset ACC Prec Rec Spec F1 BACC AUC AP UM 0.873 0.857 0.857 0.885 0.857 0.871 0.896 0.877 UP 0.800 0.765 0.929 0.636 0.839 0.783 0.844 0.882 AU 0.750 0.750 1.000 0.600 0.750 0.600 1.000 1.000

[0084] Table 3 Extrapolation performance under cold start partitioning As shown in Table 3, the model of this application can effectively transfer to new small molecule chemical spaces and adapt to new protein structure backgrounds. Furthermore, in the AU scenario, the model's various indicators are good, showing that the model still has a certain predictive ability for unseen combinations. The above results fully demonstrate that the multi-mechanism molecular glue inducibility prediction method of this application exhibits good generalization performance under conditions across chemical spaces, protein spaces, and mechanisms, and can be applied to high-throughput screening of molecular glues, identification of unknown targets, and exploration of novel ternary systems in real-world scenarios.

[0085] In addition to the aforementioned internal cross-validation, ablation experiments, and cold start validation, to further verify the generalization ability of the method in this application on cross-source data, this embodiment also selects two sets of external independent validation sets for testing. See [link to documentation]. Figure 3 , Figure 3 The ROC curves of the proposed method on an external independent validation set are presented. The results show that even when the external data sources, protein class composition, and sample distribution differ from the training data, the proposed method can still maintain good discriminative ability and stability, thus demonstrating its good cross-source robustness and practical application value.

[0086] In summary, this application provides a ternary multimodal deep learning method, its apparatus, electronic device, and storage medium for predicting the inducibility of multi-mechanism molecular adhesives. This method uses a ternary system of "small molecule—first protein—second protein" as the basic modeling unit, comprehensively utilizing one-dimensional sequence features, two-dimensional graph structure features, and three-dimensional structural features. It achieves alignment and collaborative expression of different modal information through intra-entity multimodal fusion and ternary cross-entity feature interaction, and combines a joint optimization strategy of classification learning and self-supervised constraints to improve the model's robustness and generalization ability. Verified by various evaluation settings, the method of this application maintains stable predictive performance under scenarios with conventional data partitioning and distribution, and is suitable for applications such as high-throughput virtual screening of molecular adhesives, discovery of potential target combinations, and exploration of novel ternary systems. It should be understood that the above specific embodiments are merely illustrative of this application and not limiting. Various modifications, equivalent substitutions, or combinations made by those skilled in the art to the network structure, feature construction, fusion strategy, training parameters, and loss function without departing from the spirit and substance of this application should fall within the protection scope of this application.

Claims

1. A method for predicting the inducibility of multi-mechanism molecular adhesives, characterized in that, The method includes: Obtain a molecular glue ternary complex dataset, which includes small molecule compound data, first protein data, second protein data, and corresponding inducible tag data; Multimodal feature extraction processing is performed on the small molecule compound data to obtain small molecule multimodal feature data; multimodal feature extraction processing is performed on the first protein data to obtain first protein multimodal feature data; multimodal feature extraction processing is performed on the second protein data to obtain second protein multimodal feature data; The small molecule multimodal feature data is subjected to intra-molecule feature fusion processing to obtain small molecule fusion feature data; the first protein multimodal feature data is subjected to intra-protein feature fusion processing to obtain first protein fusion feature data; the second protein multimodal feature data is subjected to intra-protein feature fusion processing to obtain second protein fusion feature data. The small molecule fusion feature data, the first protein fusion feature data, and the second protein fusion feature data are subjected to ternary cross-entity feature interaction processing to obtain small molecule interaction feature data, first protein interaction feature data, and second protein interaction feature data. Based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data, feature construction and classification are performed to obtain the predictive probability data of the inducibility of the ternary complex. Based on the inducible label data and the inducible prediction probability data of the ternary complex, a total loss function data is constructed, and the prediction model is trained using the total loss function data to obtain a trained prediction model. The data of the small molecule compound to be tested, the data of the first protein to be tested, and the data of the second protein to be tested are input into the trained prediction model for prediction processing to obtain the inducibility prediction results of the ternary system to be tested.

2. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps to obtain a dataset of molecular glue ternary complexes include: Experimentally validated ternary complex data were obtained from pre-defined publicly available literature as a positive sample dataset; Perform the following operations on the positive sample dataset: replace the first protein data while keeping the small molecule compound data unchanged, replace the second protein data while keeping the small molecule compound data unchanged, or replace both the first protein data and the second protein data simultaneously to obtain a negative sample dataset. The positive sample dataset and the negative sample dataset are combined in a preset ratio to obtain a class-balanced molecular glue ternary complex dataset.

3. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps for performing multimodal feature extraction processing on the small molecule compound data to obtain small molecule multimodal feature data include: The small molecule compound data is subjected to one-dimensional sequence encoding to obtain small molecule one-dimensional sequence feature data; The small molecule compound data is subjected to two-dimensional graph structure encoding processing to obtain small molecule two-dimensional graph feature data; The small molecule compound data is subjected to three-dimensional structure encoding processing to obtain small molecule three-dimensional structure feature data; The small molecule multimodal feature data includes the small molecule one-dimensional sequence feature data, the small molecule two-dimensional graph feature data, and the small molecule three-dimensional structure feature data.

4. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps of performing intra-molecule feature fusion processing on the multimodal feature data of the small molecules to obtain fused feature data of the small molecules include: The small molecule multimodal feature data is input into the feature remodeling unit for remodeling processing to obtain small molecule remodeling feature data; Assign trainable weight data to different modal features in the small molecule reforming feature data; The weighted small molecule reforming feature data are summed and fused to obtain the initial fusion feature data of the small molecules. The initial fusion feature data of the small molecules are subjected to feature optimization processing to obtain the fusion feature data of the small molecules.

5. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps of performing ternary cross-entity feature interaction processing on the small molecule fusion feature data to obtain small molecule interaction feature data include: Using the small molecule fusion feature data as query data and the first protein fusion feature data and the second protein fusion feature data as reference data, interactive computation processing is performed to obtain small molecule intermediate representation data. The small molecule intermediate representation data and the small molecule fusion feature data are combined to obtain small molecule combined feature data. The small molecule combination feature data are subjected to nonlinear transformation processing to obtain the small molecule interaction feature data.

6. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 5, characterized in that, The interactive computing process includes: The query data is projected onto the query subspace to obtain the projected query data. The reference data is projected onto a reference subspace to obtain projected reference data; Calculate the similarity between the projection query data and the projection reference data to obtain the original association data; The original associated data is normalized to obtain associated weight data; The reference data is weighted and combined based on the associated weight data to obtain the intermediate representation data of the small molecule.

7. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps for feature construction and classification based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction data to obtain the predictive probability data of the inducibility of the ternary complex include: Calculate the similarity between the small molecule interaction feature data and the first protein interaction feature data to obtain the first similarity feature data; Calculate the similarity between the small molecule interaction feature data and the second protein interaction feature data to obtain the second similarity feature data; The small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data are spliced ​​together to obtain ternary combination feature data. The first similarity feature data, the second similarity feature data, and the ternary combination feature data are input into the classification unit for processing to obtain the predictability probability data of the ternary complex.

8. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The steps for constructing the total loss function data based on the inducible label data and the inducible prediction probability data of the ternary complex include: Prediction error data is calculated based on the inducibility tag data and the inducibility prediction probability data of the ternary complex; Calculate structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data; The prediction error data and the structure regularization term data are weighted and combined to obtain the total loss function data.

9. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 8, characterized in that, The steps for calculating the structure regularization term data based on the small molecule interaction feature data, the first protein interaction feature data, and the second protein interaction feature data include: The small molecule interaction feature data of the same ternary complex and the first protein interaction feature data are used as positive correlation pairs; The entity interaction feature data of different ternary complexes are used as negative correlation pair data; Calculate the correlation strength data of the positively correlated pairs; The regularization value data is calculated based on the correlation strength data to obtain the structure regularization term data.

10. The method for predicting the inducibility of multi-mechanism molecular adhesives as described in claim 1, characterized in that, The method further includes: The small molecule compound data is subjected to sequence random masking to obtain enhanced small molecule compound data; Structural perturbation processing is performed on the first protein data and the second protein data to obtain enhanced first protein data and enhanced second protein data; The enhanced small molecule compound data is used to replace the original small molecule compound data, the enhanced first protein data is used to replace the original first protein data, and the enhanced second protein data is used to replace the original second protein data.

Images