A method for screening small molecule polypeptide targeting based on deep learning

Abstract

The application belongs to the technical field of deep learning and computer-aided molecular screening, and particularly relates to a small molecule polypeptide targeting screening method based on deep learning. The sequence and candidate conformation of a small molecule polypeptide to be screened and the candidate conformation of a target are obtained, a target residue graph, a polypeptide sequence graph and a polypeptide conformation graph are constructed, a graph representation is extracted by using a deep learning coding network, conditional competition of the target conformation is completed by relying on the polypeptide graph representation, and a conformation fitting weight is obtained; the weight is cross-aligned with the target and polypeptide residue representations, a residue-peptide segment interaction tensor is constructed, and candidate action residues and action segments are screened out. Through counterfactual masking of the interaction residues and associated edges, in combination with interaction contribution differences, cross-conformation action consistency and the conformation fitting weight, accurate polypeptide targeting screening results are output. The application can reduce the dependence of screening results on a single target conformation, eliminate false high scores formed by accidental spatial contact, and improve the recognition reliability of induced fitting polypeptides.

Description

Technical Field

[0001] This invention belongs to the field of deep learning and computer-aided molecular screening technology, specifically a method for targeted screening of small molecule peptides based on deep learning. Background Technology

[0002] Small molecule peptide targeted screening typically uses known target structures and candidate peptide sequences as inputs, predicting the binding probability through computational models. Existing conventional methods often employ a single target crystal conformation or a fixed conformation obtained through energy optimization. The target protein is converted into a residue map or a 3D mesh, and the peptide sequence is converted into residue sequence features. Convolutional networks, recurrent networks, graph neural networks, or attention networks are then used to extract target and peptide features, respectively. Subsequently, the model performs concatenation, inner product, or cross-attention calculations on the two types of features to obtain a binding score between the candidate peptide and the target, and then ranks the candidates according to their scores. Some methods also use the distance, contact type, electrostatic relationship, and hydrophobic relationship generated by molecular docking as additional inputs to enhance spatial feature representation. The training objective of these methods is usually established based on known binding tags or affinity values, enabling the model to learn the mapping relationship between sequence composition, local structure, and overall physicochemical properties and binding results under fixed target conformation conditions. Because the screening process is mainly based on single-encoding and single-scoring, the target conformation remains unchanged after entering the model, and the induced changes in the target's local structure by the peptide are not included in a unified discrimination process.

[0003] For targets with multiple possible conformations, some existing technologies pre-generate several candidate conformations and perform feature encoding or molecular docking with the same peptide. The predicted scores for each conformation are then calculated by taking the maximum, average, or weighted sum. The weights are typically determined based on the target conformation's energy, frequency of occurrence, or manually set rules, lacking a direct correlation with the local interaction characteristics of the peptide being screened. Another approach inputs multiple conformations as parallel samples into the same deep learning model, which outputs multiple independent binding scores, which are then aggregated externally. This approach expands the target structure coverage, but different conformations do not form a competitive relationship around the current peptide, making it impossible to determine which target state is supported by specific peptide interaction conditions. For targets with significant changes in local pocket opening and closing, side chain rearrangement, or residue exposure, different conformations may generate conflicting contact relationships. Simply taking the maximum value may retain accidentally formed high-scoring contacts, while simply averaging may dilute the true interaction signal that only appears under reasonable induced conformations, making the screening results still affected by the number of conformations generated, the initial structure, and the aggregation rules.

[0004] Existing deep learning screening schemes also interpret prediction results through attention weights, feature attribution, or residue masking. A common practice is to read high-attention residues after the model completes binding scoring, or to sequentially delete a residue feature and observe score changes to identify potential key action sites. Some schemes further compare high-response regions in different peptide conformations or use distance constraints to screen out obviously unreasonable residue pairings. These interpretation processes are usually attached to the scoring model and are not trained together with target conformation selection, residue interaction modeling, and final sequencing. High-attention responses may originate from statistically relevant features such as residue class, sequence length, and hydrophobicity distribution. Masking a single node may simultaneously alter its neighborhood propagation results, causing score changes to not directly correspond to the interaction relationship between specific residues and specific peptides. If high-response positions in different conformations are compared only by number or distance, it is difficult to identify continuous action clusters formed by local displacements, and it is impossible to distinguish between stable cross-conformation interactions and isolated contacts appearing only in a single anomalous conformation. Therefore, the interpretation results are difficult to use for candidate screening.

[0005] The selection of target candidate conformations is separated from the discrimination of residue interactions of small molecule peptides. The model cannot determine the appropriate target state based on the current local structure and sequence characteristics of the peptide, nor can it verify whether high binding scores are indeed generated by stable interactions between specific residues and specific peptides. Fixed conformation models may miss binding relationships dependent on pocket opening and closing, side chain rearrangements, or changes in local residue exposure, while multi-conformation independent scoring models may retain high-scoring conformations corresponding to accidental spatial overlap. Post-scoring attention interpretation or single-point masking does not form a closed loop with conformational competition and cross-conformational correspondences, making it difficult to exclude non-specific high scores caused by overall hydrophobicity, amino acid composition, or local anomalous geometric relationships. Therefore, the screening results remain strongly dependent on the initial target conformation and the conformational aggregation method, creating an unstable discrimination boundary between induced-fit peptides and non-specific contact peptides, resulting in the omission of reasonable candidates or the inclusion of accidental contact candidates in the high-ranking region. Summary of the Invention

[0006] The purpose of this invention is to provide a small molecule peptide targeted screening method based on deep learning, which can solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for targeted screening of small molecule peptides includes: acquiring multiple candidate conformations of the target and the sequence and multiple candidate conformations of the small molecule peptide to be screened, and constructing a target residue map, a peptide sequence map and a peptide conformation map, respectively. The graph representations of each target conformation and each peptide conformation are extracted using a deep learning coding network. Based on the current peptide graph representation, conditional conformation competition is performed on multiple target conformations to obtain target conformation adaptation weights. Based on the adaptation weights, the target residue characterization and peptide residue characterization are cross-aligned to generate a residue-peptide interaction tensor. Candidate active residues and candidate active fragments are determined from the residue-peptide interaction tensor. Counterfactual masking is applied to the candidate active residues and their corresponding interaction edges, and the interaction contribution is recalculated. Based on the difference in interaction contributions before and after masking, the consistency of candidate interaction relationships in different conformational combinations, and the adaptation weights, the target screening results for small molecule peptides are generated.

[0008] Preferably, the target residue map uses amino acid residues as nodes, and classifies edges based on main chain connections, non-covalent contact relationships within a preset spatial adjacency range, and cooperative displacement relationships within the same local secondary structure. Each node is configured with residue category, solvent exposure level, local curvature, and side chain orientation characteristics. The polypeptide sequence map uses polypeptide residues as nodes and peptide bond connections and sequence spacing relationships to form classification edges; The polypeptide conformation diagram is based on the polypeptide sequence diagram with the addition of spatial constraint edges determined by atomic distance, relative residue orientation, and local dihedral angles. The deep learning coding network propagates node features along sequence topology and spatial topology respectively, and obtains a graph representation that combines sequence continuity and conformational geometric information by fusing the index correspondence of the same residue in the two types of topologies.

[0009] Preferably, the conditional conformational competition includes: constructing conformational pairing characterization for each target conformation and each peptide conformation, wherein the conformational pairing characterization is composed of target global map characterization, peptide global map characterization, characterization of locally high-response residues of both parties, and geometric matching description between conformations; Each conformation pairing representation is input into a competitive scoring network with shared parameters to obtain the corresponding initial conformation fitting values. Normalization competition is performed among all target conformations corresponding to the same small molecule polypeptide to be screened, and the initial conformation adaptation values ​​are redistributed according to the consistency of the adaptation ranking of each target conformation under different polypeptide conformations to obtain the target conformation adaptation weight. Conformation pairs with a relative support lower than that within the same conformational competition group are removed from subsequent cross-alignments.

[0010] Preferably, the residue-peptide interaction tensor is jointly encoded by the multi-head cross-attention result between the target residue characterization and the peptide residue characterization, the inter-residue distance binning characterization, the relative orientation characterization, and the adaptation weight of the conformation to which it belongs; Candidate action fragments are identified based on high-response regions that are repeated across attention heads in the interaction tensor and are continuously distributed between adjacent polypeptide residues, and candidate action residues are identified based on the set of target residues that form a bidirectional response with the candidate action fragments. For each candidate active residue, node feature masking, related interaction edge removal, and neighborhood information freezing are performed to obtain mutually independent counterfactual inputs; Each counterfactual input is re-entered into the same cross-alignment process to obtain the masked interaction contribution corresponding to each of the candidate action residues.

[0011] Preferably, the deep learning coding network includes a local residue coding layer and a conformational convergence layer. The local residue coding layer generates subgraph representations of the same residue in the main chain connection subgraph, spatial contact subgraph, and cooperative displacement subgraph, respectively, and determines the subgraph fusion coefficient based on the identifiable increment of the central residue representation by the neighborhood features in each subgraph. The conformational aggregation layer uses locally high-response residues as anchor points to aggregate residue representations within a preset topological step size into local patch representations, and then aggregates multiple local patch representations with background residue representations not covered by anchor points in a layered manner. For different candidate conformations of the same target, the background residue representation is constrained to remain aligned according to the residue number correspondence, and the local patch representation is allowed to form differentiated representations as the conformation changes.

[0012] Preferably, the geometric matching description of the conformation pairing characterization is generated in the following manner: locally high-response residues are selected from the target conformation and the peptide conformation respectively, and a rotation- and translation-invariant local coordinate description is established with each locally high-response residue as the center. The relative distance sequence, relative orientation sequence and accessibility overlap sequence between the target local coordinate description and the peptide local coordinate description are calculated, and the three types of sequences are mapped into pairing geometric vectors through a shared encoder. The competitive scoring network simultaneously outputs the initial conformational fit value and the conformational evidence vector. Before normalization competition, it merges target conformations supporting the same local region of action according to the complementarity between the conformational evidence vectors, and separates target conformations dependent on different local regions of action according to the degree of conflict between the conformational evidence vectors.

[0013] Preferably, after applying counterfactual masking to the candidate active residues, the method further includes: replacing the masked candidate active residues with neutral reference representations obtained by polymerizing their structural neighborhoods, while keeping the characterization of the remaining target residues and the peptide conformation unchanged, thus forming a substitutional counterfactual input; The alternative counterfactual input and the counterfactual input formed by node feature masking, related interaction edge removal and neighborhood information freezing are respectively fed into the cross-alignment process to obtain multiple counterfactual interaction contributions; The confidence level of the various counterfactual interaction contributions is assigned according to the perturbation magnitude of their interaction tensors in non-candidate regions, and the difference in interaction contributions after confidence weighting constitutes the causal contribution characterization of the candidate action residues. For multiple candidate action residues belonging to the same candidate action fragment, fragment-level causal contributions are generated based on the consistency of their causal contribution characterization directions.

[0014] Preferably, the consistency of the candidate interaction relationship in different conformational combinations is determined by the following method: mapping the residue-peptide interaction tensor corresponding to each conformational combination to a unified index plane according to the target residue number and peptide residue number; For each interaction position in the unified index plane, extract its interaction contribution, local geometric relationship and the adaptation weight of its respective configuration under different configuration combinations to form a cross-configuration interaction trajectory; Multiple interaction sites that have undergone continuous spatial migration but whose residue indices remain adjacent are merged into the same action cluster, and a corresponding consistency characterization is generated based on the directional consistency of interaction contributions within the action cluster, the variable range of geometric relationships, and the joint distribution of adaptation weights. Interaction positions that occur only in a single conformational combination and cannot be incorporated into any action cluster are marked as isolated interaction positions.

[0015] Preferably, establishing a bidirectional constraint relationship between the local region representation and the conformational evidence vector includes: determining the evidence coverage region based on the response distribution of the conformational evidence vector in the target residue dimension, and feeding back the local region representation corresponding to the evidence coverage region to the competitive scoring network to update the initial conformational fit values ​​of each target conformation. The locally high-response residues are re-determined based on the updated conformational adaptation initial values, and a residue migration mapping is established between the re-determined locally high-response residues and the original locally high-response residues. When residue migration maps remain within the same local region or adjacent local regions, the corresponding conformation pairings are preserved. When residue migration maps cross non-adjacent local regions, the corresponding conformation pairings are split into multiple local evidence pairs and each participates in the normalization competition.

[0016] Preferably, the joint verification of the fragment-level causal contribution and the corresponding consistency representation includes: establishing a fragment evidence vector for each candidate action fragment, which consists of the fragment-level causal contribution, the cross-conformational interaction trajectory of the action cluster to which it belongs, and the proportion of isolated interaction positions; According to the index correspondence of the same candidate action fragment in different target conformations, the evidence vectors of multiple fragments are aggregated, and the aggregation result is multiplied by the corresponding target conformation adaptation weight to obtain the joint contribution of the fragments. When the candidate action fragment corresponds to multiple mutually separate action clusters, the joint contribution of each action cluster is calculated, and merging or independent retention is determined based on whether the action clusters share the candidate action residues. The combined contribution of all independently retained fragments and conformational fit weights are summarized to generate targeted screening results for small molecule peptides.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A target residue map, peptide sequence map, and peptide conformation map are constructed using multiple target candidate conformations and multiple peptide candidate conformations. After forming corresponding residue representations through a deep learning encoding network, conditional conformation competition reallocates adaptation weights to each target conformation based on the current peptide representation, so that subsequent interaction calculations focus on the target state supported by the current peptide features. The residue-peptide interaction tensor encodes the adaptation weights together with the distance between residues, relative orientation, and bidirectional response, which can determine continuous candidate action fragments and corresponding target residues within the conformation selection results. Node feature masking, interaction edge removal, neighborhood information freezing, and neutral reference replacement are performed on candidate action residues. Confidence is assigned based on the perturbations caused to non-candidate regions by various counterfactual inputs, which can limit the score changes to the correspondence between candidate residues and candidate peptides. Cross-conformation interaction trajectories then merge interaction positions with adjacent indices and continuous spatial migration into action clusters, and exclude isolated interaction positions that cannot be merged into action clusters. This forms a continuous discriminant chain of conformational adaptation, action localization, counterfactual verification, and cross-conformational correspondence, reducing the dependence of the screening results on a single initial conformation and external aggregation rules.

[0018] 2. Target residue maps propagate node features according to main chain connections, spatial contacts, and cooperative displacements. Peptide sequence maps and peptide conformation maps fuse sequence continuity and spatial geometry information through residue index correspondence, ensuring that the representation of the same residue in different topologies remains consistent. Local residue encoding and conformational hierarchical aggregation organize high-response residues into local patches, allowing for differential representations of local patches in different target conformations, while constraining background residues to maintain alignment. After establishing feedback constraints between the conformational evidence vector and the local patch representation, conformational adaptation weights can be recalculated based on the evidence coverage patch. Residue migrations across non-adjacent patches are split into independent local evidence competing for space, reducing the possibility of different action regions being merged into a single high score. Fragment-level causal contribution and the consistency of action cluster correspondence further form fragment evidence vectors, which are then aggregated into screening results according to target conformational adaptation weights. This ensures that the ranking of the same peptide simultaneously includes local action source, conformational support, and cross-conformational correspondence, facilitating the differentiation of stable action fragments, conformation-dependent action fragments, and isolated high-score contacts. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating the overall process of a deep learning-based small molecule peptide targeting screening method. Figure 2 This is a profile of the fusion coefficients for different types of edges; Figure 3 A flowchart for target conformation conditional competition and local evidence feedback process; Figure 4 For conformational pairing support topographic maps; Figure 5 A flowchart for counterfactual verification of candidate functional residues and summary of cross-conformation consistency. Figure 6 Input a reliable weight profile for counterfactual data; Figure 7 The bar chart shows the contribution of counterfactual causality. Figure 8 For the consistency trajectory diagram corresponding to the action cluster; Figure 9 The final filtered values ​​form a graph; Figure 10 This is a graph showing the loss curve during joint training. Detailed Implementation

[0020] In one embodiment, please refer to Figure 1 This paper provides a deep learning-based method for targeted screening of small molecule peptides, comprising: acquiring multiple candidate conformations of a target and the sequence and multiple candidate conformations of the small molecule peptide to be screened, and constructing a target residue map, a peptide sequence map, and a peptide conformation map, respectively; extracting graph representations of each target conformation and each peptide conformation using a deep learning coding network, and performing conditional conformation competition on multiple target conformations based on the current peptide graph representation to obtain target conformation fitting weights; cross-aligning the target residue representations and peptide residue representations according to the fitting weights to generate a residue-peptide interaction tensor; determining candidate active residues and candidate active fragments from the interaction tensor, performing counterfactual masking on the candidate active residues and their corresponding interaction edges, and recalculating the interaction contribution; and generating targeted screening results for small molecule peptides based on the difference in interaction contributions before and after masking, the consistency of candidate action relationships in different conformational combinations, and the fitting weights.

[0021] In this embodiment, the candidate conformations of the target are derived from the structural records formed by the same target in different local pocket opening and closing states, side chain orientation states, and residue exposure states, or from the set of structures obtained after performing conformational sampling on known target structures. The candidate conformations maintain the correspondence between amino acid residue numbers. The candidate conformations of the small molecule peptides to be screened are generated based on the peptide bond connection constraints, the allowable range of local dihedral angles, and the repulsion relationship between residues in the peptide sequence. After generation, conformations with atomic overlap, peptide bond breakage, or local configurations that do not meet the sequence constraints are deleted. Each target conformation and peptide conformation retained are converted to a unified data expression form. The unified data expression form records at least the residue identifier, residue category, residue coordinates, main chain connection, side chain direction, solvent exposure state, and local geometric adjacency relationship, so that different conformations can be compared according to the residue index, while retaining the spatial differences caused by conformational changes.

[0022] Specifically, the target residue map uses amino acid residues as nodes, and classifies edges based on main chain connections, non-covalent contacts within a preset spatial adjacency range, and cooperative displacement relationships within the same local secondary structure. Each node is configured with residue category, solvent exposure level, local curvature, and side chain orientation features. The polypeptide sequence map uses polypeptide residues as nodes, and classifies edges based on peptide bond connections and sequence spacing. The polypeptide conformation map adds spatial constraint edges determined by atomic distance, relative residue orientation, and local dihedral angles to the polypeptide sequence map. The deep learning coding network propagates node features along both sequence topology and spatial topology, and fuses them through the index correspondence of the same residue in both types of topology to obtain a graph representation that combines sequence continuity and conformational geometry information.

[0023] In the target residue map, main chain connection edges are used to record uninterrupted sequence relationships between adjacent residues, spatial contact edges are used to record residue adjacency relationships that may form local interactions in the current conformation, cooperative displacement edges are used to record residue relationships located in the same helical, folded, or flexible loop region that have a common tendency to move during conformational changes, local curvature is represented by the degree of spatial folding formed by the current residue and the preceding and following main chain residues, and side chain orientation is represented by the direction of the side chain centroid relative to the local coordinate system of the main chain. In the peptide sequence map, sequence spacer edges connect residues whose sequence positions are not adjacent but whose spacing is within a set range, so that the coding results retain short-order sequence dependence. In the peptide conformation map, spatial constraint edges are used to express residues that are far apart on the sequence but close in space after folding, thereby avoiding inferring the spatial interaction between the peptide and the target solely based on sequence adjacency relationships.

[0024] In a preferred embodiment, the deep learning encoding network performs message aggregation on the classified edges separately, and then fuses the residue representations generated from different subgraphs. The single-layer encoding process is represented as follows: ; in, This represents the feature vector of the i-th residue in the l-th layer. This represents the residue feature vector after propagation at the current layer is complete. Represents the transformation matrix of the central residue. Represents a set of edge types. This represents the set of neighboring residues connected to the i-th residue via edge r-th type. This represents the feature transformation matrix corresponding to the r-th type edge. This represents the fusion coefficient of the i-th residue with the r-th subgraph information. The nonlinear transformation function is defined as follows: when a residue is located in a flexible loop region and its spatial contact neighborhood changes significantly with conformation, the identifiable increment generated by the spatial contact subgraph is higher than that generated by the main chain subgraph. The fusion coefficient enables the residue to retain more spatial information of the current conformation. When a residue is located in a conformationally stable region, the main chain connection information and cooperative displacement information maintain a high proportion, so that the background residues in different candidate conformations can be aligned.

[0025] Please refer to Figure 2 , Figure 2 The horizontal axis represents the position of the target residue in the peptide targeting screening coding process, shown as the amino terminus, pre-terminal residue, central residue, post-terminal residue, and carboxyl terminus. The vertical axis represents the fusion coefficient, ranging from approximately 0.14 to 0.47. The three curves correspond to the information contributions of three types of edges: main chain connection, spatial contact, and cooperative displacement. The legend uses different line types to indicate the fusion ratio of the three types of edges in the same residue coding layer.

[0026] from Figure 2 As can be seen, the spatial contact curve is generally higher than the main chain connection curve, with an average fusion coefficient of approximately 0.3352, reaching a peak of approximately 0.4733 in the central residue region; the average value of the co-displacement curve is approximately 0.2752, also showing a high value of approximately 0.4133 in the same segment; the average value of the main chain connection curve is approximately 0.2152, indicating that the stable sequence topology still provides background alignment information, but is not the dominant source of the flexible pocket region. This trend corresponds to the categorical neighborhood weighted aggregation in the single-layer coding formula, indicating that the model increases the proportion of spatial contact and co-displacement information near high-response residues, preserving the locally identifiable increments generated by conformational changes.

[0027] The deep learning coding network includes a local residue coding layer and a conformational convergence layer. The local residue coding layer generates subgraph representations of the same residue in the main chain connection subgraph, spatial contact subgraph, and cooperative displacement subgraph, respectively, and determines the subgraph fusion coefficient based on the identifiable increment of the central residue representation according to the neighborhood features in each subgraph. The conformational convergence layer uses locally high-response residues as anchors to aggregate residue representations within a preset topological step size into local patch representations, and then performs hierarchical convergence of multiple local patch representations with background residue representations not covered by anchors. For different candidate conformations of the same target, the background residue representations are constrained to remain aligned according to the residue number correspondence, and the local patch representations are allowed to form differentiated representations with conformational changes.

[0028] Locally high-response residues are determined by the degree of deviation of residue characterization from global conformational characterization and its correlation with the initial peptide characterization. For each locally high-response residue, the neighborhood is expanded along the main chain connection edge, spatial contact edge, and cooperative displacement edge in the target residue map. If the neighborhoods of multiple anchor points overlap, the overlapping residues are assigned to the corresponding slices according to the response intensity of the overlapping residues to each anchor point, or the overlapping area is retained as a shared boundary between slices. Residues that do not enter any local slice form a background residue set. The background residue set mainly carries the overall folding state of the target and information on non-pocket regions. Local slice characterization mainly carries pocket opening and closing, side chain flipping, and local exposure changes. Thus, the conformational characterization simultaneously includes the cross-conformational stable part and the local part that changes with the peptide adaptation relationship.

[0029] In this embodiment, please refer to Figure 3 The conditional conformational competition includes: constructing conformational pairing representations for each target conformation and each peptide conformation, wherein the conformational pairing representations are composed of target global map representations, peptide global map representations, representations of locally high-response residues of both sides, and geometric matching descriptions between conformations; inputting each conformational pairing representation into a competition scoring network with shared parameters to obtain the corresponding initial conformational fit values; performing normalized competition among all target conformations corresponding to the same small molecule peptide to be screened, and redistributing the initial conformational fit values ​​according to the consistency of the fit ranking of each target conformation under different peptide conformations to obtain the target conformational fit weights; and removing conformational pairings with a relative support lower than that within the same conformational competition group from subsequent cross-alignment.

[0030] Conditional conformational competition does not directly fix conformational weights based on the energy of the target conformation itself. Instead, it recalculates the weights for the current peptide to be screened. Let the pairing score between the u-th target conformation and the v-th peptide conformation be... The order consistency of the u-th target conformation among all peptide conformations is: The corresponding adaptation weights are expressed as follows: ; in, This represents the fit weight of the u-th target conformation given the v-th peptide conformation, where U represents the number of candidate target conformations. The scoring network calculates the scores based on conformational pairing representations. This indicates the degree to which the u-th target conformation maintains a relatively high ranking across different peptide conformations. The coefficient representing the adjustment of the order consistency is defined as follows: If there are 3 target conformations and 2 peptide conformations, and a certain target conformation has a high pairing score for both peptide conformations, then its... By taking a positive value and increasing the corresponding adaptation weight, if a target conformation only produces an isolated high score for a single abnormal polypeptide conformation, its ranking consistency is low, which suppresses the isolated high score in the normalization competition.

[0031] Please refer to Figure 4 , Figure 4 The horizontal axis represents the pairing combination formed by the target conformation and the peptide conformation, and the vertical axis represents the support. Solid lines represent pairing scores, dashed lines represent evidence coverage, and dotted lines represent fit weights. The legend corresponds to all three line types. The horizontal axis shows the pairing source in the order of target-peptide pairing number, avoiding the use of only ordinal numbers, so that each peak corresponds to a specific conformation combination.

[0032] from Figure 4 As can be seen, the pairing scores for target 4 and target 3 both reach 0.87, the highest in the entire graph; the evidence coverage for target 3 is approximately 0.6571, the highest point of evidence coverage; the fitting weight for target 3 and target 11 is approximately 0.6011, slightly higher than the 0.5985 for target 4 and target 4. Combinations such as target 3 and target 2 also maintain high support, constituting the main candidate groups for cross-alignment. Pairings with high scores, evidence coverage, and fitting weights simultaneously indicate that the same local region of action is supported by multiple pieces of evidence; if only the score is high but the coverage is insufficient, it should not be directly used as the core input for subsequent tensor calculations.

[0033] The relative support is determined based on the distribution of adaptation weights within the same conformational competition group and the overlap between conformational evidence. A fixed absolute threshold is not used. Specifically, conformational pairs with adaptation weights lower than the median support level within the group and whose local evidence does not cover any high-response regions can be eliminated. Conformational pairs with low adaptation weights but complementary local evidence with high-weight conformations are retained. Complementary local evidence refers to two conformations that support different consecutive action fragments of the same peptide, and the corresponding target residues are located in regions that can be mutually converted through local conformational migration. This avoids missing action relationships formed in stages by retaining only a single conformation, and also avoids sending target conformations without corresponding local evidence to the current peptide into the subsequent cross-alignment process.

[0034] Furthermore, the geometric matching description of the conformation pairing characterization is generated in the following manner: locally high-response residues are selected from the target conformation and the peptide conformation respectively, and a rotation- and translation-invariant local coordinate description is established with each locally high-response residue as the center. The relative distance sequence, relative orientation sequence, and accessibility overlap sequence between the target local coordinate description and the peptide local coordinate description are calculated, and the three types of sequences are mapped into pairing geometric vectors through a shared encoder. The competitive scoring network simultaneously outputs the initial conformation fitting value and the conformation evidence vector. Before normalization competition, target conformations supporting the same local action region are merged according to the complementarity between the conformation evidence vectors, and target conformations dependent on different local action regions are separated according to the degree of conflict between the conformation evidence vectors.

[0035] The local coordinate description uses the main chain direction, side chain direction, and normal direction of the central residue as the coordinate base. The target local residues and peptide local residues are projected into their respective coordinate bases. The geometric matching description is then constructed through the relative rotation and relative displacement relationships between the two sets of local coordinate bases. The relative distance sequence records the distance interval between candidate residue pairs, the relative orientation sequence records the side chain orientation, main chain angle, and local plane orientation, and the accessibility overlap sequence records whether the entry direction from the peptide local region to the target local region is blocked by neighboring residues. The shared encoder maps the three types of sequences with the same parameters, so that the geometric evidence of different conformation pairings is in the same comparison space. Each component in the conformation evidence vector corresponds to a local action region rather than a single score of the entire target.

[0036] In a preferred embodiment, establishing a bidirectional constraint relationship between the local region representation and the conformational evidence vector includes: determining the evidence coverage region based on the response distribution of the conformational evidence vector in the target residue dimension; feeding back the local region representation corresponding to the evidence coverage region to the competitive scoring network to update the initial conformational fit values ​​of each target conformation; re-determining the locally high-response residues based on the updated initial conformational fit values; establishing a residue migration map between the re-determined locally high-response residues and the original locally high-response residues; retaining the corresponding conformational pairing when the residue migration map remains within the same local region or adjacent local regions; and splitting the corresponding conformational pairing into multiple local evidence pairs and having them participate in the normalization competition when the residue migration map crosses non-adjacent local regions.

[0037] Residue migration mapping is established based on residue number correspondence, spatial neighborhood overlap, and local characterization similarity. If the original high-response residue and the updated high-response residue have the same number, a mapping is established directly. If the numbers are different but they are located in the same flexible loop region and there are a majority of corresponding residues in their respective neighborhoods, a neighborhood migration mapping is established. If they are located in non-adjacent regions and there is no continuous neighborhood change path in between, they are not interpreted as conformational migration of the same action region. Instead, the current conformational pairing is decomposed into multiple local evidence pairings. Each local evidence pairing includes the target characterization, peptide action fragment characterization, and geometric matching description of the corresponding region. This allows the competition process to distinguish the scores generated by different pocket regions in the same conformation and prevents unrelated local responses from forming a single high fit value after global convergence.

[0038] In this embodiment, please refer to Figure 5 The residue-peptide interaction tensor is jointly encoded by the multi-head cross-attention results between target residue characterization and peptide residue characterization, residue distance binning characterization, relative orientation characterization, and adaptation weights of the corresponding conformation. Candidate action fragments are determined according to the high-response regions that appear repeatedly across attention heads and are continuously distributed between adjacent peptide residues in the interaction tensor, and candidate action residues are determined according to the set of target residues that form a bidirectional response with the candidate action fragments. For each candidate action residue, node feature masking, related interaction edge removal, and neighborhood information freezing are performed to obtain mutually independent counterfactual inputs. Each counterfactual input is fed back into the same cross-alignment process to obtain the masked interaction contribution corresponding to each candidate action residue.

[0039] Let the interaction tensor element between the i-th residue in the u-th target conformation and the j-th residue in the v-th polypeptide conformation be . Its calculation form is: ; in, This represents the total interaction amount of the corresponding residue pair under the current conformational combination. This represents the conformational adaptation weight, and M represents the number of cross-attention heads. This represents the fusion weight of the m-th attention head. and These represent the query characterization and response characterization of the target residue and peptide residue in the m-th attention head, respectively. This represents the distance binning coding function. Indicates the distance between residue pairs. This represents the relative orientation coding function. The interaction tensor describes the orientation of residue pairs. When a residue pair has a high attentional response but the distance binning and relative orientation do not support contact, the geometric coding term restricts the corresponding interaction quantity. When multiple attention heads, distance binning, and orientation relationships simultaneously support the same residue pair, the interaction tensor forms a stable response at the corresponding position.

[0040] Candidate action fragments are not determined directly based on the highest response of a single polypeptide residue. Instead, continuous high-response regions are detected by indexing along the polypeptide sequence. Specifically, the interaction amounts of each polypeptide residue with all target residues are aggregated to form a polypeptide residue response sequence. Then, it is checked whether adjacent polypeptide residues recur in different attention heads and different conformational combinations. A small number of low-response residues are allowed as structural inflection points in the continuous region, but the high-response regions on both sides of the low-response residues should correspond to the same or adjacent target fragment regions. Candidate action residues are composed of target residues that form a bidirectional response with the continuous region. Bidirectional response means that the attention distribution of the target residues pointing to the polypeptide fragment and the attention distribution of the polypeptide fragments pointing to the target residues have a common high-value region, thereby reducing false candidate sites caused by unidirectional attention concentration.

[0041] When performing node feature masking on candidate active residues, their residue category, side chain orientation, exposure state, and local geometric features are replaced with neutral occupancy characteristics, while retaining the residue number and main chain position. When performing related interaction edge removal, the cross-alignment edges between the candidate active residue and the candidate active fragment are deleted, but the internal graph connections between the residue and the target neighborhood residue are retained. When performing neighborhood information freezing, the encoding results of non-candidate residues in the unmasked state are cached, and the changes of candidate active residues are prohibited from continuing to propagate to non-candidate regions during recalculation, so that the score changes after masking mainly correspond to the relationship between candidate residues and candidate peptides. The three types of counterfactual inputs change the residue content, cross-molecular interaction relationship, and local information propagation path, respectively, which can examine the interaction contribution of the same candidate active residue from different perspectives.

[0042] Furthermore, after applying counterfactual masking to the candidate active residues, the process further includes: replacing the masked candidate active residues with neutral reference representations obtained by aggregation of their structural neighborhoods, while keeping the characterization of the remaining target residues and peptide conformation unchanged, thus forming a substitutional counterfactual input; sending the substitutional counterfactual input and the counterfactual inputs formed by node feature masking, removal of relevant interaction edges, and freezing of neighborhood information into a cross-alignment process to obtain multiple counterfactual interaction contributions; assigning credibility to the multiple counterfactual interaction contributions according to their perturbation amplitude on the interaction tensor of non-candidate regions, and using the credibility-weighted difference in interaction contributions to form the causal contribution characterization of the candidate active residues; and generating fragment-level causal contributions for multiple candidate active residues belonging to the same candidate active fragment, based on the directional consistency of their causal contribution characterizations.

[0043] Please refer to Figure 6 , Figure 6 The horizontal axis represents the combination of candidate residues and counterfactual input types, while the vertical axis represents the confidence weight, with values ​​ranging from approximately 0.30 to 0.68. The bars cover four types of inputs: node masking, edge removal, neighborhood freezing, and neutral substitution, reflecting the degree of perturbation and attributability of the same candidate residue to non-candidate regions under different counterfactual processing.

[0044] from Figure 6 As can be seen, the average confidence weight of the neutral substitution type is approximately 0.6158, which is higher than that of neighborhood freezing (0.5300), edge removal (0.4442), and node masking (0.3583). Residue position 25 reaches the highest confidence weight (0.68) under neutral substitution, the highest across the entire graph. This result indicates that after replacing candidate residues with neutral references generated from structural neighborhoods, the input remains relatively close to its original distribution, making it more suitable for measuring the contribution of the candidate residues themselves to the fragment response. The low node masking weight suggests that direct masking easily induces significant background perturbations, and the score decrease cannot be entirely attributed to this single residue.

[0045] Let the interaction contribution of candidate residue i under the original input be... The interaction contribution under the e-th type of counterfactual input is The perturbation caused by the e-th type of counterfactual input to the non-candidate region is The causal contribution of the candidate active residues is expressed as follows: ; in, Let E represent the causal contribution of the i-th candidate residue, and let E represent the number of counterfactual input types. The difference between the interaction tensor of non-candidate regions and the counterfactual input of type e is calculated. The smaller the perturbation, the higher the confidence weight of the corresponding counterfactual difference. If node occlusion reduces the interaction amount of candidate regions but also causes a large-scale change in non-candidate regions, the difference receives a lower weight. If the removal of interaction edges only changes the tensor position between candidate residues and candidate fragments, the corresponding difference receives a higher weight. When multiple residues within the same candidate fragment... When all residues are in the same direction, they are aggregated into fragment-level causal contributions. When the contribution directions of residues cancel each other out, the fragment is split into shorter, continuous sub-fragments for re-verification.

[0046] Please refer to Figure 7 , Figure 7The horizontal axis represents the position of candidate action residues, and the vertical axis represents the credible weighted contribution, with values ​​ranging from approximately 0.18 to 0.62. Each bar corresponds to the difference in weighted contribution of a candidate target residue before and after the counterfactual input. The larger the bar height, the more significant the decrease in the candidate fragment response after the residue is masked, its related interaction edge is removed, or it is replaced with a neutral reference, and the less the non-candidate region is perturbed.

[0047] from Figure 7 As can be seen, residue 22 has the highest contribution, reaching 0.62, and is the core residue of the current fragment-level contribution; residues 31, 20, 18 and 27 have contributions of around 0.425 to 0.46, forming an auxiliary contribution group; residues 26 and 30 have lower contributions, approximately 0.18 and 0.215, respectively, indicating that their original interaction response may be more affected by background propagation or adjacent residues. Figure 7 Corresponding to the credible weight summation relationship in the causal contribution formula, it is used to distinguish between key residues that truly support candidate fragments and residues that only incidentally rise in the original attention results.

[0048] Neutral reference characterization is generated from the structural neighborhood of candidate action residues. The structural neighborhood is defined as a node within the target that is adjacent to the main chain or in stable spatial contact with the candidate action residue but has not been identified as a candidate action residue. During aggregation, specific values ​​in the residue category are removed, and only the average exposure state, average local curvature, and average main chain direction of the neighborhood are retained. This avoids directly using all-zero vectors to generate inputs that exceed the training distribution. The directional consistency of fragment-level causal contributions is determined by comparing the change direction of interaction amounts before and after the masking of each candidate action residue within the same fragment. If multiple residues correspond to different target fragments, they are aggregated separately according to the action cluster affiliation, and spatially separated action relationships are not merged into the same fragment contribution.

[0049] In this embodiment, the consistency of the candidate interaction relationships across different conformational combinations is determined as follows: The residue-peptide interaction tensors corresponding to each conformational combination are mapped to a unified index plane according to the target residue number and peptide residue number; for each interaction position in the unified index plane, its interaction contribution, local geometric relationship, and conformational adaptation weight under different conformational combinations are extracted to form a cross-conformational interaction trajectory; multiple interaction positions that undergo continuous spatial migration but whose residue indices remain adjacent are merged into the same interaction cluster, and a corresponding consistency characterization is generated based on the directional consistency of the interaction contribution within the interaction cluster, the variable range of the geometric relationship, and the joint distribution of the adaptation weights; interaction positions that appear only in a single conformational combination and cannot be merged into any interaction cluster are marked as isolated interaction positions.

[0050] The unified index plane uses target residue numbering as one dimension and peptide residue numbering as another dimension. In different conformational combinations, the same residue number corresponds to the same index position. If the original active residue is replaced by a residue with an adjacent number due to local conformational change, it will appear as continuous migration in the unified index plane. Cross-conformational interaction trajectories record the interaction amount, residue pair distance, relative orientation, and conformational adaptation weight corresponding to each index position. When merging trajectories, adjacent interaction positions are required to remain continuous in the target index dimension or peptide index dimension, and their local geometric changes are required to be explained by the residue displacement within the same region. If two high-response positions are far apart on the index plane and there is no migration path in their respective regions, they will form independent action clusters to avoid merging different action regions through global similarity.

[0051] Let the set of cross-conformational interaction locations contained in the action cluster c be . The corresponding consistency representation is expressed as: ; in, This represents the correspondence consistency representation of the c-th action cluster. This indicates the adaptation weight of the conformation combination to which the interaction position p belongs. This indicates the causal contribution of candidate residues corresponding to the interaction positions. Represents a symbolic function. This represents the deviation of the interaction position from the geometric relationship of the cluster center. It is considered valid when the interaction contributions in multiple conformational combinations are aligned and the geometric deviation can be explained by local conformational changes. To maintain unidirectional accumulation, when an interaction location has a high original score but its causal contribution direction is opposite or its geometric deviation is large, its contribution to the consistency of the action cluster is suppressed, and isolated interaction locations are not included in the action cluster accumulation but are recorded separately.

[0052] Please refer to Figure 8 , Figure 8 The horizontal axis represents the position of peptide residues, and the vertical axis represents the position of target residues. The three lines correspond to the cross-conformation migration trajectories of the three functional clusters. Solid, dashed, and dotted lines are all connected by consecutive dots to the interaction positions in adjacent conformational combinations. The legend indicates different functional clusters, and the direction of the lines is used to represent the positional migration of candidate fragments on the unified index plane.

[0053] from Figure 8 As can be seen, action cluster one gradually migrates from the target position 21 at position 1 of the peptide to the target position 26 at position 10 of the peptide; action cluster two covers target positions 26 to 31; and action cluster three covers target positions 17 to 22. All three trajectories maintain local continuity, indicating that these high-response points can be explained by residue shifts in the same or adjacent local regions. Figure 8In the corresponding action cluster consistency formula, the adaptation weight, causal contribution direction, and geometric deviation attenuation term are used. Positions with continuous trajectories and stable directions are included in the action cluster accumulation, while isolated responses far from the trajectory will not directly enter the stable contribution.

[0054] Preferably, the center of the action cluster is defined as the target residue index center, peptide residue index center, and local geometric description center after adaptation weighting. The geometric deviation considers the residue pair distance difference, side chain orientation difference, and local coordinate base rotation difference simultaneously, and does not determine the correspondence based on a single distance change. Flexible peptides may undergo local flipping in different conformations. If, after flipping, adjacent peptide residues still form continuous interactions with the same target region, the corresponding trajectory can be classified into the same action cluster. If a high response appears only in a single conformation combination, and there are no indexed or geometrically continuous responses in adjacent target conformations and adjacent peptide conformations, it is marked as an isolated interaction position, and its position record is retained in the final summary but is not directly accumulated as a stable action contribution.

[0055] In a preferred embodiment, the joint verification of the fragment-level causal contribution and the corresponding consistency characterization includes: establishing a fragment evidence vector for each candidate action fragment, consisting of the fragment-level causal contribution, the cross-conformational interaction trajectory of the corresponding action cluster, and the proportion of isolated interaction positions; aggregating multiple fragment evidence vectors according to the index correspondence of the same candidate action fragment in different target conformations, and multiplying the aggregation result by the corresponding target conformational adaptation weight to obtain the joint fragment contribution; when a candidate action fragment corresponds to multiple mutually separate action clusters, calculating the joint fragment contribution of each action cluster separately, and determining whether to merge or retain independently based on whether the action clusters share candidate action residues; summing up the joint contributions of all independently retained fragments with the conformational adaptation weights to generate the target screening result of the small molecule peptide.

[0056] Let the final target screening value of the small molecule peptide to be screened be... The number of retained candidate action fragments is K, and the causal contribution of the k-th candidate action fragment is... The corresponding cluster consistency is The proportion of isolated interactive positions is The final target screening value is expressed as: ; in, This represents the convergence coefficient of the k-th candidate action fragment, determined based on its contiguous length and residue coverage. It is obtained by the aggregation of causal contributions from candidate active residues within the fragment. It is obtained by aggregating the corresponding consistency representations of the respective action clusters. This represents the ratio of the number of isolated interaction locations within the segment to the total number of interaction locations. This represents the evidence coverage coefficient for the u-th target conformation. This represents the aggregate adaptation weight of the u-th target conformation across multiple peptide conformations. When a candidate fragment has a high primitive interaction but a low counterfactual contribution... To limit its inclusion in the final screening criteria, when a fragment has a high causal contribution but appears only in a single anomalous conformation. and Together they limit their aggregate value.

[0057] Please refer to Figure 9 , Figure 9 The horizontal axis represents the candidate peptide category, and the vertical axis represents the final screening value, ranging from approximately 0.454 to 0.759. The bar height represents the comprehensive ranking result after summarizing fragment-level causal contribution, consistency of action cluster correspondence, proportion of isolated interaction positions, and conformational adaptation weight. The horizontal axis represents different peptides to be screened from candidate one to candidate ten.

[0058] from Figure 9 As can be seen, Candidate 1 has the highest final screening value, approximately 0.759; Candidates 2 and 3 have values ​​of approximately 0.710 and 0.661 respectively, forming a priority verification echelon; Candidate 5 has a value of approximately 0.631, slightly higher than Candidate 4's 0.612, indicating that the final ranking is not simply a decrease in candidate number, but is influenced by fragment contribution, cluster stability, and isolation ratio. Candidate 10 has the lowest value, approximately 0.454, indicating that although it may have local responses, the overall evidence is insufficient. Figure 9 The formula for the final selection value indicates that high-scoring candidates need to simultaneously possess continuous action segments, high-confidence contributions, stable cross-conformation trajectories, and a low proportion of isolated responses.

[0059] Whether multiple action clusters are merged is determined based on the shared candidate action residues and the continuity of the peptide fragments. If two action clusters share at least one candidate action residue and their corresponding peptide fragments are sequentially continuous, the two action clusters are considered as the same complex action region and the fragment evidence vectors are merged. If two action clusters correspond to mutually separate target fragments and do not share candidate action residues, they are retained separately so that the final screening results can represent the situation where the same peptide interacts with the target through multiple independent fragments. The summarization of conformational adaptation weights does not use a simple average, but rather calculates the evidence coverage coefficient according to the completeness of each conformational evidence covering the candidate action fragment. Conformations that cover multiple stable action clusters receive a higher summarization percentage, while conformations that only cover isolated interaction positions do not receive a higher percentage due to a single high score.

[0060] Deep learning encoding networks, competitive scoring networks, and cross-alignment networks can be trained jointly. Training data includes target candidate conformations, peptide sequences, peptide candidate conformations, and corresponding binding tags. During training, different candidate conformations of the same target are grouped into the same competitive group, and multiple candidate conformations of the same peptide are grouped into the same peptide conformation group. For data with known key action residues, residue-level supervision information is retained, while for data without residue-level labels, only binding tags and cross-conformation consistency constraints are used. The training objective consists of binding prediction error, conformation competition constraints, counterfactual consistency constraints, and action cluster continuity constraints, so that the model can still form training signals based on recurring local action relationships between conformations even when complete site labels are lacking.

[0061] The joint training loss can be expressed as: ; in, Indicates total training loss. This represents the prediction loss between the targeted screening value and the known binding label. This represents the competition ranking loss between reasonable conformation pairings and those that do not support conformation pairings. This represents the counterfactual constraint loss relating the change in interaction contribution after masking candidate active residues to the continuity of candidate active fragments. This represents the interaction direction and local geometric constraint loss of the same action cluster in different conformational combinations. , , and These represent the combination coefficients of each training objective, which can be improved in the early stages of training. and The proportion is used to establish basic combination discrimination and conformational differentiation. After the model forms a stable interactive response, the proportion is then increased. and The proportion of this allows counterfactual contributions and cross-conformational correspondences to participate in parameter updates.

[0062] Please refer to Figure 10 , Figure 10 The horizontal axis represents the training phases, in order: initial training, early convergence, mid-term constraints, late-term stabilization, and final epoch. The vertical axis represents the loss. Solid lines represent total loss, dashed lines represent combination error, dotted lines represent competition constraints, and dotted lines represent counterfactual constraints. The legend corresponds to each line type. Curves are used to illustrate the convergence relationship of different constraint terms during multi-objective joint training.

[0063] from Figure 10As can be seen, the total loss decreased from 1.1641 in the initial stage to 0.3152 in the final round; the assemblage error decreased from 0.7195 to 0.1932, the competition constraint from 0.5550 to 0.1475, and the counterfactual constraint from 0.4375 to 0.1150. All curves show a smooth downward trend, with the total loss at approximately 0.4821 in the mid-term constraint stage and decreasing to approximately 0.3696 in the later stable stage. This indicates that the model does not only optimize the assemblage of labels but converges simultaneously across conformational competition, counterfactual interpretation, and action cluster continuity. Figure 10 Based on the joint training loss formula, it is proven that multiple training objectives can form a consistent parameter update direction.

[0064] The training samples that do not support conformational pairing can be constructed from low-evidence conformations of the same target, geometrically unreasonable conformations of the same peptide, and peptide conformations that maintain the residue composition but disrupt the order of local action fragments. After construction, the target residue number and peptide residue number must still be traceable to avoid generating samples that cannot perform unified index mapping. Counterfactual training samples are selected from the candidate action residues currently identified by the model, and node feature masking, interaction edge removal, neighborhood information freezing, and neutral reference replacement are performed respectively. If a certain type of counterfactual input causes a large-scale change in non-candidate regions, its training weight is reduced according to the perturbation amount, and the corresponding score drop is not directly attributed entirely to the candidate action residues.

[0065] During the model inference phase, multiple candidate conformations are generated or read for each small molecule peptide to be screened, and they form conformation competition groups with the target candidate conformations. The encoding network can pre-calculate and cache the background residue characterization and stable region characterization of each candidate conformation of the target. When the current peptide enters, only the peptide map characterization, target local region conditional characterization and conformation pairing characterization are recalculated. After the conditional conformation competition is completed, the retained conformation pairings are sent to the cross-alignment network, which sequentially generates residue-peptide interaction tensors, candidate action fragments, candidate action residues, counterfactual contributions, cross-conformation interaction trajectories and fragment joint contributions. The output includes the final target screening value, the target conformations involved in the summary, the retained peptide conformations, candidate action fragment index, candidate action residue index and isolated interaction position records.

[0066] In a specific application, the target has 4 candidate conformations, and the peptide to be screened has 3 candidate conformations. The conformational competition network forms 12 conformational pairings, of which 3 pairs cover the same local pocket region and have continuous residue migration relationships, 2 pairs cover the auxiliary action regions adjacent to the pocket region respectively, and the conformational evidence of the remaining pairs does not cover stable local regions or only forms isolated high responses. Normalization competition retains the aforementioned 5 pairs for cross-alignment. The interaction tensor forms a high response in a continuous 4 residue region of the peptide sequence. The 4 residues form a bidirectional response with the 6 candidate action residues in the target pocket. Four types of counterfactual inputs are generated for the 6 candidate action residues respectively. Then, the causal contribution is calculated according to the perturbation amount of the non-candidate region. The masking change of one candidate action residue is accompanied by a large-scale background perturbation and its confidence is reduced. The remaining candidate residues form 2 action clusters and maintain index continuity in multiple conformational combinations, thereby forming a joint contribution of 2 independently retained fragments.

[0067] In another specific application, there are local flips between candidate peptide conformations. The active fragments before and after the flip are composed of the same peptide residues, but the side chain orientation of the peptide residues changes. Conditional conformational competition gives the pre-flip conformation a lower fit weight with the target conformation in the pocket-closed state, and gives the post-flip conformation a higher fit weight with the target conformation in the pocket-semi-open state. Cross-alignment results show that the active fragments are taken over by adjacent target residues in different target conformations. The interaction positions in the unified index plane migrate continuously along the target residue index. The local geometric deviation is within the allowable range of the same active cluster. After counterfactual masking, the interaction contribution of the corresponding fragments continues to decrease in the same direction. The interaction relationship is included in the joint contribution of fragments. If another target conformation only produces a single point high response due to spatial overlap and cannot form a continuous migration trajectory, the corresponding position is marked as an isolated interaction position.

[0068] Preferably, when some target candidate conformations lack individual side chain coordinates, a placeholder side chain representation can be generated based on the side chain direction of the same residue in adjacent candidate conformations and the local coordinates of the main chain in the current conformation. A missing label is added to the placeholder representation. When the deep learning coding network aggregates, it limits the contribution of the side chain representation to the local patch representation based on the missing label. When a peptide candidate conformation lacks some atomic-level details but the centroid of the residue and the main chain direction are complete, spatial constraint edges can be constructed using only residue-level distances, relative orientations, and local dihedral angles. Specific contacts are not directly constructed using missing atoms. Counterfactual masking and cross-conformation consistency calculations are both based on the same residue-level expression, maintaining data scale consistency between different candidate conformations.

[0069] In a preferred embodiment, candidate conformation subsets can be established for different functional states of the same target, but each subset still uses the same residue numbering and index plane. Conditional conformation competition first forms local fit weights within each conformation subset, and then the highest evidence coverage region of each subset is characterized and sent to cross-subset competition. If the peptide to be screened forms a stable functional cluster in multiple functional states, the joint contribution of the corresponding fragments is retained respectively. If the peptide forms a functional fragment with causal contribution and cross-conformation consistency only in a single functional state, the final screening result records the conformation fit weight and functional cluster information corresponding to the functional state, and does not average low-support conformations in other functional states into the current result.

[0070] Specifically, the screening results can be ranked according to the final target screening value, while saving the fragment evidence vector that constitutes the screening value, so that subsequent verification can check the candidate action fragments and candidate action residues. During the ranking process, peptides with similar final target screening values ​​can be further ranked according to the stability of fragment-level causal contribution, number of action clusters, proportion of isolated interaction positions, and conformational adaptation weight distribution. The secondary ranking still uses the data generated by the aforementioned conformational competition, counterfactual masking, and cross-conformational correspondence processes, without adding external scoring items that deviate from the screening logic. This ensures that the entire screening process maintains a continuous data closed loop from input conformation, graph representation, conformational competition, residue interaction, counterfactual verification, action cluster aggregation to candidate ranking.

Claims

1. A deep learning-based method for targeted screening of small molecule peptides, characterized in that, include: Multiple candidate conformations of the target and sequences and multiple candidate conformations of the small molecule peptides to be screened were obtained, and target residue maps, peptide sequence maps and peptide conformation maps were constructed respectively. The graph representations of each target conformation and each peptide conformation are extracted using a deep learning coding network. Based on the current peptide graph representation, conditional conformation competition is performed on multiple target conformations to obtain target conformation adaptation weights. Based on the adaptation weights, the target residue characterization and peptide residue characterization are cross-aligned to generate a residue-peptide interaction tensor. Candidate active residues and candidate active fragments are determined from the residue-peptide interaction tensor. Counterfactual masking is applied to the candidate active residues and their corresponding interaction edges, and the interaction contribution is recalculated. Based on the difference in interaction contributions before and after masking, the consistency of candidate interaction relationships in different conformational combinations, and the adaptation weights, the target screening results for small molecule peptides are generated.

2. The method for targeted screening of small molecule peptides based on deep learning according to claim 1, characterized in that, The target residue map uses amino acid residues as nodes, and classifies edges based on main chain connections, non-covalent contact relationships within a preset spatial adjacency range, and cooperative displacement relationships within the same local secondary structure. Each node is configured with residue category, solvent exposure level, local curvature, and side chain orientation characteristics. The polypeptide sequence map uses polypeptide residues as nodes and peptide bond connections and sequence spacing relationships to form classification edges; The polypeptide conformation diagram is based on the polypeptide sequence diagram with the addition of spatial constraint edges determined by atomic distance, relative residue orientation, and local dihedral angles. The deep learning coding network propagates node features along sequence topology and spatial topology respectively, and obtains a graph representation that combines sequence continuity and conformational geometric information by fusing the index correspondence of the same residue in the two types of topologies.

3. The method for targeted screening of small molecule peptides based on deep learning according to claim 1, characterized in that, The conditional conformational competition includes: constructing conformational pairing characterization for each target conformation and each peptide conformation. The conformational pairing characterization is composed of target global map characterization, peptide global map characterization, characterization of locally high-response residues of both sides, and geometric matching description between conformations. Each conformation pairing representation is input into a competitive scoring network with shared parameters to obtain the corresponding initial conformation fitting values. Normalization competition is performed among all target conformations corresponding to the same small molecule polypeptide to be screened, and the initial conformation adaptation values ​​are redistributed according to the consistency of the adaptation ranking of each target conformation under different polypeptide conformations to obtain the target conformation adaptation weight. Conformation pairs with a relative support lower than that within the same conformational competition group are removed from subsequent cross-alignments.

4. The method for targeted screening of small molecule peptides based on deep learning according to claim 1, characterized in that, The residue-peptide interaction tensor is jointly encoded by the multi-head cross-attention result between the target residue characterization and the peptide residue characterization, the inter-residue distance binning characterization, the relative orientation characterization, and the adaptation weight of the corresponding conformation. Candidate action fragments are identified based on high-response regions that are repeated across attention heads in the interaction tensor and are continuously distributed between adjacent polypeptide residues, and candidate action residues are identified based on the set of target residues that form a bidirectional response with the candidate action fragments. For each candidate active residue, node feature masking, related interaction edge removal, and neighborhood information freezing are performed to obtain mutually independent counterfactual inputs; Each counterfactual input is re-entered into the same cross-alignment process to obtain the masked interaction contribution corresponding to each of the candidate action residues.

5. The method for targeted screening of small molecule peptides based on deep learning according to claim 2, characterized in that, The deep learning coding network includes a local residue coding layer and a conformational convergence layer. The local residue coding layer generates subgraph representations of the same residue in the main chain connection subgraph, spatial contact subgraph, and cooperative displacement subgraph, respectively, and determines the subgraph fusion coefficient based on the identifiable increment of the central residue representation by the neighborhood features in each subgraph. The conformational aggregation layer uses locally high-response residues as anchor points to aggregate residue representations within a preset topological step size into local patch representations, and then aggregates multiple local patch representations with background residue representations not covered by anchor points in a layered manner. For different candidate conformations of the same target, the background residue representation is constrained to remain aligned according to the residue number correspondence, and the local patch representation is allowed to form differentiated representations as the conformation changes.

6. The method for targeted screening of small molecule peptides based on deep learning according to claim 3, characterized in that, The geometric matching description of the conformation pairing is generated in the following way: select locally high-response residues from the target conformation and the peptide conformation respectively, establish a rotation- and translation-invariant local coordinate description with each locally high-response residue as the center, calculate the relative distance sequence, relative orientation sequence and accessibility overlap sequence between the target local coordinate description and the peptide local coordinate description, and map the three types of sequences into pairing geometric vectors through a shared encoder. The competitive scoring network simultaneously outputs the initial conformational fit value and the conformational evidence vector. Before normalization competition, it merges target conformations supporting the same local region of action according to the complementarity between the conformational evidence vectors, and separates target conformations dependent on different local regions of action according to the degree of conflict between the conformational evidence vectors.

7. The method for targeted screening of small molecule peptides based on deep learning according to claim 4, characterized in that, After applying counterfactual masking to the candidate active residues, the method further includes: replacing the masked candidate active residues with neutral reference representations obtained by polymerizing their structural neighborhoods, while keeping the characterization of the remaining target residues and the peptide conformation unchanged, thus forming a substitutional counterfactual input. The alternative counterfactual input and the counterfactual input formed by node feature masking, related interaction edge removal and neighborhood information freezing are respectively fed into the cross-alignment process to obtain multiple counterfactual interaction contributions; The confidence level of the various counterfactual interaction contributions is assigned according to the perturbation magnitude of their interaction tensors in non-candidate regions, and the difference in interaction contributions after confidence weighting constitutes the causal contribution characterization of the candidate action residues. For multiple candidate action residues belonging to the same candidate action fragment, fragment-level causal contributions are generated based on the consistency of their causal contribution characterization directions.

8. The method for targeted screening of small molecule peptides based on deep learning according to claim 1, characterized in that, The consistency of the candidate interaction relationships in different conformational combinations is determined by mapping the residue-peptide interaction tensors corresponding to each conformational combination to a unified index plane according to the target residue number and peptide residue number. For each interaction position in the unified index plane, extract its interaction contribution, local geometric relationship and the adaptation weight of its respective configuration under different configuration combinations to form a cross-configuration interaction trajectory; Multiple interaction sites that have undergone continuous spatial migration but whose residue indices remain adjacent are merged into the same action cluster, and a corresponding consistency characterization is generated based on the directional consistency of interaction contributions within the action cluster, the variable range of geometric relationships, and the joint distribution of adaptation weights. Interaction positions that occur only in a single conformational combination and cannot be incorporated into any action cluster are marked as isolated interaction positions.

9. The method for targeted screening of small molecule peptides based on deep learning according to claim 5, characterized in that, Establishing a bidirectional constraint relationship between the local region representation and the conformational evidence vector includes: determining the evidence coverage region based on the response distribution of the conformational evidence vector in the target residue dimension, and feeding back the local region representation corresponding to the evidence coverage region to the competitive scoring network to update the initial conformational fit values ​​of each target conformation. The locally high-response residues are re-determined based on the updated conformational adaptation initial values, and a residue migration mapping is established between the re-determined locally high-response residues and the original locally high-response residues. When residue migration maps remain within the same local region or adjacent local regions, the corresponding conformation pairings are preserved. When residue migration maps cross non-adjacent local regions, the corresponding conformation pairings are split into multiple local evidence pairs and each participates in the normalization competition.

10. A method for targeted screening of small molecule peptides based on deep learning according to claim 7, characterized in that, Jointly verifying the fragment-level causal contribution with the corresponding consistency representation includes: establishing a fragment evidence vector for each candidate action fragment, consisting of the fragment-level causal contribution, the cross-conformational interaction trajectory of the action cluster to which it belongs, and the proportion of isolated interaction positions; According to the index correspondence of the same candidate action fragment in different target conformations, the evidence vectors of multiple fragments are aggregated, and the aggregation result is multiplied by the corresponding target conformation adaptation weight to obtain the joint contribution of the fragments. When the candidate action fragment corresponds to multiple mutually separate action clusters, the joint contribution of each action cluster is calculated, and merging or independent retention is determined based on whether the action clusters share the candidate action residues. The combined contribution of all independently retained fragments and conformational fit weights are summarized to generate targeted screening results for small molecule peptides.

Images