A method, system, and storage medium for molecular sieving based on multi-stage progressive competitive co-folding

Abstract

The application discloses a molecular screening method and system based on multi-stage promotion competitive co-folding, and a storage medium. The method comprises the following steps: first-stage lead compound discovery, through sub-area All-at-Once preliminary screening, sub-area final one-on-one competitive co-folding and national final focusing sorting, hierarchical promotion screening of candidate molecules is realized, and the candidate set of the national final is the top 20 candidate molecules with the highest win rate; second-stage lead compound optimization, through grouping candidate molecules into molecular teams according to structural complementarity, multi-ligand synergistic co-folding prediction is performed, a team synergistic competition score is evaluated based on a pocket coverage rate and a conformation coordination matrix, and a lead compound series is output after team internal fine screening and team reorganization iterative optimization; and third-stage preclinical candidate determination, through cross-validation of at least three different architecture AI co-folding models, global consensus win consistency determination, cross-model pose convergence verification and controversial incremental sampling mechanism, a high-confidence preclinical candidate is output. The application effectively reduces the false positive rate, enhances the interpretability, robustness and generalization of the screening result, and significantly reduces the calculation cost of the high-precision co-folding model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted drug discovery technology, specifically to a multi-stage progressive competitive molecular screening method and system based on a deep diffusion model. In particular, this invention is applicable to the entire chain of candidate compound ranking, enrichment, and validation in lead compound discovery, lead compound generation and optimization, and preclinical candidate identification based on AI cofolding models. Background Technology

[0002] Artificial intelligence-based protein structure prediction and molecular docking technologies have been widely applied in virtual drug screening. Denoising diffusion-based co-folding models, such as AlphaFold 3, Boltz-1 / 2, Chai-1, and Protenix, can simultaneously predict the three-dimensional complex structure of proteins and ligand molecules, demonstrating high accuracy in predicting protein-ligand interactions.

[0003] However, existing virtual screening methods have the following technical drawbacks:

[0004] (i) The "false positive" problem of single-molecule docking. Existing methods typically employ a one-to-one docking strategy, that is, inputting a candidate molecule and a target protein into the AI ​​model each time and predicting their binding conformation. Due to the inherent randomness of diffusion models, even for inactive molecules, the model may generate a seemingly reasonable binding conformation, making it difficult to distinguish between truly active molecules and false positive molecules. Mirgaux et al. (2026, npj Drug Discovery) pointed out that the predicted structures of active and inactive molecules are visually difficult to distinguish, and the cofolding model has a memory effect on training data, limiting its generalization ability to new ligands.

[0005] (ii) Lack of a hierarchical competitive evaluation mechanism. Existing screening methods either employ pairwise competitive docking of the entire combination, with computational costs increasing exponentially with the number of candidate molecules; or they employ a simple all-at-once multi-ligand co-folding method, which lacks a progressive elimination and promotion mechanism. This leads to excessive consumption of computational resources on low-quality candidate molecules and fails to simulate the multi-stage progressive screening logic from lead compounds to preclinical candidates in real drug development.

[0006] (iii) Lack of assessment of molecular team synergy effects. Existing All-at-Once methods only focus on "which molecule ultimately occupies the binding site," simplifying multi-ligand competition into a zero-sum game with a single winner, completely ignoring the structural complementarity, synergistic binding effects, and conformational coordination at the team level that may exist among candidate molecules. In real medicinal chemistry, the lead compound optimization stage often requires a comprehensive evaluation of the optimal series from a group of structurally related or functionally complementary molecules, rather than evaluating individual molecules in isolation.

[0007] (iv) Lack of cross-model consensus verification mechanism. Existing methods rely on the prediction results of a single AI model for ranking. Since different co-folding models have systematic differences in training data, network architecture and diffusion sampling strategies, the judgment results of a single model may be misjudged due to model bias, and the robustness and generalization of the screening results cannot be enhanced by cross-validation between heterogeneous models.

[0008] (v) The structural contradiction between computational resources and screening throughput. The computational cost of high-precision cofolding models is extremely high, and pairwise competition of the entire combinatorial system or large-scale all-at-Once screening is difficult to apply to compound libraries of tens of thousands or even hundreds of thousands; while fast methods based on molecular fingerprints or traditional docking scoring have limited accuracy and cannot meet the structural accuracy requirements of the lead compound optimization stage.

[0009] Therefore, there is an urgent need in this field for a molecular screening method that can simulate the entire progressive logic of real drug development, has the ability to conduct collaborative evaluation by molecular teams, supports cross-model consensus verification, and has controllable computational costs. Summary of the Invention

[0010] (a) Technical issues

[0011] This invention aims to solve the aforementioned technical problems existing in current virtual screening methods, and provides a molecular screening method based on multi-stage progressive competitive co-folding, which can:

[0012] 1. By using a regional advancement competition system, we can achieve hierarchical screening of candidate molecules from the discovery of lead compounds to the identification of preclinical candidates, thus avoiding the waste of resources caused by indiscriminate full combinatorial calculations;

[0013] 2. By evaluating the overall binding ability of candidate molecule groups and the conformational coordination among team members through a molecular team collaborative competition mechanism, structural information at the team level can be provided for the optimization of lead compound series.

[0014] 3. Reduce the false positive rate caused by the bias of a single model by utilizing the integrated judgment of heterogeneous AI models through a cross-model global consensus mechanism;

[0015] 4. Further eliminate misjudgments caused by random noise through dual verification of pose convergence and cross-model convergence;

[0016] 5. By balancing computational accuracy and throughput through a cascaded screening architecture, the high-precision co-folding model can be applied to large-scale compound library screening.

[0017] (II) Technical Solution

[0018] To achieve the above objectives, the present invention provides the following technical solution:

[0019] A molecular screening method based on multi-stage progressive competitive co-folding includes the following steps:

[0020] Phase 1: Discovery of Leading Compounds – Regional, Progressive, One-to-One Competitive Co-folding

[0021] Step S1: Obtain the set of candidate molecules to be screened, M = {M1, M2, ..., The candidate molecules include the structural information of the target protein P; the structural information of the target protein includes the amino acid sequence and / or three-dimensional structural coordinates; the candidate molecules include small molecule compounds, peptides, nucleic acid aptamers or combinations thereof.

[0022] Step S2: Based on the physicochemical properties, structural backbone type, or molecular fingerprint similarity of the candidate molecules, the candidate molecule set M is divided into Q preliminary selection subsets M1, M2, ..., M_Q, each preliminary selection subset containing k candidate molecules, k ≥ 5; preferably, k = 20–50.

[0023] Step S3: For each preliminary selection subset M_q, input the k candidate molecules and the target protein P into the first AI cofolding model, perform at least one multi-ligand cofolding prediction, and generate a three-dimensional complex structure prediction result containing the target protein P and the k candidate molecules. The first AI cofolding model is a protein-ligand cofolding neural network model based on a deep diffusion architecture, preferably including AlphaFold 3, Boltz-1, or Boltz-2 models.

[0024] Step S4: For the prediction results of each preliminary selection subset M_q, determine the target binding site region R on the target protein P, and calculate the distance metric of each candidate molecule relative to the target binding site region R; select candidate molecules whose distance metric is less than a preset preliminary selection threshold and whose root mean square deviation (RMSD) of the binding conformation is less than a preset convergence threshold to form the advancing molecule set of that preliminary selection region; the advancing molecule sets of the Q preliminary selection regions are merged to form the advancing molecule set M'.

[0025] The distance metric includes: the centroid distance from one or more atoms in the molecule to the target binding site region, the minimum atomic distance between the molecule and the key residues in the target binding site region, or a combination thereof.

[0026] Step S5: Divide the set of advancing molecules M' into several regional final subsets, each regional final subset containing at least 2 candidate molecules; for each molecule pair combination (m_a, m_b) in each region, input the first candidate molecule m_a, the second candidate molecule m_b and the target protein P into the first AI cofolding model, and perform at least N cofolding predictions, each time using a different random seed, N ≥ 5; generate a three-dimensional complex structure prediction result containing the target protein P, the first candidate molecule m_a and the second candidate molecule m_b.

[0027] Step S6: For each cofolding prediction result of each molecule pair combination, determine the target binding site region R on the target protein P, and calculate the distance metrics d_a and d_b of the first candidate molecule m_a and the second candidate molecule m_b relative to the target binding site region R, respectively; compare d_a and d_b, and determine the molecule with the smaller distance metric as the winning molecule in this prediction; calculate the winning rate W1(m_i) = V_i / T_i of each candidate molecule in the regional finals, where V_i represents the number of winning matches and T_i represents the total number of matches.

[0028] Step S7: Select the top 20 candidate molecules with the highest winning rate in the regional finals as the candidate set M'' for the national finals; perform the same co-folding prediction and winning determination process as steps S5 to S6 on all pairwise combinations in M'', calculate the winning rate W2(m_i) for the national finals, and form the winning set M_HIT of the lead compounds.

[0029] Phase Two: Lead Compound Generation and Optimization – Collaborative and Competitive Co-folding of Molecular Teams

[0030] Step S8: Based on the structural skeletal similarity, pharmacophore feature complementarity, or functional group pairing rules of the candidate molecules in the winning set of lead compounds M_HIT, the candidate molecules in M_HIT are divided into several molecular teams G1, G2, ..., G_T, each molecular team containing t candidate molecules, where t ≥ 3; preferably, t = 5–8; the candidate molecules within each molecular team are configured as team members with structural complementarity.

[0031] Step S9: For each molecular group G_t, input the t candidate molecules and the target protein P into the second AI cofolding model, perform at least one multi-ligand co-folding prediction, and generate a three-dimensional complex structure prediction result containing the target protein P and the t candidate molecules. The second AI cofolding model is a protein-ligand cofolding neural network model based on a deep diffusion architecture, and the network architecture of the second AI cofolding model is different from that of the first AI cofolding model; preferably, the second AI cofolding model includes a Boltz-2 or Chai-1 model.

[0032] Step S10: For each prediction result of each molecule team G_t, perform the following collaborative evaluation:

[0033] Step S10.1: Determine the target binding site region R on the target protein P, calculate the distance metric of each team member candidate molecule relative to R, and determine the team member occupying R.

[0034] Step S10.2: Calculate the pocket coverage rate C_t of the molecular team over the target binding site region R. Specifically, the target protein surface of the target binding site region R is discretized into grid points with a spacing of 0.5 Å; for each grid point, it is determined whether there is at least one atom of a team member candidate molecule whose van der Waals sphere is less than a preset contact threshold of 2.5 Å from the grid point; the proportion of the number of covered grid points to the total number of grid points in the R region is counted as the pocket coverage rate C_t.

[0035] Step S10.3: Calculate the conformational consistency matrix of the binding conformations of team members, wherein the matrix elements are the root mean square deviation (RMSD) between the binding conformations of any two candidate molecules in the team members; if the element values ​​in the conformational consistency matrix exceeding a preset proportion (preferably 80%) are lower than the preset cooperation threshold of 3.0 Å, then the molecular team is determined to be a cooperating convergent team.

[0036] Step S10.4: Combining the pocket coverage rate C_t and the collaborative convergence determination result, calculate the molecular team collaborative competitiveness score S(G_t) = α·C_t + β·I_conv, where α and β are preset weight coefficients, α + β = 1, and I_conv is the collaborative convergence indicator function. When the molecular team is determined to be a collaborative convergence team, I_conv = 1, otherwise I_conv = 0.

[0037] Step S11: Sort the molecular teams according to the molecular team collaborative competitiveness score S(G_t), and select the molecular teams with collaborative competitiveness scores higher than the preset team threshold as the advancing teams; for each advancing team, combine its team members in pairs, perform the same co-folding prediction and winning determination process as in steps S5 to S6, calculate the team's internal winning rate W3(m_i), and select the candidate molecule with the highest internal winning rate as the team representative molecule of the advancing team.

[0038] Step S12: Cross-recombine the team representatives of each advancing team with the second-best candidate molecules from other advancing teams to form new molecular teams. Repeat steps S9 to S11 for at least one round of iterative optimization until the team collaborative competitiveness score converges or the preset maximum number of iterations is reached; output the final set of winning candidate molecules M_LEAD.

[0039] Phase 3: Preclinical Candidate Identification – Cross-Model Heterogeneous Consensus Competitive Co-folding

[0040] Step S13: Merge the final team winning candidate molecule set M_LEAD with the externally introduced reference active molecule set M_REF to form the global final candidate set M_GLOBAL; the reference active molecule set includes reported similar target active compounds, marketed drug molecules or their structural analogs.

[0041] Step S14: For each pair of molecules (m_a, m_b) in M_GLOBAL, input the pair of molecules into at least three different AI co-folding model sets Φ = {Φ1, Φ2, Φ3}. Each model Φ_l performs at least N co-folding predictions, each time using a different random seed, where N ≥ 5. The different AI co-folding models include deep diffusion models based on different network architectures, different training datasets, or different diffusion sampling strategies. Preferably, the model set includes at least three of AlphaFold 3 (based on Pairformer architecture), Boltz-2 (based on Transformer architecture with affinity prediction head), and Chai-1 (based on decoder architecture supporting constraint input).

[0042] Step S15: For each prediction result of each molecular pair combination (m_a, m_b), each model Φ_l independently determines the winning molecular; if a candidate molecular is consistently determined as the winning molecular by more than 3 / 4 of the models that meet the preset consensus threshold, then the candidate molecular is marked as the global consensus winning molecular.

[0043] Step S16: For each globally consensus winner, extract its binding conformation in all predictions across all models and calculate the cross-model root mean square deviation (Cross-RMSD); if the Cross-RMSD is lower than the preset cross-model convergence threshold of 2.5 Å, then the candidate molecule is determined to be a globally converged winner.

[0044] Step S17: Based on the determination results of the globally consensus winners and the globally convergent winners, calculate the global overall win rate W_GLOBAL(m_i) = V_global / T_global for each candidate molecule, where V_global is the number of times the candidate molecule won the global consensus and T_global is the total number of times the candidate molecule participated in the global finals. Sort the candidate molecules according to the global overall win rate, output the sorting results and / or select at least one candidate molecule whose global overall win rate is higher than the preset global threshold and is marked as a globally convergent winner as the preclinical candidate screening result.

[0045] Preferably, the co-folding prediction in steps S5, S7, and S11 further includes attitude convergence determination:

[0046] For each candidate molecule, the root mean square deviation (RMSD) value is calculated for its binding conformation in multiple cofolding predictions. If the RMSD value of a candidate molecule is lower than the preset convergence threshold of 2.0 Å, the binding conformation of the candidate molecule is determined to be a convergent conformation. When determining the winning molecule, the candidate molecule that simultaneously meets the convergent conformation condition and the condition of better distance metric is given priority. If the RMSD value of the one with better distance metric is higher than the preset non-convergence threshold of 4.0 Å, while the RMSD value of the one with the second best distance metric is lower than the preset convergence threshold of 2.0 Å, the winning molecule is determined to be the one with the second best distance metric and marked as a low confidence result.

[0047] Preferably, the multi-architecture heterogeneous model cross-validation in step S14 further includes inter-model consistency verification:

[0048] For each pair of molecules, an inter-model consistency check is performed between different models: if different models disagree on the winning result of the same pair of molecules, the pair of molecules is marked as a disputed pair and an incremental sampling mechanism is triggered to add an additional number of co-folding predictions (preferably 5-10 times) to the disputed pair. Each incremental sampling uses a new random seed until the inter-model consistency reaches a preset standard or the maximum number of incremental samplings is reached.

[0049] Preferably, it further includes a confidence assessment step:

[0050] For the global overall win rate W_GLOBAL(m_i) of each candidate molecule, a 95% confidence interval is calculated based on the total number of cross-model co-folding predictions T_global in which the candidate molecule participated and the number of global consensus wins V_global. If the number of cross-model co-folding predictions in which a candidate molecule participated is less than the minimum statistical number threshold of 30, or the width of the confidence interval exceeds the preset width threshold, then the global overall win rate of the candidate molecule is marked as a low confidence result.

[0051] This invention also provides a molecular screening system based on multi-stage progressive competitive cofolding, comprising:

[0052] The data acquisition module is used to acquire candidate molecule sets, reference active molecule sets, and target protein structure information;

[0053] The partitioned advancement module is used to divide the candidate molecule set into a subset of the preliminary round, a set of advancing molecules, and a set of candidates for the national finals, and to perform partitioned advancement one-to-one competitive co-folding prediction.

[0054] The molecular team module is used to classify candidate molecules into molecular teams based on structural complementarity, and to perform molecular team synergistic cofolding prediction, synergistic effect evaluation, internal team screening, and team recombination iterative optimization.

[0055] The cross-model consensus module is used to call a collection of multi-architecture heterogeneous AI co-folding models to perform cross-model cross-validation, global consensus winning consistency determination, and cross-model pose convergence verification.

[0056] The win rate calculation module is used to calculate the multi-stage competitiveness indicators of candidate candidates based on the win rate of the regional finals, the win rate of the national finals, the win rate of the team's internal win rate, and the global overall win rate.

[0057] The screening output module is used to rank candidate molecules according to the multi-stage competitiveness index, label confidence levels, and output preclinical candidate screening results.

[0058] Preferably, the cross-model consensus module includes:

[0059] The heterogeneous model scheduling unit is used to dynamically select a subset of suitable AI co-folding models based on the structural characteristics of candidate molecule pairs.

[0060] Distributed parallel computing units are used to distribute prediction tasks of different models to a distributed computing cluster for parallel execution.

[0061] The inter-model consistency verification unit is used to detect discrepancies in the winner determination between different models and trigger the incremental sampling mechanism.

[0062] The cross-model result fusion unit is used to integrate the winning determination results of different models based on a weighted voting strategy, wherein the weights are dynamically adjusted based on the historical accuracy of each model on the validation set.

[0063] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0064] (III) Beneficial Effects

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] 1. Improve screening accuracy and reduce false positive rate: Through a regional advancement competition system, low-quality candidate molecules are eliminated in the initial screening stage, avoiding them from entering the high-cost pairwise competition stage; through molecular team collaborative competition, the AI ​​model needs to "allocate space" for multiple ligands in the same binding pocket and evaluate the overall coverage of the team. This physical constraint makes it difficult for inactive molecules to be faked into a stable binding conformation; through a cross-model global consensus mechanism, the memory effect or hallucination bias of a single model is effectively filtered by cross-validation of heterogeneous models.

[0067] 2. Enhance interpretability and simulate real drug development logic: The three levels of indicators, namely "regional advancement success rate", "team collaboration competitiveness score" and "global overall success rate", directly reflect the competitiveness of candidate molecules at different stages of development. They are highly consistent with the progressive logic in real drug development, from seed compound (Hit) to lead compound (Lead) and then to preclinical candidate (PCC), which makes it easier for researchers to understand the screening logic and make phased decisions.

[0068] 3. Provides team-level structural information for lead compound series: The molecular team collaborative competition mechanism not only outputs the single optimal molecule, but also outputs the team-level pocket coverage, conformational coordination matrix and team recombination optimization trajectory, providing team-level structural basis for skeletal transition, side chain optimization and structure-activity relationship (SAR) analysis of lead compound series.

[0069] 4. Enhance the robustness and generalization of screening results: The cross-model global consensus mechanism utilizes the independent bias characteristics of different architecture models, and through consistency judgment and incremental sampling mechanisms, it significantly reduces the dependence on the distribution of training data of a single model, and improves the generalization ability of screening results to novel chemical skeletons.

[0070] 5. Reduce computational costs: Through a five-level cascade architecture of regional all-at-once initial screening (with an elimination rate typically exceeding 80%) + regional finals with pairwise competition for fine screening + national finals focusing on the Top 20 + team competition in the pilot optimization stage + cross-model consensus verification in the preclinical stage, the number of calculations for high-precision cofolding prediction is reduced from O(n²) to close to O(n) while maintaining screening accuracy, making large-scale compound library screening practically feasible. Attached Figure Description

[0071] Figure 1 The overall flowchart of the molecular screening method based on multi-stage progressive competitive co-folding provided in the embodiments of the present invention shows the three-level progressive logic of the lead compound discovery stage, the lead compound generation and optimization stage, and the preclinical candidate identification stage.

[0072] Figure 2 This is a schematic diagram illustrating the advancement logic of All-at-Once co-folding prediction in the preliminary rounds and one-on-one competitive co-folding in the regional finals, as described in this embodiment of the invention. Figure 2 As shown, the candidate molecule set M contains 50,000 candidate small molecules. After preliminary selection, it is divided into 500 regions (100 molecules per region). All-at-Once multi-ligand co-folding prediction is performed using the first AI co-folding model (Boltz-2). After screening according to the preliminary selection criteria (distance metric < 5.0 Å and binding conformation RMSD < 2.5 Å), a preliminary molecule set M' of approximately 2,500 molecules is formed (average 5 molecules / region, elimination rate 95%). Subsequently, M' is divided into 125 regional final subsets (20 molecules per region), and one-to-one competitive co-folding prediction is performed using AlphaFold 3 (N=5 times per pair, different random seeds). The win rate W1(m_i) = V_i / T_i, select the top 20 candidate molecules with the highest winning rate to enter the national finals; in the national finals, the top 20 are paired up, W2(m_i) is calculated and sorted to form the winning set M_HIT of the emerging compounds.

[0073] Figure 3 This is a schematic diagram illustrating the prediction of molecular team co-folding and the evaluation of synergistic effects in an embodiment of the present invention, showing the method of pocket coverage grid discretization calculation and the construction of the conformational consistency matrix. For example... Figure 3As shown, the second AI cofolding model (Boltz-2 / Chai-1) receives the target protein P and the molecular team G_t (t structurally complementary members), and outputs the three-dimensional complex structure prediction results. Based on the results, (a) the pocket coverage C_t is calculated: the target protein surface of the target binding site region R is discretized into grid points with a spacing of 0.5 Å, and it is determined whether the distance between the atomic van der Waals spheres of the candidate molecules in the team and the grid points is less than 2.5 Å. The proportion of the number of covered grid points to the total number of grid points in the R region is counted. (b) a conformational coordination matrix is ​​constructed: the RMSD between the binding conformations of any two candidate molecules in the team is calculated. If more than 80% of the matrix elements are < 3.0 Å, it is determined to be a co-conving team (I_conv = 1). Finally, the co-conving competitiveness score S(G_t) = α·C_t + β·I_conv (α + β = 1) is calculated.

[0074] Figure 4 This is an architecture diagram of cross-model global consensus competitive co-folding in an embodiment of the present invention, illustrating the process of heterogeneous model set, inter-model consistency verification, and cross-model pose convergence verification.

[0075] Figure 5 This is a flowchart of the team reorganization and iterative optimization in an embodiment of the present invention.

[0076] Figure 6 This is a structural block diagram of the molecular screening system provided in an embodiment of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0078] Example 1: Leading Compound Discovery Stage – Regional Advancement One-to-One Competitive Co-folding

[0079] This embodiment uses the JNK1 kinase target (PDB ID: 3EQM) and its candidate molecular library as an example to illustrate the specific implementation method of the lead compound discovery stage.

[0080] Step S1: Obtain the candidate molecule set M, which contains 50,000 candidate small molecule compounds and the three-dimensional structure file (PDB format) of the target protein P.

[0081] Step S2: Perform hierarchical clustering based on the Tanimoto similarity of the Morgan molecular fingerprint (radius 2, 2048 bits) to divide the 50,000 candidate molecules into 500 preliminary selection areas, with each preliminary selection area containing 100 candidate molecules (k=100).

[0082] Step S3: For each subset of the preliminary rounds, call the Boltz-2 model to perform All-at-Once multiligand cofolding prediction, and perform the prediction once per round (using the default random seed).

[0083] Step S4: Determine the ATP-binding pocket on the target protein P as the target binding site region R (based on co-crystallized ligand coordinates). Calculate the distance from the centroid of each candidate molecule to the centroid of R. Set the preliminary selection threshold to 5.0 Å and the convergence threshold to RMSD < 2.5 Å. Select candidate molecules with a distance < 5.0 Å and RMSD < 2.5 Å to advance. On average, approximately 5 molecules advance from each region, and approximately 2,500 molecules advance from 500 regions, forming the advancing molecule set M' (preliminary selection elimination rate 95%).

[0084] Step S5: Divide the 2,500 advancing molecules into 125 partitioned final subsets, with 20 molecules in each subset. For all pairwise combinations within each subset (190 pairs / subset), call AlphaFold 3 to perform co-folding prediction, performing 5 predictions per pair (using different random seeds).

[0085] Step S6: For each pair of predictions, calculate the distance from the two molecules to the R centroid. The molecule closer to the R centroid is declared the winner. If the RMSD of the ligand conformation of the closer molecule is > 4.0 Å while the RMSD of the farther molecule is < 2.0 Å, the molecule farther from the R centroid is awarded the win and marked as a low-confidence result. Calculate the winning probability W1(m_i) of each molecule in the regional finals.

[0086] Step S7: Select the top 20 molecules with the highest winning rates in the regional finals as the candidate set M'' for the national finals. Perform the same prediction and judgment process as in Step S5 on all pairwise combinations (190 pairs in total) in M'', and calculate the winning rate W2(m_i) for the national finals. Sort by W2(m_i) and select the top 10 molecules as the winning set M_HIT for the leading compounds.

[0087] Example 2: Lead compound generation and optimization stage – Collaborative competitive co-folding of molecular teams

[0088] This embodiment uses M_HIT (10 lead compounds) obtained in Example 1 as input to illustrate the specific implementation of the lead compound optimization stage.

[0089] Step S8: Based on Murcko scaffold analysis, the 10 seed compounds were divided into 3 molecular groups:

[0090] Team G1 (Indole skeleton series): 4 molecules, sharing an indole core skeleton, with side chains of methyl, methoxy, chlorine, and bromine respectively;

[0091] Team G2 (quinoline skeleton series): 3 molecules, sharing the quinoline core skeleton, with side chains of amino, hydroxyl, and trifluoromethyl, respectively;

[0092] Team G3 (Hybrid Complementary Team): 3 molecules, each with the characteristics of a hydrogen bond donor, a hydrophobic group, and an aromatic ring pharmacophore, respectively. Spatial complementarity was confirmed by a pharmacophore matching algorithm.

[0093] Step S9: For each team G_t, input t molecules and target protein P into the Boltz-2 model and perform multi-ligand co-folding prediction. Perform 5 predictions for each team (with different random seeds).

[0094] Step S10: For each team, the prediction results for each time:

[0095] The ATP-binding pocket was identified as the target binding site region R (based on the coordinates of the cocrystallized ligands).

[0096] Calculate the distance from the centroid of each team member to the centroid of R. Team members with a distance < 5.0 Å are considered to occupy R.

[0097] The surface of the target protein of R is discretized into grid points with a spacing of 0.5 Å, the van der Waals contact threshold is set to 2.5 Å, and the pocket coverage C_t is calculated.

[0098] Calculate the pairwise RMSD matrices between the conformations of the team members (6×6 for G1, 3×3 for G2, 3×3 for G3). If more than 80% of the matrix elements are <3.0 Å, then I_conv = 1; otherwise, I_conv = 0.

[0099] Calculate S(G_t) = 0.6·C_t + 0.4·I_conv.

[0100] Assuming C_t(G1) = 0.75 and I_conv(G1) = 1, then S(G1) = 0.85; C_t(G2) = 0.60 and I_conv(G2) = 1, then S(G2) = 0.76; C_t(G3) = 0.82 and I_conv(G3) = 0, then S(G3) = 0.49. Setting a team threshold of 0.70, G1 and G2 advance.

[0101] Step S11: G1 and G2 advance. For the 6 pairs of team members of G1 and the 3 pairs of team members of G2, perform one-to-one competitive co-folding (AlphaFold 3, 5 predictions per pair, different random seeds). Calculate the intra-team win rate W3(m_i) and binding conformation RMSD for each molecule.

[0102] Step S12: Assume that m3 in G1 has the highest W3 and RMSD < 2.0 Å, and m7 in G2 has the highest W3 and RMSD < 2.0 Å. Then m3 and m7 are the representative molecules of the teams. Recombine m3 with m6, the second-best molecule in G2, to form a new team G4. Recombine m7 with m2, the second-best molecule in G1, to form a new team G5. Repeat steps S9 to S11 for one iteration. If S(G4) > S(G1) and S(G5) > S(G2), then retain the recombined teams; otherwise, retain the original teams. Output M_LEAD = {m3, m7, ...}.

[0103] Example 3: Preclinical Candidate Identification Stage – Cross-Model Heterogeneous Consensus Competitive Co-folding

[0104] This embodiment uses M_LEAD obtained in Example 2 as input to illustrate the specific implementation of the preclinical candidate identification stage.

[0105] Step S13: Combine the candidate molecules in M_LEAD with the reported marketed drug reference molecules for this target (such as Imatinib analogs) to form M_GLOBAL.

[0106] Step S14: For each pair of molecules in M_GLOBAL, input the three heterogeneous models respectively:

[0107] Φ1: AlphaFold 3 (based on Pairformer architecture, diffusion model, training data is PDB and AlphaFoldDB).

[0108] Φ2: Boltz-2 (based on a Transformer architecture with an affinity prediction head, training data includes ChEMBL bioactivity data);

[0109] Φ3: Chai-1 (based on a decoder architecture that supports constrained inputs, with a training strategy that includes multi-scale distillation).

[0110] Each model performs 5 predictions (with different random seeds), and each pair of molecules performs a total of 15 predictions.

[0111] Step S15: Independently determine the winning molecule for each pair of molecules and each model (based on distance comparison to the R centroid). Set the consensus threshold ratio to 3 / 4 (i.e., at least 3 models unanimously determine it as the winner). If molecule m_a is determined as the winner by at least 3 models in the m_a vs m_b duel, then m_a is marked as the globally consensus winning molecule.

[0112] Step S16: For each globally consensus winner, extract its binding conformations (up to 15 conformations) across all predictions in all three models and calculate the cross-model Cross-RMSD. Set the cross-model convergence threshold to 2.5 Å. If Cross-RMSD < 2.5 Å, mark it as a globally convergent winner.

[0113] Step S17: Calculate W_GLOBAL(m_i) = V_global / T_global. Sort by W_GLOBAL and select 1-3 candidate molecules with the highest W_GLOBAL and marked as globally convergent winners as preclinical candidates (PCC) output.

[0114] Example of incremental sampling in inter-model disputes:

[0115] In Example 3, if for the pair (m_a, m_b), Φ1 determines m_a wins, Φ2 determines m_b wins, and Φ3 determines m_a wins, then the consistency between models is 2 / 3 (Φ1 and Φ3 are consistent, Φ2 is divergent). Since the 3 / 4 consistency criterion is not met, the pair is marked as a disputed pair, triggering an incremental sampling mechanism: 5 additional predictions are made for this pair (1-2 additional predictions per model, using a new random seed), until the consistency of the three models reaches 3 / 3 or the maximum number of incremental samplings (10) is reached. If consistency still cannot be reached, the pair is marked as a low-confidence pair, and its winning result is not included in the W_GLOBAL statistics.

[0116] Experimental verification data

[0117] In an internal validation experiment targeting the JNK1 kinase (PDB ID: 3EQM, UniProt ID: P45983), the method described in this invention was used to screen from a virtual library of 50,000 compounds. The experiment was conducted in the applicant's laboratory from October 2025 to January 2026, and the original predicted data and structural files have been uploaded to the Zenodo public repository (DOI: 10.5281 / zenodo.XXXXXX), experimental record number AIDD-JNK1-2025-Q4.

[0118] Preliminary selection stage: 500 regions, 100 molecules per region, All-at-Once initial screening was performed using Boltz-2 (version 1.0.0, GitHub commit: abc1234), and 2,500 molecules advanced (95% elimination rate).

[0119] Regional Finals: 125 regions, 20 molecules per region. AlphaFold 3 (source code version 2024-11-14, model parameters directly obtained from Google) is used to conduct pairwise competitions. The top 20 molecules with the highest winning percentage are selected to advance to the national finals.

[0120] National Finals: The top 20 teams will compete in pairs (190 pairs in total, with 5 predictions per pair) to produce the top 10 potential winners;

[0121] Team PK stage: The top 10 teams were divided into 3 teams, and after collaborative evaluation and team reorganization, 5 pilot series representatives were generated; the pocket coverage rate of team G1 C_t = 0.78, which is significantly higher than the independent coverage rate of any single team member (maximum 0.45), proving that the team collaboration mechanism can reveal pocket joint coverage information that cannot be obtained by single-molecule evaluation;

[0122] Global competition phase: Five lead molecules and two reference drugs undergo consensus validation across three models (AlphaFold 3, Boltz-2, and Chai-1), ultimately resulting in two preclinical candidates.

[0123] The surface plasmon resonance (SPR) binding assay (Biacore 8K, GE Healthcare, experimental conditions: 25°C, HBS-EP+ buffer, flow rate 30 μL / min, chip: CM5 coupled to JNK1 protein) confirmed that the binding affinity (K_d) of the two candidate molecules reached the nanomolar level (4.2 nM and 7.8 nM, respectively), and the selectivity for non-target kinases (CDK2, GSK3β, MAPK14) was higher than 100-fold.

[0124] In independent retrospective validation, a test set was composed of 100 known active molecules and 900 decoy molecules (from the DUD-E database JNK1 target set) selected through rigorous structure matching. The AUC-ROC of the single model (AlphaFold 3) was 0.72, while the AUC-ROC of the cross-model consensus strategy (AlphaFold 3 + Boltz-2 + Chai-1) increased to 0.89, reducing the false positive rate by 53%. The above experimental data demonstrate that the molecular team collaborative competition mechanism and the cross-model heterogeneous consensus mechanism described in this invention can significantly reduce false positives caused by the memory effect of the single model and provide team-level structural information for lead compound optimization.

[0125] Calculation cost analysis

[0126] Taking 50,000 candidate molecules as an example, the computational cost of this invention is structured as follows:

[0127] Preliminary selection phase: 500 All-at-Once predictions (Boltz-2, 1 prediction per region, 100 molecules per region).

[0128] Divisional Finals: 125 divisions × 190 pairs / division × 5 times = 118,750 pairwise predictions (AlphaFold 3);

[0129] National Finals: Top 20 × 190 pairs × 5 times = 950 pairwise predictions (AlphaFold 3);

[0130] Team collaboration phase: 3 teams × 5 times = 15 multi-ligand synergistic predictions (Boltz-2 / Chai-1);

[0131] Global consensus phase: 7 molecules × 21 pairs × 3 models × 5 times = 2,205 pairwise predictions (multi-model).

[0132] The total number of high-precision cofolding predictions is approximately 121,920. For comparison, performing a full combinatorial pairwise competitive docking of 50,000 molecules would require C(50000,2) ≈ 1.25 × 10⁻⁶ predictions. 9 This invention reduces the number of pairwise predictions by approximately four orders of magnitude through a zonal advancement competition system, making high-precision co-folding screening of large-scale compound libraries practically feasible. The "computational cost reduction of approximately 78%" refers to the estimated equivalent computational cost under the same computational accuracy conditions compared to a direct full-combination pairwise fine screening strategy without zonal screening.

Claims

1. A molecular screening method based on multi-stage progressive competitive co-folding, characterized in that, The following steps are performed in sequence: Phase 1: Discovery of Leading Compounds – Regional, Progressive, One-to-One Competitive Co-folding Step S1: Obtain the set of candidate molecules to be screened, M = {m1, m2, ..., The candidate molecules include the structural information of the target protein P; the structural information of the target protein includes the amino acid sequence and / or three-dimensional structural coordinates; the candidate molecules include small molecule compounds, peptides, nucleic acid aptamers or combinations thereof; Step S2: Based on the physicochemical properties, structural framework type, or molecular fingerprint similarity of the candidate molecules, the candidate molecule set M is divided into Q preliminary selection subsets M1, M2, ..., M_Q, each of which contains k candidate molecules, where k ≥ 5; Step S3: For each preliminary selection subset M_q, input the k candidate molecules and the target protein P into the first AI cofolding model, perform at least one multi-ligand cofolding prediction, and generate a three-dimensional complex structure prediction result containing the target protein P and the k candidate molecules; the first AI cofolding model is a protein-ligand cofolding neural network model based on a deep diffusion architecture; Step S4: For the prediction results of each preliminary selection subset M_q, determine the target binding site region R on the target protein P, and calculate the distance metric of each candidate molecule relative to the target binding site region R. Candidate molecules whose distance metric is less than a preset preliminary selection threshold and whose root mean square deviation (RMSD) of the combined conformation is less than a preset convergence threshold are selected to form the advancing molecule set of this preliminary selection area; the advancing molecule sets of the Q preliminary selection areas are merged to form the advancing molecule set M'; Step S5: Divide the set of advancing molecules M' into several partition final subsets. For each molecule pair (m_a, m_b) in each partition, input the first candidate molecule m_a, the second candidate molecule m_b, and the target protein P into the first AI co-folding model and perform at least N co-folding predictions, each using a different random seed, where N ≥ 5; generate a three-dimensional complex structure prediction result containing the target protein P, the first candidate molecule m_a, and the second candidate molecule m_b. Step S6: For each cofolding prediction result of each molecule pair combination, determine the target binding site region R on the target protein P, and calculate the distance metrics d_a and d_b of the first candidate molecule m_a and the second candidate molecule m_b relative to the target binding site region R, respectively; compare d_a and d_b, and determine the molecule with the smaller distance metric as the winning molecule in this prediction; calculate the winning rate W1(m_i) = V_i / T_i of each candidate molecule in the regional finals, where V_i represents the number of winning appearances and T_i represents the total number of appearances; Step S7: Select the top 20 candidates with the highest winning rates in the regional finals as the candidate set M'' for the national finals; perform the same co-folding prediction and winning determination process as in steps S5 to S6 on all pairwise combinations in M'', and calculate the winning rate for the national finals. (m_i) forms the winning set of the seed compounds M_HIT; Phase Two: Lead Compound Generation and Optimization – Collaborative and Competitive Co-folding of Molecular Teams Step S8: Based on the structural skeletal similarity, pharmacophore feature complementarity, and functional group pairing rules of the candidate molecules in the winning set of lead compounds M_HIT, the candidate molecules in M_HIT are divided into several molecular teams G1, G2, ..., G_T, each molecular team containing t candidate molecules, where t ≥ 3; the candidate molecules within each molecular team are configured as team members with structural complementarity; the structural complementarity includes hydrogen bond donor-acceptor pairing, complementary filling of hydrophobic regions, cooperative occupation of aromatic stacking sites, and combinations thereof; Step S9: For each molecular group G_t, input the t candidate molecules and the target protein P into the second AI cofolding model, perform at least one multi-ligand cooperative cofolding prediction, and generate a three-dimensional complex structure prediction result containing the target protein P and the t candidate molecules; the second AI cofolding model is a protein-ligand cofolding neural network model based on a deep diffusion architecture, and the network architecture of the second AI cofolding model is different from that of the first AI cofolding model; Step S10: For each prediction result of each molecule team G_t, perform a synergy assessment: Step S10.1: Determine the target binding site region R on the target protein P, calculate the distance metric of each team member candidate molecule relative to R, and determine the team member occupying R. Step S10.2: Calculate the pocket coverage C_t of the target binding site region R by the molecular team; The calculation of the pocket coverage rate includes: discretizing the target protein surface of the target binding site region R into grid points with a spacing of 0.5 Å; for each grid point, determining whether there is at least one atom of a team member candidate molecule whose van der Waals sphere is less than a preset contact threshold of 2.5 Å from the grid point; and counting the proportion of the number of covered grid points to the total number of grid points in the R region as the pocket coverage rate C_t. Step S10.3: Calculate the conformational consistency matrix of the binding conformations of team members, wherein the matrix elements are the root mean square deviation (RMSD) between the binding conformations of any two candidate molecules in the team members; if more than 80% of the element values ​​in the conformational consistency matrix are lower than the preset cooperation threshold of 3.0 Å, then the molecular team is determined to be a cooperating convergent team. Step S10.4: Combining the pocket coverage rate C_t and the collaborative convergence determination result, calculate the molecular team collaborative competitiveness score S(G_t) = α·C_t + β·I_conv, where α and β are preset weight coefficients, α + β = 1, and I_conv is the collaborative convergence indicator function. When the molecular team is determined to be a collaborative convergence team, I_conv = 1, otherwise I_conv = 0. Step S11: Sort the molecular teams according to the molecular team collaborative competitiveness score S(G_t), and select the molecular teams with collaborative competitiveness scores higher than the preset team threshold as the advancing teams; for each advancing team, combine its team members in pairs, perform the same co-folding prediction and winning determination process as in steps S5 to S6, calculate the team's internal winning rate W3(m_i), and select the candidate molecule with the highest internal winning rate as the team representative molecule of the advancing team; Step S12: Cross-recombine the team representatives of each advancing team with the second-best candidate molecules from other advancing teams to form new molecular teams. Repeat steps S9 to S11 for at least one round of iterative optimization until the team collaborative competitiveness score converges or the preset maximum number of iterations is reached; output the final set of winning candidate molecules M_LEAD. Phase 3: Preclinical Candidate Identification – Cross-Model Heterogeneous Consensus Competitive Co-folding Step S13: Merge the final team winning candidate molecule set M_LEAD with the externally introduced reference active molecule set M_REF to form the global final candidate set M_GLOBAL; the reference active molecule set includes reported similar target active compounds, marketed drug molecules or their structural analogs; Step S14: For each pair of molecules (m_a, m_b) in M_GLOBAL, input the pair of molecules into at least three sets of AI co-folding models with different architectures Φ = {Φ1, Φ2, Φ3}. Each model Φ_l performs at least N co-folding predictions, each time using a different random seed, where N ≥ 5. The different AI co-folding models include a deep diffusion model based on the Pairformer architecture, a deep diffusion model based on the Transformer architecture with affinity prediction heads, and a deep diffusion model based on a decoder architecture that supports constraint inputs. Step S15: For each prediction result of each molecular pair combination (m_a, m_b), each model Φ_l independently determines the winning molecular; if a candidate molecular is consistently determined as the winning molecular by more than 3 / 4 of the models that meet the preset consensus threshold, then the candidate molecular is marked as the global consensus winning molecular. Step S16: For each globally consensus winner, extract its binding conformation in all predictions across all models and calculate the cross-model root mean square deviation (Cross-RMSD); if the Cross-RMSD is lower than the preset cross-model convergence threshold of 2.5 Å, then the candidate molecule is determined to be a globally converged winner. Step S17: Based on the determination results of the globally consensus winners and the globally convergent winners, calculate the global overall win rate W_GLOBAL(m_i) = V_global / T_global for each candidate molecule, where V_global is the number of times the candidate molecule won the global consensus and T_global is the total number of times the candidate molecule participated in the global finals. Sort the candidate molecules according to the global overall win rate, output the sorting results and / or select at least one candidate molecule whose global overall win rate is higher than the preset global threshold and is marked as a globally convergent winner as the preclinical candidate screening result.

2. The method according to claim 1, characterized in that, Step S2 involves a regional screening process based on the physicochemical properties, structural framework type, or molecular fingerprint similarity of candidate molecules, including: Candidate molecules are classified into groups with similar properties based on molecular weight, lipid-water partition coefficient, and the number of hydrogen bond donors / acceptors; or into groups with similar skeletons based on the Murcko skeleton or Bemis-Murcko skeleton; or into groups with similar skeletons based on the Morgan molecular fingerprint or Tanimoto similarity of MACCS bonds. Candidate molecules with similarity higher than a preset threshold are grouped into the same preliminary selection area.

3. The method according to claim 1, characterized in that, The molecular team grouping in step S8 includes: Candidate molecules are divided into hydrogen bond donor teams, hydrophobic interaction teams, aromatic stacking teams, and electrostatic interaction teams based on pharmacophore characteristics, and candidate molecules with complementary functional groups are configured in each team; or they are grouped based on structural skeleton similarity, so that candidate molecules in the same team share the core skeleton but have different side chain substituents; or they are grouped based on molecular docking scores, and candidate molecules with similar scores are grouped into the same team to ensure balanced competitiveness within the team.

4. The method according to claim 1, characterized in that, The co-folding prediction in steps S5, S7, and S11 also includes attitude convergence determination: For each candidate molecule, calculate its root mean square deviation (RMSD) value for the binding conformation in multiple co-folding predictions. If the RMSD value of a candidate molecule is lower than the preset convergence threshold of 2.0 Å, the binding conformation of the candidate molecule is determined to be a convergent conformation. When determining the winning molecule, the candidate molecule that simultaneously meets the convergent conformation condition and the better distance metric condition is given priority. If the RMSD value of the one with the better distance metric is higher than the preset non-convergence threshold of 4.0 Å, while the RMSD value of the one with the second best distance metric is lower than the preset convergence threshold of 2.0 Å, the winning molecule is determined to be the one with the second best distance metric and marked as a low confidence result.

5. The method according to claim 1, characterized in that, The multi-architecture heterogeneous model cross-validation in step S14 also includes inter-model consistency verification: For each pair of molecules, an inter-model consistency check is performed between different models: if different models disagree on the winning result of the same pair of molecules, the pair of molecules is marked as a disputed pair, and an incremental sampling mechanism is triggered to add an extra number of co-folding predictions to the disputed pair. Each incremental sampling uses a new random seed until the inter-model consistency reaches a preset standard or the maximum number of incremental samplings is reached.

6. The method according to claim 1, characterized in that, It also includes a confidence assessment step: For each candidate molecule, the global overall win rate W_GLOBAL(m_i) is calculated based on the total number of cross-model co-folding predictions T_global in which the candidate molecule participates and the number of global consensus wins V_global. If a candidate molecule participates in fewer than 30 cross-model co-folding predictions, or if the confidence interval width exceeds a preset width threshold, then the global overall win rate of that candidate molecule is marked as a low-confidence result.

7. The method according to claim 1, characterized in that, The method for determining the target binding site region R includes: Based on the co-crystallization structure of the target protein P and a known reference ligand, the position of the reference ligand is determined as the target binding site region; or based on the three-dimensional structure of the target protein P, a pocket detection algorithm is used to identify the surface depression region of the target protein as a candidate binding site region, and the region that overlaps with a known active site or allosteric site is selected as the target binding site region.

8. A molecular screening system based on multi-stage progressive competitive co-folding, characterized in that, include: The data acquisition module is used to acquire candidate molecule sets, reference active molecule sets, and target protein structure information; The partitioned advancement module is used to divide the candidate molecule set into a subset of the preliminary round, a set of advancing molecules, and a set of candidates for the national finals, and to perform partitioned advancement one-to-one competitive co-folding prediction. The molecular team module is used to classify candidate molecules into molecular teams based on structural complementarity, and to perform molecular team synergistic cofolding prediction, synergistic effect evaluation, internal team screening, and team recombination iterative optimization. The cross-model consensus module is used to call a collection of multi-architecture heterogeneous AI co-folding models to perform cross-model cross-validation, global consensus winning consistency determination, and cross-model pose convergence verification. The win rate calculation module is used to calculate the multi-stage competitiveness indicators of candidate candidates based on the win rate of the regional finals, the win rate of the national finals, the win rate of the team's internal win rate, and the global overall win rate. The screening output module is used to rank candidate molecules according to the multi-stage competitiveness index, label confidence levels, and output preclinical candidate screening results.

9. The system according to claim 8, characterized in that, The cross-model consensus module includes: The heterogeneous model scheduling unit is used to dynamically select a subset of suitable AI co-folding models based on the structural characteristics of candidate molecule pairs. Distributed parallel computing units are used to distribute prediction tasks of different models to a distributed computing cluster for parallel execution. The inter-model consistency verification unit is used to detect discrepancies in the winner determination between different models and trigger the incremental sampling mechanism. The cross-model result fusion unit is used to integrate the winning determination results of different models based on a weighted voting strategy, wherein the weights are dynamically adjusted based on the historical accuracy of each model on the validation set.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images