Polymer chain automatic construction and stable conformation generation method based on distributed isomerism
By automatically completing the topological assembly and stable conformation generation of polymer chains through a distributed heterogeneous computing method, the problems of low efficiency and poor scalability in existing technologies are solved, and high-throughput screening and reproducible stable conformation generation are achieved.
Patent Information
- Application Number
- CN202511486507.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-02-03
AI Technical Summary
Existing technologies rely on expert experience and experimental methods in the process of polymer molecular conformation generation and stabilization, which are inefficient and costly, making it difficult to achieve high-throughput screening and topology construction, and difficult to achieve elastic scaling and load balancing in multi-node, cross-center, and CPU+GPU heterogeneous environments.
Employing a distributed heterogeneous computing approach, the system receives task information via a message queue and automatically completes topology assembly, force field and charge parameter assignment, molecular dynamics minimization, and stabilization evaluation. It supports multi-node collaborative computing and uses a three-stage relaxation strategy of short-range NVT preheating, multi-round annealing, and final-state NVT balancing to generate stable conformations.
It achieves high-throughput stable conformation generation, supports high-throughput screening of thousands to tens of thousands of candidate formulations/topologies, generates conformations with strong physical consistency, meets the needs of material data-driven modeling, the results are reproducible, the task success rate is stable, and it solves the problems of low efficiency and poor scalability in traditional methods.
Smart Images
Figure CN121459994A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a method for automatically constructing polymer chains and generating stable conformations based on distributed heterogeneity, belonging to the field of computational chemistry and high-performance computing. BACKGROUND
[0002] The generation and stabilization of polymer molecular conformations is a key link in the design and performance prediction of high polymer materials. Stable and physically reliable three-dimensional conformations are directly related to glass transition temperature (Tg), mechanical modulus, dielectric constant, permeability, and swelling behavior, etc. For shape memory polymers, self-healing polymers and other intelligent polymer systems, the mapping of conformation-performance is more sensitive and nonlinear. At present, the construction of stable conformations still largely relies on expert inference based on empirical rules, supplemented by experimental methods such as electron microscopy, scattering, and NMR for verification. This process is time-consuming, labor-intensive, and costly, and it is difficult to support high-throughput screening of thousands to tens of thousands of candidate formulations / topologies.
[0003] To address the above challenges, the academic and industrial communities have explored multiple types of computational methods: one type is a material informatics / heuristic rule-driven constructor (based on manual templates and local rules to automatically splice segments), which can quickly give a usable initial state, but is limited by expert knowledge and template coverage, making it difficult to generate sufficient conformation diversity; another type is molecular dynamics (MD) simulation, which obtains stable structures through energy minimization and short-range dynamics, with stronger physical consistency, but the traditional workflow requires researchers to manually complete topology modeling, force field parameter assignment, charge calculation, and input script writing, which is time-consuming and requires high skills. More critically, existing tool chains cannot achieve elastic expansion and load balancing in a multi-node, cross-center, CPU+GPU heterogeneous environment, and I / O and scheduling overheads often become bottlenecks, making it difficult to provide scalable, fine-grained, and reproducible conformation samples for deep generative models or material screening platforms at a reasonable cost.
[0004] Therefore, there is an urgent need for an end-to-end method that can automatically complete topology assembly, force field and charge assignment, molecular dynamics minimization and stability evaluation starting from monomer SMILES; this method should natively support distributed heterogeneous computing and mixed precision strategies, with high throughput, low human intervention, strong reproducibility, and scalability to meet the needs of high-throughput polymer screening and data-driven modeling. SUMMARY
[0005] The present application is to solve the problem of relying on conventional empirical inference and experimental means for verification in the calculation of stable conformation of polymer segments, which is low in efficiency, high in cost, and difficult to form screening support for a large number of candidate formulations / topologies, and further proposes a method for automatically constructing polymer chains and generating stable conformations based on distributed heterogeneity.
[0006] The technical scheme adopted by the present application to solve the above problems is as follows: Step 1: receiving task information from a message queue, forwarding or calculating the incoming task according to the affinity information in the task information and the current load status, and configuring a dead letter queue for archiving and counting failed information; Step 2: parsing the monomer PSMILES and topological rules in the task information, identifying anchor atoms, leaving groups and protecting groups, calculating the initial conformation of the monomer based on the specified force field type, and calculating the atomic charge distribution by quantum chemical method; Step 3: combining the topological rule library to set the connection mode of the initial conformation of the monomer, aligning according to the anchor points, constructing a local coordinate system for each anchor point, merging redundant atoms, and splicing the initial conformation of the monomer into an initial segment of a specified length; Step 4: automatically generating an input file for a molecular dynamics simulation tool and verifying constraint conditions such as charge conservation; Step 5: sequentially performing short-range NVT preheating, annealing and final NVT dynamics relaxation to obtain a stable segment conformation; Step 6: dividing the stable segment conformation into multiple monomers according to the anchor points, and aligning the divided monomers to the local coordinates to calculate the rigid body transformation of the monomers to the chain; Step 7: checking the division result, completing task callback, and uploading the output to a distributed storage platform.
[0007] Further, the task message received in step 1 is in JSON format, and the task message includes fields, transaction ID and affinity information. The transaction ID is used for end-to-end tracking, and the fields include but are not limited to monomer PSMILES, target aggregation degree and topological rules, force field type, simulation temperature / pressure / time step, relaxation phase configuration, convergence threshold, random seed, node affinity, callback URL and data storage strategy. The affinity information includes GPU / CPU load and precision information.
[0008] Further, step 2 specifically includes: The valence state of the monomer PSMILES in the task information is standardized and the aromaticity is checked. If the check fails, the task is directly rejected and transmitted to the dead letter queue, and the error log is counted. After the check is completed, the PSMILES rules are used to identify anchor atoms, leaving groups and protecting groups, and a local coordinate system is constructed for each anchor point. In the constructed local coordinate system, two connection point atoms and one neighbor atom in the monomer PSMILES are set as Star1, Neighbor1, Star2, Neighbor2, and the Neighbor1 atom is set as the coordinate origin. The calculation of the initial conformation of the monomer is completed; The reasonable monomer initial conformation is screened using the fast molecular mechanics optimization in the rdkit tool, wherein the fast molecular mechanics optimization includes UFF / MMFF pre-drop, charge distribution using the PSI4 tool, DFT calculation of electrostatic potential, and generation of RESP / AM1-BCC charges.
[0009] Further, step 3 specifically includes: The connection mode of the monomer initial conformation is set in combination with a topology rule library, wherein the connection mode includes but is not limited to linear and block connection, a plurality of monomers are sampled by fixed length or user-defined polymerization degree distribution, rigid body alignment is performed based on the normal and tangential directions of the anchor point, redundant atoms and bonds are combined, and boundary bond orders are repaired; heuristic energy scanning is used for the rotatable bond with high tension to eliminate unreasonable conformations, initial segments of the target polymerization degree are generated, and the simulation box boundary conditions are set according to the simulation parameters provided by the user.
[0010] Further, step 4 specifically includes: The data and input scripts required by the molecular dynamics simulation tool are automatically generated, wherein the generated data and input scripts include but are not limited to atom types, bond angle dihedral angle parameters, non-bonding interactions, neighbor list and thermodynamic output; and before running the simulation power, the charge conservation, valence consistency, geometric extreme value and initial energy are checked, if the check fails, the task is rejected, and the task is transmitted to the dead letter queue.
[0011] Further, step 5 specifically includes: The molecular dynamics simulation tool is run, and short-range NVT preheating, multiple rounds of annealing temperature rise and fall, and terminal NVT equilibrium are sequentially performed, high-energy contacts of the monomer initial conformation are eliminated through NVT preheating, local energy barriers are crossed through multiple rounds of annealing temperature rise and fall, and stable conformations are obtained through terminal NVT equilibrium at the target temperature.
[0012] Further, step 6 specifically includes: For the triplet (s1, n1, s2) composed of Star1 atom, Neighbor1 atom and Star2 atom in the monomer PSMILES, n1 is set as the origin of the local coordinate system, two vectors starting from n1 are obtained and and the standard coordinate system and the rigid body transformation matrix from the monomer local coordinates to the corresponding part on the segment after relaxation are obtained through GramSchmidt normalization The monomer PSMILES after coordinate localization are respectively stored in the corresponding SDF file, and the corresponding relaxed rigid body transformation matrix is saved and stored in the numpy array.
[0013] Further, step 7 specifically includes: Read the SDF file of the local coordinates of the storage unit and the rigid body transformation matrix, reassemble the polymer chain, and use rdkit to compare with the original polymer chain, if the RMSD threshold is higher than the specified value, the result is rejected, then the corresponding task is stored in the dead letter queue, if it is lower than, it is uploaded to the distributed storage platform and the information callback is confirmed.
[0014] The beneficial effects of the present application are: 1. The present application can reach about 3500 conformation tasks per day in a single card 4090 node (double floating point precision, 16 core 32 thread CPU) environment, far exceeding the traditional manual modeling and single machine MD simulation of hundreds of processing capacity, and can quickly complete the conformation calculation of large-scale single sample.
[0015] 2. The present application supports multi-remote node collaborative computing, realizes intelligent task distribution through RabbitMQ message queue, and collects results through MinIO distributed storage, effectively solves the cross-node I / O and scheduling overhead bottleneck, meets the high-throughput screening demand of thousands to tens of thousands of candidate formulations / topologies, and fills the gap of traditional tool chain which is difficult to scale computing.
[0016] 3. Compared with the traditional modeling relying on expert experience, the present application stabilizes the task success rate at 68% through full-process automatic verification such as PSMILES valence state / aromaticity inspection, charge conservation verification, geometric extreme verification, RMSD result comparison, and at the same time, through the dead letter queue, the failed tasks are archived and counted, which is convenient for problem tracing and process optimization.
[0017] 4. The present application adopts the three-stage relaxation strategy of short-range NVT preheating, multiple rounds of annealing and final state NVT balance, combines DFT calculation of atomic charge and heuristic energy scanning to eliminate high-tension conformations, and generates stable conformations with strong physical consistency, which can be directly associated with glass transition temperature, mechanical modulus and other macroscopic properties, avoiding the problems of insufficient conformation diversity and lack of physical meaning in traditional template splicing method.
[0018] 5. The present application realizes full-process automation, without the need for experts to manually complete steps such as topology modeling, charge calculation and script writing, and through transaction ID association of the whole life cycle of the task, combined with distributed storage to save intermediate results, ensures that the generation process of each stable conformation is traceable and the results are reproducible, meeting the strict requirements of material data-driven modeling for sample reproducibility. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 The flowchart of the polymer chain automatic construction and stable conformation generation method based on distributed heterogeneous architecture is shown. Figure 2 The framework diagram of the polymer chain automatic construction and stable conformation generation method based on distributed heterogeneous architecture is shown. Figure 3 This is a schematic diagram showing the correspondence between rigid body transformation matrices and individual PSMILES. Detailed Implementation
[0020] Combination Figures 1-3 This implementation method is described as follows: Figure 2 As shown, in this embodiment, the framework of the polymer chain automatic construction and stable conformation generation method includes a distributed message queue, a distributed storage structure, a chain segment conformation generation component, and chain segment conformation relaxation calculation results. Figure 1 As shown, the steps of the method for automatic construction and stable conformation generation of polymer chains based on distributed heterogeneity described in this embodiment include: S1: Receive task information, forward or calculate tasks based on the task information, and configure a dead letter queue for archiving and statistical analysis of failed information. The system receives task messages from message queues, such as RabbitMQ. The messages are in JSON format and include fields such as singleton PSMILES, target aggregation degree and topology, force field type, simulated temperature / pressure / time step, relaxation phase configuration, convergence threshold, random seed, node affinity, callback URL and data storage strategy, and are accompanied by a transaction ID for end-to-end tracking.
[0021] Based on affinity information attached to the message, such as GPU / CPU load, accuracy information, and current load status, the node determines whether to perform computation or forward the incoming task, and configures a dead letter queue to archive and statistically analyze failed messages.
[0022] S2: Analyze task information to identify anchor atoms, leaving groups and protecting groups, calculate the initial conformation of monomers based on the specified force field type, and calculate atomic charge distribution through quantum chemical methods; The system performs valence normalization and aromaticity specification checks on the input PSMILES. Non-compliant inputs are directly rejected, sent to the dead-letter queue, and recorded in the error log. Subsequently, the PSMILES rules are used to identify anchor atoms, leaving groups, and protecting groups, and a local coordinate system is constructed for each anchor point. The two connection point atoms in the PSMILES and one of their neighbor atoms are designated as Star1, Neighbor1, Star2, and Neighbor2, respectively, with Neighbor1 atom set as the origin.
[0023] Reasonable initial states were screened using the rapid molecular mechanics optimization (UFF / MMFF pre-descent) tool in rdkit, and charge distribution was performed using the PSI4 tool. Electrostatic potential was calculated using DFT (such as B3LYP / 6-31G*) and RESP / AM1-BCC charge was generated.
[0024] S3: Registration according to anchor points, merging redundant atoms, splicing into initial chain segments of specified length; In combination with the topology rule library, linear, block and other connection modes are declared, and multiple monomers are sampled through fixed-length or user-defined aggregation degree distribution, and rigid body alignment is performed based on anchor point normal and tangent, redundant atoms and bonds are merged, and boundary bond order is repaired. For high-tension rotatable bonds, heuristic energy scanning is used to eliminate unreasonable conformations. Then the initial chain segment of the target aggregation degree is generated, and the simulation box boundary conditions are set according to the simulation parameters provided by the user.
[0025] S4: Automatically generate input files for molecular dynamics simulation tools and verify constraints such as charge conservation; Automatically generate data and input scripts required by molecular dynamics simulation tools, in this embodiment, the molecular dynamics simulation tool selected is LAMMPS / OPENMM, covering atom type, bond angle dihedral angle parameters, non-bonding interaction, neighbor list and thermodynamic output, etc. The system verifies charge conservation, valence consistency, geometric extreme (such as the shortest non-bonding distance and the maximum bond stretching) and initial energy before running, and if the verification fails, the task is rejected and the task is transferred to the dead letter queue.
[0026] S5: Sequentially perform short-range NVT preheating, annealing and final NVT dynamics relaxation to obtain stable chain segment conformation; Molecular dynamics relaxation conformation is divided into three stages: first, short-range NVT preheating (e.g. 10-50 ps, 50-100 K to target temperature) is performed to eliminate high-energy contacts; then, through multiple rounds of annealing temperature rise and fall (e.g. 300 ↔ 600 K, single round 20-50 ps) to cross local energy barriers; finally, at the target temperature, the final NVT equilibrium (50-200 ps) is performed to obtain the stable conformation.
[0027] S6: Divide the chain segment into multiple monomers according to the anchor points, align to the local coordinates, and calculate the rigid body transformation of the monomer to the chain; For each monomer, the embodiment has a three-tuple (s1, n1, s2) consisting of Star1 atom, Neighbor1 atom, and Star2 atom. We set n1 as the origin of the local coordinate system, and obtain two vectors and from n1 as the starting point, and through GramSchmidt normalization, we get the standard coordinate system and the rigid body transformation matrix from the local coordinates of the monomer to the corresponding part of the chain segment The corresponding relationship between the rigid body transformation matrix and the PSMILES of the monomer is shown in Figure 3 . The monomers after coordinate localization are stored in the corresponding SDF file, and the corresponding rigid body transformation after relaxation is stored in the numpy array.
[0028] S7: Check the division result, complete the task callback, and upload the output to the distributed storage platform; Read in the previous monomer local coordinate alignment SDF and the rigid body transformation matrix, reassemble the polymer chain, and use rdkit to compare with the original polymer chain. If the RMSD threshold is higher than the specified threshold, the result is rejected and sent to the dead letter queue. If the result is correct, it is uploaded to the distributed storage platform (such as MinIO) and the information callback is confirmed.
[0029] After completing the above polymer chain automatic construction and stable conformation generation method setting, to verify the effect of the polymer chain automatic construction and stable conformation generation method, the present embodiment carries out the following simulation verification: Dataset: The present embodiment calculates the stable conformation based on the PSMILES provided by PI1M, as shown in Table 1, wherein there are 995799 PSMILES in total. Based on the PI1M dataset, 100000 organic polymer PSMILES are screened, the number of polymer monomer repeating units is set to 44, and a total of 97584 polymer chain segment stable conformations are calculated.
[0030] Table 1
[0031] Evaluation index: The present application adopts single machine throughput, task success rate as performance index, and component scalability as index, and the evaluation results are shown in Table 2.
[0032] Table 2
[0033] From Table 2, the single machine throughput test is on a single card 4090 node, the test environment is double floating point precision, and the 16 core 32 thread CPU cluster throughput test is on 4 different nodes, the test environment is A100 / 4090 / CPU heterogeneous hybrid computing cluster, RabbitMQ is used between clusters to distribute tasks, and MINIO collects results. The present application supports two open source molecular dynamics tools as the molecular dynamics simulation backend, and supports three force fields covering small molecules, polymer chain macromolecules and various types.
[0034] The above merely describes preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with preferred embodiments as above, it is not intended to limit the present application. Any person skilled in the art, without departing from the technical solution of the present application, can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the technical solution of the present application, and is within the spirit and principle of the present application, any simple modification, equivalent replacement and improvement of the above embodiments are still within the protection scope of the technical solution of the present application.
Claims
1. A method for automatic construction of stable conformations of polymer chains based on distributed heterogeneous aggregation, characterized in that, Comprise: Step 1: receive task information from the message queue, forward or calculate the incoming task according to the affinity information in the task information and the current load condition, and configure the dead letter queue for archiving statistics of failed information; Step 2: parse the monomer PSMILES and topology rules in the task information, identify anchor atoms, leaving groups and protecting groups, calculate the initial conformation of the monomer based on the specified force field type, and calculate the atomic charge distribution by quantum chemistry method; Step 3: combine the topology rule library to set the connection mode of the initial conformation of the monomer, align according to the anchor points, and construct a local coordinate system for each anchor point, merge redundant atoms, and splice the initial conformation of the monomer into an initial segment of a specified length; Step 4: automatically generate the input file of the molecular dynamics simulation tool, and verify the charge conservation and other constraint conditions; Step 5: sequentially perform short-range NVT preheating, annealing and final NVT dynamics relaxation to obtain a stable segment conformation; Step 6: divide the stable segment conformation into multiple monomers according to the anchor points, and align the divided monomers to the local coordinates to calculate the rigid body transformation of the monomer to the chain; Step 7: check the division result, complete the task callback, and upload the output to the distributed storage platform.
2. The method of claim 1, wherein the method is based on the automatic construction of stable conformations of polymer chains with distributed heterogeneity. The task message received in step 1 is in JSON format, and the task message includes fields, transaction ID and affinity information. The transaction ID is used for end-to-end tracking, and the fields include but are not limited to monomer PSMILES, target aggregation degree and topology rules, force field type, simulation temperature / pressure / time step, relaxation phase configuration, convergence threshold, random seed, node affinity, callback URL and data storage strategy. The affinity information includes GPU / CPU load and precision information.
3. The method of claim 1, wherein the method is based on the automatic construction of the stable conformations of the polymer chain with the distribution of the isomers. Step 2 specifically includes: The valence state of the monomer PSMILES in the task information is standardized and the aromaticity is checked. If the check fails, the task is directly rejected and transmitted to the dead letter queue, and the error log is counted. After the check is completed, the PSMILES rules are used to identify anchor atoms, leaving groups and protecting groups, and a local coordinate system is constructed for each anchor point. In the constructed local coordinate system, two connection point atoms and one neighbor atom in the monomer PSMILES are set as Star1, Neighbor1, Star2, Neighbor2, respectively, and the Neighbor1 atom is set as the coordinate origin. The calculation of the initial conformation of the monomer is completed; The reasonable initial conformation of the monomer is screened using the fast molecular mechanics optimization in the rdkit tool, which includes UFF / MMFF pre-drop, charge distribution using PSI4 tool, and DFT calculation of electrostatic potential and generation of RESP / AM1-BCC charge.
4. The method for automatic construction and stable conformation generation of polymer chains based on distributed heterogeneous according to claim 1, characterized in that, Step 3 specifically includes: The connection mode of the initial conformation of the monomer is set in combination with a topology rule library, the connection mode includes but is not limited to linear and block connection, a plurality of monomers are sampled by fixed length or user-defined aggregation degree distribution, rigid body alignment is performed based on the normal and tangential direction of the anchor point, redundant atoms and bonds are combined, and the boundary bond order is repaired; the rotatable bond with high tension is subjected to heuristic energy scanning to eliminate unreasonable conformations, initial segments of the target aggregation degree are generated, and the simulation box boundary conditions are set according to the simulation parameters provided by the user.
5. The method for automatic construction of stable conformations of polymer chains based on distributed heterogeneous architecture according to claim 1, characterized in that, Step 4 specifically comprises: The data and input scripts required by the molecular dynamics simulation tool are automatically generated, wherein the generated data and input scripts include but are not limited to atom types, bond angle dihedral parameters, non-bond interactions, neighbor lists and thermodynamic outputs; and before running the simulation power, the charge conservation, valence consistency, geometric extreme value and initial energy are checked, if the check fails, the task is rejected, and the task is transmitted to the dead letter queue.
6. The method for automatic construction of stable conformations of polymer chains based on distributed heterogeneous architecture according to claim 1, characterized in that, Step 5 specifically comprises: The molecular dynamics simulation tool is run, and short-range NVT preheating, multiple rounds of annealing temperature rise and fall, and final state NVT balance are sequentially performed, high-energy contacts of the initial conformation of the monomer are eliminated through NVT preheating, local energy barriers are crossed through multiple rounds of annealing temperature rise and fall, and stable conformations are obtained through final state NVT balance at the target temperature.
7. The method for automatic construction of stable conformations of polymer chains based on distributed heterogeneous architecture according to claim 1, characterized in that, Step 6 specifically comprises: For the triplet (s1, n1, s2) of Star1 atom, Neighbor1 atom, Star2 atom in monomer PSMILES, set n1 as the origin of local coordinate system, get two vectors with n1 as the starting point and And get the standard coordinate system and the rigid body transformation matrix from monomer local coordinate to the corresponding part on the chain segment after relaxation through GramSchmidt normalization Store the monomer PSMILES after coordinate localization to the corresponding SDF file, save the corresponding relaxed rigid body transformation matrix and store it in numpy array.
8. The method for automatic construction of stable conformations of polymer chains based on distributed heterogeneous architecture according to claim 1, characterized in that, Step 7 specifically comprises: The SDF file of the local coordinate alignment of the monomer and the rigid body transformation matrix are read, the polymer chain is reassembled, and the rdkit is used to compare with the original polymer chain, if the specified RMSD threshold is higher, the result is rejected, the corresponding task is stored in the dead letter queue, and if it is lower, it is uploaded to the distributed storage platform and the information callback is confirmed.