Stabilizer screening process for improving low-temperature stability of chlorantraniliprole and lufenuron suspending agent

Abstract

The invention relates to the field of pesticide composition screening, and discloses a stabilizer screening process for improving the low-temperature stability of a chlorantraniliprole and lufenuron suspending agent, which comprises the following steps: a) analyzing the low-temperature behavior of chlorantraniliprole and lufenuron based on molecular simulation, constructing a candidate stabilizer molecule library, and preliminarily screening candidate stabilizers; b) applying low-temperature stress to the chlorantraniliprole and lufenuron suspending agent sample containing the candidate stabilizer by using a microfluidic platform, and performing in-situ dynamic characterization to obtain dynamic data related to the low-temperature stability of the sample; and c) processing the dynamic data, and establishing an evaluation index representing the low-temperature stability performance of the candidate stabilizer. According to the invention, a cognitive bridge from a micromechanism to a mesoscopic phenomenon is constructed by integrating molecular simulation pre-screening and microfluidic high-throughput dynamic characterization. The problems of long research and development cycle and resource consumption caused by a large number of trial and error and experience judgment in a traditional screening method are effectively solved.

Description

Technical Field

[0001] The invention relates to the technical field of pesticide composition screening, in particular to a stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrates. Background Art

[0002] Chlorantraniliprole and lufenuron are both insecticides widely used in modern agricultural production. Chlorantraniliprole belongs to the diamide insecticide class and plays an important role in the control of various crop pests due to its high efficiency, broad spectrum, and relative safety against non-target organisms. Lufenuron is a benzoylurea insect growth regulator that kills larvae and eggs mainly by inhibiting the synthesis of chitin in insects. Compounding these two active ingredients and preparing them into a suspension concentrate (SC) can broaden the insecticide spectrum, delay the development of resistance, and facilitate user use. This type of preparation is widely used in the integrated pest management of field crops such as rice, corn, and cotton, as well as cash crops such as vegetables and fruit trees.

[0003] However, chlorfenapyr and lufenuron suspension concentrates, especially high-concentration or compound formulations, often face severe stability challenges during low-temperature storage, transportation, and use. Existing stabilizer screening methods often struggle to accurately predict and effectively address issues such as low-temperature precipitation, particle agglomeration, or Ostwald ripening, resulting in a time-consuming and labor-intensive screening process with a low success rate. Traditional screening methods lack sufficient understanding of the complex interaction mechanisms between stabilizers and pesticide particles, making the selection of stabilizers more reliant on experience and lacking in-depth theoretical guidance. Summary of the Invention

[0004] The purpose of the present invention is to provide a stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrates, thereby solving the problems in the prior art that stabilizer screening is difficult to solve low-temperature precipitation, particle agglomeration or Ostwald ripening, and often relies on empirical trial and error and lacks in-depth insight into and accurate prediction of the stabilization mechanism.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: A stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrates is characterized by comprising the following steps: a) Based on molecular simulation, the low-temperature behavior of chlorfenapyr and lufenuron was analyzed, a molecular library of candidate stabilizers was constructed, and preliminary screening of candidate stabilizers was performed; b) applying low temperature stress to samples of chlorfenapyr and lufenuron suspension concentrate containing the candidate stabilizer using a microfluidic platform, and performing in situ dynamic characterization to obtain dynamic data related to the low temperature stability of the samples; c) processing the dynamic data to establish an evaluation index characterizing the low-temperature stability performance of the candidate stabilizer; d) using a machine learning method to construct a prediction model based on the structural information, concentration information and / or molecular simulation information of the candidate stabilizer and the low-temperature stability performance evaluation index; e) Based on the prediction model, iteratively optimizing the selection, concentration, combination or experimental conditions of the candidate stabilizers through an intelligent feedback mechanism until the target stabilizer or candidate solution is screened out.

[0006] The innovation of this process lies in its systematic integration of multiple advanced technical links such as multi-scale computational simulation, high-throughput microfluidic experiments, data-driven evaluation index establishment, machine learning predictive modeling, and intelligent feedback iterative optimization, forming a closed-loop and efficient screening process. Specifically, the process first starts from the molecular level through step a), using molecular simulation technology to deeply analyze the unstable behavior of chlorfenapyr and / or lufenuron under low temperature conditions, and analyze their tendency and microscopic mechanism of aggregation, crystallization or phase change. At the same time, a library of candidate stabilizer molecules with diverse structures is constructed, and computational chemistry methods (molecular docking or preliminary molecular dynamics simulation) are used to conduct preliminary evaluation and screening of the interactions between these candidate stabilizers and pesticide molecules, so as to focus on high-potential molecules before entering the experimental stage. The mechanism of this step is to gain preliminary insight into the instability mode of pesticide molecules and the mode of action of stabilizers through theoretical calculations, avoiding a large number of blind attempts in traditional screening and improving the pertinence and efficiency of screening from the source.

[0007] Preferably, the preliminary screening of candidate stabilizers based on molecular simulation in step a) comprises: The molecular docking method was used to evaluate the binding affinity of the candidate stabilizer to chlorfenapyr and lufenuron particles or their surfaces, and the molecular dynamics simulation method was used to analyze the dynamic adsorption behavior of the candidate stabilizer on the surface of chlorfenapyr and lufenuron particles and its effect on the interaction between particles.

[0008] By using simulation methods, the interaction strength and pattern between stabilizers and pesticide particles, as well as the macroscopic impact of this interaction on the inter-particle forces, can be quantitatively or semi-quantitatively predicted at the atomic / molecular scale. This provides a theoretical basis for screening molecules that can effectively adsorb and produce stabilization effects. This is a significant innovation in traditional experience-based screening methods.

[0009] Preferably, the construction of the candidate stabilizer molecule library in step a) comprises at least one of the following methods: Virtual derivatization of known stabilizers; Based on the low-temperature behavior analysis of chlorfenapyr and lufenuron, candidate stabilizers were designed de novo.

[0010] One method explores the potential of the existing chemical space through systematic structural modification, while the other, based on the principle of "treating the disease with the right medicine", directly designs new molecular structures that can theoretically interact specifically with key destabilizing sites of pesticide molecules or effectively intervene in the destabilization process. Rational design based on mechanistic understanding is an important innovative approach to improve screening success rates and discover new and efficient stabilizers.

[0011] Preferably, the microfluidic platform in step b) includes a microchannel structure for sample mixing, gradient concentration generation and precise temperature control; the in situ dynamic characterization includes using at least one of microscopic imaging technology, spectral analysis technology and particle size analysis technology to monitor the physicochemical changes of the sample under low temperature stress in real time.

[0012] Microfluidic technology enables precise control of experimental conditions (temperature, concentration, mixing) and high-throughput parallel operation, greatly shortening the experimental cycle and reducing sample consumption; more importantly, in situ dynamic characterization can capture and record the dynamic details of the unstable process in real time without disturbing the sample, which provides rich information that is difficult to obtain with traditional macroscopic and endpoint detection methods for in-depth understanding of the mechanism of action of stabilizers and the establishment of more accurate stability evaluation models.

[0013] Preferably, the microchannel structure of the microfluidic platform further includes at least one of the following structures designed to simulate the specific low-temperature instability mode of the suspension: A structure for simulating the micro-region gradient concentration environment of Ostwald ripening; Microstructures for simulating ice crystal formation and squeezing effects during freeze-thaw cycles; Structures used to achieve specific flow fields or confined spaces that are sensitive to inter-particle interactions or early aggregation behavior.

[0014] By constructing a simulated, controllable, specific instability environment on a microfluidic chip, it is possible to more specifically study and evaluate the stability performance of candidate stabilizers under specific challenging conditions, rather than simply conducting general low-temperature tests. This helps to screen out stabilizers that can address key instability problems in specific practical application scenarios (repeated freeze-thaw cycles, long-term low-temperature storage).

[0015] Preferably, the low-temperature stability performance evaluation index in step c) is a low-temperature stability digital fingerprint or comprehensive score formed by integrating multiple characteristic parameters in the dynamic data, and the characteristic parameters are selected from at least one of crystallization kinetic parameters, agglomeration kinetic parameters, particle size distribution change parameters or specific spectral feature evolution over time parameters.

[0016] By converting complex, multi-dimensional dynamic experimental phenomena into quantifiable, standardized numerical indicators, we can objectively and accurately evaluate the low-temperature stabilization effects of different candidate stabilizers. This "digital fingerprint" or comprehensive score not only facilitates direct comparison and ranking of different experimental results, but more importantly, it provides a clear and consistent output target (Y value) for the subsequent construction of machine learning models.

[0017] Preferably, the machine learning method in step d) adopts a model capable of fusing multi-scale data, wherein the multi-scale data includes: Microscopic molecular descriptors or interaction parameters of candidate stabilizers derived from molecular simulations, and mesoscopic temporal data or their derived features derived from microfluidic in situ dynamic characterization; The objective function of the machine learning method is expressed as follows: in: θ is the model parameter; N is the number of training samples; (x i ,y i ) is the i-th training sample, x i are input features: molecular descriptors of candidate stabilizers, y i It is the target output: low temperature stability evaluation index; f(x i ; θ) is the model based on the input x i and the predicted value given by the parameter θ; L(y i ,f(x i ; θ)) is the loss function, which is used to measure the difference between the predicted value and the true value; R(θ) is a regularization term used to prevent overfitting; λ is the regularization coefficient.

[0018] Leveraging the powerful fitting capabilities of machine learning, this approach learns and establishes a quantitative structure-activity relationship (QSAR) or quantitative structure-property relationship (QPR) model between the candidate stabilizer's "structure / properties / conditions" and its "low-temperature stability" from high-dimensional, complex input features. By integrating multi-scale data, the model captures the inherent connections from molecular microscopic properties to the system's mesoscopic dynamic behavior, thereby establishing a more comprehensive and robust predictive capability, surpassing traditional analytical methods that rely on single-scale information.

[0019] Preferably, the intelligent feedback mechanism in step e) includes using a Bayesian optimization algorithm, which includes: The objective function g(x) is used to model the distribution of the low temperature stability performance evaluation index.t ,y t )}, the prediction of the new point x follows a Gaussian distribution: where μ t (x) is the posterior mean, is the posterior variance, D 1:t It is the existing observation data.

[0020] Transforming the predictive model from a passive information output tool into an intelligent engine that proactively guides experimental design and optimization. Through this closed-loop iteration of "prediction-experimentation-learning-re-prediction," efficient data-driven searches across vast chemical and parameter spaces are possible, significantly accelerating the discovery of high-performance stabilizers and uncovering counterintuitive optimization solutions.

[0021] Preferably, the iterative optimization guided by the intelligent feedback mechanism includes at least one of the following: Optimize the structural modification direction of candidate stabilizers or the chemical space area to be investigated in the next round of molecular simulation; Dynamically adjust the candidate stabilizer types, concentration gradient ranges, combinations, or parameters of the low-temperature stress program in the next round of microfluidic experiments.

[0022] The intelligent feedback system dynamically and precisely adjusts subsequent computational simulations and experimental design based on existing data and model predictions. This adaptive adjustment enables the entire screening process to continuously optimize itself based on "learned" knowledge, focusing on the most promising areas and conditions, thereby avoiding redundant investment in inefficient areas and further improving screening efficiency and success rate.

[0023] Preferably, the chlorfenapyr and lufenuron suspension sample is a mixture of a basic suspension comprising chlorfenapyr, lufenuron, a dispersant, a wetting agent and water and the candidate stabilizer, and the low temperature stress includes at least one of uniform cooling, constant low temperature maintenance or freeze-thaw cycles.

[0024] By defining a relatively standardized basic suspension system, it can be ensured that different candidate stabilizers are evaluated in comparable matrices. The diverse low-temperature stress methods are intended to simulate the various low-temperature environmental challenges encountered by pesticide products during actual storage, transportation, and use, thereby ensuring that the screened stabilizers are practical under real-world conditions. Using these precisely controlled experimental condition parameters (stabilizer concentration, stress type, and intensity) as one of the input features of the machine learning model helps the model learn more detailed structure-activity relationships and reveal the differences in the responses of different stabilizers to different types of low-temperature stress.

[0025] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This invention builds a bridge from microscopic mechanisms to mesoscopic phenomena by integrating molecular simulation pre-screening with microfluidic high-throughput dynamic characterization. This significantly improves the initial hit rate of stabilizer screening and the targeted nature of subsequent experiments, effectively eliminating the lengthy R&D cycles and resource-intensive nature of traditional screening methods, which rely on extensive trial-and-error and empirical judgment.

[0026] 2. The process of the present invention deeply integrates machine learning prediction models with intelligent feedback iterative optimization mechanisms. This data-driven closed-loop screening strategy empowers the R&D process with self-learning and dynamic adjustment capabilities, intelligently guiding experimental direction and exploring a broader chemical and parameter space. Compared to the more rigid, linear screening processes of the prior art, the present invention addresses the limitations of the prior art, which have limited ability to rapidly respond to complex structure-activity relationships and optimization efficiency.

[0027] 3. This invention innovatively introduces a multi-dimensional "low-temperature stability digital fingerprint" based on in-situ dynamic data as an evaluation criterion. This approach quantifies the complex destabilization process into objective, comparable indicators, providing a more detailed and comprehensive picture of stabilizer performance. It overcomes the limitations of previous stability assessments that primarily relied on macroscopic endpoint observations or single parameters, thereby improving evaluation accuracy and insight into early signs of instability.

[0028] 4. This invention utilizes microfluidics technology to simulate and accelerate the low-temperature instability of suspension concentrates in a controlled microenvironment, and combines this with molecular simulation to reveal the interaction mechanism between stabilizers and pesticide particles. This multi-scale approach enables researchers to gain a deeper understanding of the stabilization mechanism, enabling more predictive stabilizer molecular design or formulation optimization. This overcomes the technical bottleneck of traditional methods, which often struggle to accurately analyze the stabilizer's mode of action at the molecular level and rely more on inferring mechanisms based on macroscopic phenomena.

[0029] 5. The present invention uses a systematic screening process to ultimately obtain a target low-temperature stabilizer or candidate solution that can significantly enhance the physical stability of chlorfenapyr and lufenuron suspension concentrates in a low-temperature environment. After experiencing a low-temperature challenge, the formulation can effectively inhibit or delay the irreversible aggregation and crystallization of pesticide particles, thereby avoiding quality problems such as stratification, agglomeration, or precipitation in the product. Compared with the prior art, which suffers from a decline in product performance and uneven distribution of active ingredients due to insufficient low-temperature stability, the present invention ensures that the screened reagents can still maintain their proper dispersion state and uniformity of efficacy under low-temperature conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0031] The following is combined with Figure 1 , the present invention is described in further detail.

[0032] The present invention provides a stabilizer screening process for improving the low temperature stability of chlorfenapyr and lufenuron suspension concentrates, such as Figure 1 As shown, the screening process includes the following steps: Step S1: Theoretical analysis and candidate stabilizer screening based on molecular simulation In this example, molecular dynamics (MD) simulations were first performed on the molecular behaviors of chlorfenapyr as a single component, lufenuron as a single component, and a binary mixture of chlorfenapyr and lufenuron in a low-temperature aqueous environment.

[0033] In general, MD simulations aim to track the trajectories of atoms in the system in phase space by solving the Newtonian equations of motion for all atoms in the system. At its core is Newton's second law: Among them: F i (t) represents the total force acting on atom i at time t; m i Represents atom m i The quality of a i (t) represents the acceleration of atom i at time t; r i (t) represents the position vector of atom i at time t.

[0034] The force F acting on the atom i By finding the negative gradient of the atomic coordinates of the total potential energy function U of the system, we can obtain: In this embodiment, an empirical force field is used to describe the interaction potential U between molecules. Common force field function forms can include bonding terms and non-bonding terms: Alternatively, the Lennard-Jones term can be expressed as: Among them: The first term is the bond stretching potential: K b is the bond stretching force constant, r is the actual bond length, and r0 is the equilibrium bond length.

[0035] The second term is the bond angle bending potential: K θ is the bond angle bending force constant, θ is the actual bond angle, and θ0 is the equilibrium bond angle.

[0036] The third term is the dihedral angle torsion potential: V n is the torsion barrier height, n is the periodic parameter, φ is the actual dihedral angle, and γ is the phase angle.

[0037] The fourth term is the non-bonded interactions, including: Van der Waals interaction (Lennard-Jones potential): ∈ ij is the potential well depth between atomic pair i and j; is the interatomic distance when the potential energy reaches its minimum, σ ij is the collision diameter between atom pair i and j r ij is the actual distance between atoms i and j. Electrostatic interaction (Coulomb potential): q i and q j are the point charges of atoms i and j respectively; ∈0 is the dielectric constant of vacuum; ∈ r is the relative dielectric constant or effective dielectric constant of the system.

[0038] Specifically, the Generalized Amber Force Field (GAFF / GAFF2) or OPLS-AA force field parameters were selected for small organic molecules (chlorfenapyr and lufenuron) in the simulations. The explicit solvent model, TIP3P, SPC / E, or TIP4P-EW, was used for the simulation of water molecules.

[0039] The simulation system is usually constructed as a periodic boundary box containing a small number (1 to 10) of original drug molecules and a large number (5000 to 20000) of water molecules. The box size is adjusted according to the size of the system, and the side length is generally in the range of 4 to 8 nanometers (nm).

[0040] Simulations were performed in the NPT ensemble (constant number of atoms, pressure, and temperature) or the NVT ensemble (constant number of atoms, volume, and temperature). Simulation temperatures were controlled within the low-temperature range of 263 K (-10°C), 268 K (-5°C), 273 K (0°C), and 278 K (5°C) to examine low-temperature effects. Pressure was typically maintained at 1 standard atmosphere (atm).

[0041] The integration time step is typically set to 1 to 2 femtoseconds (fs). To ensure system equilibrium and adequate sampling, the total simulation time for each system can be 100 to 500 nanoseconds (ns), or even longer.

[0042] Long-range electrostatic interactions are treated using the particle mesh Ewald summation method (PME) or its variants, with a cutoff radius typically between 1.0 and 1.4 nm. Similar cutoff radii are used for van der Waals interactions.

[0043] Extract key information by analyzing the simulation trajectory. Calculate the radial distribution function (RDF) between the original drug molecules to characterize their aggregation degree and characteristic distance. Analyze the root mean square displacement (MSD) to evaluate the diffusion behavior of the molecules. Study the changes in the hydrogen bond network (intra- and inter-molecular hydrogen bonds with water molecules). Observe the size, morphology, and stability of molecular aggregates. Use order parameters (Steinhardt order parameters) or tactoid cluster analysis methods to evaluate potential nucleation trends and identify the dominant crystal faces of the precipitated crystals.

[0044] These analyses help to understand, at the molecular scale, the intrinsic driving forces and specific manifestations of aggregation, crystallization, or precipitation of the original drug at low temperatures, and provide key structural and mechanistic information for the subsequent targeted design or screening of stabilizers.

[0045] In this example, based on the above simulation analysis and existing chemical knowledge, a molecular library containing a variety of potential low-temperature stabilizers was constructed.

[0046] In one possible implementation, known common or potential low-temperature stabilizers or antifreeze agents in pesticide formulations are collected, including small molecule polyols such as ethylene glycol, propylene glycol, glycerol, sorbitol, and urea; high molecular weight compounds such as polyvinyl pyrrolidone (PVP, with different K values ​​K15, K30, K90), polyvinyl alcohol (PVA, with different degrees of hydrolysis and polymerization), sodium carboxymethyl cellulose (CMC, with different degrees of substitution and viscosity), xanthan gum, and guar gum derivatives; and some block copolymers with interface activity (poloxamer-type EO-PO block copolymers).

[0047] Alternatively, computational chemistry tools can be used to perform virtual functional group modifications or fragment combinations on the backbones of known stabilizers, introducing hydroxyl, carboxyl, amino, or ether groups, or varying the length and degree of branching of the alkyl chain, to systematically expand the chemical diversity of candidate molecules. This process can generate hundreds of derivative molecules.

[0048] In one possible implementation, based on the low-temperature instability mechanism of the original drug revealed by the aforementioned MD simulation (specific crystal face exposure, hydrophobic-driven aggregation), de novo design or fragment-based drug design (FBDD) strategies can be used to design novel molecular structures that can theoretically interact specifically with key sites of the original drug (hydrogen bonding, π-π stacking) or provide efficient steric hindrance. This approach can generate dozens to hundreds of novel molecular concepts.

[0049] Ultimately, the molecules from the above sources are integrated to construct a database containing approximately 200 to 1,000 or even more candidate molecules.

[0050] Subsequently, a series of molecular descriptors are calculated for all candidate stabilizer molecules in the molecular library.

[0051] In general, molecular descriptors are numerical representations of molecular structure and properties.

[0052] Specifically, the calculated descriptors may include: Physical and chemical properties: molecular weight (MW), molar volume, density, octanol-water partition coefficient (logP or clogP), water solubility (logS).

[0053] Topological descriptors: Wiener index, Balaban J index, molecular connectivity index (Chi index).

[0054] Geometric descriptors: molecular surface area, solvent accessible surface area (SASA), sphericity, principal moments of inertia.

[0055] Electronic descriptors: dipole moment, highest occupied molecular orbital energy (HOMO), lowest unoccupied molecular orbital energy (LUMO), HOMO-LUMO energy gap, and parameters related to electrostatic potential distribution.

[0056] Hydrogen bond related descriptors: number of hydrogen bond donors, number of hydrogen bond acceptors.

[0057] Prediction of pharmacokinetic-related properties (initial screening of ADMET properties): compliance with Lipinski's Rule of Five, topological polar surface area (TPSA).

[0058] In this example, approximately 50 to 200 different descriptors are calculated for each candidate molecule. These descriptors will provide input features for subsequent quantitative structure-activity relationship (QSAR) studies or machine learning modeling.

[0059] In this example, molecular docking technology was used to perform preliminary high-throughput virtual screening on the constructed candidate stabilizer molecule library.

[0060] Molecular docking aims to predict the binding mode and affinity of small molecule ligands (candidate stabilizers) to macromolecular receptors (molecules of chlorfenapyr and / or lufenuron or their aggregates / crystal surface models).

[0061] The receptor model can be constructed based on the stable dimer of the original drug, the small molecule cluster structure, or the dominant crystal face structure of its precipitated crystals observed in the aforementioned MD simulation.

[0062] The ligand library is the candidate stabilizer molecule library constructed and characterized above.

[0063] In one possible implementation, the molecular docking process includes: Define the docking box: Set a three-dimensional search space around the active site of the receptor or the surface area of ​​interest. The box size should be large enough to accommodate different orientations and conformations of the ligand molecule, with a side length of 20 to 40 angstroms.

[0064] Conformational search: Use efficient conformational search algorithms such as genetic algorithms, Monte Carlo simulated annealing algorithms, or systematic search algorithms to explore multiple binding conformations and orientations of the ligand molecule within the docking box.

[0065] Scoring and ranking: A scoring function is used to evaluate the affinity of each binding mode and rank the candidate molecules.

[0066] Scoring functions are the core of molecular docking, whose goal is to quickly estimate the ligand-receptor binding free energy or quantities related to it. A general scoring function can be expressed as a linear combination of multiple energy terms: Score(G bind )=∑ k w k E k ; Among them: Score (G bind ) is the predicted binding score, which is related to the binding free energy. k is the weight coefficient of the kth energy term, which is usually obtained by regression analysis of known binding data.

[0067] E k represents various energy terms that contribute to binding affinity, including: E vdw : Van der Waals interaction term, describing the attraction and repulsion between molecules. E elec : Electrostatic interaction term, describing the interaction between charged or polar groups. E hbond : Hydrogen bond term, specifically describing the formation of hydrogen bonds. E desolv : Desolvation effect term, which estimates the energy change of ligand and receptor transferred from the solvent to the binding interface, can be calculated based on solvent accessible surface area (SASA) or more complex implicit solvent model. entropy : Entropy contribution term, taking into account the entropy disadvantage caused by the loss of flexibility after ligand binding. E torsion : Changes in the internal torsion energy or conformational energy of the ligand. In some embodiments, the scoring function may further include a knowledge-based potential function term, which statistically obtains the preferential contact frequencies of atom pairs from a database of known protein-ligand complex structures.

[0068] Based on the docking scoring results, a screening threshold is set. Molecules whose docking scores are superior to those of a known reference antifreeze (ethylene glycol) by a certain percentage (10% to 30%) are selected. At the same time, it is also necessary to check whether the predicted binding mode is reasonable and whether the stabilizer molecules form favorable interactions with key regions of the original drug molecule (hydrophobic surfaces, groups that are prone to hydrogen bonding). Through this high-throughput virtual screening, the number of candidate molecules can be initially reduced from hundreds or thousands to approximately 50 to 200.

[0069] Detailed evaluation and re-screening based on molecular dynamics simulation In this example, more detailed molecular dynamics simulations were performed on candidate stabilizer-derivative complexes with near-optimal molecular pairs to further validate their binding stability and potential mechanism of action. Short-duration MD simulations of 10 to 50 nanoseconds (ns) were performed on each selected stabilizer-derivative complex. The simulation conditions (force field, water model, temperature, pressure, and integration step) were essentially the same as those used in the aforementioned simulations of the derivative behavior.

[0070] In one possible implementation, the analysis includes: Calculation of binding free energy: Using the molecular mechanics / Poisson-Boltzmann surface area (MM / PBSA) or molecular mechanics / generalized Boltzmann surface area (MM / GBSA) method, the binding free energy (ΔGbindΔGbind) of the candidate stabilizer to the parent drug is estimated from the MD trajectory. This method takes into account molecular mechanics energy, solvation free energy (including polar and nonpolar contributions), and entropic contributions (which are often approximated or ignored to simplify the calculation).

[0071] ΔG bind = <E complex >- <E receptor >- <E ligand > E=E MM +G solvation -TS; E MM =E bonded +E nonbonded =E bonded +(E vdw +E elec ); G solvation =G polar +G nonpolar ; in: <E complex >, <E receptor >, <E ligand > are the average energies of the complex, receptor, and ligand in the solvent, respectively. E MM is the molecular mechanics energy, including the bonding term E bonded and the non-bonded term Enonbonded (Van der Waals E vdw and electrostatic E elec ). G solvation is the solvation free energy, including the polar part G polar (usually calculated using the Poisson-Boltzmann equation or the generalized Boltzmann model) and the non-polar part G nonpolar (Usually proportional to the solvent accessible surface area.) TS is the contribution from conformational entropy.

[0072] Step S2: Microfluidic experiment and in situ dynamic characterization In this example, a basic formula of a mixed suspension concentrate of chlorfenapyr and lufenuron without the low-temperature stabilizer to be screened was first prepared as a benchmark sample for subsequent testing with the addition of candidate stabilizers.

[0073] Generally, the basic suspension contains the following components (by weight percentage): chlorfenapyr technical, 5.0% to 15.0%; lufenuron technical, 2.0% to 10.0%; one or more high-efficiency dispersants (lignin sulfonates, naphthalene sulfonate formaldehyde condensates, polycarboxylates or EO-PO block copolymer dispersants), with a total amount of 3.0% to 8.0%; one or more wetting agents (alkylphenol polyoxyethylene ethers, fatty alcohol polyoxyethylene ethers or silicone wetting agents), with a total amount of 1.0% to 4.0%; a basic thickener (a low addition amount of xanthan gum or magnesium aluminum silicate), 0.05% to 0.2%, to provide initial pourability but not excessively affect the subsequent observation of low-temperature behavior; and an appropriate amount of defoaming agent and pH regulator (to adjust the pH to a neutral or weakly acidic range of 5.0 to 7.5), and the balance is deionized water.

[0074] Specifically, after the above components are thoroughly mixed, they are ground using a wet sand milling process. A laboratory-grade horizontal sand mill can be used as the sand milling equipment, and the grinding media is typically high-density zirconia beads with a particle size of 0.3 to 0.8 millimeters (mm). The speed during sand milling can be controlled between 2000 and 4000 revolutions per minute (rpm), and the grinding time is adjusted according to actual conditions until the particle size of the active ingredient in the suspension reaches the predetermined requirements, with a volume average particle size D[4,3] of less than 5 microns (μm) and a D90 (the diameter of 90% of the volume of particles less than this value) of less than 5 to 10 microns (μm). The prepared base suspension should have good fluidity and uniformity.

[0075] Specifically, in order to more closely follow the specific instability mode of the suspension under actual low-temperature conditions, a special microstructure can be designed in the observation area of ​​the microfluidic chip.

[0076] In some embodiments, to simulate the Ostwald ripening process, a micro-region with controllable solvent evaporation or penetration can be designed, or a stable chemical potential gradient or supersaturation gradient can be formed in the observation area by finely controlling the flow rate ratio of different inlets.

[0077] In some embodiments, in order to simulate the formation of ice crystals during freeze-thaw cycles and their physical squeezing effect on suspended particles, a microcolumn array (column diameter 5 to 20 microns, column spacing 10 to 50 microns) or a periodic microchannel contraction-expansion structure can be designed in the observation area to study the behavior of particles in confined space and under the influence of ice front advancement.

[0078] In this example, a variety of in situ dynamic characterization techniques were combined with a microfluidic chip platform to monitor the changes in the physicochemical state of samples under low temperature stress in real time and in multiple dimensions.

[0079] Typically, an integrated characterization system includes: Optical microscopy: Typically, an inverted optical microscope is used, equipped with a high-resolution CCD or sCMOS camera. Brightfield, darkfield, phase contrast, or differential interference contrast (DIC) imaging can be performed to observe and record in real time the macroscopic and microscopic morphological changes of the original drug particles in the suspension, including crystallization, agglomeration, and stratification (phase separation). The image acquisition frequency can be set from 0.1 to 10 frames per second, or even higher.

[0080] Spectral Analysis Techniques: Micro-Raman Spectroscopy: Combining a Raman spectrometer with a microscope, laser light is focused on a specific area within a microfluidic chip (with a spot diameter of up to 1-2 microns). This can be used to in situ identify transitions in the original drug's crystal form (from metastable to stable), the formation of new crystalline phases, and the adsorption state or distribution of candidate stabilizer molecules on the particle surface (if the stabilizer has a characteristic Raman signal and is present at a sufficient concentration). The laser wavelength can be selected at 532nm or 785nm, and the spectral acquisition time is adjusted based on signal intensity, ranging from 0.5 to 30 seconds per spectrum.

[0081] Micro-FTIR Spectroscopy: Similarly, it is used to monitor changes in molecular functional groups and assist in determining changes in chemical composition or changes in intermolecular interactions.

[0082] Particle size analysis technology: Micro-dynamic light scattering (μ-DLS): In some embodiments, a miniaturized DLS probe or optical fiber can be coupled to the observation area of ​​a microfluidic chip to monitor the changes in the size distribution of suspended particles in real time. DLS technology obtains Brownian motion information of particles by analyzing the fluctuations in scattered light intensity, and then calculates the hydrodynamic radius (R) using the Stokes-Einstein equation. h ): Where: D t is the translational diffusion coefficient; k B is the Boltzmann constant; T is the absolute temperature; η is the viscosity of the dispersion medium.

[0083] Particle Tracking Analysis (PTA): By processing a sequence of images continuously acquired by a microscope, the Brownian motion trajectory of individual particles is identified and tracked, thereby calculating their diffusion coefficient and particle size. This method is particularly suitable for the characterization of polydisperse systems and non-spherical particles.

[0084] In this embodiment, the high-potential candidate low-temperature stabilizers (about 20 to 50 types) screened in step S1 are evaluated in microfluidic experiments one by one or in combination (binary or ternary formulation).

[0085] First, each candidate stabilizer or combination thereof is formulated into a series of aqueous solutions or dispersions at different concentrations, ranging from 0.1% to 20.0% (w / w) relative to the total amount of active ingredients in the base suspension, depending on its expected effective concentration.

[0086] A high-precision microinjection pump (Harvard syringe pump) then pumps the prepared base suspension and candidate stabilizer solution (or its solvent as a blank control) into the corresponding inlets of the microfluidic chip at preset volume flow rates and mixing ratios. The total volume flow rate can be controlled within a range of 0.1 to 10.0 microliters per minute (μL / min). After mixing in the micromixer within the chip, the sample enters the observation and detection area.

[0087] Next, programmed low-temperature stress is applied to the samples in the observation area of ​​the microfluidic chip.

[0088] Specifically, the low temperature program can be designed as follows: Constant cooling rate: Cool down from room temperature (20-25°C) to the target low temperature (-15°C, -10°C, -5°C, 0°C) at a rate of 0.5 to 5.0°C / min.

[0089] Constant low temperature maintenance: After reaching the target low temperature, maintain a constant temperature for a period of time, from 30 minutes to 24 hours, or even longer, to observe changes under long-term low temperature storage.

[0090] Freeze-thaw cycle: simulates the temperature fluctuations encountered in actual storage and transportation. The sample is cooled to -10°C and held for 1 to 4 hours, then warmed to room temperature (20°C) and held for 30 minutes to 1 hour. This cycle is repeated 3 to 10 times.

[0091] Throughout the cold stress process, the previously described integrated in-situ dynamic characterization system (microscopy, spectroscopy, and particle size analysis) simultaneously and continuously collected data from samples within each parallel unit of the microfluidic chip. The collected image sequences, spectral data over time, and particle size distribution evolution data constitute the original dynamic data set for evaluating the low-temperature stabilization effect of candidate stabilizers. This step, utilizing microfluidic technology, enables efficient, parallel testing of multiple candidate solutions under precisely controlled conditions, yielding a wealth of in-situ dynamic information.

[0092] Step S3: Establishment of low temperature stability evaluation index In this embodiment, targeted feature parameter extraction is performed on various types of in-situ dynamic data collected in step S2.

[0093] In general, image data processing aims to quantify the macroscopic and microscopic physical changes that occur in suspensions at low temperatures.

[0094] Specifically, for image sequences acquired through microscopy, image processing algorithms (using ImageJ, CellProfiler open source software, or self-developed algorithm modules) can be used to analyze the following features: Quantification of precipitates (crystallization or agglomeration): the number density of precipitates per unit field of view (number / mm 2 ), average equivalent circular diameter or maximum length (μm), aspect ratio or circularity shape factor, and the percentage of total area covered by precipitates (%). The rate of change of these parameters with time (precipitation rate, growth rate) is also an important dynamic characteristic.

[0095] Quantification of phase separation or stratification: If obvious stratification occurs, the rate of change of the clear liquid layer height or sedimentation layer height over time can be measured (μm / min or % / hr).

[0096] For in-situ spectral data (Raman spectroscopy or infrared spectroscopy), the main focus is on analyzing the changes in spectral features over time and temperature to reveal the transition of chemical composition or phase.

[0097] In some embodiments, the intensity, peak position (wavenumber), and full width at half maximum (FWHM) of characteristic vibrational peaks of chlorfenapyr or lufenuron (carbonyl C=O stretching vibration peak, benzene ring skeleton vibration peak, or lattice vibration peak corresponding to a specific crystal form) can be monitored over time. A decrease in peak intensity indicates a change in solubility or degradation, while a shift in peak position or a change in FWHM is associated with intermolecular interactions or environmental changes.

[0098] The appearance of new characteristic peaks and their changes in intensity over time can indicate the formation of new crystal forms or the generation of chemical transformation products. The time, rate, and ultimate extent of these changes can all be used as characteristic parameters.

[0099] For in-situ particle size analysis data (DLS or PTA results), focus on the evolution of particle size distribution over time. Specifically, the characteristic parameters that can be extracted include: volume average particle size (D[4,3] or Z-average), number average particle size (D[1,0]), and specific percentiles of the distribution (D10, D50, i.e. median diameter, D90) over time. Calculate the growth rate of the average particle size (nm / hr or μm / hr). Monitor changes in the particle size distribution width index, changes in the polydispersity index (PDI) (for DLS), or changes in the span value: Where D10, D50, and D90 are the particle sizes at 10%, 50%, and 90% of the cumulative volume distribution, respectively. An increase in Span usually means a broadening of the particle size distribution, which is associated with agglomeration or Ostwald ripening.

[0100] In this embodiment, the multiple (5 to 20) characteristic parameters extracted from the different types of dynamic data are integrated and converted to form one or a few low-dimensional "digital fingerprints" or comprehensive scores that can comprehensively reflect the low-temperature stability performance of the stabilizer system.

[0101] In general, the extracted original feature parameters need to be normalized first (minimum-maximum normalization or Z-score normalization) to eliminate the differences in dimensions and numerical ranges between different parameters and make them comparable. The minimum-maximum normalization formula is: where f is the original eigenvalue, f min and f max are the minimum and maximum values ​​of the feature in all samples, respectively, and f ′ is the normalized value (usually in the range [0,1]).

[0102] Alternatively, principal component analysis (PCA) can be used to reduce the dimensionality of the normalized multidimensional feature dataset. PCA is a commonly used linear dimensionality reduction method that transforms a set of correlated variables into a set of linearly uncorrelated variables through an orthogonal transformation. These uncorrelated variables are called principal components. The first few principal components with the highest variance contribution (the first 1 to 3 principal components with a cumulative variance contribution of 80%-95%) are selected as the "digital fingerprint" that comprehensively reflects low-temperature stability.

[0103] The core of PCA is to find the eigenvalues ​​and eigenvectors of the feature covariance matrix. If X is an m×n data matrix (m samples, n features, and centered), its covariance matrix Cx for: Then solve for C x The characteristic equation of : C x w=λw; Where λ is the eigenvalue and w is the corresponding eigenvector (the load of the principal component). Sort the eigenvalues ​​from large to small and select the eigenvectors corresponding to the first k eigenvalues ​​to form the transformation matrix W(n×k). The data after dimensionality reduction Y(m×k) is: Y=XW; In a possible implementation, the importance weight (w i , and ∑w i =1), and then a comprehensive score is obtained by linear weighted summation: Where N is the number of characteristic parameters selected, f i ′ is the normalized value of the i-th characteristic parameter. Careful weighting is required to ensure that the score objectively reflects stability. Systems with rapid particle size growth and a large number of precipitates will have negative weights for their corresponding characteristic parameters or, after transformation, contribute negatively to the overall score.

[0104] In some embodiments, the extracted feature parameter set may be used as input to output a comprehensive low-temperature stability index using a relatively simple, pre-trained small neural network (single-layer or double-layer perceptron).

[0105] The resulting "digital fingerprint," or comprehensive score, serves as a quantitative indicator for evaluating the low-temperature stabilization effectiveness of a candidate stabilizer (at a specific concentration and low-temperature conditions). This metric lays the foundation for subsequent horizontal comparisons among numerous candidate solutions and provides a standardized training target for machine learning models.

[0106] Step S4: Prediction model construction based on machine learning In this example, we first integrated information from step S1 (molecular simulation and candidate stabilizer characterization), step S2 (microfluidic experimental conditions), and step S3 (low-temperature stability evaluation indicators) to construct a comprehensive dataset for training, validation, and testing of machine learning models.

[0107] Generally, each record (sample) in the dataset corresponds to a specific low-temperature stability experiment (or a specific candidate stabilizer system).

[0108] Input features (X vectors) usually contain the following types of information: The molecular description of the candidate stabilizer from step S1: The calculated molecular descriptors include molecular weight (MW), topological polar surface area (TPSA), number of rotatable bonds, number of hydrogen bond donors, number of hydrogen bond acceptors, octanol-water partition coefficient (logP), and dipole moment.

[0109] Parameters related to the candidate stabilizer-drug interaction obtained by molecular docking or molecular dynamics simulation, such as docking score, predicted binding free energy (value calculated by MM / PBSA or MM / GBSA method), and the number or strength of key interaction sites.

[0110] Experimental condition information from step S2: Information on the type of candidate stabilizer (which can be numerically processed through one-hot encoding or embedding layers), the concentration value used (usually numerical), the batch information of the base suspension (if different), and the key parameters of the low-temperature stress program (minimum temperature, cooling rate, holding time, number of freeze-thaw cycles).

[0111] The output target (Y value) is the low temperature stability "digital fingerprint" or comprehensive score established in step S3.

[0112] Specifically, if the "digital fingerprint" is a single comprehensive score (a continuous value between 0 and 1, with higher values ​​indicating greater stability), the model building task is regression prediction. If the "digital fingerprint" is a multidimensional vector (the first few principal components obtained by PCA), a multi-output regression model must be built or predictions must be made for each dimension separately.

[0113] The size of the dataset, that is, the number of sample points, generally needs to be hundreds to thousands to ensure that the model can learn effective rules and have good generalization capabilities.

[0114] In this embodiment, a suitable machine learning algorithm is selected to construct a prediction model based on the characteristics of the data set (feature dimension, sample size, data type) and the nature of the prediction task (mainly regression task).

[0115] As an alternative, classic machine learning algorithms can be used: Gradient Boosting Machines (GBM), XGBoost, and LightGBM. These algorithms iteratively train a series of weak learners (usually decision trees) and perform weighted combinations. They have good fitting capabilities for nonlinear relationships and are generally insensitive to feature scaling.

[0116] Random Forest (RF) constructs a large number of decision trees and averages their predictions (for regression tasks) or votes (for classification tasks). It can effectively process high-dimensional data and has good resistance to overfitting.

[0117] Support Vector Regression (SVR): It is based on the principle of support vector machine and seeks a regression hyperplane that can make all sample points as close to the function as possible (within the allowed ∈ deviation).

[0118] In one possible implementation, considering that the input features come from different modalities (molecular structure descriptors, molecular simulation parameters, experimental condition parameters), a model architecture that can effectively fuse multimodal data can be adopted.

[0119] Design specific deep neural network (DNNs) structures.

[0120] Intermediate Fusion or Layered Fusion: Design a neural network with multiple input branches, each responsible for processing features from a specific modality (using a convolutional neural network for molecular graph structure information and an MLP for tabular features). These branches fuse feature representations in the network's intermediate layers, and then perform joint learning and prediction through subsequent shared layers. This approach allows the model to learn representations within each modality as well as interactive information between modalities.

[0121] The core of these is to minimize an objective function through a learning algorithm, which usually consists of a loss term and a regularization term: in: θ is the model parameter; N is the number of training samples; (x i ,y i ) is the i-th training sample, x i are input features: molecular descriptors of candidate stabilizers, y i Is the target output: low temperature stability evaluation index; f(x i ; θ) is the model based on the input x i and the predicted value given by the parameter θ; R(θ) is the regularization term used to prevent overfitting; λ is the regularization coefficient; L(y i ,f(x i ; θ)) is the loss function, which is used to measure the difference between the predicted value and the true value. For regression tasks, commonly used loss functions are: Mean Squared Error (MSE): Mean Absolute Error (MAE): Common regularization terms are: L1 regularization (Lasso): R(θ) = ||θ||1 = ∑ j |θ j |, which tends to produce sparse parameters (i.e., many parameters are zero), thus achieving the effect of feature selection.

[0122] L2 regularization (Ridge or weight decay): It tends to make parameter values ​​smaller and more evenly distributed.

[0123] In this embodiment, after the model architecture is selected, the model needs to be trained and verified.

[0124] Generally, the constructed dataset is first divided into training set, validation set and test set.

[0125] The dataset can be randomly divided into 70% (training set), 15% (validation set), and 15% (test set). The training set is used to fit the model parameters; the validation set is used to monitor model performance, perform hyperparameter tuning, and implement early stopping strategies to avoid overfitting during training; the test set is not involved in the training process at all and is only used for the final, unbiased evaluation of the model's generalization ability after training is complete.

[0126] In one possible implementation, k-fold cross-validation can be used to more robustly evaluate model performance and select hyperparameters, especially when the dataset size is relatively small.

[0127] Model training involves iteratively adjusting the model parameters θ via an optimization algorithm (gradient descent and its variants Adam, RMSprop) to minimize the loss function on the training set.

[0128] Hyperparameter tuning is a key step in building high-performance models. Hyperparameters are parameters that need to be set before model training, such as the learning rate, the regularization coefficient λ, the number of neural network layers and the number of neurons in each layer, the maximum depth of the decision tree, and the number of trees in the gradient boosting tree.

[0129] Specifically, grid search (GridSearch), random search (RandomSearch) or more advanced Bayesian optimization (BayesianOptimization) methods can be used to find the optimal hyperparameter combination on the validation set.

[0130] The performance of the model is usually evaluated using the following metrics commonly used in regression tasks, calculated on the validation set and the test set: Root Mean Squared Error (RMSE): Mean Absolute Error (MAE): Coefficient of Determination (R 2 score): where N eval is the number of samples in the evaluation dataset (validation set or test set), y i is the true value, is the model prediction value, is the average value of the true value. RMSE and MAE measure the size of the prediction error, the smaller the value, the better. 2 The value range is usually between 0 and 1. The closer the value is to 1, the better the model fits the data.

[0131] Step S5: Intelligent feedback iterative optimization and target stabilizer screening In this embodiment, an intelligent algorithm is used to determine the most worthwhile experiment for the next round, so as to obtain the maximum information gain or the fastest approach to the optimal solution with the least number of experiments.

[0132] In general, an active learning (AL) strategy can be used. Active learning aims to select, from a large number of unlabeled candidate samples, those samples that would have the greatest impact on model performance if labeled (i.e., subjected to experiments and obtained stability data).

[0133] A commonly used active learning sampling strategy is uncertainty-based sampling (UncertaintySampling).

[0134] Specifically, if the prediction model constructed in step S4 can provide an uncertainty measure for its prediction results (if the model is a Gaussian process regression, the prediction variance can be directly obtained; if the ensemble model is a random forest, it can be estimated by the divergence or variance of the member prediction values; if it is a single model XGBoost, the uncertainty can be estimated by the bootstrap method or by constructing quantile regression), then the model can be selected for its prediction stability. The most uncertain stabilizer candidate x * (and its corresponding experimental conditions) for the next round of experiments: Where: x represents an unlabeled candidate stabilizer and its related parameters (molecular structure, concentration). Unlabeled Candidate Pool The unlabeled candidate pool refers to the set of candidate stabilizers constructed in step S1 but not yet experimentally evaluated in step S2, or a newly generated candidate point in the parameter space. is the predicted value of the stability score of candidate x by the current prediction model. Model S4 / S5 is a measure of the variance or uncertainty of the model's prediction stability for candidate x. Model S4 / S5 refers to the prediction model initially constructed in step S4 and continuously updated during the iterative process in step S5.

[0135] As another option, a Bayesian Optimization (BO) strategy can be used. Bayesian Optimization is particularly suitable for finding the global optimum on an expensive black-box function, where the black-box function is the true, unknown mapping relationship between candidate stabilizer parameters and low-temperature stability.

[0136] Bayesian optimization consists of two core components: Probabilistic Surrogate Model: Gaussian Process (GP) is usually used to model the objective function g(x)g(x) (i.e., the actual low-temperature stability evaluation index). The GP model is based on the observed experimental data points D1:t={(x1,y1),...,(x t ,y t )}(where x i is the input parameter of the i-th experiment, y i is the stability score obtained in the corresponding step S3), and a probability distribution prediction is given for the true stability g(x) of any untested point x, usually a Gaussian distribution: Where: μ t (x) is the posterior mean prediction of the true stability g(x) at point x. μ is the posterior variance prediction of the true stability g(x) at point x, reflecting the uncertainty of the prediction. t (x) and The calculation of depends on the selected kernel function (covariance function), the commonly used radial basis function (RBF) kernel or Matern kernel, and the existing observation data D 1:t The machine learning model constructed in step S4 can provide a good initial prior information for GP, or its predicted value can be used as an input feature of the GP model.

[0137] Acquisition Function: The predicted mean μ based on the surrogate model (GP) t (x) and the prediction variance The acquisition function is used to evaluate the "value" of conducting the next experiment at the candidate point x, thereby guiding the next sampling position x. t+1 The commonly used acquisition functions are: Expected Improvement (EI): Assume that the currently observed optimal stability score is y best =max(y1,...,y t ) (Assuming that higher scores are better). EI calculates the amount by which we can expect to gain above the current optimal value when performing the experiment at point x: EI(x)=E[max(0,g(x)-y best )|D t ]=(μ t (x)-y best )Φ(Z)+σ t (x)φ(Z); in Φ(·) is the cumulative distribution function (CDF) of the standard normal distribution, and φ(·) is the probability density function (PDF) of the standard normal distribution. Select the point x that maximizes EI(x) t+1 As the next experimental point.

[0138] Upper Confidence Bound (UCB): UCB uses an adjustable parameter κ to balance the exploitation of high-expectation areas and the exploration of high-uncertainty areas: UCB(x)=μ t (x)+κσ t (x); Where κ≥0 is a trade-off parameter. Select the point x that maximizes UCB(x) t+1 As the next experimental point.

[0139] In this embodiment, a closed-loop iterative screening process is performed based on the guidance of the above-mentioned intelligent feedback algorithm.

[0140] Specifically, each cycle of the iterative loop typically consists of the following steps: Intelligent decision-making: Based on the currently used intelligent algorithm (uncertainty sampling of active learning or acquisition function maximization of Bayesian optimization), the next most valuable candidate solution x is selected from the candidate stabilizer pool (including molecules not tested in step S1, new concentrations and new combinations of known molecules, or completely new parameter combinations suggested by the algorithm) next .

[0141] Experiment execution: for the selected candidate solution x next Experiment. If x next If it involves a completely new molecular structure, it is necessary to first perform the partial molecular simulation evaluation in step S1; if it involves new conditions or combinations of known molecules, the microfluidic experiment in step S2 is mainly performed.

[0142] Data acquisition and processing: Perform data processing and feature extraction in step S3 on the experimental results. next Corresponding low temperature stability evaluation index y next .

[0143] Model update: The newly obtained experimental data points (x next ,y next ) to the existing dataset. Then, update or retrain the prediction model.

[0144] If an active learning strategy is adopted, the machine learning model (XGBoost, random forest or neural network) constructed in step S4 is updated (retrained) so that it learns new information.

[0145] If the Bayesian optimization strategy is used, the new data point is added to the observation data of the Gaussian process model and the posterior mean μ of the GP is updated. i (x) and the posterior variance If the model constructed in step S4 is not GP, after accumulating a certain amount of new data, the model of step S4 type can be retrained regularly with all data (including historical data and new data) to maintain a main prediction model with good global performance.

[0146] This iterative process is repeated.

[0147] Generally, the termination conditions of the iteration can be preset, when a preset number of experimental iterations (5 to 20 rounds) is reached; the performance of the prediction model no longer improves significantly on the validation set (converges); one or more candidate solutions have been found, and their low-temperature stability scores have reached a preset satisfaction threshold; or the total budget (time or cost) of the experiment has been exhausted.

[0148] In this embodiment, after the iterative screening cycle is terminated, a final comprehensive evaluation and priority ranking of all candidate stabilizer solutions explored and evaluated throughout the entire process is required.

[0149] Specifically, the assessment is based on: The low-temperature stability prediction score given by the prediction model after final iterative optimization.

[0150] The “low temperature stability digital fingerprint” or comprehensive score actually observed in the microfluidic experiment and calculated according to step S3.

[0151] In some embodiments, other considerations may also be introduced, such as the estimated production cost of the candidate stabilizer, known toxicological data, environmental friendliness, and compatibility with other formulation components.

[0152] By comprehensively considering these multiple factors (using multi-objective optimization method or weighted scoring method), all tested stabilizer systems (including single stabilizer and compound stabilizer solutions) are finally ranked by performance.

[0153] Ultimately, the top N (3 to 10) candidate low-temperature stabilizer solutions with the best overall performance are screened out. These solutions will serve as the final output of the process of this invention and can be used for subsequent large-scale formulation verification, pilot production, and ultimately commercial application development.

[0154] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrates, characterized in that: It is characterized in that it includes the following steps: a) Based on molecular simulation, the low-temperature behavior of chlorfenapyr and lufenuron was analyzed, a molecular library of candidate stabilizers was constructed, and preliminary screening of candidate stabilizers was performed; b) applying low temperature stress to samples of chlorfenapyr and lufenuron suspension concentrate containing the candidate stabilizer using a microfluidic platform, and performing in situ dynamic characterization to obtain dynamic data related to the low temperature stability of the samples; c) processing the dynamic data to establish an evaluation index characterizing the low-temperature stability performance of the candidate stabilizer; d) using a machine learning method to construct a prediction model based on the structural information, concentration information and / or molecular simulation information of the candidate stabilizer and the low-temperature stability performance evaluation index; e) Based on the prediction model, iteratively optimizing the selection, concentration, combination or experimental conditions of the candidate stabilizers through an intelligent feedback mechanism until the target stabilizer or candidate solution is screened out.

2. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The preliminary screening of candidate stabilizers based on molecular simulation in step a) includes: The molecular docking method was used to evaluate the binding affinity of the candidate stabilizer to chlorfenapyr and lufenuron particles or their surfaces, and the molecular dynamics simulation method was used to analyze the dynamic adsorption behavior of the candidate stabilizer on the surface of chlorfenapyr and lufenuron particles and its effect on the interaction between particles.

3. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The construction of the candidate stabilizer molecule library in step a) includes at least one of the following methods: Virtual derivatization of known stabilizers; Based on the low-temperature behavior analysis of chlorfenapyr and lufenuron, candidate stabilizers were designed de novo.

4. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The microfluidic platform in step b) includes a microchannel structure for sample mixing, gradient concentration generation and precise temperature control; the in situ dynamic characterization includes using at least one of microscopic imaging technology, spectral analysis technology and particle size analysis technology to monitor the physical and chemical changes of the sample under low temperature stress in real time.

5. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 4, characterized in that: The microchannel structure of the microfluidic platform further includes at least one of the following structures designed to simulate a specific low-temperature instability mode of the suspension: A structure for simulating the micro-region gradient concentration environment of Ostwald ripening; Microstructures for simulating ice crystal formation and squeezing effects during freeze-thaw cycles; Structures used to achieve specific flow fields or confined spaces that are sensitive to inter-particle interactions or early aggregation behavior.

6. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The low-temperature stability performance evaluation index in step c) is a low-temperature stability digital fingerprint or comprehensive score formed by integrating multiple characteristic parameters in the dynamic data, and the characteristic parameters are selected from at least one of crystallization kinetic parameters, agglomeration kinetic parameters, particle size distribution change parameters, or specific spectral feature evolution parameters over time.

7. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The machine learning method in step d) uses a model capable of fusing multi-scale data, wherein the multi-scale data includes: Microscopic molecular descriptors or interaction parameters of candidate stabilizers derived from molecular simulations, and mesoscopic temporal data or their derived features derived from microfluidic in situ dynamic characterization; The objective function of the machine learning method is expressed as follows: in: θ is the model parameter; N is the number of training samples; (x i ,y i ) is the i-th training sample, x i are input features: molecular descriptors of candidate stabilizers, y i It is the target output: low temperature stability evaluation index; f(x i ; θ) is the model based on the input x i and the predicted value given by the parameter θ; L(y i ,f(x i ; θ)) is the loss function, which is used to measure the difference between the predicted value and the true value; R(θ) is a regularization term used to prevent overfitting; λ is the regularization coefficient.

8. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The intelligent feedback mechanism in step e) includes using a Bayesian optimization algorithm, which includes: The objective function g(x) is used to model the distribution of the low temperature stability performance evaluation index. When the data set D1:t={(x1,y1),...,(x t ,y t )}, the prediction of the new point x follows a Gaussian distribution: where μ t (x) is the posterior mean, is the posterior variance, D 1:t It is the existing observation data.

9. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 8, characterized in that: The iterative optimization guided by the intelligent feedback mechanism includes at least one of the following: Optimize the structural modification direction of candidate stabilizers or the chemical space area to be investigated in the next round of molecular simulation; Dynamically adjust the candidate stabilizer types, concentration gradient ranges, combinations, or parameters of the low-temperature stress program in the next round of microfluidic experiments.

10. The stabilizer screening process for improving the low-temperature stability of chlorfenapyr and lufenuron suspension concentrate according to claim 1, characterized in that: The chlorfenapyr and lufenuron suspension sample is a mixture of a basic suspension comprising chlorfenapyr, lufenuron, a dispersant, a wetting agent and water and the candidate stabilizer, and the low temperature stress includes at least one of uniform cooling, constant low temperature maintenance or freeze-thaw cycles.

Images