A protein phase separation prediction method and system based on physical parameter perturbation

Abstract

The application discloses a protein phase separation prediction method and system based on physical parameter perturbation, and relates to the technical field of bioinformatics. The method comprises the following steps: obtaining an amino acid sequence of a to-be-tested protein and environmental conditions, and constructing input features; inputting the input features into a first neural network to generate physical parameters of interaction parameters; inputting the input features into a second neural network to output a perturbation correction amount for the physical parameters; superimposing the perturbation correction amount on the physical parameters through a limited scale coefficient; calculating interaction parameters according to the corrected physical parameters; and outputting a natural logarithm value of a dilute phase volume fraction to obtain a saturation concentration prediction value of the to-be-tested protein. The method can guarantee rigid constraints of physical laws on a prediction process, and improve prediction accuracy and physical consistency of a model under complex environmental conditions.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and in particular relates to a method and system for predicting protein phase separation based on physical parameter perturbation. Background Technology

[0002] Liquid-liquid phase separation (LLPS) is a core physicochemical mechanism by which intracellular biomolecules (especially proteins and RNA) spontaneously aggregate and form membraneless organelles through multivalent weak interactions. Recent studies have shown that LLPS is widely involved in key cellular processes such as transcriptional regulation, stress response, and signal transduction. Its dysregulation is closely related to neurodegenerative diseases and tumors. Therefore, accurately predicting the phase separation ability of proteins and elucidating their molecular driving mechanisms is of great significance for understanding the principles of cellular compartmentalization and discovering novel drug targets.

[0003] Currently, there are two main technical approaches for the computational prediction of protein phase separation ability: one is a pure physical model based on mean-field theory such as Flory-Huggins. Such models have clear physical interpretability, but they rely heavily on idealized assumptions and are difficult to cover naturally disordered protein sequences and complex environmental conditions (such as changes in temperature and salt concentration), resulting in weak fitting ability; the other is a pure AI model based on deep learning such as fully connected neural networks. Such models usually extract features such as the inherent disorder of the sequence and the content of aromatic residues for binary classification prediction, which has strong nonlinear fitting ability and performs well on known datasets.

[0004] However, existing technologies have insurmountable drawbacks: First, pure AI models are mostly end-to-end "black box" architectures, which are prone to over-reliance on inherent disorder during training. When faced with changes in environmental conditions (such as temperature and salt concentration), they will produce serious false positives and other non-physical predictions, and lack the interpretability of intrinsic physical mechanisms. Second, in existing physics-artificial intelligence hybrid models (such as the traditional PINN), the high-capacity AI correction branches often cover and suppress the physical backbone branches during the optimization process, leading to the phenomenon of "AI hijacking," which causes the physical model parameters to become invalid and gradients to annihilate. The model eventually degenerates into a pure black box, losing the extrapolation reliability brought by physical constraints. Third, the simple weighted fusion of physical formulas and AI outputs cannot structurally ensure that the final prediction results must be rigorously verified by physical laws. Summary of the Invention

[0005] In view of this, the present invention provides a protein phase separation prediction method, system, electronic device and storage medium based on physical parameter perturbation, which solves the problems of AI correction branches in hybrid models easily covering physical laws and causing physical distortion, as well as the poor extrapolation ability of the model under complex environmental conditions and easy generation of non-physical predictions.

[0006] To achieve the above objectives, in a first aspect, the technical solution of this application to solve the technical problem is to provide a protein phase separation prediction method based on physical parameter perturbation, comprising: acquiring the amino acid sequence and environmental conditions of the protein to be tested; constructing input features including at least the protein chain length, wherein the environmental conditions include at least temperature and salt concentration; inputting the input features into a first neural network, wherein the first neural network generates physical parameters for constituting the Flory-Huggins interaction parameters, wherein the physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters; inputting the input features into a second neural network, wherein the second neural network outputs a perturbation correction amount for the physical parameters, wherein the perturbation correction amount serves as a correction term for the physical parameters and is used to constrain the physical parameters; superimposing the perturbation correction amount onto the physical parameters through a constrained scaling coefficient to obtain the corrected physical parameters, wherein the constrained scaling coefficient α is a globally learnable parameter obtained through pre-training; the pre-training process includes: applying a constraint penalty term to the constrained scaling coefficient α during the training of the first neural network and the second neural network, such that it satisfies: |α·Δθ|≤λ(|Θ phys |+ε θ And the corrected physical parameters satisfy Θ final =Θ phys +α·Δθ; where, Θ phys Here, Δθ is the physical parameter, α is the perturbation correction amount, α is the confined scale coefficient, λ is a preset scaling factor greater than 0 and less than 1, and ε is the physical parameter. θ Θ is a preset constant greater than 0. final For the corrected physical parameters, the first neural network, the second neural network, and the restricted scaling coefficient α are pre-trained models and parameters. Based on the corrected physical parameters, Flory-Huggins interaction parameters are calculated. These interaction parameters include a chain length critical term and a physical contribution term determined by the salt response parameter, temperature response parameter, and direct interaction parameter. The natural logarithm of the protein chain length and the interaction parameters are input into a proxy network pre-trained and frozen based on thermodynamic phase equilibrium data. The proxy network outputs the natural logarithm of the dilute phase volume fraction. The natural logarithm of the dilute phase volume fraction is directly mapped to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted saturation concentration of the protein to be tested.

[0007] In one possible implementation, the input features include at least protein chain length, sequence composition features, charge distribution features, inherent disorder features, and environmental condition features. Before constructing the input features, the method further includes: determining whether the amino acid sequence length of the protein to be tested is greater than a preset threshold; if so, performing the step of constructing input features; otherwise, outputting a prediction result that does not meet the phase separation condition; wherein, the preset threshold is 30 amino acids; explicitly encoding the temperature range and salt concentration range, and incorporating the encoded environmental condition features into the input features.

[0008] In one possible implementation, the specific process of the second neural network outputting the perturbation correction amount for the physical parameters is as follows: The residue embeddings of the protein to be tested are extracted using a protein language model, and a sequence semantic vector is obtained through masked mean pooling and projection; wherein, the protein language model is a self-supervised neural network model pre-trained with large-scale protein sequence data, used to output vectorized embedding representations of amino acid residues; the input features are concatenated and input into an environment modulation network, which outputs scaling modulation parameters and translation modulation parameters; wherein, the environment modulation network is a neural network containing at least one fully connected layer, the input of which is the concatenated feature vector, and the outputs scaling modulation parameters for feature-level scaling of the sequence semantic vector and translation modulation parameters for translation; for each element in the sequence semantic vector, the operation of multiplying the element with the corresponding scaling modulation parameter is performed, and the multiplication result is added to the corresponding translation modulation parameter to obtain the environment-modulated sequence representation; based on the environment-modulated sequence representation, the perturbation correction amount for correcting the physical parameters is output.

[0009] In one possible implementation, the preset proportionality coefficient λ ranges from 0.01 to 0.30.

[0010] In one possible implementation, the Flory-Huggins interaction parameter χ is calculated as follows: χ=χ crit +χ direct +χ salt +χ temp ; Among them, the chain length critical term χ crit =(1+√N) 2 / (2N), where N is the protein chain length; χ direct For direct interaction terms; salt response parameter χ salt The square root ion intensity attenuation term is used, based on a combination of Softplus and logarithmic functions; the temperature response parameter χ... temp With 1 / (T+ε T It is directly proportional to ε, where T is the temperature in the environmental conditions. TThis is the preset minimum positive temperature constant.

[0011] In one possible implementation, the proxy network pre-trained and frozen based on thermodynamic phase equilibrium data is constructed as follows: a grid is constructed for protein chain lengths and interaction parameters within a preset range; the corresponding dilute phase volume fraction is obtained by numerical solution using the Flory-Huggins phase equilibrium method; a multilayer perceptron proxy network is trained using the natural logarithm of the protein chain lengths in the grid and the interaction parameters as inputs, and the natural logarithm of the corresponding dilute phase volume fraction as output, with a monotonicity penalty that decreases monotonically as the interaction parameters increase during training; after training, the weights and normalized metadata are saved to obtain the proxy network pre-trained and frozen based on thermodynamic phase equilibrium data.

[0012] In one possible implementation, after obtaining the predicted saturation concentration of the protein, the method further includes a mutation effect assessment: mutating specific key residues in the sequence of the protein to be tested into non-crosslinked residues, locking the environmental conditions and the inherent disorder characteristics unchanged, inputting the mutant sequence into the first neural network and the second neural network, obtaining the mutant saturation concentration prediction value, and comparing the difference with the predicted saturation concentration value of the protein to be tested, so as to isolate and assess the physical contribution of specific key residues or environmental factors in driving the phase separation process.

[0013] Secondly, this application provides a protein phase separation prediction system based on physical parameter perturbation, including a feature vector construction module for acquiring the amino acid sequence and environmental conditions of the protein to be tested, and constructing input features including at least the protein chain length, wherein the environmental conditions include at least temperature and salt concentration; a physical parameter generation module for inputting the input features into a first neural network, wherein the first neural network generates physical parameters for constituting the Flory-Huggins interaction parameters, wherein the physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters; a perturbation output module for inputting the input features into a second neural network, wherein the second neural network outputs a perturbation correction amount for the physical parameters, wherein the perturbation correction amount serves as a correction term for the physical parameters, and is used to constrain the correction of the physical parameters; and a constrained correction module for superimposing the perturbation correction amount onto the physical parameters through a constrained scaling coefficient to obtain the corrected physical parameters, wherein the constrained scaling coefficient α is a globally learnable parameter obtained through pre-training; the pre-training process includes: when training the first neural network and the second neural network, applying a constraint penalty term to the constrained scaling coefficient α, such that it satisfies: |α·Δθ|≤λ(|Θ phys |+ε θ And the corrected physical parameters satisfy Θ final =Θphys +α·Δθ; where, Θ phys Here, Δθ is the physical parameter, α is the perturbation correction amount, α is the confined scale coefficient, λ is a preset scaling factor greater than 0 and less than 1, and ε is the physical parameter. θ Θ is a preset constant greater than 0. final The corrected physical parameters are provided, where the first neural network, the second neural network, and the restricted scaling coefficient α are pre-trained models and parameters. A parameter synthesis module calculates the Flory-Huggins interaction parameters based on the corrected physical parameters. These interaction parameters include a chain length critical term and a physical contribution term determined by the salt response parameter, temperature response parameter, and direct interaction parameter. A surrogate mapping module inputs the natural logarithm of the protein chain length and the interaction parameters into a surrogate network pre-trained and frozen based on thermodynamic phase equilibrium data, and the surrogate network outputs the natural logarithm of the dilute phase volume fraction. A concentration prediction module directly maps the natural logarithm of the dilute phase volume fraction to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted saturation concentration of the protein to be tested.

[0014] Thirdly, this application provides an electronic device, including: an input unit, a memory, at least one processor, and an output unit, wherein the memory stores program instructions that can be executed on the processor, and the processor can execute a protein phase separation prediction method based on physical parameter perturbation by calling the program instructions.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a protein phase separation prediction method based on physical parameter perturbation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a three-tiered structure of "physical parameters generated first, AI only perturbs, and thermodynamic proxy remapping," the scope of the AI ​​branch is strictly limited to the numerical correction of physical parameters, preventing it from directly determining the final prediction result. This overcomes the problems of "AI hijacking," physical distortion, and gradient annihilation caused by the AI ​​branch easily overriding physical laws in traditional physics-AI hybrid models at the architectural level. At the same time, by using the restricted scale coefficient α and its mathematical constraints, the amount of AI perturbation correction is limited at a small scale, realizing a hard structural constraint of "physics-dominated, AI perturbation." This ensures that the weight of physical parameters is not suppressed by AI from the mathematical level, improving the training stability of the model. In addition, by constructing a physical closed loop through a proxy network pre-trained and frozen based on thermodynamic phase equilibrium data, the saturation concentration prediction value output by the model must be verified by strict physical laws. This overcomes the defects of traditional pure AI models that are prone to non-physical predictions and lack of physical interpretability when extrapolating under complex environmental conditions such as temperature and salt concentration. This improves the model's extrapolation ability and physical consistency in sparse regions, providing a reliable computational tool for high-precision prediction of protein phase separation behavior. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of a protein phase separation prediction method based on physical parameter perturbation provided according to an embodiment of the invention. Figure 2 A comparison chart of the prediction performance of different model architectures provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the prediction of protein upper critical dissolution temperature phase behavior in an embodiment of this application. Figure 4 This is a schematic diagram illustrating the prediction of charge shielding effect provided in an embodiment of this application; Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] like Figure 1 As shown, the protein phase separation prediction method based on physical parameter perturbation provided in the first embodiment of the present invention includes the following steps: S100: Obtain the amino acid sequence and environmental conditions of the protein to be tested, and construct input features that include at least the protein chain length and environmental conditions that include at least temperature and salt concentration. Specifically, the amino acid sequence of the protein to be tested is extracted from public databases such as LLPSDB (Liquid-Liquid Phase Separation Database). Environmental conditions include, but are not limited to, temperature and salt concentration. The constructed input feature vector includes at least the protein chain length N (number of amino acids), as well as sequence composition features, charge distribution features, and environmental condition features, and is processed using Z-score normalization or Min-Max normalization algorithms to eliminate the influence of different physicochemical properties' dimensions and numerical magnitudes on the neural network weight updates.

[0020] S200, the input features are input into the first neural network, and the first neural network generates physical parameters for constituting the Flory-Huggins interaction parameters. The physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters. Specifically, the first neural network corresponds to the physical backbone branch structure, which generates physical parameters based on the input sequence features and environmental features through a parameterized expression with clear physical meaning. Among them, the salt response parameter is used to simulate the charge shielding effect, and adopts a square root ion strength decay term based on a combination of Softplus and logarithmic functions to stably simulate nonlinear effects; the temperature response parameter adopts a temperature-dependent form to simulate the phase behavior of the upper critical dissolution temperature; and the direct interaction parameters represent the fundamental physical contributions such as hydrophobic interactions and electrostatic interactions between molecules.

[0021] S300, the input features are input into the second neural network, and the second neural network outputs a perturbation correction amount for the physical parameters. The perturbation correction amount is used as a correction term for the physical parameters to perform constrained correction on the physical parameters. Specifically, the second neural network corresponds to the AI ​​correction branch. This branch extracts protein residue embeddings using a protein language model, obtains a sequence semantic vector through pooling, and modulates the sequence semantic vector using the environmental context. Based on the modulated sequence semantic output, it applies a perturbation correction to the physical parameters. This artificial intelligence branch is structurally limited to outputting only small-scale correction terms for the physical parameters and cannot directly generate the final predicted value of protein saturation concentration.

[0022] S400, the perturbation correction is superimposed on the physical parameters through a restricted scaling coefficient to obtain the corrected physical parameters. The restricted scaling coefficient α is a globally learnable parameter obtained through pre-training. The pre-training process includes: when training the first neural network and the second neural network, applying a constraint penalty term to the restricted scaling coefficient α, such that it satisfies: |α·Δθ|≤λ(|Θ phys |+ε θ And the corrected physical parameters satisfy Θ final =Θ phys +α·Δθ; Specifically, the first neural network and the second neural network are pre-trained models and parameters. Wherein, Θ phys Here, θ is the physical parameter, Δθ is the perturbation correction, α is the confined scaling factor, and Θ is the physical parameter. final For the corrected physical parameters, the constrained scaling factor α controls the overall scaling ratio of the perturbation correction. Its initial value ranges from 0.01 to 0.10, preferably 0.05, and it is adaptively adjusted through backpropagation during training. It should be noted that the above constraint condition (i.e., |α·Δθ|≤λ(|Θ)) applies. phys |+ε θ The penalty term used during the training phase participates in backpropagation to guide the constrained scale coefficient α to converge to a value that conforms to physical constraints. During the inference and prediction phase, α is a fixed global parameter determined after training, based on the model's generalization ability, thus maintaining physical dominance in the prediction.

[0023] The physical parameter Θphys is a three-dimensional vector composed of the salt response parameter, temperature response parameter, and direct interaction parameter; the perturbation correction Δθ is a three-dimensional vector with the same dimension as the physical parameter, and its three components correspond to the perturbation corrections of the salt response parameter, temperature response parameter, and direct interaction parameter, respectively; the correction is added element by element, and the constraint is also satisfied element by element.

[0024] λ is a preset scaling factor greater than 0 and less than 1, used to limit the maximum magnitude of the perturbation correction, that is, to limit the absolute value of the AI ​​perturbation correction to no more than λ times the absolute value of the physical parameter, thereby ensuring that the AI ​​correction branch only produces small-scale corrections to the physical parameter and does not cover the dominant role of the physical backbone branch. ε θ This is a preset, dimensionless constant greater than 0, used to prevent numerical errors caused by zero denominators or meaningless logarithmic operations in formulas. The preferred value range is 10. -6 Up to 10 -8 In this embodiment, 10 is preferred. -6 .

[0025] By using this mathematical constraint, physical parameters and artificial intelligence corrections are superimposed using a restricted scaling coefficient, achieving a hard structural constraint of "physics-driven, AI perturbation," thus preventing the risk of AI overriding the physical model from the mathematical foundation.

[0026] S500 calculates the Flory-Huggins interaction parameters based on the corrected physical parameters. The interaction parameters include the chain length critical term and the physical contribution term determined by the salt response parameter, temperature response parameter and direct interaction parameter. Specifically, the Flory-Huggins interaction parameter is calculated as follows: Interaction Parameter = Critical Chain Length Term + Direct Interaction Term + Salt Response Term + Temperature Response Term. The critical chain length term is calculated from the protein chain length N. The Flory-Huggins theory is a classic standard theoretical model in polymer physics and phase separation. Its interaction parameter measures the strength of intermolecular interactions in polymer-solvent mixtures. When this parameter exceeds a critical value, phase separation occurs in the system. Only then are the relevant parameters input into the proxy network for concentration prediction; if the critical condition is not met, the prediction result indicating no phase separation occurs is directly output, thus forming a strictly thermodynamically closed loop within the model.

[0027] S600 takes the natural logarithm of the protein chain length and the interaction parameters as input to a proxy network that has been pre-trained and frozen based on thermodynamic phase equilibrium data, and outputs the natural logarithm of the dilute phase volume fraction from the proxy network. Specifically, the surrogate network is pre-solved numerically using the Flory-Huggins phase equilibrium equation to calculate the parameter space consisting of protein chain lengths and interaction parameters, generating a large number of phase equilibrium data points. This data is then used to train a multilayer perceptron (MLP) surrogate network. After training, the weights of this surrogate network are frozen and do not participate in subsequent parameter updates during the prediction process.

[0028] S700 directly maps the natural logarithm of the dilute phase volume fraction to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted value of the saturation concentration of the protein to be tested. Under phase separation equilibrium, the volume fraction of the dilute phase corresponds strictly to the protein saturation concentration in physical terms. Therefore, after the proxy network outputs the natural logarithm of the volume fraction of the dilute phase, it can be directly mapped to the natural logarithm of the protein saturation concentration as the final prediction result.

[0029] By establishing a three-level structure of first generating physical parameters, AI perturbing only, and thermodynamic proxy remapping, the scope of the artificial intelligence branch is limited to the numerical correction of physical parameters, so that artificial intelligence cannot directly determine the final prediction result. At the same time, the perturbation correction amount is mathematically constrained by the restricted scale coefficient, ensuring the dominance of physics. This fundamentally overcomes the problem that traditional AI models are prone to producing non-physical predictions when extrapolating under complex environmental conditions, improves the extrapolation ability and physical interpretability of the model under complex environmental conditions, and provides a high-confidence computational tool for predicting protein phase separation behavior.

[0030] In another optional embodiment, to prevent gradient annihilation of weak signal branches during backpropagation, a gain compensation of a preset multiple (such as 1.5 times) is applied to the temperature response parameter in the physical parameters to achieve balanced training of multiple physical modes.

[0031] In an optional embodiment of this application, the input features include at least protein chain length, sequence composition features, charge distribution features, inherent disorder features, and environmental condition features; before constructing the input features, the following are also included: Determine whether the amino acid sequence length of the protein to be tested is greater than a preset threshold. If so, proceed with the step of constructing input features; otherwise, output the prediction result that does not meet the phase separation condition. The preset threshold is 30 amino acids.

[0032] Specifically, the acquired raw protein sequence data is first filtered based on sequence length. Protein liquid-liquid phase separation depends on the formation of a sufficient number of weak multivalent interactions between molecules; amino acid sequences that are too short cannot thermodynamically form effective multivalent cross-linking networks. Therefore, if the sequence length is less than or equal to 30 amino acids, the system directly determines that the protein does not meet the phase separation conditions, outputs the corresponding prediction result, and does not perform further calculations, thus saving computational resources and avoiding invalid operations.

[0033] The temperature range and salt concentration range are explicitly encoded, and the encoded environmental condition features are incorporated into the input features. Specifically, for protein sequences filtered by length, temperature and salt concentration data ranges are explicitly encoded, distinguishing labels such as precise measurements, interval observations, lower boundary values, and upper boundary values. This enables the model to understand the reliability differences in experimental data and incorporate the encoded environmental condition features into the subsequently constructed input features.

[0034] When constructing input features, the input features include at least protein chain length, sequence composition features, charge distribution features, intrinsic disorder features, and environmental condition features. The protein chain length is the number of amino acids in the sequence, denoted as N. Sequence composition features include global features based on amino acid type statistics, such as average hydrophobicity and aromatic residue density. Charge distribution features calculate the distribution of local net charge using a sliding window to capture charge patching effects. Intrinsic disorder features quantify the tendency of proteins to not form stable tertiary structures under physiological conditions using a disorder prediction algorithm; this feature reflects conformational flexibility and provides the necessary topological background for phase separation. Environmental condition features include temperature, salt concentration, divalent ion concentration, and congestion agent identification. The constructed input feature vector contains the above multidimensional features and is processed using Z-score normalization or Min-Max normalization algorithms to eliminate the influence of different physicochemical properties' dimensions and numerical magnitudes on the neural network weight updates.

[0035] In an optional embodiment of this application, the specific process by which the second neural network outputs a perturbation correction for the physical parameters includes: The protein language model is used to extract the residue embeddings of the protein to be tested, and the sequence semantic vector is obtained by mask mean pooling and projection. The protein language model is a self-supervised neural network model pre-trained with large-scale protein sequence data, which is used to output the vectorized embedding representation of amino acid residues. Specifically, the protein language model is a self-supervised neural network model pre-trained on massive amounts of natural protein sequences, capable of capturing the evolutionary information and physicochemical patterns contained within the sequences. When the amino acid sequence of the protein to be tested is input into the model, it outputs a high-dimensional vectorized embedding representation of each amino acid residue. Subsequently, masked mean pooling is performed on these residue embeddings to transform the variable-length residue sequence into a fixed-length vector, which is then projected to obtain a sequence semantic vector of appropriate dimension. This sequence semantic vector contains global semantic information of the protein sequence, providing a fundamental representation for subsequent environmental modulation.

[0036] The input features are concatenated and then fed into the environment modulation network, which outputs scaling modulation parameters and translation modulation parameters. The environment modulation network is a neural network containing at least one fully connected layer. The input of the network is the concatenated feature vector, and the output is the scaling modulation parameters used to scale the sequence semantic vector at the feature level and the translation modulation parameters used for translation. Specifically, the constructed input features (including protein chain length, sequence composition features, charge distribution features, intrinsic disorder features, environmental condition features, etc.) are concatenated into a one-dimensional feature vector. This concatenated feature vector is then input into an environment modulation network, which contains at least one fully connected layer. Through forward propagation, the network outputs two parameter vectors with the same dimension as the sequence semantic vector: scaling modulation parameters and translation modulation parameters. The scaling modulation parameters are used to perform feature-level scaling (multiplication operation) on the sequence semantic vector, while the translation modulation parameters are used for translation (addition operation). Since the input features include environmental condition features, the modulation parameters learned by this network can reflect the influence of the external environment (such as temperature and salt concentration) on the semantic features of the protein sequence.

[0037] For each element in the sequence semantic vector, perform the operation of multiplying the element with the corresponding scaling modulation parameter, and add the result of the multiplication to the corresponding translation modulation parameter to obtain the sequence representation after environmental modulation; Specifically, the sequence semantic vector is multiplied by the scaling modulation parameter element-wise, and then the result is added to the translation modulation parameter. This operation mathematically achieves feature-level linear modulation, ensuring that the final environmentally modulated sequence representation not only contains the protein's own sequence semantic information but also incorporates feature information from the current environmental conditions. The same protein sequence will yield different environmentally modulated sequence representations under different temperatures and salt concentrations, realizing dynamic modulation of sequence semantics by the environmental context.

[0038] Based on the environmentally modulated sequence characterization, a perturbation correction amount is output to correct the physical parameters; The environmentally modulated sequence representation is input into the output layer of the second neural network. After linear or nonlinear transformation, it outputs a perturbation correction for the physical parameters. This perturbation correction is used to make limited corrections to the generated physical parameters. It only serves as a correction term for the physical parameters, rather than directly generating the final predicted value of protein saturation concentration. This ensures, in terms of architecture, that the AI ​​correction branch plays an auxiliary role in correcting the main physical branch.

[0039] By constructing a dynamic environment-aware semantic representation extraction process, the artificial intelligence branch can not only perceive the inherent characteristics of protein sequences, but also perceive changes in the external environment, thereby generating perturbation correction quantities with clear environmental orientation, improving the rationality and robustness of the model's predictions in sparse regions under conditions such as temperature and salt concentration.

[0040] In an optional embodiment of this application, the preset scaling factor λ ranges from 0.01 to 0.30.

[0041] Specifically, the preset scaling factor λ is a key parameter used to limit the maximum magnitude of the perturbation correction. This factor is determined by the mathematical constraint |α·Δθ|≤λ(|Θ_phys|+ε θ The role of this is to ensure that the perturbation correction output by the artificial intelligence branch is always limited to a certain proportion of the magnitude of the physical parameter itself, thereby ensuring the dominant position of the physical backbone branch in the model.

[0042] In this embodiment, the value of λ is preferably between 0.01 and 0.30, more preferably 0.10. If λ is greater than 0.30, the AI ​​correction amount may be too large, which may cause the AI ​​correction branch to cover or distort the true laws of physical parameters during training, causing the model to degenerate into a black box and lose physical interpretability. If λ is less than 0.01, the AI ​​correction amount is excessively suppressed, causing the AI ​​correction branch to be unable to effectively capture subtle features in the sequence and environment, thereby significantly reducing the model's fitting accuracy and extrapolation ability in complex scenarios. Limiting λ to the range of 0.01 to 0.30 can simultaneously take into account the robustness of physical constraints and the flexibility of AI correction, so that the model still has a strong data fitting ability without destroying physical laws.

[0043] By limiting the preset scaling factor λ to a specific range of 0.01 to 0.30, a dynamic balance between AI perturbation correction and physical parameter dominance is achieved. This eliminates the risk of AI overriding the physical model from a mathematical perspective, while ensuring the model's prediction accuracy and robustness under various environmental conditions.

[0044] In an optional embodiment of this application, the formula for calculating the Flory-Huggins interaction parameter χ is as follows: χ=χ crit +χ direct +χ salt +χ temp ; Among them, the chain length critical term χ crit =(1+√N) 2 / (2N), where N is the protein chain length; direct interaction term χ direct Determined by direct interaction parameters, these parameters represent the physical contributions of fundamental intermolecular interactions such as hydrophobic and electrostatic interactions. Their values ​​are determined by the amino acid composition and sequence of the protein sequence itself, and are independent of specific environmental conditions (such as temperature and salt concentration). They characterize the intrinsic phase separation tendency of protein molecules in the absence of environmental disturbances. The salt response parameter χ²... salt The square root ion intensity attenuation term is used, based on a combination of Softplus and logarithmic functions; the temperature response parameter χ... temp With 1 / (T+ε TIt is directly proportional to ε, where T is the temperature in the environmental conditions. T This is a preset minimum positive temperature constant, with units of K and the same dimensions as T.

[0045] Specifically, by decomposing the Flory-Huggins interaction parameters into chain length critical terms, direct interaction terms, salt response terms, and temperature response terms, decoupled modeling of environmental factors and sequence physical properties was achieved.

[0046] Among them, the critical chain length term is calculated from the protein chain length N, which reflects the contribution of the polymer chain length to the critical condition for phase separation. The longer the chain length, the smaller the critical interaction parameter, and the easier it is for the system to undergo phase separation.

[0047] Direct interaction terms are determined by the inherent properties of the protein sequence itself and are not affected by external conditions, providing a stable benchmark for the model to describe the intrinsic driving forces of molecules.

[0048] The salt response parameter employs a square root ion intensity decay term based on a combination of Softplus and logarithmic functions to simulate the Debye-Hückel charge shielding effect. This combined function form effectively avoids the numerical errors caused by the traditional power function when normalizing negative inputs, and its function shape itself constitutes a strong soft constraint on the target physical law (i.e., high sensitivity in low-salt regions and charge shielding effect tending towards saturation in high-salt regions).

[0049] Salt response parameter χ salt Using the square root of the ionic strength as the independent variable, a decay term form combining the Softplus and logarithmic functions is employed. The Softplus function is a monotonically increasing smooth function; the logarithmic function is used to combine the square root of the ionic strength, causing the salt response parameter to monotonically decrease with increasing ionic strength. This combination makes the salt response parameter more sensitive to changes in ionic strength in the low-salinity region, while in the high-salinity region, the salt response parameter gradually approaches saturation, thus simulating the charge shielding effect.

[0050] Temperature response parameters and 1 / (T+ε) T It is proportional to the upper critical dissolution temperature (UCST) phase behavior, which enables the model to accurately predict the physical trend of protein solubility monotonically increasing with increasing temperature.

[0051] ε T For a preset minimum positive temperature constant (e.g., 10), -6 K) is used to prevent the denominator from becoming zero when the temperature approaches absolute zero, thus preventing calculation overflow and ensuring the numerical stability of the formula.

[0052] By explicitly parameterizing and modeling the real thermodynamic laws (chain length criticality, intrinsic molecular interaction, charge shielding, and temperature response), the model output has a clear physical causal chain, which lays a solid theoretical foundation for subsequent analysis of the influence of different environmental factors and the inherent properties of the protein sequence itself on the phase separation tendency, and improves the extrapolation ability and physical interpretability of the model under changes in conditions such as temperature and salt concentration.

[0053] In an optional embodiment of this application, the proxy network, pre-trained and frozen based on thermodynamic phase equilibrium data, is constructed as follows: A grid was constructed for protein chain lengths and interaction parameters within a preset range, and the corresponding dilute phase volume fraction was obtained by numerical solution of the Flory-Huggins phase equilibrium. Specifically, the training of the proxy network is completed independently before subsequent model training. First, a grid is constructed for protein chain lengths and interaction parameters within a preset range. In this embodiment, the preset range includes the range of protein chain length values ​​(e.g., 10 to 2000) and the range of interaction parameter values ​​(e.g., 0.5 to 2.5). This range is discretized at predetermined step sizes to form a series of grid points. For each grid point, the Flory-Huggins phase equilibrium equation is substituted, and the corresponding dilute phase volume fraction is calculated through numerical iteration, thereby generating a large number of high-quality thermodynamic phase equilibrium training data points. This numerical solution process does not rely on any experimental data and fully follows the fundamental laws of thermodynamics, ensuring the physical accuracy of the training data.

[0054] Using the natural logarithm of protein chain length and interaction parameters in the grid as input, and the natural logarithm of the corresponding dilute phase volume fraction as output, a multilayer perceptron agent network is trained, and a monotonicity penalty that decreases monotonically with the increase of interaction parameters is added during training. Specifically, the generated phase equilibrium data points are converted into training samples, where the input is the natural logarithm of the protein chain length in the grid and the interaction parameters, and the output is the natural logarithm of the corresponding dilute phase volume fraction. Subsequently, a multilayer perceptron (MLP) surrogate network is trained using this dataset. To ensure that the surrogate network strictly follows thermodynamic physical laws during the mapping process, a monotonicity penalty that decreases monotonically with increasing interaction parameters is added during training. This penalty constrains the trend of the network output with the interaction parameters, ensuring that the natural logarithm of the dilute phase volume fraction monotonically decreases with increasing interaction parameters, i.e., satisfying... The physical constraints completely prevent the proxy network from outputting prediction results that violate thermodynamic principles.

[0055] Where χ is the Flory-Huggins interaction parameter. It represents the volume fraction of the dilute phase. This is the natural logarithm of the volume fraction of the dilute phase. The natural logarithm of the dilute phase volume fraction is the derivative of the interaction parameter with respect to the natural logarithm. A derivative less than or equal to zero indicates that the natural logarithm of the dilute phase volume fraction monotonically decreases as the interaction parameter increases.

[0056] After training is completed, the weights and normalized metadata are saved to obtain the agent network that was pre-trained and frozen based on thermodynamic phase equilibrium data. Specifically, once the agent network has been trained and reached a preset accuracy (e.g., MSE is 2.58e-05, R...), 2 (Greater than 0.999), its weights and the metadata required for normalization (such as mean and standard deviation) are saved to obtain a surrogate network pre-trained and frozen based on thermodynamic phase equilibrium data. During subsequent model training, the weights of this surrogate network remain frozen and do not participate in gradient backpropagation updates. This "frozen" design constructs an immutable physical mapping standard, ensuring that the interaction parameters output by the subsequent prediction model must be calculated by this physically trained surrogate network to obtain the final predicted value of the dilute phase volume fraction.

[0057] By constructing an independent physical proxy network and freezing its weights, an unavoidable physical closed loop is formed. Since the proxy network uses high-precision physical numerical solutions and incorporates monotonicity penalties during training, the network structure ensures that any final prediction result must pass through an output channel verified by physical laws. This eliminates the possibility of the neural network escaping physical penalties by outputting extreme values, greatly improving the physical consistency and robustness of the model when extrapolating under complex environmental conditions.

[0058] In an optional embodiment of this application, after obtaining the predicted protein saturation concentration, a mutation effect assessment is further included: Specific key residues in the protein sequence to be tested are mutated to non-crosslinked residues, while keeping the environmental conditions and inherent disorder characteristics unchanged. The mutant sequence is then input into a first neural network and a second neural network to obtain the predicted value of the mutant saturation concentration. The difference between the mutant sequence and the predicted value of the saturation concentration of the protein to be tested is compared to isolate and evaluate the physical contribution of specific key residues or environmental factors in driving the phase separation process.

[0059] Specifically, to accurately quantify the physical contribution of specific residues to phase separation, controlled variable calculation experiments were performed. First, the wild-type sequence of the test protein and its corresponding predicted saturation concentration were obtained. Then, specific key residues in the test protein sequence were mutated to non-crosslinked residues. These specific key residues refer to residues capable of providing intermolecular multivalent crosslinking, such as aromatic residues providing π-π stacking interactions (e.g., phenylalanine, tyrosine, tryptophan) or charged residues providing electrostatic attraction; the non-crosslinked residues refer to residues whose side chains do not possess the aforementioned crosslinking interaction capabilities, such as alanine. By mutating key residues to non-crosslinked residues, the physical state of a locally severed protein intermolecular interaction network was simulated.

[0060] When constructing the input features for the mutant sequence, the environmental conditions and the inherent disorder characteristics are forcibly locked unchanged. Specifically, the environmental conditions (such as temperature and salt concentration) and the inherent disorder characteristics of the mutant sequence are set to be exactly the same as those of the wild-type sequence. The purpose of this is to completely shield the interference of background noise by changing only the local cross-linking sites while maintaining the overall conformational flexibility of the protein (i.e., the disorder background) and the absolute consistency of the external thermodynamic environment.

[0061] Subsequently, the mutant sequence is input into the first neural network and the second neural network to obtain the predicted mutant saturation concentration. The predicted mutant saturation concentration is then compared with the predicted saturation concentration of the target protein (i.e., wild type) to obtain the difference between the two. This difference directly reflects the physical contribution of the mutated key residues in driving phase separation.

[0062] By performing the above controlled variable calculation experiments, we achieved “isolated measurement” of the driving force of specific residues, which can accurately locate the core site driving phase separation and provide a high-throughput, high-precision computational tool for discovering novel membraneless organelles and drug targets.

[0063] The training and construction process of the model used in this application and its predictive application are explained in detail with a specific example: (1) Data preparation and preprocessing: This example is based on protein phase separation experimental data from LLPSDB (Liquid-Liquid Phase Separation DataBase). The dataset covers measured values ​​of saturated solubility (Csat) at a temperature range of -14°C to 80°C and a NaCl concentration gradient of 0 to 2000 mM, containing a total of 1288 samples, including 836 positive samples and 452 negative samples.

[0064] In the data preprocessing stage, protein sequences shorter than 30 amino acids were first removed to ensure that the sequences were thermodynamically capable of forming multivalent cross-linked networks. Subsequently, environmental conditions such as temperature and salt concentration were explicitly coded, and different label types were distinguished, including Exact, Interval, Lower Bound, and Upper Bound, so that the model could understand the reliability differences in the experimental data.

[0065] For the initial negative sample set, transmembrane and secretory proteins were excluded using UniProt annotation information. Transmembrane proteins naturally contain a very high proportion of hydrophobic amino acids and aromatic residues (used for anchoring in the lipid bilayer). If they are retained in negative samples where phase separation does not occur, they will severely interfere with the model's judgment of feature boundaries. Secretory proteins, due to their specific folding and modification environments, are also unsuitable as baseline references for aqueous phase separation. This preprocessing step established a pure "cytoplasmic folding protein baseline," ensuring that subsequent positive and negative samples for neural network learning are all within the same aqueous thermodynamic reference frame.

[0066] (2) Extraction of multidimensional physicochemical features: For each preprocessed protein sequence, the system extracts multidimensional core physicochemical measures to construct a feature vector. The feature vector includes at least the protein chain length N (number of amino acids), sequence composition features, charge distribution features, intrinsic disorder features, and environmental condition features. Specifically, the protein chain length N represents the number of amino acids; sequence composition features include the density of aromatic residues (the proportions of phenylalanine (F), tyrosine (Y), and tryptophan (W)) and average hydrophobicity (calculated based on the Kyte-Doolittle scale); the charge distribution feature is the standard deviation of the charge distribution, calculated by sliding a window of 10 amino acids along the sequence, calculating the local net charge within each window (arginine R and lysine K are denoted as +1, aspartic acid D and glutamate E as -1), and calculating the standard deviation of the net charge across all windows; the intrinsic disorder feature uses the IUPred2A algorithm to quantify structural flexibility, characterizing the protein's tendency not to form stable tertiary structures under physiological conditions; environmental condition features include temperature, salt concentration, divalent ion concentration, and congestion agent identification. The extracted features are Z-score normalized and projected onto a zero-mean, unit-variance normal distribution space.

[0067] (3) Generation of physical parameters: The input features are fed into the first neural network (the physical backbone branch), which generates three types of physical parameters through a parameterized representation with explicit physical meaning: 1. Salt response parameters: The square root ion intensity decay term based on the combination of Softplus and logarithmic functions is used to simulate the charge shielding Debye-Hückel effect.

[0068] 2. Temperature response parameters: Using 1 / (T+ε) T The relevant forms are used to simulate the phase behavior of UCST (Upper Critical Solution Temperature) type.

[0069] 3. Direct interaction parameters: These represent the contributions of fundamental hydrophobic and electrostatic interactions between molecules.

[0070] (4) AI perturbation correction: The input features are fed into a second neural network (artificial intelligence branch). This branch uses a protein language model to extract protein residue embeddings, which are then processed through masked mean pooling and projection to obtain a sequence semantic vector. Simultaneously, the physical features, environmental context, and chain length are concatenated and input into an environmental modulation network. This network outputs scaling and translation modulation parameters. For each element in the sequence semantic vector, the element is multiplied by its corresponding scaling parameter, and the result is added to the corresponding translation modulation parameter to obtain the environmentally modulated sequence representation. Based on this sequence representation, a perturbation correction for the physical parameters is output. This perturbation correction is used solely as a correction term for the physical parameters, employing it for constrained correction of the physical parameters.

[0071] (5) Constrained perturbation coupling and parameter synthesis: The perturbation correction is superimposed onto the physical parameters by the restricted scaling factor α to obtain the corrected physical parameters: Θ final =Θ phys +α·Δθ; where, Θ phys Let λ be a physical parameter, Δθ be a perturbation correction, and α be a learnable parameter during neural network training, with an initial value ranging from 0.01 to 0.10. During training, a constraint penalty is imposed on α to ensure that: |α·Δθ|≤λ(|Θ|θ ... phys |+ε θ ), where λ is greater than 0 and less than 1, preferably 0.01 to 0.30, ε θ It is a preset constant greater than 0.

[0072] The Flory-Huggins interaction parameter χ is calculated based on the corrected physical parameters: χ = χ crit +χ direct +χ salt +χ temp Among them, the chain length critical term χ crit =(1+√N) 2 / (2N). When χ is greater than this critical value, phase separation will occur in the system.

[0073] (6) Freeze proxy network mapping: The natural logarithm of the protein chain length and interaction parameters are input into a proxy network that has been pre-trained and frozen based on thermodynamic phase equilibrium data. The proxy network outputs the natural logarithm of the dilute phase volume fraction. The natural logarithm of the dilute phase volume fraction is directly mapped to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted value of the saturation concentration of the protein to be tested.

[0074] The proxy network was constructed as follows: First, a grid was constructed for protein chain lengths (e.g., 10 to 2000) and interaction parameters (e.g., 0.5 to 2.5) within a preset range. The natural logarithm of the dilute phase volume fraction was obtained using Flory-Huggins phase equilibrium numerical solutions, generating 51653 sets of numerical solutions. Using the natural logarithm of the protein chain lengths and interaction parameters in the grid as inputs, and the corresponding natural logarithm of the dilute phase volume fraction as outputs, a multilayer perceptron proxy network was trained. A monotonicity penalty, which decreases monotonically with increasing interaction parameters, was added during training. After training, the proxy network achieved an MSE of 2.58e-05 and an R0... 2 The fitting accuracy is greater than 0.999. During the subsequent training of the first and second neural networks, the weights of this proxy network are frozen and do not participate in gradient updates.

[0075] (7) Training process and effect verification: Specifically, during the training of the first and second neural networks, the system executes a dynamic environment sampling strategy, dynamically and uniformly sampling and projecting interval-type temperature and salt concentration. The training loss employs a hybrid loss mechanism, specifically including: interval loss with out-of-bounds penalty for interval labels; negative sample constraint loss with one-sided ReLU penalty for negative samples that have not undergone phase separation; and a dynamic weight compensation mechanism for tail samples with extremely low and extremely high concentrations. During the gradient update phase, a staged gradient update strategy is adopted, freezing the surrogate network weights and updating the parameters of the first neural network, the second neural network, the environment modulation network, and the constrained scale coefficient α, ensuring the physical consistency and stability of the model training.

[0076] Comparison Example 1: Compared with pure linear regression and pure neural network models: This example uses five random data partitions to compare the prediction accuracy of three models:

[0077] like Figure 2As shown, linear regression cannot characterize nonlinear relationships; pure neural networks have moderate accuracy but large fluctuations; the MAE of this scheme (PIDL model) is reduced by approximately 13.8% compared to the physical baseline (0.889), exhibiting lower variance and higher robustness while maintaining high accuracy. This scheme mathematically constrains the perturbation correction of the AI ​​branch output through the aforementioned restricted scaling coefficient, effectively avoiding the problem of AI hijacking the physical model, thus reactivating the main physical branches, and reducing the weights of key physical parameters from <10. -3 Upgraded to 10 -2 Magnitude.

[0078] Comparative Example 2: Physical Consistency of Model's Temperature Response like Figure 3 As shown, the proposed model successfully recovered the trend of the UCST phase diagram, i.e., solubility increases monotonically with temperature, which is consistent with the prediction of the Flory-Huggins thermodynamic theory. Although the pure neural network model can fit the data, it cannot explain the intrinsic reasons for the solubility changes and cannot maintain the rationality of the trend in the sparse region.

[0079] Comparative Example 3: Charge shielding effect of the model's response to salt concentration Figure 4 The results show that the proposed model correctly captures the nonlinear charge shielding characteristics in the salt concentration response, namely, the rapid increase in solubility in low-salt regions and the tendency to saturate in high-salt regions. This characteristic is consistent with the predictions of the Debye-Hückel electrostatic shielding theory. While pure physical models possess interpretability, they struggle to fit complex experimental data. This proposed model achieves a balance between fitting accuracy and physical interpretability by combining physical parameterization modeling with AI-constrained correction.

[0080] (8) Assessment of mutation effects: Specifically, after obtaining the predicted protein saturation concentration, the process further includes evaluating the mutation effect. Specific key residues in the test protein sequence (e.g., aromatic residues providing π-π stacking interactions such as phenylalanine, tyrosine, and tryptophan) are mutated to non-crosslinked residues (e.g., alanine). While maintaining the environmental conditions and inherent disorder, the mutant sequence is input into the first and second neural networks to obtain the mutant saturation concentration prediction. This prediction is then compared with the predicted saturation concentration of the test protein to isolate and evaluate the physical contribution of specific key residues in driving phase separation. This step, through controlled variable computation experiments, achieves isolated measurement of the driving force of specific residues, thereby accurately locating the core sites driving phase separation.

[0081] By establishing a three-tiered structure of "physical parameters generated first, AI perturbing only, and thermodynamic proxy remapping," the scope of the artificial intelligence branch is limited to the numerical correction of physical parameters, preventing AI from directly determining the final prediction result. Simultaneously, a restricted scale coefficient is used to mathematically constrain the perturbation correction amount, ensuring the dominant role of physics. This scheme fundamentally overcomes the limitations of traditional physical models in fitting complex experimental data and the problem of pure AI models easily producing non-physical predictions during conditional extrapolation. It successfully recovers physical laws such as UCST-type phase diagram trends and charge shielding effects, significantly improving the model's extrapolation ability and physical interpretability under complex environmental conditions.

[0082] A second embodiment of the present invention provides a protein phase separation prediction system based on physical parameter perturbation, comprising: The feature vector construction module is used to obtain the amino acid sequence and environmental conditions of the protein to be tested, and to construct input features that include at least the protein chain length, wherein the environmental conditions include at least temperature and salt concentration. A physical parameter generation module is used to input the input features into a first neural network, and the first neural network generates physical parameters for constituting the Flory-Huggins interaction parameters. The physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters. The perturbation output module is used to input the input features into the second neural network, and the second neural network outputs a perturbation correction amount for the physical parameters. The perturbation correction amount is used as a correction term for the physical parameters to perform limited correction on the physical parameters. The constrained correction module is used to superimpose the perturbation correction amount onto the physical parameters using a constrained scaling coefficient to obtain the corrected physical parameters. The constrained scaling coefficient α is constrained to satisfy: |α·Δθ|≤λ(|Θ) phys |+ε θ And the corrected physical parameters satisfy Θ final =Θ phys +α·Δθ; where, Θ phys The physical parameters are: Δθ is the perturbation correction, α is the confined scale coefficient, λ is a preset scaling factor less than 1, and ε is the physical parameter. θ Θ is a preset constant greater than 0. final The corrected physical parameters are the first neural network, the second neural network, and the restricted scale coefficient α, which are pre-trained models and parameters. The parameter synthesis module is used to calculate the Flory-Huggins interaction parameters based on the corrected physical parameters. The interaction parameters include a chain length critical term and a physical contribution term determined by the salt response parameters, temperature response parameters, and direct interaction parameters. The proxy mapping module is used to input the natural logarithm of the protein chain length and the interaction parameters into a proxy network that has been pre-trained and frozen based on thermodynamic phase equilibrium data, and the proxy network outputs the natural logarithm of the dilute phase volume fraction. The concentration prediction module is used to directly map the natural logarithm of the dilute phase volume fraction to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted saturation concentration value of the protein to be tested.

[0083] like Figure 5 As shown, a third embodiment of the present invention provides an electronic device 800, which includes a memory 801, a processor 802, and an input / output (I / O) interface 803. The memory 801 is used to store instructions. The processor 802 is used to execute the protein phase separation prediction method based on physical parameter perturbation according to the embodiments of this application by calling the instructions stored in the memory 801. The processor 802 is connected to the memory 801 and the I / O interface 803, for example, through a bus system and / or other forms of connection mechanism (not shown). The memory 801 can be used to store programs and data, including the program of the protein phase separation prediction method based on physical parameter perturbation involved in the embodiments of this application. The processor 802 executes various functional applications and data processing of the electronic device 800 by running the program stored in the memory 801.

[0084] In this embodiment, the processor 802 can be implemented using at least one of the following hardware forms: digital signal processor (DSP), field programmable gate array (FPGA), and programmable logic array (PLA). The processor 802 can be one or a combination of several of the following: central processing unit (CPU) or other processing units with data processing capability and / or instruction execution capability.

[0085] The memory 801 in this embodiment may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0086] In this embodiment, the I / O interface 803 can be used to receive input instructions (such as numeric or character information, and to generate key signal inputs related to user settings and function control of the electronic device 800), and can also output various information (such as images or sounds) to the outside. In this embodiment, the I / O interface 803 may include one or more of the following: a physical keyboard, function keys (such as volume control keys, power buttons, etc.), a mouse, a joystick, a trackball, a microphone, a speaker, and a touch panel.

[0087] The fourth embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0088] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one of relational and non-relational databases. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these. The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0089] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A protein phase separation prediction method based on physical parameter perturbation, characterized in that, Including the following steps: Obtain the amino acid sequence and environmental conditions of the protein to be tested, and construct an input feature that includes at least the protein chain length, wherein the environmental conditions include at least temperature and salt concentration; The input features are input into a first neural network, which generates physical parameters for constituting the Flory-Huggins interaction parameters. The physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters. The input features are input into a second neural network, which outputs a perturbation correction amount for the physical parameters. The perturbation correction amount is used as a correction term for the physical parameters to perform constrained correction on the physical parameters. The perturbation correction is superimposed on the physical parameters by a restricted scaling coefficient to obtain the corrected physical parameters. The restricted scaling coefficient α is a globally learnable parameter obtained through pre-training. The pre-training process includes: when training the first neural network and the second neural network, applying a constraint penalty term to the restricted scaling coefficient α, such that it satisfies: |α·Δθ|≤λ(|Θ) phys |+ε θ And the corrected physical parameters satisfy Θ final =Θ phys +α·Δθ; Where, Θ phys Here, Δθ is the physical parameter, α is the perturbation correction amount, α is the confined scale coefficient, λ is a preset scaling factor greater than 0 and less than 1, and ε is the physical parameter. θ Θ is a preset constant greater than 0. final The corrected physical parameters are obtained by pre-training models for the first and second neural networks. The Flory-Huggins interaction parameters are calculated based on the corrected physical parameters. These interaction parameters include a chain length critical term and a physical contribution term determined by the salt response parameters, temperature response parameters, and direct interaction parameters. The natural logarithm of the protein chain length and the interaction parameters are input into a proxy network that has been pre-trained and frozen based on thermodynamic phase equilibrium data, and the proxy network outputs the natural logarithm of the dilute phase volume fraction. The natural logarithm of the dilute phase volume fraction is directly mapped to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted value of the saturation concentration of the protein to be tested.

2. The protein phase separation prediction method based on physical parameter perturbation as described in claim 1, characterized in that: The input features include at least protein chain length, sequence composition features, charge distribution features, inherent disorder features, and environmental condition features; Before constructing the input features, the following is also included: Determine whether the amino acid sequence length of the protein to be tested is greater than a preset threshold; If so, proceed with the step of constructing the input features; otherwise, output the prediction result that does not satisfy the phase separation condition. The preset threshold is 30 amino acids; The temperature range and salt concentration range are explicitly encoded, and the encoded environmental condition features are incorporated into the input features.

3. The protein phase separation prediction method based on physical parameter perturbation as described in claim 1, characterized in that, The specific process by which the second neural network outputs the perturbation correction for the physical parameters is as follows: The protein language model is used to extract the residue embeddings of the protein to be tested, and the sequence semantic vector is obtained by mask mean pooling and projection; wherein, the protein language model is a self-supervised neural network model pre-trained with large-scale protein sequence data, which is used to output the vectorized embedding representation of amino acid residues. The input features are concatenated and then input into an environment modulation network, which outputs scaling modulation parameters and translation modulation parameters. The environment modulation network is a neural network containing at least one fully connected layer. The input of the network is the concatenated feature vector, and the output is scaling modulation parameters for feature-level scaling of the sequence semantic vector and translation modulation parameters for translation. For each element in the sequence semantic vector, the operation of multiplying the element with the corresponding scaling modulation parameter is performed, and the result of the multiplication is added to the corresponding translation modulation parameter to obtain the sequence representation after environmental modulation; Based on the environmentally modulated sequence characterization, a perturbation correction amount is output to correct the physical parameters.

4. The protein phase separation prediction method based on physical parameter perturbation as described in claim 1, characterized in that: The preset proportional coefficient λ ranges from 0.01 to 0.

30.

5. The protein phase separation prediction method based on physical parameter perturbation as described in claim 1, characterized in that: The formula for calculating the Flory-Huggins interaction parameter χ is as follows: x=x crit +x direct +x salt +x temp ; Among them, the chain length critical term χ crit =(1+√N) 2 / (2N), where N is the protein chain length; χ direct For direct interaction terms; salt response parameter χ salt The square root ion intensity attenuation term is used, based on a combination of Softplus and logarithmic functions; the temperature response parameter χ... temp With 1 / (T+ε T It is directly proportional to ε, where T is the temperature in the environmental conditions. T This is the preset minimum positive temperature constant.

6. The protein phase separation prediction method based on physical parameter perturbation as described in claim 1, characterized in that, The proxy network, which is pre-trained and frozen based on thermodynamic phase equilibrium data, is constructed as follows: A grid was constructed for protein chain lengths and interaction parameters within a preset range, and the corresponding dilute phase volume fraction was obtained by numerical solution of the Flory-Huggins phase equilibrium. Using the natural logarithm of the protein chain length in the grid and the interaction parameters as inputs, and the natural logarithm of the corresponding dilute phase volume fraction as outputs, a multilayer perceptron agent network is trained, and a monotonicity penalty that monotonically decreases as the interaction parameters increase is added during training. After training is completed, the weights and normalized metadata are saved to obtain the agent network that was pre-trained and frozen based on thermodynamic phase equilibrium data.

7. The protein phase separation prediction method based on physical parameter perturbation as described in claim 2, characterized in that, After obtaining the predicted protein saturation concentration, the evaluation of mutation effects is also included: Specific key residues in the protein sequence to be tested are mutated to non-crosslinked residues, while the environmental conditions and inherent disorder characteristics are kept unchanged. The mutant sequence is input into the first neural network and the second neural network to obtain the mutant saturation concentration prediction value, and the difference is compared with the saturation concentration prediction value of the protein to be tested, so as to isolate and evaluate the physical contribution of specific key residues or environmental factors in driving the phase separation process.

8. A protein phase separation prediction system based on physical parameter perturbation, characterized in that, include: The feature vector construction module is used to obtain the amino acid sequence and environmental conditions of the protein to be tested, and to construct input features that include at least the protein chain length, wherein the environmental conditions include at least temperature and salt concentration. A physical parameter generation module is used to input the input features into a first neural network, and the first neural network generates physical parameters for constituting the Flory-Huggins interaction parameters. The physical parameters include at least salt response parameters, temperature response parameters, and direct interaction parameters. The perturbation output module is used to input the input features into the second neural network, and the second neural network outputs a perturbation correction amount for the physical parameters. The perturbation correction amount is used as a correction term for the physical parameters to perform limited correction on the physical parameters. The constrained correction module is used to superimpose the perturbation correction amount onto the physical parameters through a constrained scaling coefficient to obtain the corrected physical parameters. The constrained scaling coefficient α is a globally learnable parameter obtained in advance through training. The pre-training process includes: when training the first neural network and the second neural network, applying a constraint penalty term to the restricted scaling coefficient α, such that it satisfies: |α·Δθ|≤λ(|Θ) phys |+ε θ And the corrected physical parameters satisfy Θ final =Θ phys +α·Δθ; where, Θ phys Here, Δθ is the physical parameter, α is the perturbation correction amount, α is the confined scale coefficient, λ is a preset scaling factor greater than 0 and less than 1, and ε is the physical parameter. θ Θ is a preset constant greater than 0. final The corrected physical parameters are obtained by pre-training models for the first and second neural networks. The parameter synthesis module is used to calculate the Flory-Huggins interaction parameters based on the corrected physical parameters. The interaction parameters include a chain length critical term and a physical contribution term determined by the salt response parameters, temperature response parameters, and direct interaction parameters. The proxy mapping module is used to input the natural logarithm of the protein chain length and the interaction parameters into a proxy network that has been pre-trained and frozen based on thermodynamic phase equilibrium data, and the proxy network outputs the natural logarithm of the dilute phase volume fraction. The concentration prediction module is used to directly map the natural logarithm of the dilute phase volume fraction to the natural logarithm of the protein saturation concentration, thereby obtaining the predicted saturation concentration value of the protein to be tested.

9. An electronic device, characterized in that, include: The method includes an input unit, a memory, at least one processor, and an output unit, wherein the memory stores program instructions that can be executed on the processor, and the processor can execute the protein phase separation prediction method based on physical parameter perturbation as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the protein phase separation prediction method based on physical parameter perturbation as described in any one of claims 1 to 7.

Images