A protein phase separation prediction method and system based on a physical information neural network, an electronic device, and a storage medium

Abstract

This application discloses a protein phase separation prediction method, system, electronic device, and storage medium based on a physical information neural network, relating to the field of bioinformatics. The method includes: acquiring the amino acid sequence of the protein to be tested and extracting multidimensional physicochemical features of the amino acid sequence to construct a feature vector; inputting the feature vector into a pre-trained physical information neural network; the physical information neural network includes a shared hidden layer and a first output branch and a second output branch respectively connected to the shared hidden layer; outputting the predicted probability of phase separation of the protein to be tested through the first output branch and outputting a continuous scalar value as a thermodynamic score through the second output branch; and outputting the phase separation prediction result of the protein to be tested based on the predicted probability. This method effectively overcomes the excessive reliance of traditional sequence prediction models on inherent disorder, reducing the false positive rate while improving the interpretability of the results through physical mechanism constraints.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and particularly relates to a protein phase separation prediction method, system, electronic device and storage medium based on physical information neural network. Background Technology

[0002] Liquid-liquid phase separation (LLPS) is a core physicochemical mechanism by which intracellular biomolecules (especially proteins and RNA) spontaneously aggregate through multivalent weak interactions to form membraneless organelles (MLOs). Recent studies have shown that LLPS participates in various key cellular processes, including transcriptional regulation, stress response, and signal transduction, and its dysregulation is closely related to neurodegenerative diseases, tumors, and metabolic diseases. 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, the identification of protein phase-separation ability mainly relies on intracellular or in vitro remodeling experiments based on fluorescence microscopy, such as fluorescence recovery after photobleaching (FRAP) and droplet formation observation. While these methods are accurate, they suffer from limitations such as low throughput, long processing times, high costs, and difficulty in systematically resolving sequence-phase transition relationships. To overcome these bottlenecks, researchers have begun to explore computational prediction methods. For example, publicly available LLPS databases (such as PhaSepPro) compile experimentally validated annotations of phase-separating proteins and their disordered region (IDR). Building upon this, deep learning models using fully connected neural networks (MLPs) have been developed to perform binary classification predictions of protein phase separation based on features such as aromatic residue content, disorder, and hydrophobicity extracted from amino acid sequences. These models have shown some classification accuracy on known datasets.

[0004] However, existing deep learning-based phase separation prediction models have fundamental flaws. First, these models are mostly end-to-end "black box" architectures, making them prone to over-reliance on the "intrinsic disorder" of sequences during training. This leads to the misclassification of many structurally flexible but naturally disordered proteins lacking multivalent cross-linking sites (such as aromatic residues) as phase-separating proteins, resulting in severe false positives (i.e., "illusion predictions"). Furthermore, the abundance of aromatic residues in transmembrane proteins in traditional negative sample sets exacerbates feature confusion. Second, existing models only output binary classification results, lacking interpretability for intrinsic physical mechanisms (such as thermodynamic stability driven by π-π interactions), failing to distinguish between different types of non-membrane organelles (such as nucleoli and stress granules), and struggling to quantify the contribution of key residues to phase separation through computer-simulated site-directed mutagenesis. Therefore, there is an urgent need for a protein phase separation prediction method based on physically-informed neural networks that can integrate biophysical prior knowledge, eliminate false positive predictions, and possess physical interpretability. 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 information neural network, which solves the problems of high false positive rate, lack of physical mechanism constraint and poor interpretability of prediction results in phase separation prediction.

[0006] To achieve the above objectives, in a first aspect, the technical solution of the present invention to solve the technical problem is to provide a protein phase separation prediction method based on a physical information neural network, comprising: acquiring the amino acid sequence of the protein to be tested, and extracting multidimensional physicochemical features of the amino acid sequence to construct a feature vector; inputting the feature vector into a pre-trained physical information neural network; the physical information neural network includes a shared hidden layer, and a first output branch and a second output branch respectively connected to the shared hidden layer; outputting the predicted probability of phase separation of the protein to be tested through the first output branch, and outputting a continuous scalar value as a thermodynamic score through the second output branch; outputting the phase separation prediction result of the protein to be tested according to the predicted probability; wherein, the physical information neural network is constrained by an asymmetric physical loss term and an auxiliary physical anchor point during the training phase; the asymmetric physical loss term is constructed such that: when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output, otherwise zero is output; the constraint process of the auxiliary physical anchor point includes: using an adaptive learning parameter constrained to be positive by a physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and including the deviation between the thermodynamic score and the reference score as an anchor point constraint term in the total loss function.

[0007] In some embodiments, the multidimensional physicochemical characteristics include aromatic residue density, intrinsic disorder, average hydrophobicity, and standard deviation of charge distribution.

[0008] In some embodiments, the extraction process of the standard deviation of charge distribution includes: sliding a sliding window of predetermined length along the amino acid sequence, calculating the local net charge within each window, and calculating the standard deviation of the net charge of all windows.

[0009] In some embodiments, the first output branch uses a Sigmoid activation function to map the network output to a predicted probability in the (0,1) interval; the second output branch uses a Linear activation function to output an unbounded scalar value.

[0010] In some embodiments, the physical information neural network introduces adversarial trap sequences as negative samples during the training phase. The adversarial trap sequences meet the following feature filtering criteria: the inherent disorder of the sequence is greater than a first predetermined threshold, and the aromatic density is less than a second predetermined threshold.

[0011] In some embodiments, before extracting the multidimensional physicochemical features of the amino acid sequence, data preprocessing is further included, specifically: for the initial negative sample set, UniProt annotation information is used to exclude transmembrane proteins and secretory proteins in order to establish a baseline of cytoplasmic folding proteins.

[0012] In some embodiments, after outputting the phase separation prediction result of the test protein based on the predicted probability, a mutation effect evaluation step is further included, specifically: obtaining the wild-type amino acid sequence of the test protein, and the mutant amino acid sequence by mutating specific key residues that provide multivalent cross-linking to non-cross-linking residues or by performing deletion mutations; extracting the multidimensional physicochemical features of the mutant amino acid sequence, and when constructing the feature vector, forcibly locking the input value of the inherent disorder feature to be consistent with the inherent disorder feature value of the wild-type amino acid sequence; inputting the constructed feature vector into the physical information neural network, obtaining the mutant prediction probability output by the first output branch, and comparing the difference with the wild-type prediction probability to isolate and evaluate the physical contribution of specific key residues in driving the phase separation process.

[0013] Secondly, this invention provides a protein phase separation prediction system based on a physical information neural network, comprising: a feature vector construction module for acquiring the amino acid sequence of the protein to be tested and extracting multidimensional physicochemical features of the amino acid sequence to construct a feature vector; an output module for inputting the feature vector into a pre-trained physical information neural network; the physical information neural network includes a shared hidden layer and a first output branch and a second output branch respectively connected to the shared hidden layer; the first output branch outputs the predicted probability of phase separation of the protein to be tested, and the second output branch outputs a continuous scalar value as a thermodynamic score; and the result prediction... The module is used to output the phase separation prediction result of the protein to be tested based on the predicted probability. During the training phase, the physical information neural network is constrained by an asymmetric physical loss term and auxiliary physical anchors. The asymmetric physical loss term is constructed such that when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output; otherwise, zero is output. The constraint process of the auxiliary physical anchors includes: using adaptive learning parameters constrained to positive numbers by a physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and incorporating the deviation between the thermodynamic score and the reference score as an anchor constraint term into the total loss function.

[0014] Thirdly, the present invention provides an electronic device comprising: an input unit, a memory, at least one processor, and an output unit, wherein the memory stores program instructions executable on the processor, and the processor can execute the protein phase separation prediction method based on physical information neural networks by calling the program instructions.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the protein phase separation prediction method based on a physical information neural network.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: By introducing a ReLU physical penalty term and adversarial trap data characterized by "high disorder and low viscosity," the model is forced to learn the physical nature of π-π interactions, fundamentally overcoming the shallow dependence of traditional algorithms on disorder features and significantly reducing the false positive rate. Simultaneously, it outputs not only statistical probabilities but also thermodynamic scores reflecting the system's energy state, upgrading phase separation prediction from simple binary classification to physical modeling with thermodynamic insights. Furthermore, by combining manifold learning clustering and computational mutation scanning, it can accurately quantify the thermodynamic contribution of specific residues to phase separation, providing a high-confidence computational tool for discovering novel non-membrane organelles and drug targets. Attached Figure Description

[0017] Figure 1A flowchart illustrating the steps of the protein phase separation prediction method based on a physical information neural network provided in the first embodiment of the present invention; Figure 2 This is a spatial clustering diagram of the thermodynamic fingerprint of candidate proteins under adaptive physical anchor constraints provided in an embodiment of this application. Figure 3 Aromatic residue density distribution diagram of phase separation positive sample and folded protein negative sample provided in the embodiments of this application; Figure 4 A schematic diagram showing the distribution of the anti-trap sequence provided in this application embodiment in the feature space composed of disorder degree and aromatic density; Figure 5 A schematic diagram comparing the probability density of trap sequence predictions using the conventional model provided in this application and the physical information neural network of this application; Figure 6 The receiver operating characteristic curve of the physical information neural network provided in this application embodiment on an independent validation set; Figure 7 This is a schematic diagram comparing the phase separation prediction probabilities of wild-type and mutant Bmal1 protein provided in an embodiment of this application. Figure 8 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 a physical information neural network provided in the first embodiment of the present invention includes the following steps: S100: Obtain the amino acid sequence of the protein to be tested, and extract the multidimensional physicochemical features of the amino acid sequence to construct a feature vector; In this step, the amino acid sequence string of the protein to be tested is received first. In order for the neural network to learn the underlying physical driving mechanism of the protein sequence, feature engineering is required on this one-dimensional character sequence to extract multi-dimensional numerical features that can characterize the physicochemical properties of the protein molecule. These numerical features are then combined and concatenated to construct an input feature vector that can be read by the neural network.

[0020] S200, the feature vector is input into a pre-trained physical information neural network; the physical information neural network includes a shared hidden layer, and a first output branch and a second output branch respectively connected to the shared hidden layer; The physical information neural network employs a base layer such as a multilayer perceptron (MLP) combined with a dual-branch parallel topology. Shared hidden layers can comprise multi-level neural networks used to extract deep nonlinear mappings from the input feature vectors. By sharing underlying weights, the network can simultaneously capture statistical distribution patterns and physical driving mechanisms within its high-dimensional manifold space. After passing through the shared hidden layers, the network splits into a first output branch and a second output branch to perform functionally decoupled task prediction.

[0021] S300 outputs the predicted probability of phase separation of the protein to be tested through the first output branch, and outputs a continuous scalar value as a thermodynamic score through the second output branch. The first output branch maps the network output to a value in the interval (0,1), which is the predicted probability, representing the statistically significant possibility that the protein to be tested will undergo phase separation. The second output branch outputs a continuous scalar value as a thermodynamic score.

[0022] It should be noted that the thermodynamic score is not a numerical value with absolute energy dimensions as in classical physics, but rather a dimensionless proxy variable for network adaptive learning. This proxy variable characterizes the thermodynamic stability of the system assembly: when the sequence conformation is extremely soft but lacks viscous sites, the real physical system is unstable, and the thermodynamic score is mapped to an extremely low value. Through this dimensionless mapping, the predicted probability and the thermodynamic score are projected into the same numerical space, allowing for direct numerical comparison between the two.

[0023] S400 outputs the phase separation prediction results of the protein to be tested based on the predicted probability; During the model deployment and actual inference phase, the predicted probability obtained from the first output branch is compared with the preset classification threshold. If the predicted probability meets the condition, the prediction result "the protein to be tested has a tendency to phase separation" is output; at this time, the thermodynamic score can be used as a stability reference index within the system and no longer determines the final classification result.

[0024] In this embodiment, the physical information neural network is constrained during the training phase by an asymmetric physical loss term and auxiliary physical anchor points; The asymmetric physical loss term is constructed such that when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output; otherwise, zero is output. The constraint process for the auxiliary physical anchor points includes: using adaptive learning parameters that are positively constrained by the physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and incorporating the deviation between the thermodynamic score and the reference score as an anchor point constraint term into the total loss function.

[0025] During the model training phase, the operating logic is as follows: when the network outputs an extremely high phase separation prediction probability, but its own physical characteristics result in an extremely low thermodynamic score calculated by the network (i.e., the predicted probability value > the thermodynamic score value), it indicates that the network has fallen into statistical illusion. At this time, physical inconsistency is triggered, and a loss value greater than zero is output as a gradient penalty, forcing backpropagation to correct the network weights. Conversely, if the predicted probability value ≤ the thermodynamic score value, it means that the prediction given by the network conforms to the thermodynamic physical laws, and is judged as physically consistent, and the system does not impose additional penalties. Through the above asymmetric truncation logic, this invention successfully encodes the interaction laws into the neural network.

[0026] Specifically, the asymmetric truncation logic, which determines whether to generate a loss value by comparing the predicted probability with the thermodynamic score, can be implemented at the underlying mathematical level by introducing a rectified linear unit (ReLU).

[0027] Asymmetric physical loss term The calculation formula can be constructed as follows: ; Or, equivalently: ; Where P is the predicted probability value and Thermo is the thermodynamic score value.

[0028] Binary cross-entropy loss combined with regular training samples The basic loss function used in the training phase of a physical information neural network can be defined as: ; in, The physical weighting coefficient is set (preferably between 0.5 and 0.7).

[0029] In actual training, relying solely on the aforementioned asymmetric physical loss term may be insufficient to constrain the behavior of the second output branch: the network can infinitely increase the output value of the thermodynamic score, thus... Always negative, thus Reduced to zero. To avoid this "shortcut collapse," this embodiment adds an auxiliary physical anchor point to the thermodynamic score during training. Specifically, for each training sample, an auxiliary physical anchor point is introduced. Physical gating mechanism calculation reference score The calculation formula is as follows: ; in, , , , These correspond to aromatic residue density, standard deviation of charge distribution, intrinsic disorder, and average hydrophobicity, respectively. , , , These are adaptively learned parameters that are automatically optimized during backpropagation of the network. for Activation functions, at a mathematical level, enforce that the final combination coefficients of all features must be strictly positive. This is a truncation function used to restrict the result to the interval [0, 5.0].

[0030] The significance of introducing this physical gating mechanism lies in forcibly maintaining the physical baseline of phase separation, ensuring that aromatic density (representing π-π interactions), charge distribution polarization (representing electrostatic interactions), and intrinsic disorder (representing conformational entropy driving force) are all strictly locked as positive driving forces. Under this constraint, the convergence weights learned autonomously by the network from massive amounts of data are: aromatic density +1.293, intrinsic disorder +0.949, charge polarization +0.944, and hydrophobicity +0.882.

[0031] Finally, after introducing auxiliary physical anchors during training, the complete total loss function was explicitly defined as follows: ; in, Mean square error, The constraint weight for the auxiliary physical anchor point is set to 0.1. This mechanism forcibly anchors the thermodynamic score of the network's second branch within a reasonable range of physical dimensions, fundamentally eliminating the possibility of the network escaping physical penalties by outputting extreme values.

[0032] To further verify the effectiveness of the quantitative physical anchoring mechanism, the optimized model was used to output the latest thermodynamic fingerprint spatial clustering graph (e.g., Figure 2(As shown). The results show that, under the constraints of this adaptive anchoring formula, known types of physical benchmark proteins (such as the nucleolar marker NPM1 / FBL and the stress granulation marker FUS / G3BP1) are precisely and stably located at the core gravitational centers of their respective candidate protein clusters. Simultaneously, the clustering diagram objectively presents transitional buffer zones representing "dual localization / intermediate states," rejecting hard cut boundaries lacking biological basis. This demonstrates that the thermodynamic score output by the model has extremely high smoothness and continuous physical meaning, confirming the success of the "preventing shortcut collapse" mechanism in this scheme.

[0033] In some implementations, multidimensional physicochemical characteristics include aromatic residue density, intrinsic disorder, average hydrophobicity, and standard deviation of charge distribution. Aromatic residue density refers to the proportion of phenylalanine (F), tyrosine (Y), and tryptophan (W) amino acid residues with aromatic ring structures in the sequence of the protein being tested. Physically, aromatic residues constitute the core "sticky sites" of multivalent cross-linking between protein molecules by providing π-π stacking interactions and cation-π interactions. By treating this as an independent feature dimension, neural networks can learn intermolecular interaction potential beyond flexibility.

[0034] Intrinsic disorder is used to characterize the tendency of a protein sequence not to form a stable tertiary structure under physiological conditions, i.e., conformational flexibility. Intrinsic disorder regions (IDRs) typically contain a large number of polar or charged residues, providing the necessary conformational space and entropy contribution for phase separation. Disorder values ​​are used as input features to establish the basic topological background for liquid-liquid phase separation (LLPS) of proteins.

[0035] Average hydrophobicity reflects the tendency of amino acid residues to avoid water molecules, and hydrophobic interactions are important non-covalent forces driving protein folding and assembly. In this embodiment, the Kyte-Doolittle scale is preferably used to score the hydrophobicity of each residue in the sequence and calculate the global average value. This feature helps the model identify local concentration increases caused by hydrophobic collapse.

[0036] Since the arrangement of charged residues in a protein sequence (not just the total net charge) has a significant impact on phase separation, the standard deviation of charge distribution is introduced. This feature captures the distribution characteristics of charge blocks in the sequence by measuring the local dispersion of the sequence charge distribution, effectively replacing the calculation of complex sequence charge parameters (such as SCD) and improving the computational efficiency of feature extraction.

[0037] After extracting the values ​​from the four dimensions mentioned above, they are concatenated to construct a four-dimensional initial feature vector. To eliminate the influence of different physicochemical properties and numerical magnitudes (for example, density is usually between 0 and 0.2, while hydrophobicity has a larger numerical range) on the neural network weight updates, Z-score normalization or Min-Max normalization algorithms can be used to preprocess the feature vector.

[0038] By combining and mapping the aforementioned four-dimensional specific features, protein sequences can be transformed from a simple "character stream" into a "physicochemical feature stream" with physical connotations. This specific combination of dimensions can establish a clear physical boundary between phase-separated positive and negative samples in the feature space, thus providing a high-quality input benchmark for the subsequent learning of "thermodynamic scores" by the dual-branch neural network.

[0039] In some implementations, the process of extracting the standard deviation of charge distribution includes: sliding a sliding window of predetermined length along the amino acid sequence, calculating the local net charge within each window, and calculating the standard deviation of the net charge across all windows.

[0040] When processing the input one-dimensional amino acid sequence, a sliding window of predetermined length is used to traverse the entire sequence with a step size of 1. At each window position, the total net charge within that local window is calculated (e.g., arginine R and lysine K are denoted as +1, aspartic acid D and glutamic acid E as -1, and the rest as 0). After traversing the entire sequence, a sequence set containing the net charge values ​​of all local windows is obtained, and finally, the standard deviation of this set is calculated and input as the standard deviation feature of the charge distribution into the physical information neural network.

[0041] Preferably, the predetermined length is 10 amino acid residues.

[0042] In polymer physics, the total net charge of a sequence cannot reflect the true electrostatic interaction pattern. For example, the total net charge of an "alternating charge arrangement (+-+-+-)" and a "block charge arrangement (++++----)" may both be 0, but the latter (charge patches) is much more likely to drive phase separation.

[0043] In intrinsically disordered proteins (IDPs), the typical physical span of transient interaction-binding motifs (MoRFs) or locally charged sticky patches that mediate liquid-liquid phase separation is usually between 8 and 12 residues. A window of length 10 can capture this microscopic feature most precisely, avoiding feature smoothing blurring caused by an overly large window or high-frequency noise introduced by an overly small window.

[0044] By applying the above constraints, not only is the local electrostatic topological information of the sequence accurately preserved, but the O(N) value calculated using traditional SCD parameters is also reduced. 2 The time complexity has been reduced from O(N) to linear time complexity. This enables real-time online computation for processing ultra-large-scale proteomics sequences (such as genome-wide mutation scans) during feature extraction, greatly improving its industrial applicability.

[0045] In some implementations, the first output branch uses a Sigmoid activation function to map the network output to a predicted probability in the (0,1) interval; the second output branch uses a Linear activation function to output an unbounded scalar value.

[0046] The task of the first output branch is to perform classification prediction. In this embodiment, the branch ends with the Sigmoid activation function, whose mathematical expression is f(x)=1 / (1+e^(-x)). -x This function smoothly and nonlinearly compresses the unbounded real-domain features (Logits) passed from the hidden layer into the (0,1) interval. In a biological sense, this output value is strictly defined as the "predicted probability P", which directly represents the probability of liquid-liquid phase separation of the test sequence in a statistical distribution, so as to facilitate subsequent classification and comparison with a threshold (such as 0.5).

[0047] Unlike the first output branch, the second output branch's task is to construct a proxy thermodynamic model. In this embodiment, the branch terminates with a Linear activation function (i.e., a linear identity mapping f(x) = x). The physical reason for not using saturating activation functions like Sigmoid or Tanh is that the thermodynamic stability of real physical systems (such as free energy change) is a physical quantity with a wide dynamic range; there is a huge energy potential difference between extremely unstable sequences and highly stable sequences. The Linear activation function allows this branch to output an unbounded, continuous scalar value, Thermo. This unbounded output ensures that the network can output extremely large or small scalar values ​​when facing extreme trap sequences (such as purely disordered flexible peptides), thereby generating a backpropagation penalty signal with sufficient gradient in the subsequent ReLU (P-Thermo) physical loss calculation, avoiding the gradient vanishing problem caused by activation function saturation.

[0048] In some implementations, the physical information neural network introduces adversarial trap sequences as negative samples during the training phase. The adversarial trap sequences meet the following feature filtering criteria: the inherent disorder of the sequence is greater than a first predetermined threshold, and the aromatic density is less than a second predetermined threshold.

[0049] In traditional protein phase separation prediction models based on physical information neural networks, the neural network is prone to learning a statistical "spurious correlation": that is, mistakenly assuming that as long as the sequence has extremely high conformational flexibility (inherent high disorder), liquid-liquid phase separation will inevitably occur. This overfitting caused by data distribution bias is the root cause of a large number of false positive predictions in existing end-to-end black-box models. To overcome this spurious correlation, this embodiment deliberately injects adversarial trap sequences into the training set (especially the negative sample set).

[0050] Preferably, the first predetermined threshold is 0.6 and the second predetermined threshold is 0.08.

[0051] In this embodiment, an intrinsic disorder greater than 0.6 defines the selected polypeptide sequences as being in a highly random coil state, i.e., possessing extremely high conformational flexibility. In traditional deep learning models that do not introduce physical constraints, the apparent distribution characteristics of such sequences in the disorder dimension highly overlap with those of real phase-separating kinesin proteins, making them highly susceptible to acting as confounding features that induce the model to fall into local extrema, thereby outputting false positive predictions.

[0052] The aromatic density of less than 0.08 ensures that the sequence is extremely lacking in non-covalent cross-linking sites such as phenylalanine and tyrosine, which can provide π-π interactions. In a non-equilibrium thermodynamic system, the lack of such multivalent cross-linking networks means that the system cannot effectively reduce the Gibbs free energy barrier during self-assembly. Such peptides do not possess the thermodynamic spontaneity to drive liquid-liquid phase separation (LLPS) under objective physical laws.

[0053] By combining thresholds where the inherent disorder of the column is greater than 0.6 and the aromatic density is less than 0.08, a precise "highly flexible, low-viscosity" trap is constructed. During the training phase, when the model encounters such adversarial trap sequences, its first output branch, due to the disturbance of disorder, often outputs an extremely high prediction probability P; however, the second output branch, unable to detect the viscous aromatic characteristics, outputs an extremely low thermodynamic score, Thermo.

[0054] This numerical difference triggers the asymmetric physical loss term (i.e., the positive error generated when calculating ReLU(P-Thermo)) and imposes a severe gradient penalty on the network. Through this linkage between specially crafted data and the underlying constraint formula, the neural network is forced to abandon the strategy of "relying on disorder" and truly learn the objective truth that it must rely on a physically multivalent cross-linked network to drive phase separation, fundamentally eliminating the illusion of false positives in the model.

[0055] In some implementations, a physical adversarial perturbation strategy is introduced during the training phase to further improve the model's stability in complex environments. This strategy involves superimposing a small perturbation vector δ, simulating fluctuations in the physical environment (such as pH changes or ionic strength changes), onto the input feature vector x, and optimizing a stability constraint to minimize the change in the model's thermodynamic score output before and after the perturbation. This stability constraint can be a loss function in the form of mean squared error or other similarity measures.

[0056] Specifically, a perturbation vector x+δ ​​is constructed by superimposing a small perturbation vector δ, simulating fluctuations in the physical environment (such as pH changes and ionic strength fluctuations), onto the input feature vector x. Subsequently, the model optimizes a stability-constrained loss function to minimize the change in the model's output before and after the perturbation.

[0057] This constraint forces the model not only to identify static sequence features, but also to learn the robustness of phase separation behavior under physical disturbances, ensuring that real phase-separating proteins maintain the robustness of their thermodynamic predictions during simulated physical disruption.

[0058] In some implementations, data preprocessing is included before extracting the multidimensional physicochemical features of the amino acid sequence, specifically: For the initial negative sample set, UniProt annotation information was used to exclude transmembrane proteins and secretory proteins in order to establish a baseline of cytoplasmic folding proteins.

[0059] After obtaining the initial negative sample set (i.e., the set of protein sequences that clearly do not undergo phase separation) and before performing physicochemical feature extraction, the computer device automatically traverses and removes protein sequences marked as "transmembrane" and "secreted / extracellular" by calling the subcellular location annotation information from the UniProt database. After this filtering step, the remaining negative sample set constitutes a pure "cytoplasmic folded protein baseline".

[0060] In real biophysical environments, transmembrane proteins, in order to stably integrate into hydrophobic lipid bilayer structures, inherently possess a high proportion of hydrophobic amino acids and specific aromatic residues (for anchoring at gas-liquid / solid-liquid interfaces). However, predicted liquid-liquid phase separation (LLPS) typically occurs in an aqueous phase environment such as the cytoplasm or nucleoplasm. If transmembrane proteins are retained in negative samples where phase separation does not occur, their "high aromatic density" and "high hydrophobicity" characteristics will serve as significant noise input to the neural network. This can severely interfere with the model's judgment of feature boundaries, and may even lead the model to a physical inversion conclusion that "high aromatic density corresponds to non-phase separation." Similarly, secretory proteins, due to their specific folding and modification environments, are also unsuitable as baseline references for aqueous phase separation.

[0061] By eliminating interference from transmembrane and secretory proteins, structural noise from the lipid environment was successfully eliminated. This preprocessing step ensured that both positive and negative samples entering the neural network were within the same aqueous thermodynamic reference frame, forcing the model to classify only by recognizing physical laws such as real π-π crosslinking networks. This significantly improved the specificity of the model's feature extraction and the accuracy of its physical boundary learning.

[0062] In some implementations, after outputting the phase separation prediction result of the protein to be tested based on the predicted probability, a mutation effect assessment step is further included, specifically: Obtain the wild-type amino acid sequence of the protein to be tested, as well as the mutant amino acid sequence by mutating specific key residues that provide multivalent cross-linking to non-cross-linking residues or by performing deletion mutations; Multidimensional physicochemical features of mutant amino acid sequences are extracted. When constructing feature vectors, the input values ​​of the inherent disorder feature are forced to be consistent with the inherent disorder feature values ​​of wild-type amino acid sequences. The constructed feature vector is input into the physical information neural network to obtain the mutant prediction probability output by the first output branch, and the difference is compared with the wild type prediction probability to isolate and evaluate the physical contribution of specific key residues in the driving phase separation process.

[0063] In the mutation effect assessment step, the wild-type amino acid sequence of the protein to be tested is first obtained. Subsequently, based on the biophysical driving mechanism, the corresponding mutant amino acid sequence is constructed.

[0064] "Specific key residues that provide multivalent cross-linking" refer to core sites that mediate protein-molecule interactions through physical mechanisms, thereby reducing the thermodynamic barrier to self-assembly of the system. Preferably, key residues include, but are not limited to: aromatic residues (such as phenylalanine, tyrosine, and tryptophan) capable of providing π-π stacking interactions, charged residues capable of providing electrostatic attraction, and specific amino acid combinations capable of providing cation-π interactions. By mutating these key residues to "non-cross-linked residues" (i.e., residues whose side chains do not possess the aforementioned interaction capabilities, such as aliphatic residues like alanine) or by directly performing "deletion mutations," the physical state of a partially severed protein-molecule interaction network can be simulated.

[0065] When extracting the physicochemical characteristics of mutant sequences and constructing feature vectors, the feature value of the "intrinsic disorder" dimension of mutant sequences is forcibly modified and locked to make it completely consistent with the intrinsic disorder value of wild-type sequences.

[0066] In real biophysical systems, while point mutations at key sites can drastically weaken intermolecular crosslinking, they typically do not macroscopically disrupt the overall conformational flexibility of proteins (i.e., the disorder background). However, traditional deep learning models suffer from severe overfitting to disorder features, and their output mutation effect predictions are often mixed with algorithmic interference caused by "minor fluctuations in disorder."

[0067] By implementing a variable control method of "locking disorder" at the computational level, background noise was successfully masked, enabling "isolated measurement" of the physical contribution of specific residues.

[0068] The constructed mutant feature vector is input into a network trained with physical constraints, and the difference in predicted probability before and after the mutation (ΔP=P) is compared. WT -P Mutant This allows for the quantification of the substantial impact of the mutation site on the phase separation driving force.

[0069] The specific implementation of forced locking is as follows: after calculating the multidimensional physicochemical characteristics of the mutant amino acid sequence, the inherent disorder feature value is directly replaced with the pre-calculated and stored inherent disorder feature value of the wild-type sequence, thereby generating a corrected feature vector. This corrected vector serves as the input to the physical information neural network to ensure that the conformational flexibility features before and after the mutation are completely consistent, achieving the purpose of isolated evaluation of the contribution of specific residues.

[0070] In some implementations, after outputting the phase separation prediction result of the protein to be tested based on the predicted probability, the method further includes: extracting the feature vector of the shared hidden layer of the physical information neural network, performing unsupervised analysis using UMAP dimensionality reduction and Leiden clustering, and screening candidate proteins based on the distribution of known non-membrane organelle marker proteins.

[0071] After obtaining the convergent physical information neural network, the output features of its shared hidden layers (e.g., the 128-dimensional feature vector of the penultimate layer) are extracted as the "thermodynamic fingerprint" of each protein. The UMAP (Uniform Manifold Approximation and Projection) algorithm is used to reduce the high-dimensional features to a two-dimensional manifold space, with parameters set to: n_neighbors=15, min_dist=0.1, and metric='euclidean'. Subsequently, the Leiden community detection algorithm is applied to perform unsupervised clustering of the two-dimensional projected points, with a resolution parameter set to 1.0.

[0072] Each cluster was labeled based on known phase-separating proteins and their membraneless organelle (MLO) types (e.g., nucleolar proteins, stress granule proteins, P-body proteins, etc.). Proteins not annotated but located within known functional clusters were identified as high-confidence candidate MLO-forming proteins. The distribution of the model's output thermodynamic scores among the candidate proteins was compared to select the top 10 proteins with the highest thermodynamic stability for wet-scale experimental validation.

[0073] It should be noted that this method uses the thermodynamic score of the second output branch, rather than the predicted probability of the first branch, for candidate protein screening. Experiments revealed a significant saturation effect in the predicted probability region near 1.0, where the probability values ​​of many candidate proteins were too similar to be easily distinguishable. In contrast, the original output thermodynamic score exhibited significant heterogeneity; a higher score indicated stronger thermodynamic stability of the phase-separated core, providing higher resolution for distinguishing high-confidence candidate proteins.

[0074] In a specific example, based on the above thermodynamic score ranking, the Top 10 candidate proteins with high thermodynamic stability selected include: Foot protein 3 variant 11, Phosphoprotein P, Alpha / beta-gliadin MM1, Small ribosomal subunit protein bS18c, Gamma M1 crystallin, WW domain binding protein 2, Gamma M8b crystallin, Gamma-crystallin D, Histone acetyltransferase type B catalytic subunit, and Rab GDP dissociation inhibitor.

[0075] Taking the core biological clock protein Bmal1 as an example, experimental data show that the predicted probability of the wild-type sequence is 0.9905. After mutating its aromatic residues and locking the disorder feature (constantly 0.506588), the predicted probability of the mutant given by the model in this embodiment drops sharply to 0.1049.

[0076] To better illustrate this embodiment, the method flow of the present invention will be illustrated below with specific datasets, including StrictBio-LLPS-V12, underlying calculation formulas, and comparative verification: The first step is data preprocessing and trap sequence injection (building a physical baseline). The computer equipment first acquires an initial protein dataset containing 2,876 sequences. To eliminate data bias, the system executes an automated cleaning script: 1. Pure baseline extraction: Using UniProt annotation, transmembrane proteins and secretory proteins that are deeply embedded in the lipid bilayer and naturally rich in aromatic residues are forcibly removed.

[0077] 2. Trap Sequence (Negative Sample) Generation: The system selectively mined and synthesized 83 "trap sequences" with perfect adversarial value. The stringent filtering criteria for inclusion in the trap were set as follows: intrinsic disorder of the sequence > 0.6 and aromatic residue density < 0.08.

[0078] like Figures 3 to 4 As shown, Figure 3 The results showed that after washing with transmembrane proteins, the peak aromatic density of the negative control group (folded negative sample) was successfully reduced to below approximately 0.08, creating a significant physical boundary with the positive group (above approximately 0.10). Figure 4 The established trap sequence precisely fills the "high disorder, low viscosity" quadrant in two-dimensional space. This step completely eliminates the possibility of subsequent AI cheating by relying solely on "disorder degree" for classification, forcing it to establish a thermodynamic boundary at Y=0.08.

[0079] The second step is the precise extraction of multidimensional physicochemical features. For each of the processed sequences, the system extracts 4-dimensional core physicochemical measures to construct a feature vector: 1. Aromatic residue density: The proportions of phenylalanine (F), tyrosine (Y), and tryptophan (W) in the sequence are calculated to characterize the π-π stacking crosslinking ability. The basic calculation formula is: R F / Y =(N F +N Y ) / L Where, N F and N Y ...

[0080] 2. Inherent disorder: The structural flexibility is quantified using the IUPred2A algorithm.

[0081] 3. Average hydrophobicity: calculated based on the Kyte-Doolittle scale.

[0082] 4. Charge distribution standard deviation: Set a sliding window of length 10 to calculate local net charge fluctuations, replacing the highly complex SCD parameter.

[0083] The extracted features are standardized by Z-score and then projected into a normal distribution space with zero mean and unit variance.

[0084] The third step is the forward propagation of the two-branch physical information neural network (PINN). The standardized feature vectors are input into the multilayer perceptron (MLP) base.

[0085] 1. Shared hidden layer: sequentially through 256, 128, and 64 neurons (activation function is ReLU, dropout=0.3 after the first layer); 2. Heterogeneous output: The network is split into two branches. Branch 1 outputs the predicted probability P(0,1) through the Sigmoid function; Branch 2 outputs an unbounded continuous scalar value, i.e., the thermodynamic score Thermo, through the Linear function.

[0086] Step 4: Calculation of asymmetric physical penalty and weight update (core training rule) ; in, This is the binary cross-entropy loss. In this embodiment, physical weight coefficients are set. =0.7, the asymmetric physical loss term is constructed as follows: ; Suppose the network reads a pre-defined trap sequence. Due to its disorder degree > 0.6, the network initially experiences a hallucination, with the first branch outputting an extremely high phase separation probability (assuming P = 0.90). However, due to aromaticity < 0.08 and a lack of viscous sites, the second branch calculates an extremely low thermodynamic score (assuming Thermo = -0.5). At this point, the physical constraint L is triggered. phys =ReLU(0.90-(-0.5))=1.40.

[0087] The penalty gradient 0.7 x 1.40 = 0.98 is generated and propagated back to the network, forcing the network to modify the underlying weights and abandon its dependence on disorder.

[0088] like Figures 5 to 6 As shown, Figure 5 The traditional model (old model) without this constraint produces a severe high-probability illusion (right-side peak) when encountering a trap sequence; while the model in this embodiment (new model) successfully forces the probability distribution to shift left to the safe zone. Figure 6 This indicates that after applying After applying a one-way penalty of 0.7, the model did not collapse, and its independent validation set AUC remained as high as 0.9781, achieving an optimal balance between taking into account physical laws and extremely high classification accuracy.

[0089] Step 5: Downstream industrial applications (stress testing of computer-aided mutation scanning) To verify that the model has fully grasped the physical driving mechanism, the system was deployed for virtual mutation testing of the core rhythmic protein Bmal160AA fragment: 1. Benchmark test: The Bmal1 wild-type (WT) sequence was extracted, and its inherent disorder was objectively calculated to be 0.506588. The model output predicted probability PWT=0.9905.

[0090] 2. Mutation construction: The system performs an alanine scan, mutating all 6 aromatic amino acids (F / Y / W) in the sequence to alanine (A) which has no cross-linking ability.

[0091] 3. Feature Locking and Secondary Prediction: When extracting mutant features, the computer forcibly locks the disorder input value to a constant 0.506588, completely eliminating the interference of conformational flexibility changes. This is then input into the model, outputting a mutant prediction probability PMT = 0.1049.

[0092] 4. Isolated quantitative assessment: Calculate the phase separation capability attenuation ΔP = P WT -P Mutant .

[0093] like Figure 7 As shown, under the control condition of maintaining the absolute invariance of macroscopic conformational flexibility (disorder), simply stripping away local viscous sites resulted in a precipitous drop in the predicted probability, with a value as high as ΔP = 0.8856. This extreme numerical value quantitatively confirms that the proposed method successfully stripped away surface statistical features, achieving accurate measurement of the physical contribution of the π-π crosslinking network dominated by specific residues.

[0094] The second embodiment of the present invention provides a protein phase separation prediction system based on a physical information neural network, which includes: The feature vector construction module is used to obtain the amino acid sequence of the protein to be tested and extract the multidimensional physicochemical features of the amino acid sequence to construct the feature vector; The output module is used to input the feature vector into a pre-trained physical information neural network. The physical information neural network includes a shared hidden layer and a first output branch and a second output branch connected to the shared hidden layer, respectively. The first output branch outputs the predicted probability of phase separation of the protein to be tested, and the second output branch outputs a continuous scalar value as a thermodynamic score. The results prediction module is used to output the phase separation prediction results of the protein to be tested based on the prediction probability. Among them, the physical information neural network is constrained by an asymmetric physical loss term and auxiliary physical anchor points during the training phase; The asymmetric physical loss term is constructed such that when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output; otherwise, zero is output. The constraint process for the auxiliary physical anchor points includes: using adaptive learning parameters that are positively constrained by the physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and incorporating the deviation between the thermodynamic score and the reference score as an anchor point constraint term into the total loss function.

[0095] like Figure 8As shown, a third embodiment of the present invention provides an electronic device 700, which includes an input unit 710, a memory 720, at least one processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the protein phase separation prediction method based on the physical information neural network in the aforementioned embodiments by calling the program instructions. This electronic device can be a mobile terminal device such as a mobile phone or a computer.

[0096] 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.

[0097] 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.

[0098] 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 a physical information neural network, characterized by, include: The amino acid sequence of the protein to be tested is obtained, and the multidimensional physicochemical features of the amino acid sequence are extracted. The multidimensional physicochemical features are processed by Z-score normalization or Min-Max normalization algorithm to construct a feature vector. The multidimensional physicochemical features include aromatic residue density, intrinsic disorder, average hydrophobicity, and standard deviation of charge distribution. The feature vector is input into a pre-trained physical information neural network; the physical information neural network includes a shared hidden layer, and a first output branch and a second output branch respectively connected to the shared hidden layer; The first output branch outputs the predicted probability of phase separation of the protein to be tested, and the second output branch outputs a continuous scalar value as a thermodynamic score. In the training phase, the physical information neural network introduces adversarial trap sequences as negative samples. The adversarial trap sequences meet the following feature filtering criteria: the intrinsic disorder of the sequence is greater than 0.6 and the aromatic density is less than 0.

08. Furthermore, the physical information neural network employs an asymmetric physical loss term and auxiliary physical anchor point constraints during the training phase. The asymmetric physical loss term is constructed such that when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output; otherwise, zero is output. The constraint process of the auxiliary physical anchor point includes: using adaptive learning parameters constrained to positive numbers by a physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and incorporating the deviation between the thermodynamic score and the reference score as an anchor point constraint term into the total loss function. The physical gating mechanism adopts The activation function, wherein the reference score is represented as: ； in, For reference scoring, , , , These correspond to the aromatic residue density, charge distribution standard deviation, intrinsic disorder, and average hydrophobicity after Z-score normalization or Min-Max normalization, respectively. , , , These are adaptive learning parameters that are automatically optimized during backpropagation of the network. for Activation function The deviation is calculated using the mean square error as the cutoff function. The phase separation prediction results of the target protein are output based on the predicted probability. After outputting the phase separation prediction result of the protein to be tested based on the predicted probability, the method further includes a mutation effect assessment step, specifically: Obtain the wild-type amino acid sequence of the protein to be tested, as well as the mutant amino acid sequence by mutating specific key residues that provide multivalent cross-linking to non-cross-linking residues or by performing deletion mutations; Multidimensional physicochemical features of mutant amino acid sequences are extracted. When constructing feature vectors, the input values ​​of the inherent disorder feature are forced to be consistent with the inherent disorder feature values ​​of wild-type amino acid sequences. The constructed feature vector is input into the physical information neural network to obtain the mutant prediction probability output by the first output branch, and the difference is compared with the wild type prediction probability to isolate and evaluate the physical contribution of specific key residues in the driving phase separation process.

2. The protein phase separation prediction method based on a physical information neural network as described in claim 1, characterized in that: The process of extracting the standard deviation of the charge distribution includes: sliding a sliding window of a predetermined length along the amino acid sequence, calculating the local net charge within each window, and calculating the standard deviation of the net charge of all windows.

3. The protein phase separation prediction method based on a physical information neural network as described in claim 1, characterized in that: The first output branch uses the Sigmoid activation function to map the network output to a predicted probability in the (0,1) interval; the second output branch uses the Linear activation function to output an unbounded scalar value.

4. The protein phase separation prediction method based on a physical information neural network as described in claim 1, characterized in that, Before extracting the multidimensional physicochemical features of the amino acid sequence, data preprocessing is also included, specifically: For the initial negative sample set, UniProt annotation information was used to exclude transmembrane proteins and secretory proteins in order to establish a baseline of cytoplasmic folding proteins.

5. A protein phase separation prediction system based on a physical information neural network, characterized in that, include: The feature vector construction module is used to obtain the amino acid sequence of the protein to be tested and extract the multidimensional physicochemical features of the amino acid sequence. The multidimensional physicochemical features are processed by Z-score normalization or Min-Max normalization algorithm to construct feature vectors. The multidimensional physicochemical features include aromatic residue density, intrinsic disorder, average hydrophobicity and standard deviation of charge distribution. The output module is used to input the feature vector into a pre-trained physical information neural network. The physical information neural network includes a shared hidden layer and a first output branch and a second output branch connected to the shared hidden layer, respectively. The first output branch outputs the predicted probability of phase separation of the protein to be tested, and the second output branch outputs a continuous scalar value as a thermodynamic score. In the training phase, the physical information neural network introduces adversarial trap sequences as negative samples. The adversarial trap sequences meet the following feature filtering criteria: the intrinsic disorder of the sequence is greater than 0.6 and the aromatic density is less than 0.

08. Furthermore, the physical information neural network employs an asymmetric physical loss term and auxiliary physical anchor point constraints during the training phase. The asymmetric physical loss term is constructed such that when the predicted probability is greater than the thermodynamic score, a loss value greater than zero is output; otherwise, zero is output. The constraint process of the auxiliary physical anchor point includes: using adaptive learning parameters constrained to positive numbers by a physical gating mechanism to weight and truncate the multidimensional physicochemical features to obtain a reference score; and incorporating the deviation between the thermodynamic score and the reference score as an anchor point constraint term into the total loss function. The physical gating mechanism adopts The activation function, wherein the reference score is represented as: ； in, For reference scoring, , , , These correspond to the aromatic residue density, charge distribution standard deviation, intrinsic disorder, and average hydrophobicity after Z-score normalization or Min-Max normalization, respectively. , , , These are adaptive learning parameters that are automatically optimized during backpropagation of the network. for Activation function The deviation is calculated using the mean square error as the cutoff function. The results prediction module is used to output the phase separation prediction results of the protein to be tested based on the prediction probability. The mutation effect assessment module is used to output the phase separation prediction result of the protein to be tested based on the predicted probability: Obtain the wild-type amino acid sequence of the protein to be tested, as well as the mutant amino acid sequence by mutating specific key residues that provide multivalent cross-linking to non-cross-linking residues or by performing deletion mutations; Multidimensional physicochemical features of mutant amino acid sequences are extracted. When constructing feature vectors, the input values ​​of the inherent disorder feature are forced to be consistent with the inherent disorder feature values ​​of wild-type amino acid sequences. The constructed feature vector is input into the physical information neural network to obtain the mutant prediction probability output by the first output branch, and the difference is compared with the wild type prediction probability to isolate and evaluate the physical contribution of specific key residues in the driving phase separation process.

6. 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 a physical information neural network as described in any one of claims 1 to 4 by calling the program instructions.

7. 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 a physical information neural network as described in any one of claims 1 to 4.

Images