A multi-model cascade prediction and safety screening method for polypeptide sequences

Abstract

This invention discloses a multi-model cascade prediction and safety screening method for peptide sequences, applicable to local or server-side analysis of single or batch peptide sequences. The method includes: standardizing the input sequence, isolating anomalies, and splitting data; prioritizing the calculation of low-cost physicochemical properties and performing first-level safety inference; triggering Early-Stopping when hemolytic, toxic, or sensitizing indicators exceed thresholds, terminating subsequent time-consuming model inference; continuing to perform functional class and propensity model predictions for sequences that have not triggered the stopping condition; dynamically fusing and scoring the safety score, functional score, propensity score, and physicochemical property penalty term, and outputting screening conclusions, ranking results, and a unified results table. This method can reduce unnecessary computational power consumption, improve high-throughput screening efficiency, and enhance availability in a shared server environment.

Description

Technical Field

[0001] This invention belongs to the field of computational screening and safety evaluation technology for peptide drugs, and particularly relates to a multi-model cascade prediction and safety screening method for peptide sequences. Background Technology

[0002] Early screening of functional peptides, bioactive peptides, and candidate therapeutic peptides typically requires simultaneous consideration of multiple biological and safety properties, including but not limited to anticancer, antibacterial, anti-inflammatory, antifungal, antiviral, cell-penetrating, antioxidant, DPP-IV inhibition, immune binding tendency, as well as hemolysis, toxicity, and sensitization. In existing research and application scenarios, common technical solutions mainly fall into the following categories:

[0003] The first category is single-task prediction tools, which build independent models for only a certain peptide function or a certain safety attribute;

[0004] The second category involves multiple decentralized online platforms or local scripts, requiring researchers to manually switch between different tools and manually summarize the results.

[0005] The third category is analysis methods that only output the raw scores of the model, lacking a unified screening, sorting, and derivation logic;

[0006] The fourth category consists of analysis methods that only operate on a single sequence, making them difficult to adapt to high-throughput FASTA or CSV batch input scenarios;

[0007] The fifth method involves running all models on all sequences first, and then manually reviewing the results table afterwards to make selections.

[0008] For large-scale candidate sequence screening, existing technologies often suffer from a long-overlooked problem: computational resources are heavily consumed by low-value samples. Specifically, in a batch of sequence files, a significant proportion of sequences clearly do not meet the requirements for subsequent experiments in terms of safety attributes, such as toxicity scores, hemolytic scores, or sensitization scores that far exceed thresholds. However, existing technologies still continue to run subsequent time-consuming functional prediction models or deep inference sub-models for these sequences, resulting in the ineffective use of CPU, GPU, memory, and disk I / O resources.

[0009] Furthermore, existing technologies often fail to effectively integrate model outputs with the physicochemical properties of the sequence. Researchers frequently need to manually review multiple model scores, net charge, isoelectric point, hydrophobicity, hydrophobic moment, and other data simultaneously, and then rely on experience to determine candidate priorities. This approach is not only inefficient, but also makes it difficult for different researchers to maintain consistent judgment standards.

[0010] In a shared server environment, existing technologies still suffer from insufficient availability and fault tolerance. When dealing with batch FASTA or CSV input, if individual sequences are empty, contain illegal characters, have missing fields, or a local model call fails, the entire batch processing task is often interrupted, making it difficult to support unattended operation and multi-user reuse.

[0011] Early screening of functional peptides, bioactive peptides, and candidate therapeutic peptides typically requires simultaneous consideration of multiple biological and safety properties, including but not limited to anticancer, antibacterial, anti-inflammatory, antifungal, antiviral, cell-penetrating, antioxidant, DPP-IV inhibition, immune binding tendency, as well as hemolysis, toxicity, and sensitization. In existing research and application scenarios, common technical solutions mainly fall into the following categories:

[0012] The first category is single-task prediction tools, which build independent models for only a certain peptide function or a certain safety attribute;

[0013] The second category involves multiple decentralized online platforms or local scripts, requiring researchers to manually switch between different tools and manually summarize the results.

[0014] The third category is analysis methods that only output the raw scores of the model, lacking a unified screening, sorting, and derivation logic;

[0015] The fourth category consists of analysis methods that only operate on a single sequence, making them difficult to adapt to high-throughput FASTA or CSV batch input scenarios;

[0016] The fifth method involves running all models on all sequences first, and then manually reviewing the results table afterwards to make selections.

[0017] For large-scale candidate sequence screening, existing technologies often suffer from a long-overlooked problem: computational resources are heavily consumed by low-value samples. Specifically, in a batch of sequence files, a significant proportion of sequences clearly do not meet the requirements for subsequent experiments in terms of safety attributes, such as toxicity scores, hemolytic scores, or sensitization scores that far exceed thresholds. However, existing technologies still continue to run subsequent time-consuming functional prediction models or deep inference sub-models for these sequences, resulting in the ineffective use of CPU, GPU, memory, and disk I / O resources.

[0018] Furthermore, existing technologies often fail to effectively integrate model outputs with the physicochemical properties of the sequence. Researchers frequently need to manually review multiple model scores, net charge, isoelectric point, hydrophobicity, hydrophobic moment, and other data simultaneously, and then rely on experience to determine candidate priorities. This approach is not only inefficient, but also makes it difficult for different researchers to maintain consistent judgment standards.

[0019] In a shared server environment, existing technologies still suffer from insufficient availability and fault tolerance. When dealing with batch FASTA or CSV input, if individual sequences are empty, contain illegal characters, have missing fields, or a local model call fails, the entire batch processing task is often interrupted, making it difficult to support unattended operation and multi-user reuse. Summary of the Invention

[0020] This invention solves the following technical problems: 1. How to establish a hierarchical inference architecture suitable for single or batch peptide sequences, so that obviously high-risk sequences are automatically intercepted in the early stage, thereby reducing the use of ineffective computing power;

[0021] 2. How to construct a dynamic fusion scoring mechanism that combines multi-model results with physicochemical properties, rather than simply splicing scores or static weighted averaging;

[0022] 3. How to manage the flow of valid sequences, abnormal sequences, sequences to be retried, and terminated sequences during batch processing to improve fault tolerance and throughput in a server environment;

[0023] 4. How to achieve a unified closed loop of input standardization, model scheduling, security screening, comprehensive scoring, batch sorting, and result export within the same technical process.

[0024] To achieve the objective of this invention, the technical solution adopted is as follows:

[0025] The first invention provides a multi-model cascade prediction and safety screening method for peptide sequences, comprising the following steps:

[0026] Step S1: Receive input data and construct task records;

[0027] The input data is one or more polypeptide sequences input by the user;

[0028] Step S2: Sequence normalization, anomaly isolation, and data stream splitting; input records should be split into at least the following queues or pools:

[0029] Q_valid: A valid sequence queue used for subsequent reasoning;

[0030] Q_error: An exception sequence queue used to record illegal or unusable sequences;

[0031] Q_retry: A local failure retry queue used to temporarily store a sequence of local calls that have failed but can be retried;

[0032] Q_done: The completion queue, used to summarize sequences that have completed prediction or screening;

[0033] Step S3: Low-cost physical and chemical profiling and primary task allocation;

[0034] For sequences in Q_valid, low-cost computation modules are executed first.

[0035] Step S4: Level 1 secure reasoning and Early-Stopping computational resource optimization mechanism;

[0036] For each valid sequence Q_valid, the safety class model is run first.

[0037] Step S5: Secondary functional / propensity reasoning;

[0038] For sequences that have not triggered Early-Stopping, continue with secondary inference, invoking the remaining functional class model and the propensity model; this secondary inference is only performed on candidates that have passed the primary security gating.

[0039] The aforementioned Level 1 safety gating refers to the system performing Level 1 safety inference based on low-cost physicochemical profiling and safety models (such as hemolytic HEMO, toxic TOX, and sensitizing ALG) before the peptide sequence enters subsequent time-consuming functional prediction. If any safety indicator exceeds a preset threshold, or the overall safety risk reaches the stopping condition, Early-Stopping is immediately triggered, terminating subsequent complex model inference for that sequence and retaining only its safety screening conclusion and risk profile. This mechanism not only achieves pre-interception of high-risk sequences but also reduces the consumption of CPU, GPU, and memory resources by invalid samples, improving the throughput of batch screening and the operating efficiency in a shared server environment.

[0040] Step S6: Introduce a dynamic fusion scoring mechanism for adjusting physicochemical properties;

[0041] For sequences that have completed the second-level reasoning, dynamic fusion scoring under the adjustment of physical and chemical features is performed; when the sequence has triggered the safety hard threshold in step S4, the safety screening conclusion is directly output, and its comprehensive score is limited to the profile score that depends only on the safety retention score, or it is directly marked as "risk elimination";

[0042] Step S7: Screening status determination and batch sorting;

[0043] Based on the Early-Stopping results, the comprehensive score C_i, the safety retention score S_i, and the risk triggering items, the sequence is assigned a screening status;

[0044] Step S8: Unify result export and server sharing;

[0045] For a single sequence, a unified structured result object is output; for batch sequences, a unified result table is generated.

[0046] Secondly, the present invention provides a multi-model cascade prediction and safety screening system for polypeptide sequences that implements the above-mentioned methods of the present invention, comprising:

[0047] The input adapter module is used to receive one or more polypeptide sequences input by the user;

[0048] The standardization and validity verification module is used to standardize the input peptide sequences, isolate anomalies, and process data stream splitting.

[0049] The low-cost physical and chemical profiling module prioritizes low-cost computation of legitimate sequences used for subsequent reasoning in the standardization and legality verification module;

[0050] The Level 1 security inference and Early-Stopping gating module is used to prioritize the execution of security models for each legitimate sequence;

[0051] The secondary function / prejudice reasoning module is used to continue performing secondary reasoning on sequences that have not triggered Early-Stopping, and to call the remaining function class model and prejudice model;

[0052] The dynamic comprehensive scoring module is used to perform dynamic fusion scoring under the adjustment of physicochemical characteristics for sequences that have completed secondary reasoning.

[0053] The batch sorting and export module assigns a screening status to the sequences based on the Early-Stopping results and exports the unified results.

[0054] Furthermore, the present invention provides a computer storage medium storing a computer program that implements the method of the present invention.

[0055] Compared with the prior art, the present invention has at least the following advantages and technical effects:

[0056] (1) Significantly reduce the computational resource consumption of invalid candidates.

[0057] Existing technologies typically execute the entire model on all sequences, resulting in clearly high-risk sequences still consuming significant time-consuming inference resources in subsequent steps. This invention reduces the unnecessary consumption of CPU, GPU, memory, and I / O resources by triggering Early-Stopping after Level 1 safe inference, preventing ineligible candidates from entering the subsequent deep inference process.

[0058] (2) Improve the high-throughput screening throughput of the server

[0059] Since high-cost inference tasks are only assigned to sequences that pass the initial security screening, this invention can improve concurrent processing capabilities and the effective screening volume per unit time in batch sequence scenarios, making it more suitable for deployment in server shared environments.

[0060] (3) Improve batch processing fault tolerance and high availability

[0061] By employing a diversion mechanism consisting of a valid queue, an exception queue, and a retry queue, this invention can continue running the remaining sequences even when a single record is abnormal, a partial model call fails, or a field is incomplete, thereby improving task stability in unattended environments.

[0062] (4) Forming an automatic sorting logic that is more stable than manual timekeeping.

[0063] This invention incorporates a mathematical comprehensive scoring mechanism that integrates functional scores, safety scores, preference scores, and physicochemical property penalty items into the algorithm framework. This avoids subjective fluctuations when manually reviewing Excel spreadsheets line by line and improves the consistency of screening criteria among different users.

[0064] (5) Improve the interpretability and designability of results

[0065] The system simultaneously outputs physicochemical parameters such as net charge, isoelectric point, GRAVY, hydrophobic moment, and molecular weight, enabling researchers to not only know "which sequence scores high" but also understand "why the sequence is retained, downgraded, or eliminated."

[0066] (6) Enhance the engineering deployability of the method

[0067] This invention covers single sequence analysis, batch high-throughput screening, on-demand model calling, and unified export. It can be run as an algorithm method or stably deployed in the form of a server-shared workflow. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention.

[0069] Figure 2 This is a diagram of the asynchronous data stream and queue splitting architecture of the present invention.

[0070] Figure 3 This is the Early-Stopping gate logic diagram for the first-level security reasoning of this invention.

[0071] Figure 4 This is a diagram illustrating the reasoning structure of the secondary function and tendency model of this invention.

[0072] Figure 5 This is a flowchart of the dynamic fusion scoring process of the present invention.

[0073] Figure 6 This is a flowchart of the batch sorting and unified export process of this invention. Detailed Implementation

[0074] The multi-model cascade prediction and security screening method for peptide sequences provided by this invention is suitable for local computing environments or shared server environments. This method not only solves the problem of integrating multi-model results, but also incorporates security screening as a pre-emptive underlying computational resource optimization mechanism, and introduces a dynamic fusion scoring mechanism adjusted by physicochemical properties. See appendix. Figure 1 The method includes at least the following steps:

[0075] Step S1: Receive input data and build task records

[0076] Receive one or more polypeptide sequences input by the user. The input methods include at least:

[0077] Directly input a single polypeptide sequence;

[0078] Enter a single FASTA text;

[0079] Upload or read batch FASTA files;

[0080] Upload or read batch CSV files.

[0081] When the input is a CSV file, the system automatically recognizes the sequence field, prioritizing column names such as sequence, seq, peptide, peptide_sequence, peptide sequence, or sequence; if it cannot be automatically recognized, the user is allowed to explicitly specify the sequence column name.

[0082] The system constructs a unified task object Record_i for each input record, and the task object includes at least the following fields:

[0083] record_id: Record identifier;

[0084] raw_sequence: The original input sequence;

[0085] normalized_sequence: The normalized sequence;

[0086] status: Current status;

[0087] error_code: Exception code;

[0088] feature_ref: Reference to physical and chemical features;

[0089] light_result_ref: Lightweight model result reference;

[0090] heavy_result_ref: Reference to the results of a time-consuming model;

[0091] screen_result_ref: A reference to the screening results.

[0092] Step S2: Normalize the input peptide sequence, isolate anomalies, and offload the data stream (see Appendix). Figure 2 )

[0093] The following preprocessing was performed on the input polypeptide sequence:

[0094] Remove whitespace and newline characters;

[0095] Remove FASTA header information;

[0096] Convert the sequence to uppercase;

[0097] Determine whether it contains only standard amino acid characters;

[0098] Calculate the basic length information;

[0099] Anomalies are marked for empty sequences, sequences containing illegal characters, and sequences that do not meet the preset length condition.

[0100] In a preferred embodiment, the present invention employs a producer-consumer asynchronous data stream structure, which diverts input records to at least the following queues or pools:

[0101] Q_valid: A valid sequence queue used for subsequent reasoning;

[0102] Q_error: An exception sequence queue used to record illegal or unusable sequences;

[0103] Q_retry: A local failure retry queue used to temporarily store a sequence of local calls that have failed but can be retried;

[0104] Q_done: The completion queue, used to summarize sequences that have completed prediction or screening.

[0105] At the memory management level, valid sequences are preferentially entered into the valid tensor pool T_valid or the valid feature pool F_valid, while abnormal sequences only retain metadata and error reasons, without occupying the high-cost tensor resources required for subsequent model inference.

[0106] This diversion mechanism ensures that a single abnormal sequence will not block the entire batch of tasks, and can maintain high availability and fault tolerance in unattended server scenarios.

[0107] Step S3: Low-cost physical and chemical profiling and primary task triage

[0108] Prioritize low-cost computation modules for sequences in Q_valid, including:

[0109] Calculation of basic physical and chemical properties;

[0110] Statistical feature extraction can be completed quickly;

[0111] Lightweight security prediction model inference.

[0112] The basic physicochemical properties include at least the following:

[0113] Sequence length;

[0114] Molecular weight;

[0115] Average residue count;

[0116] Net charge at pH 5, pH 7 and pH 9 conditions;

[0117] Charge density;

[0118] Isoelectric point;

[0119] GRAVY;

[0120] Hydrophobic moment;

[0121] Aliphatic index;

[0122] The proportion of hydrophobic residues;

[0123] The percentage of aromatic residues;

[0124] Percentage of polar residues.

[0125] In a preferred embodiment, the above-mentioned low-cost physicochemical calculation results and lightweight security model results together form a first-level screening feature vector z_i, which is used for subsequent Early-Stopping determination.

[0126] Step S4: Level 1 secure inference and Early-Stopping computational resource optimization mechanism (see appendix) Figure 3 )

[0127] For each valid sequence Q_valid, the safety class model is run first, including at least:

[0128] Hemolytic model (HEMO);

[0129] Toxicity model (TOX);

[0130] Allergenicity model (ALG).

[0131] Their scores are recorded as follows:

[0132] h_i: HEMO score;

[0133] t_i: TOX fraction;

[0134] a_i: ALG score.

[0135] In one implementation, a hard security threshold is set:

[0136] θ_H = 0.46;

[0137] θ_T = 0.38;

[0138] θ_A = 0.30.

[0139] Simultaneously define a security risk normalization function:

[0140] r_H = max(0, (h_i - θ_H) / (1 - θ_H))

[0141] r_T = max(0, (t_i - θ_T) / (1 - θ_T))

[0142] r_A = max(0, (a_i - θ_A) / (1 - θ_A))

[0143] Further define the comprehensive security risk score:

[0144] R_i = α_H * r_H + α_T * r_T + α_A * r_A

[0145] Where α_H, α_T, and α_A are security risk weights, and satisfy the following:

[0146] α_H + α_T + α_A = 1

[0147] In a preferred embodiment, Early-Stopping is triggered when any of the following conditions are met:

[0148] h_i >= θ_H;

[0149] t_i >= θ_T;

[0150] a_i >= θ_A;

[0151] R_i >= R_stop.

[0152] Once Early-Stopping is triggered, the system immediately performs the following low-level scheduling actions:

[0153] Remove the current sequence from the queue of time-consuming models to be executed;

[0154] Release or no longer allocate the high-cost tensor cache corresponding to this sequence;

[0155] Write the sequence into the termination result queue Q_done;

[0156] Generate a screening status of "risk elimination" or "only profile formation" for the sequence;

[0157] The inference process of subsequent time-consuming functional models, deep attention models, ensemble sub-models, or bias models will no longer be initiated for this sequence.

[0158] In other words, security screening in this invention is not only a business-level "result filtering" but also a fundamental "computing resource scheduling gating mechanism." Through this mechanism, the invalid use of server CPU, GPU, and memory resources by obviously high-risk sequences can be significantly reduced, thereby improving throughput during high-throughput concurrent screening.

[0159] Step S5: Secondary Functional / Tendency Reasoning (see Appendix) Figure 4 )

[0160] Secondary inference is only performed on sequences that have not triggered Early-Stopping, i.e., the remaining functional class model and the tendency model are invoked.

[0161] In a preferred embodiment, the functional class model includes at least one or more of the following models:

[0162] Anticancer peptide model;

[0163] Antimicrobial peptide model;

[0164] Antifungal peptide model;

[0165] Anti-inflammatory peptide model;

[0166] Antioxidant peptide model;

[0167] Antiviral peptide model;

[0168] Cell-penetrating peptide model;

[0169] DPP-IV inhibitory peptide model.

[0170] The propensity model includes at least one or more of the following models:

[0171] MHC-I binding propensity model;

[0172] Other receptor binding tendency models;

[0173] Other immune-related propensity models.

[0174] The second-level inference is only performed on candidates that pass the first-level security gating, thereby reducing the number of meaningless deep inference calls.

[0175] The first-level safety gating described in this invention refers to the system performing first-level safety inference based on low-cost physicochemical profiling and safety models (such as hemolytic HEMO, toxic TOX, and sensitizing ALG) before the peptide sequence enters subsequent time-consuming functional prediction. If any safety indicator exceeds a preset threshold, or the overall safety risk reaches the stopping condition, Early-Stopping is immediately triggered, terminating subsequent complex model inference for that sequence and retaining only its safety screening conclusion and risk profile. This mechanism not only achieves pre-interception of high-risk sequences but also reduces the consumption of CPU, GPU, and memory resources by invalid samples, improving the throughput of batch screening and the operating efficiency in a shared server environment.

[0176] Step S6: Introduce a dynamic fusion scoring mechanism for adjusting physicochemical properties (see appendix). Figure 5 )

[0177] For sequences that have completed secondary reasoning, the system does not use a simple static weighted average, but instead performs a dynamic fusion score adjusted by physicochemical features.

[0178] First, the output of the functional models is oriented and normalized. Let the normalization score of the m-th functional model be... If _(i,m), then the functional basic score is defined as:

[0179] F_i = (Σ(w_m * _(i,m))) / Σ(w_m), where m ∈ M_f

[0180] in:

[0181] M_f represents the set of functional class models;

[0182] w_m represents the base weights of the corresponding functional model.

[0183] For the output of the propensity score model, let the normalized score of the nth propensity score model be... (i,n), then the tendency score is defined as:

[0184] T_i = (Σ(u_n * (i,n))) / Σ(u_n), where n ∈ M_t

[0185] in:

[0186] M_t represents the set of propensity models;

[0187] u_n represents the bias model weights.

[0188] Based on the security risk score R_i obtained in step S4, the security retention score is defined as follows:

[0189] S_i = 1 - R_i

[0190] Furthermore, a physicochemical property penalty function P_i is introduced to automatically reduce the weight of sequences with high functional class scores but whose physicochemical characteristics deviate from the expected range. Preferably:

[0191] P_i = λ_1 * max(0, μ_i - μ_thr) + λ_2 * max(0, G_i - G_thr) + λ_3 *max(0, |q_i - q_opt| - δ)

[0192] in:

[0193] μ_i is the hydrophobic moment;

[0194] μ_thr is the hydrophobicity moment threshold;

[0195] G_i is the GRAVY value;

[0196] G_thr is the GRAVY threshold;

[0197] q_i represents the net charge at pH 7.

[0198] q_opt is the target net charge center value;

[0199] δ represents the net charge tolerance range;

[0200] λ_1, λ_2, and λ_3 are the physicochemical penalty coefficients.

[0201] Based on this definition, the adjusted functional divisions are as follows:

[0202] F'_i = F_i * (1 - min(P_i, P_max))

[0203] P_max is the upper limit of the penalty to prevent negative scores or extremely unstable scores.

[0204] The final comprehensive score is defined as follows:

[0205] C_i = β_F * F'_i + β_T * T_i + β_S * S_i

[0206] in:

[0207] β_F, β_T, and β_S are the weights of the comprehensive score;

[0208] Satisfies β_F + β_T + β_S = 1.

[0209] In a preferred embodiment, when the sequence has already triggered the security hard threshold in step S4, the complete secondary comprehensive score is no longer calculated. Instead, the security screening conclusion is directly output, and its comprehensive score is limited to the profile score that only depends on the security retention score, or it is directly marked as "risk elimination".

[0210] Through this design, the present invention does not achieve a simple splicing of 13 or more models, but rather a "cross-model feature fusion method based on dynamic adjustment of physicochemical characteristics".

[0211] Step S7: Screening Status Determination and Batch Sorting

[0212] Based on the Early-Stopping results, the comprehensive score C_i, the safety retention score S_i, and the risk triggering item, the sequence is assigned a screening status.

[0213] In one implementation, the screening status includes at least:

[0214] Recommended to keep;

[0215] Continue observation;

[0216] Only an image is created;

[0217] Risk elimination.

[0218] in:

[0219] If any security hard threshold is triggered, it is preferable to mark it as "risk elimination" or "only create profile";

[0220] If the safety hard threshold is not triggered and C_i is higher than the recommended threshold, then it is marked as "recommended to keep".

[0221] If the safety hard threshold is not triggered, but C_i is only in the middle range, it is marked as "remaining under observation".

[0222] For batch sequences, the system sorts them according to the following priority:

[0223] 1. Has the security screening been passed?

[0224] 2. Screening status level;

[0225] 3. Overall score or safety score;

[0226] 4. Preset auxiliary sorting fields.

[0227] Step S8: Unified result export and server-shared call (see appendix) Figure 6 )

[0228] For a single sequence, a unified structured result object is output; for batch sequences, a unified result table is generated.

[0229] Preferably, the batch unified result table includes at least:

[0230] sequence;

[0231] Record status;

[0232] length;

[0233] HEMO score;

[0234] TOX score;

[0235] ALG score;

[0236] Safety screening results;

[0237] Overall score or security score;

[0238] Has the security been approved?

[0239] Risk item not passed;

[0240] Physicochemical properties field.

[0241] In one preferred embodiment, the system runs all local models by default and outputs all model results; in another embodiment, it also supports calling a single model or multiple models; in yet another embodiment, it supports executing only the security screening workflow.

[0242] Through a unified command-line interface, script interface, or service interface, this invention can be deployed in a public directory on a server for reuse by multiple users.

[0243] The methods of using this invention include, but are not limited to, the following:

[0244] Input a single polypeptide sequence, run all models by default, and output all model results, physicochemical properties results, and screening conclusions;

[0245] Input a single polypeptide sequence and specify one or more models; only the results of the corresponding model will be output.

[0246] Input a single polypeptide sequence and invoke the security screening mode to run only the security model and low-cost profiling module;

[0247] Upload batch FASTA files and perform cascaded prediction and safety screening on all sequences;

[0248] Upload batch CSV files and perform standardized batch processing, splitting, early-stopping, comprehensive scoring, and sorting export on all sequences;

[0249] In the server's shared directory, one model, multiple models, or all models can be invoked via a unified command line or a unified interface.

[0250] The principle behind this invention is:

[0251] The principle of this invention is not simply to call multiple models in parallel, but to adopt a cascaded architecture of "input standardization - anomaly diversion - low-cost profiling - first-level security inference - early-stopping gating - second-level high-time-consuming inference - dynamic fusion scoring - batch sorting and export".

[0252] Its core principles include:

[0253] Security screening should be implemented as an underlying computational scheduling strategy, rather than just a business logic judgment.

[0254] Data stream isolation for batch sequences is achieved through valid queues, exception queues, retry queues, and completion queues;

[0255] High-risk samples can be quickly identified through low-cost first-level inference, preventing them from continuing to consume high-cost computing power.

[0256] The functional scores are dynamically adjusted by using a physicochemical property penalty function to generate a comprehensive ranking result that better meets the purpose of experimental screening.

[0257] Standardized calls are achieved in server-shared scenarios by unifying the result structure and export format.

[0258] The technical solution of the present invention will be further illustrated below through specific examples.

[0259] Example 1: Cascade prediction and security screening of single polypeptide sequences

[0260] Input a single polypeptide sequence KWKLFK.

[0261] The system first performs standardization and validity checks on the sequence, then writes it into the valid queue. Subsequently, it prioritizes calculating low-cost physicochemical properties and invokes three security models: HEMO, TOX, and ALG.

[0262] If none of the three security models trigger the security threshold, the system will continue to assign functional class model and propensity model reasoning tasks to the sequence, and output the final comprehensive score and screening conclusion according to the comprehensive scoring formula.

[0263] If any item in the Level 1 security reasoning exceeds the threshold, the scheduling module will immediately terminate the task allocation of subsequent high-time-consuming models in that sequence, and only retain its physical and chemical profile, security profile and risk conclusion.

[0264] This embodiment illustrates that the present invention can perform hierarchical inference with Early-Stopping capability on a single sequence without unconditionally calling the entire model.

[0265] Example 2: Asynchronous Splitting and Unified Screening of Batch CSV Files

[0266] The input file is a CSV file containing multiple peptide sequences, with the sequence column named "sequence".

[0267] The system reads each record sequentially and constructs a Record_i object, then performs standardization and validity checks on each record.

[0268] For legitimate records:

[0269] Write to the valid queue Q_valid;

[0270] Entering the stage of physicochemical properties and first-level safety reasoning;

[0271] Determine whether to proceed to the second-level high-time-consuming reasoning stage based on the safety threshold.

[0272] Finally, the write operation is completed in the queue Q_done.

[0273] For illegal records or partially failed records:

[0274] It can be written directly to the exception queue Q_error; or it can first enter the retry queue Q_retry, and then be transferred to Q_error after the maximum number of retries is reached.

[0275] The system ultimately generates a batch screening result table and exports it sorted by "Whether it passed safely - Screening status - Comprehensive score".

[0276] This embodiment illustrates that the present invention can not only improve the efficiency of batch screening, but also maintain the continuous operation of the entire batch of tasks when abnormal data exists.

[0277] Example 3: Complex comparative screening based on physicochemical property adjustment

[0278] Input two candidate sequences A and B.

[0279] in:

[0280] The original score for the antibacterial model of sequence A was 0.93, and the original score for the anticancer model was 0.81, indicating a relatively high functional baseline score.

[0281] However, the sensitization model score for sequence A, a_A = 0.44, is higher than the sensitization threshold θ_A = 0.30;

[0282] Meanwhile, its hydrophobic moment μ_A is higher than the preset hydrophobic moment threshold μ_thr, indicating that it has a high risk of non-specific interactions.

[0283] According to the cascading scheduling rule of the present invention, sequence A has triggered the ALG hard threshold in the first-level security inference stage. Therefore, the system performs Early-Stopping and no longer assigns it subsequent high-time-consuming bias model or additional function model tasks, and directly sets its screening status to "risk elimination" or "only form a profile".

[0284] In comparison:

[0285] The original score for the antibacterial model of sequence B was 0.71, and the original score for the anticancer model was 0.62. The functional scores alone were only moderate.

[0286] However, the HEMO, TOX, and ALG values ​​of sequence B were all below their respective thresholds.

[0287] Furthermore, its pH 7 net charge, isoelectric point, and GRAVY are within the preset preferred range;

[0288] Its hydrophobic moment does not trigger the penalty term, so P_B is close to 0;

[0289] In the overall score, although the functional basis score of Sequence B is lower than that of Sequence A, the overall score C_B is higher than the safety profile score of Sequence A due to its higher safety retention score and lower physical and chemical penalty.

[0290] Therefore, the present invention automatically sorts sequence B before sequence A.

[0291] This embodiment illustrates that the value of the present invention is not merely "listing scores in Excel," but rather that it achieves automated screening that is more stable, consistent, and computationally efficient than manual table reading at the underlying level through threshold gating, physicochemical adjustment, and dynamic fusion algorithms.

[0292] Example 4: Customized invocation of a specified model set

[0293] In some research scenarios, users are only concerned with the results of specific models, for example:

[0294] Only concerned with anti-cancer, antibacterial, and cell membrane penetration properties;

[0295] Or they may only be concerned with hemolytic, toxic, and sensitizing results.

[0296] In this embodiment, the user specifies the model set through a unified call entry point, and the system only schedules the specified model to perform inference, but still retains the unified input standardization, unified result structure and unified export format.

[0297] For example:

[0298] When the user specifies ACP, AMP, and CPP, the system returns the corresponding function result;

[0299] When the user specifies HEMO, TOX, and ALG, the system automatically enters the security screening workflow and enables Level 1 gating and risk assessment.

[0300] This embodiment illustrates that the present invention supports both the default full-model cascaded workflow and the customized scheduling of models for specific scientific research problems.

[0301] Comparative Example 1: Without Early-Stopping Cascade Gating

[0302] In Comparative Example 1, all models were called without discrimination for the same batch of polypeptide sequences, and early termination was not performed after the safe model.

[0303] Under this scheme, even if a sequence has clearly exceeded the threshold on HEMO, TOX, or ALG, it will still continue to occupy the inference resources of subsequent high-time-consuming functional models and bias models.

[0304] In contrast, this invention, through Early-Stopping gating, can prevent high-risk samples from continuing into the deep inference stage, thus being significantly more beneficial in reducing ineffective computing power usage in high-throughput batch screening scenarios.

[0305] Comparative Example 2: Dynamic fusion scoring without adjustment using physicochemical properties

[0306] In Comparative Example 2, only the functional score, safety score, and tendency score are statically weighted averaged, without introducing a physicochemical property penalty term.

[0307] This approach cannot handle cases where the functional score is high but the physicochemical properties are poor or the functional score is moderate but the physicochemical properties are significantly superior. It may lead to screening results that are biased towards high functional scores, while ignoring the exploitability of the sequence and the potential risk of nonspecificity.

[0308] In contrast, this invention dynamically adjusts the hydrophobic moment, GRAVY, and net charge shift using a penalty function P_i, which makes the overall ranking more in line with the actual needs of subsequent experimental screening.

[0309] It needs to be further explained that:

[0310] The core of this invention does not lie in a single publicly disclosed model itself, but in the overall technical solution composed of multi-model cascaded scheduling, security gating, physical and chemical regulation fusion, and abnormal isolation data stream.

[0311] The Early-Stopping in this invention is not only a security screening rule, but also a computing resource optimization mechanism, which can be described as a low-level scheduling and control method.

[0312] The comprehensive score in this invention is not a simple static weighting, but a dynamic fusion method that introduces physicochemical property adjustment terms and safety risk scores.

[0313] The data stream splitting structure in this invention can be implemented using asynchronous message queues, task pools, tensor pools, and buffer structures. As long as it can achieve the isolation of legal samples from abnormal samples and the splitting of early-stopped samples from retry samples, it falls within the protection scope of this invention.

Claims

1. A multi-model cascade prediction and safety screening method for peptide sequences, characterized in that, Includes the following steps: Step S1: Receive input data and construct task records; The input data is one or more polypeptide sequences input by the user; Step S2: Sequence normalization, anomaly isolation, and data stream splitting; input records should be split into at least the following queues or pools: Q_valid: A valid sequence queue used for subsequent reasoning; Q_error: An exception sequence queue used to record illegal or unusable sequences; Q_retry: A local failure retry queue used to temporarily store a sequence of local calls that have failed but can be retried; Q_done: The completion queue, used to summarize sequences that have completed prediction or screening; Step S3: Low-cost physical and chemical profiling and primary task allocation; For sequences in Q_valid, the low-cost computation module is executed first. Step S4: Level 1 secure reasoning and Early-Stopping computational resource optimization mechanism; For each valid sequence Q_valid, the safety class model is run first. Step S5: Secondary functional / propensity reasoning; For sequences that have not triggered Early-Stopping, continue with secondary inference, invoking the remaining functional class model and the propensity model; this secondary inference is only performed on candidates that have passed the primary security gating. Step S6: Introduce a dynamic fusion scoring mechanism for adjusting physicochemical properties; For sequences that have completed the second-level reasoning, dynamic fusion scoring under the adjustment of physical and chemical features is performed; when the sequence has triggered the safety hard threshold in step S4, the safety screening conclusion is directly output, and its comprehensive score is limited to the profile score that depends only on the safety retention score, or it is directly marked as "risk elimination"; Step S7: Screening status determination and batch sorting; Based on the Early-Stopping results, the comprehensive score C_i, the safety retention score S_i, and the risk triggering items, the sequence is assigned a screening status; Step S8: Unify result export and server sharing; For a single sequence, a unified structured result object is output; for batch sequences, a unified result table is generated.

2. The method according to claim 1, characterized in that, step The input methods in S1 include at least: Directly input a single polypeptide sequence; Enter a single FASTA text; Upload or read batch FASTA files; Upload or read batch CSV files.

3. The method according to claim 1, characterized in that, In step S2, the input polypeptide sequence undergoes the following preprocessing: Remove whitespace and newline characters; Remove FASTA header information; Convert the sequence to uppercase; Determine whether it contains only standard amino acid characters; Calculate the basic length information; Anomalies are marked for empty sequences, sequences containing illegal characters, and sequences that do not meet the preset length condition.

4. The method according to claim 1, characterized in that, Step S3 prioritizes the execution of a low-cost computation module on the sequences in Q_valid, including: Calculation of basic physical and chemical properties; Statistical feature extraction can be completed quickly; Lightweight security prediction model inference; The basic physicochemical properties include at least the following: Sequence length; Molecular weight; Average residue count; Net charge at pH 5, pH 7 and pH 9 conditions; Charge density; Isoelectric point; GRAVY; Hydrophobic moment; Aliphatic index; The proportion of hydrophobic residues; The percentage of aromatic residues; Percentage of polar residues.

5. The method according to claim 1, characterized in that, In step S4, the security class model is run first for each valid sequence Q_valid, including: HEMO, a hemolytic model; Toxicity model TOX; ALG, a model of sensitization; Early-Stopping is triggered when any of the following conditions are met: h_i>=θ_H; t_i>=θ_T; a_i>=θ_A; R_i>=R_stop; Where h_i: HEMO score; t_i: TOX score; a_i: ALG score; R_i: security risk score; θ_H=0.46; θ_T=0.38; θ_A=0.

30.

6. The method according to claim 1, characterized in that, The functional class model in step S5 shall include at least one or more of the following models: Anticancer peptide model; Antimicrobial peptide model; Antifungal peptide model; Anti-inflammatory peptide model; Antioxidant peptide model; Antiviral peptide model; Cell-penetrating peptide model; DPP-IV inhibitory peptide model; The propensity model includes at least one or more of the following models: MHC-I binding propensity model; Other receptor binding tendency models; Other immune-related propensity models.

7. The method according to claim 1, characterized in that, The screening status in step S7 includes at least the following: Recommended to keep; Continue observation; Only an image is created; Risk elimination; in: If any security hard threshold is triggered, it is preferable to mark it as "risk elimination" or "only create profile"; If the safety hard threshold is not triggered, and the comprehensive score C_i is higher than the recommended threshold, it is marked as "recommended to keep". If the safety hard threshold is not triggered and the composite score C_i is only in the middle range, it is marked as "retain observation".

8. The method according to claim 1, characterized in that, The unified result table in step S8 shall include at least: sequence; Record status; length; HEMO score; TOX score; ALG score; Safety screening results; Overall score or security score; Has the security been approved? Risk item not passed; Physicochemical properties field.

9. A multi-model cascade prediction and safety screening system for polypeptide sequences implementing the method of any one of claims 1-8, characterized in that, include: The input adapter module is used to receive one or more polypeptide sequences input by the user; The standardization and validity verification module is used to standardize the input peptide sequences, isolate anomalies, and process data stream splitting. The low-cost physical and chemical profiling module prioritizes low-cost computation of legitimate sequences used for subsequent reasoning in the standardization and legality verification module; The Level 1 security inference and Early-Stopping gating module is used to prioritize the execution of security models for each legitimate sequence; The secondary function / prejudice reasoning module is used to continue performing secondary reasoning on sequences that have not triggered Early-Stopping, and to call the remaining function class model and prejudice model; The dynamic comprehensive scoring module is used to perform dynamic fusion scoring under the adjustment of physicochemical characteristics for sequences that have completed secondary reasoning. The batch sorting and export module assigns a screening status to the sequences based on the Early-Stopping results and exports the unified results.

10. A computer storage medium, characterized in that, The storage medium stores a computer program that implements the method of any one of claims 1-8.

Images