Method and system for identifying coding potential of lincRNA small peptide

Abstract

The invention discloses a lincRNA small peptide coding potential identification method and a lincRNA small peptide coding potential identification system, and belongs to the technical field of computer technology and bioinformatics. The method comprises the following steps: firstly, processing a lincRNA sequence into a standardized sequence set; identifying and integrating the deduplicated potential coding fragments to obtain a candidate coding region set; after extracting multi-dimensional feature codes of each region, performing fusion through a Mama model and a machine learning model to obtain a fusion prediction confidence coefficient and a label; and finally, judging the coding potential, if not, outputting a conclusion, if yes, deducing an amino acid sequence, calculating comprehensive sequencing confidence by combining multiple indexes, sequencing, and screening small peptide sets with high and low confidence according to a threshold value. According to the method, an efficient sequence state space modeling structure is introduced, dependence on ribosome sequencing or mass spectrum data is not needed, high-precision, low-complexity and extensible prediction of the potential small peptide coding capacity in the lincRNA sequence is achieved, and the technical defect that accuracy and calculation efficiency are difficult to consider at the same time in an existing method is overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology and bioinformatics, and particularly relates to a lincRNA small peptide coding potential identification method and system. BACKGROUND

[0002] With the rapid development of high-throughput sequencing technology and multi-omics research, more and more studies have found that long intergenic non-coding RNA (lincRNA) traditionally considered to have no coding ability can be translated to produce small peptides or micro-proteins with biological functions under certain conditions. Such small peptides derived from non-coding transcripts are considered an important part of the "hidden proteome" and play a potential role in tumor development, cell signal transduction, and immune regulation. Related studies have shown that systematic mining of lincRNA coding small peptides is of great significance for expanding genome function annotation and discovering new disease-related molecules.

[0003] However, current computational prediction of lincRNA small peptide coding potential still faces many technical challenges, mainly in the following aspects: Most existing algorithms are derived from linear mRNA or circRNA coding potential evaluation framework, and lack of special modeling for lincRNA structure and translation characteristics. Current commonly used coding potential prediction methods are mostly based on open reading frame (ORF) length, codon usage bias or simple sequence statistical features, and their original design is mainly aimed at classical protein coding transcripts. It is difficult to accurately describe the characteristics of short ORF, multiple start sites and non-classical translation mode in lincRNA, resulting in insufficient sensitivity and specificity of small peptide identification.

[0004] Traditional feature engineering methods have limited expression ability for high-dimensional sequence information, and it is difficult to capture the complex patterns of lincRNA coding potential. Existing methods usually rely on a small number of sequence features or rule thresholds designed by humans, and have insufficient modeling ability for potential long-range dependencies, local pattern combinations and context information in RNA sequences, which limits the generalization performance of the model in complex transcriptome background.

[0005] Some methods rely heavily on ribosome sequencing or mass spectrometry data, which has high application cost and limited scope of application. Although ribosome binding evidence and mass spectrometry identification are considered important means to verify translation events, the related experimental techniques have high cost and sample acquisition is difficult, and it is difficult to apply them systematically in large-scale transcriptome data, which limits the high-throughput prediction and screening of lincRNA small peptides.

[0006] The existing deep learning model has obvious bottleneck in long sequence processing and computing efficiency. Although the deep learning method introduced in recent years improves the accuracy of encoding potential prediction to some extent, the mainstream model is based on recurrent neural network or self-attention mechanism, which has high computational complexity, slow training and reasoning speed when processing long lincRNA sequence, and is difficult to meet the requirements of efficiency and stability for large-scale data analysis.

[0007] In summary, the prior art still lacks an algorithm scheme for lincRNA small peptide coding characteristics, considering the prediction accuracy and computing efficiency, and suitable for large-scale transcriptome data. Therefore, it is necessary to propose a new technical scheme to introduce a more efficient deep learning sequence modeling structure to systematically, quickly and accurately predict the potential small peptide coding ability of lincRNA. SUMMARY

[0008] In order to overcome the above shortcomings, the purpose of the present application is to provide a lincRNA small peptide coding potential identification method and system to solve the problems of insufficient feature expression ability, low long sequence modeling efficiency and poor adaptability to large-scale transcriptome data in the prediction of lincRNA small peptide coding potential in the prior art.

[0009] To achieve the above-mentioned purpose, the technical scheme of the present application is as follows: The present application provides a lincRNA small peptide coding potential identification method and system, which comprises the following steps: S1, obtaining a standardized lincRNA sequence set with uniform format, length and coding mode; S2, integrating, deduplicating and merging all potential coding fragments identified from the standardized lincRNA sequence set to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility; S3, extracting multi-class and multi-dimensional sequence features of each candidate small peptide coding region in the candidate small peptide coding region set, encoding them into vector or tensor form feature representation that can be learned by the model, and outputting the corresponding multi-dimensional sequence feature representation set; S4, inputting the multi-dimensional sequence feature representation set of the candidate small peptide coding region into the trained Mamba deep learning network, outputting the deep sequence representation vector and the Mamba single model coding probability, then inputting the multi-dimensional sequence feature, the deep sequence representation vector or the combination of the two into at least one trained machine learning model to output the machine learning single model coding probability, subsequently linearly summing the fusion prediction confidence according to the preset weight, outputting the numerical value and / or the coding-non-coding binary classification judgment label, and finally marking the candidate region with fusion prediction confidence lower than the threshold as a to-be-verified sample and returning to step S3 for re-feature encoding; S5, input the fusion prediction confidence of each candidate small peptide coding region and the coding-non-coding binary classification judgment label, comprehensively judge the small peptide coding potential of the corresponding lincRNA according to the judgment result, output no coding potential conclusion if there is no, deduce the small peptide amino acid sequence from the coding positive region if there is, then calculate the comprehensive ranking confidence according to the fixed weight of the fusion prediction confidence and the multi-class evaluation index, sort the multiple positive candidate small peptides from high to low after weighting, and then put the candidate small peptides with the comprehensive ranking confidence higher than the threshold into the high confidence set, and put the rest into the low confidence set.

[0010] The application also provides a lincRNA small peptide coding potential identification system, comprising: A standardized lincRNA sequence acquisition module is configured to acquire a standardized lincRNA sequence set with unified format, length and coding mode. A candidate small peptide coding region construction module is configured to integrate, remove duplicates and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility. A multi-dimensional sequence feature encoding module is configured to extract multi-class multi-dimensional sequence features of each candidate small peptide coding region in the candidate small peptide coding region set, encode the features into a vector or tensor form feature representation that can be learned by a model, and output a corresponding multi-dimensional sequence feature representation set. An encoding potential fusion prediction module is configured to input the multi-dimensional sequence feature representation set of the candidate small peptide coding region into a trained Mamba deep learning network, output a deep sequence representation vector and a Mamba single model coding probability, then input the multi-dimensional sequence features, the deep sequence representation vector or a combination of the two into at least one trained machine learning model to output a machine learning single model coding probability, subsequently linearly sum the coding probabilities according to a preset weight to obtain a fusion prediction confidence, output the value and / or a coding-non-coding binary classification judgment label, and finally mark the candidate region with a fusion prediction confidence lower than a threshold as a to-be-verified sample and return to step S3 for feature encoding again. A small peptide coding potential judgment and screening module is configured to input the fusion prediction confidence of each candidate small peptide coding region and the coding-non-coding binary classification judgment label, comprehensively judge the small peptide coding potential of the corresponding lincRNA according to the judgment result, output no coding potential conclusion if there is no, deduce the small peptide amino acid sequence from the coding positive region if there is, then calculate the comprehensive ranking confidence according to the fixed weight of the fusion prediction confidence and the multi-class evaluation index, sort the multiple positive candidate small peptides from high to low after weighting, and then put the candidate small peptides with the comprehensive ranking confidence higher than the threshold into the high confidence set, and put the rest into the low confidence set.

[0011] The present application takes the lincRNA full-length nucleotide sequence as input, first standardizes the sequence to obtain a standardized lincRNA sequence set with uniform format, length and encoding mode; integrates, de-duplicates and merges the potential coding fragments identified therein to construct a candidate small peptide coding region set. Further, multi-dimensional sequence features are extracted from each candidate small peptide coding region, and a model-learnable feature representation set is formed. On this basis, the feature representation set is input into the Mamba deep learning network based on the state space model to obtain a deep sequence representation vector and a single model coding probability; the multi-dimensional sequence features, the deep sequence representation vector or their combination are input into at least one machine learning model to obtain the corresponding single model coding probability; the fusion prediction confidence is obtained by fusing the multi-model results, and the coding-non-coding binary classification judgment label is output, and the low confidence sample is returned to the feature coding step for iterative optimization. Finally, the fusion prediction confidence and the classification label are comprehensively judged to determine the small peptide coding potential of the lincRNA, the corresponding small peptide amino acid sequence is derived from the coding positive region, and the high confidence and low confidence candidate small peptide sets are divided according to the comprehensive index weighted ranking, and the lincRNA coding potential conclusion and the corresponding high confidence small peptide candidate product are output. While ensuring the prediction accuracy, the modeling efficiency of long lincRNA sequence and the stability and generalization ability of the algorithm in large-scale transcriptome data analysis are improved. BRIEF DESCRIPTION OF DRAWINGS

[0012] FIG. 1 is a flow chart of the lincRNA small peptide coding potential identification method provided by the embodiment of the present application; FIG. 2 is a flow chart of the sub-step of step S1 provided by the embodiment of the present application; FIG. 3 is a flow chart of the sub-step of step S3 provided by the embodiment of the present application; FIG. 4 is a flow chart of the sub-step of step S4 provided by the embodiment of the present application; FIG. 5 is a flow chart of the sub-step of step S5 provided by the embodiment of the present application; FIG. 6 is a structure chart of the lincRNA small peptide coding potential identification system provided by the embodiment of the present application; FIG. 7 is a sub-module structure chart of the standardized lincRNA sequence acquisition module 1 provided by the embodiment of the present application; FIG. 8 is a sub-module structure chart of the multi-dimensional sequence feature coding module 3 provided by the embodiment of the present application; FIG. 9 is a sub-module structure chart of the coding potential fusion prediction module 4 provided by the embodiment of the present application; FIG. 10 is a sub-module structure chart of the small peptide coding potential determination and screening module 5 provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application is illustrated by the following examples.

[0014] Gene annotation information refers to structured annotation data used to describe the identity, category and genomic structure positioning of genes and their transcripts, which can be derived from public annotation databases or equivalent annotation files (e.g. annotation sets of GENCODE, Ensembl, RefSeq, GTF / GFF3 format files). The gene annotation information includes one or more of the following information: gene and transcript identification (gene_id, transcript_id and / or gene_name), transcript type / gene type (used to distinguish lincRNA from protein-coding genes, etc.), chromosome number, start and end coordinates of genes or transcripts on the reference genome and strand information, exon structure information (number of exons and coordinates of each exon), and annotation version and reference genome version information; the gene annotation information is used to screen and determine the lincRNA sequences to be analyzed, establish the correspondence between the sequences and the genomic coordinates, and support the position backtracking and result annotation of the predicted candidate small peptide coding regions.

[0015] The present embodiment provides a model training and saving method for unknown lincRNA sequence coding potential prediction. First, the training data is processed according to steps S1-S3: obtain and organize the labeled lincRNA sequences and their candidate small peptide coding regions, complete sequence standardization, candidate small peptide coding region construction and feature extraction, and form a training sample set that can be used for supervised learning.

[0016] In S606, the trained Mamba model parameters, machine learning model parameters and multi-model fusion strategy are saved for subsequent prediction analysis of unknown IncRNA sequences. In this embodiment, the input target length threshold of the Mamba model is consistent with S105, which is fixed at 1024 nt; for sequences longer than 1024 nt, a sliding window splitting method is used (window length 1024 nt, step length 512 nt), and for short sequences, padding is used to fill up to 1024 nt, and an effective length mask is generated to ignore the padding area. The Mamba model structure parameters are fixed as follows: embedding dimension 256, network layer number 6 layers, hidden dimension 512, dropout=0.1, and output layer is a binary classification fully connected layer; the training hyperparameters are fixed as follows: optimizer AdamW, learning rate 3x10^-4, batch_size=64, training rounds 50, early stopping strategy is to stop when the validation set index does not improve for 10 consecutive epochs, and the trained network weight and configuration file are saved. The machine learning model uses random forest (RF), and its parameters are fixed as follows: learning_rate=0.05, n_estimators=1000, max_depth=8, num_leaves=64, subsample=0.8, colsample_bytree=0.8, and the trained model file is saved. The multi-model fusion strategy is consistent with S404, and the weighted combination prediction score is used as the main fusion method, the fusion weight is determined by the validation set optimization and is fixed and saved, the Mamba weight is 0.60, the RF weight is 0.40, and the fusion judgment threshold is 0.50; at the same time, the voting arbitration rule is saved as a backup: when the two model binary classification results are inconsistent, the one with the larger deviation from its own threshold (the one with higher confidence) is selected as the final judgment. The above model parameters, fusion weights, thresholds and arbitration rules are all stored in a callable form, and are used for prediction output of unknown IncRNA sequences after processing consistent with S1-S3.

[0017] In step S105, in order to meet the fixed input length requirement of the depth model, length adaptation processing is performed on the IncRNA sequence, and the target input length threshold is fixed at 1024 nt (nucleotides). When the sequence length exceeds 1024 nt, the sequence is split using a sliding window method, the window length is fixed at 1024 nt, and the step is fixed at 512 nt, so that the long sequence is covered by multiple overlapping subsequences; when the sequence length does not exceed 1024 nt, the sequence is padded to 1024 nt using padding, and an effective length mask is generated at the same time to distinguish between real bases and padding positions, so as to ignore the influence of the padding area in subsequent feature statistics and depth network modeling. The split subsequence retains its start and end position index in the original sequence to support subsequent candidate small peptide coding region positioning and result backtracking.

[0018] In step S6, a training data set with positive and negative samples is constructed, and the Mamba deep learning network (which can be briefly described as the Mamba model) and at least one machine learning model are trained after performing the same processing procedures as S1-S3. In this embodiment, the structure parameters of the Mamba deep learning network are fixed as follows: input embedding dimension 256, network layer number 6, hidden dimension 512, dropout=0.1, and the output layer is a binary classification fully connected layer and outputs the confidence of the candidate small peptide coding region; the training hyperparameters are fixed as follows: using AdamW optimizer, learning rate 3x10^-4, batch_size=64, maximum training rounds 50, and setting the early stopping strategy as terminating training when the evaluation index of the validation set does not improve for 10 consecutive epochs, and saving the network weight file and model configuration file after training. The machine learning model uses random forest (RF), and the key parameters are fixed as follows: learning_rate=0.05, n_estimators=1000, max_depth=8, num_leaves=64, subsample=0.8, colsample_bytree=0.8, and the corresponding model file is saved after training.

[0019] In step S404, the outputs of the Mamba model and the machine learning model are fused. In this embodiment, the weighted combination prediction score is used as the main fusion method, the fusion weights are determined and fixed by validation set optimization, the Mamba weight is 0.60, the RF weight is 0.40, the fusion decision threshold is fixed at 0.50, which is used to convert the fusion score into a binary classification result of coding / non-coding; at the same time, the voting arbitration rule is saved as a backup fusion rule: when the binary classification results of the two models are inconsistent, the deviation of the output probability of each model relative to the fusion decision threshold is compared, and the label corresponding to the larger deviation is selected as the final decision.

[0020] In step S606, the trained Mamba deep learning network parameters, RF machine learning model parameters, and multi-model fusion strategy (including fusion weights, fusion threshold, and voting arbitration rules) are stored in a callable form. Subsequently, when predicting unknown lincRNA sequences, the input sequence is first subjected to length adaptation, candidate small peptide coding region construction, and feature extraction consistent with S105, S1-S3, then the saved Mamba deep learning network and RF machine learning model are called to obtain the prediction probability / score of the candidate small peptide coding region, and the fusion prediction confidence is calculated according to the saved fusion weights and threshold to output the final coding potential prediction result, thereby realizing the coding potential analysis and candidate small peptide prioritization of unknown lincRNA sequences.

[0021] As shown in Figure 1 The embodiment of the present application provides a lincRNA small peptide coding potential identification method, which comprises: Step S1, obtaining a standardized lincRNA sequence set with unified format, length, and coding mode; The source of lincRNA sequences includes public transcriptome databases, gene annotation databases, and transcript set data obtained by high-throughput sequencing. In the acquisition process, transcripts belonging to long intergenic non-coding RNA types can be selected according to gene annotation information to exclude protein-coding RNA or other non-target RNA types that may interfere with subsequent analysis.

[0022] After the sequence acquisition is completed, the lincRNA sequences are subjected to uniform preprocessing, which includes quality filtering, standardization processing, and length adaptation processing, specifically: Sequence quality filtering: removing sequences containing uncertain bases or abnormal characters to ensure the stability of subsequent feature coding and model calculation.

[0023] Sequence standardization processing: uniformly processing the sequence direction and converting the base characters to a unified coding format to ensure consistency in the representation of sequences from different sources.

[0024] Sequence length adaptation processing: setting the target input length threshold to 1024 nt, splitting the lincRNA sequences longer than the threshold using a sliding window with a window length of 1024 nt and a step length of 512 nt, and performing end padding or end alignment and cutting on the residual fragments at the end of the split. For short sequences shorter than the threshold, fill them to 1024 nt according to a fixed alignment rule, and generate an effective length label for all filled sequences to ignore the filled region. The split and filled sub-sequences retain their start and end position indexes in the original sequence.

[0025] Through the above preprocessing steps, the lincRNA sequence of any length is converted into a fixed-length subsequence input unit, so that all input sequences have consistency in format, length and encoding method, thereby ensuring that the subsequent deep learning model can stably and efficiently perform calculations.

[0026] Step S2, all potential coding fragments identified from the standardized lincRNA sequence set are integrated, de-duplicated and merged to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility.

[0027] For each preprocessed standardized lincRNA sequence, without pre-setting its classic protein coding ability, one or more combinations of multi-reading frame scanning, division based on sequence local features, and covering window scanning are used to identify potential coding fragments (where multi-reading frame scanning needs to identify short open reading frames with a length corresponding to ≥20 amino acids). After identifying all potential coding fragments, they are integrated, de-duplicated and merged to construct multiple candidate small peptide coding regions, forming a candidate small peptide coding region set to comprehensively cover potential small peptide coding events under non-classical initiation, non-standard length or special translation modes.

[0028] The construction of candidate small peptide coding regions is not limited to a single rule, but is achieved by one of the following methods or their combinations: Multi-reading frame scanning method: scanning the lincRNA sequence under different reading frames to identify short open reading frames with a length within a predetermined range to capture potential small peptide coding segments.

[0029] Division based on sequence local features: dividing the sequence into multiple fragments with potential coding possibilities according to local base arrangement features, composition characteristics or statistical feature changes in the sequence.

[0030] Covering window scanning method: using fixed or variable window strategies to continuously scan the entire lincRNA sequence, so that any sequence fragment at any position has the opportunity to be included in the candidate small peptide coding region set.

[0031] By constructing multiple candidate small peptide coding regions, it is avoided to rely solely on a single ORF rule and miss small peptide coding events under non-classical initiation, non-standard length or special translation modes.

[0032] Step S3, extract multi-class and multi-dimensional sequence features from each candidate small peptide coding region in the candidate small peptide coding region set, encode them into vector or tensor form feature representation that can be learned by the model, and output the corresponding multi-dimensional sequence feature representation set.

[0033] For example, Figure 3As shown, the extracted multi-class multi-dimensional sequence features include base level, codon or quasi-codon, sequence complexity and information, and context-related features. The context-related features are constructed by fixing the upstream 128 nt and downstream 128 nt window range, padding the insufficient part, and generating an effective length label. The feature encoding of the Mamba deep learning network is spliced and padded to 1024 nt in the form of a tensor, and the feature encoding of the machine learning model is spliced to form a vector form feature representation. Finally, the multi-dimensional sequence feature representation set corresponding to all candidate regions is output.

[0034] Multi-dimensional sequence feature extraction and encoding operation standards: For the candidate small peptide encoding region, four types of multi-dimensional sequence features are extracted, including base level, codon / quasi-codon, sequence complexity and information, and context-related features, and are encoded into vector or tensor form model learnable representation. All features are not determined by artificial fixed threshold, and the model learns the feature pattern related to small peptide coding potential in training. The specific extraction standards are as follows: Base level features: Base composition ratio: Calculate the proportion of A, C, G, and T (U) in the candidate region to form a 4-dimensional vector, and add GC content as a scalar feature; Local arrangement pattern: Calculate the k-mer frequency of di-nucleotide (16 kinds) or tri-nucleotide (64 kinds) to generate a corresponding dimension pattern vector; Distribution characteristics: Divide the candidate region into fixed segments, and splice the base composition ratio vector or k-mer frequency vector of each segment after calculation.

[0035] Codon or quasi-codon related features: Quasi-codon is a continuous base combination (3-mer) with a length of 3, without translation reading frame and start / stop codon restriction. Extract overlapping triplets with a step size of 1 or non-overlapping triplets with a step size of 3, count the normalized frequency of 64 triplets, and generate a 64-dimensional vector to reflect the sequence's potential translation bias.

[0036] Sequence complexity and information features: Low complexity ratio: Calculate the total length of continuous ≥6 nt identical base fragments in the candidate region as a proportion of the total length of the region; Repetition index: Calculate the ratio of the number of unique triplets to the total number of triplets; Information entropy: Calculate the information entropy based on single base distribution (4 base proportions) and triplet distribution (64 triplet frequencies) respectively to represent the difference in sequence structure and composition.

[0037] Context-related features: Relative position: the proportion of the start and end positions of the candidate region to the full length of the IncRNA as a scalar feature; Upstream and downstream windows: fixed windows of 128 nt upstream and 128 nt downstream of the candidate region, with padding and effective length mask generated when the window length is insufficient; Feature fusion: the Mamba model input is padded to 1024 nt after splicing the upstream 128 nt + candidate small peptide coding region + downstream 128 nt; the machine learning model input is a vector formed by splicing the candidate region features + upstream and downstream window features + position features.

[0038] Multi-dimensional sequence feature encoding of the candidate small peptide coding region: Multi-dimensional sequence feature encoding is performed on the candidate small peptide coding region to form an input representation that can be processed by a deep learning model. The core of the encoded sequence features includes the following four types, and each type of feature is input into the model in the form of a vector / tensor: Base-level features: including the base composition proportion, local arrangement pattern, and base distribution characteristics of the candidate region; Codon or quasi-codon related features: encoding the usage features of consecutive base combinations in the candidate region, reflecting the potential translation preference of the sequence; Sequence complexity and information features: including low complexity proportion, repetition index, and information entropy, describing the structure and composition differences of the candidate region; Context-related features: combining the relative position of the candidate region in the IncRNA full length and the sequence information constructed by the fixed upstream and downstream windows, reflecting the potential translation-related context features.

[0039] Step S4: input the multi-dimensional sequence feature representation set of the candidate small peptide coding region into the trained Mamba deep learning network, output the deep sequence representation vector and Mamba single model encoding probability, then input the multi-dimensional sequence features, deep sequence representation vector, or a combination of the two into at least one trained machine learning model to output machine learning single model encoding probability, subsequently linearly sum the fusion prediction confidence according to the preset weight, output the numerical value and / or coding-non-coding binary classification judgment label, finally mark the candidate regions with fusion prediction confidence lower than the threshold as verification samples and return to step S3 for re-feature encoding.

[0040] As Figure 3As shown, in the present embodiment, the Mamba deep learning network adopts a trained model, and the structure parameters thereof are embedding dimension 256, network layer number 6 layers, hidden dimension 512, dropout=0.1, and the output layer is a binary classification full connection layer; the machine learning model selected is a trained random forest (RF), and the parameters thereof are learning_rate=0.05, n_estimators=1000, max_depth=8, num_leaves=64, subsample=0.8, colsample_bytree=0.8; the model fusion adopts a weighted combination prediction score mode, the preset Mamba weight is 0.60, the RF weight is 0.40, the fusion judgment threshold is set to 0.50, and a voting arbitration rule is reserved as a backup, that is, when the binary classification results of the two models are inconsistent, the label corresponding to the one with a relatively larger deviation from the threshold is selected as the final judgment, the candidate region with a fusion prediction confidence lower than 0.50 is marked as a to-be-verified sample, and step S3 is returned to re-perform feature encoding.

[0041] In step S4, in order to further improve the accuracy and stability of the lincRNA small peptide coding potential prediction and overcome the problem that the generalization ability of a single model is limited in a complex transcriptome background, the present application further fuses a machine learning model on the basis of introducing a Mamba deep learning network structure based on a state space model, and constructs a multi-model collaborative prediction framework.

[0042] (1) Deep sequence representation learning based on the Mamba deep learning network Firstly, the multi-dimensional sequence features of the candidate small peptide coding region obtained in step S3 are input into the Mamba deep learning network.

[0043] The Mamba deep learning network is based on a state space modeling mechanism and recursively updates the sequence features, and models the internal and context information of the candidate small peptide coding region through the continuous evolution of the hidden state, thereby achieving the following technical effects: capturing long-range dependencies in the candidate small peptide coding region; effectively expressing non-local sequence patterns; and reducing the computational resource consumption required for long sequence modeling under the premise of maintaining linear time complexity.

[0044] In the above manner, the Mamba deep learning network outputs a deep sequence representation vector for representing the potential translation characteristics of the candidate small peptide coding region.

[0045] (2) Discriminative feature modeling based on a machine learning model While obtaining the deep sequence representation of Mamba deep learning network output, the present application further introduces at least one traditional machine learning model to model the discriminative features of candidate small peptide coding regions to supplement the discriminative ability of deep learning model under specific data distribution.

[0046] The machine learning model includes but is not limited to: random forest model; gradient boosting decision tree model; support vector machine model; logistic regression model. The input features of the machine learning model can include: The partial statistical features extracted in step S3; the deep sequence representation output by the Mamba deep learning network; or the combination representation of the above two types of features. By dividing the feature space through tree model or interval maximization model, the model has stronger discriminative robustness when processing candidate small peptide coding regions of different lengths and different structure types.

[0047] (3) Multi-model collaborative prediction and result fusion After completing the independent prediction of Mamba deep learning network and machine learning model, in order to accurately fuse the results of the two types of models, the present application adopts one or more of the following fusion methods, and all fusion rules are determined and fixedly saved through validation set or cross-validation in the training stage, for subsequent inference stage: The weight allocation operation standard of multi-model fusion is as follows: 1. Weighted combination prediction score weight: Validation set optimization determination: traverse the weight in the range of 0-1 with a step of 0.05 or 0.1, select the weight that makes the AUC, F1, accuracy or PR-AUC optimal and fix it; Performance index normalization determination: allocate weight according to the AUC or F1 proportion of each model on the validation set, and the model with better performance gets higher weight.

[0048] 2. Voting rules: Majority voting: when there are 3 or more models, if more than half are positive, output 1, otherwise output 0; Flat vote arbitration: decision according to confidence priority (compare the distance between model output probability and threshold) or preset model priority.

[0049] One, weighted combination prediction score: linearly weighted sum of the prediction probability / score output by each model to obtain the fusion score, the weight is preferably obtained by at least one of "validation set optimization determination" or "performance index normalization determination"; wherein the validation set optimization determination refers to searching the weight on the validation set (for example, enumerating in the range of 0 to 1 with a step of 0.05 or 0.1), selecting the weight that optimizes the preset evaluation index (the evaluation index is preferably one or more of AUC, F1, accuracy or PR-AUC), and fixing the weight as the fusion weight; the performance index normalization determination refers to first calculating the performance index value of each base model on the validation set, and then determining the weight according to the relative performance proportion (for example, using the proportion of AUC or F1 of each model as the weight), so that the model with better performance obtains higher weight.

[0050] Second, voting integration decision result: first convert the probability / score output by each model into a binary classification label through a preset threshold, and then obtain the final decision according to the voting rule; the voting rule includes majority voting and confidence priority arbitration: 1. Majority voting: when there are three or more base models, if more than half of the models output positive labels, the final decision is 1, otherwise 0; if there is a tie with the same number of votes, the decision is made according to the confidence priority arbitration or the pre-set model priority order; 2. Confidence priority arbitration: when the labels of each model are inconsistent, compare the distance between the output probability of each model and the decision threshold (or compare the maximum probability value), and select the label corresponding to the higher confidence as the final output; when the binary classification results of two models are inconsistent, the probability corresponding to the one with larger deviation threshold can be used as the prediction confidence of the candidate small peptide coding region for arbitration.

[0051] Third, secondary prediction of upper layer model: input the output of each base model as new features into an upper layer fusion model for secondary learning and prediction, wherein the new features at least include the prediction probability / score and the binary classification label output by each model, and can optionally include the confidence of each model and the candidate small peptide coding region level statistics (such as the proportion of positive windows in the same transcript, maximum / average score, etc.) as enhanced features; the type of the upper layer model is preferably one or more of logistic regression, linear discriminant model, support vector machine, random forest (GBDT / XGBoost / RF) or shallow multilayer perceptron (MLP), and the parameters of the upper layer model are learned based on the validation set or cross-validation in the training stage to minimize the classification loss and maximize the preset evaluation index, and finally the parameters of the upper layer model, the fusion weight and the voting rule are saved together for prediction analysis of unknown lincRNA sequences.

[0052] Through the above multi-model fusion strategy, the influence of single model prediction deviation on the final result can be effectively reduced, and the identification accuracy and stability of the real small peptide coding event are improved, thereby providing a reliable basis for subsequent threshold determination of the candidate small peptide coding region.

[0053] The result fusion mode includes but is not limited to: the fusion prediction confidence output by different models is weighted and combined to obtain a prediction score; a voting mechanism is used to integrate the coding potential determination results; the outputs of multiple models are used as new features and input into an upper-level determination model for secondary prediction.

[0054] Through the multi-model collaborative prediction mechanism, the influence of single model prediction deviation on the final result can be effectively reduced, thereby improving the identification accuracy and stability of the real small peptide coding event.

[0055] In the present application, the "confidence" includes two types of evaluation indexes of candidate coding regions and candidate small peptides, and the definition and application scenarios of each type of confidence are as follows: (1) Single model coding probability: respectively refers to the coding probability of the candidate small peptide coding region output by the Mamba deep learning network, and the coding probability of the same candidate coding region output by the machine learning model (such as RF), both of which are the original results of independent prediction of each model.

[0056] (2) Fusion prediction confidence: refers to the fusion output probability obtained by fusing the above two types of single model coding probability according to a preset weight, which is used for threshold determination in step S205; when the coding / non-coding determination results of the two models are inconsistent, the probability corresponding to the larger deviation threshold can be used as the prediction confidence of the candidate coding region for arbitration.

[0057] (3) Comprehensive ranking confidence: when multiple candidate small peptides correspond to the same lincRNA, in order to realize the ranking and priority output of the candidate small peptides, sequence feature scores, conservation statistics and small peptide length mapping scores and other multi-source information are further introduced on the basis of the fusion prediction confidence, and a comprehensive score is obtained by weighting according to a preset weight, which is used for candidate small peptide ranking and Top-N result output in step S505.

[0058] In the above confidence metrics, the single model coding probability and the fusion prediction confidence are used for coding / non-coding binary classification determination of the candidate small peptide coding region, and the comprehensive ranking confidence is used for priority ranking and final result output of the candidate small peptides.

[0059] (4) Prediction output module construction Based on the fused prediction result, an encoding potential prediction output module is constructed for outputting small peptide encoding potential fusion prediction confidence or binary classification judgment result of the candidate small peptide encoding region. In this step, the output form can be selected according to the actual application requirements, including but not limited to: continuous type encoding potential fusion prediction confidence; encoding and non-encoding judgment label; or a combination of the two.

[0060] (5) Technical effect By introducing a machine learning model into the Mamba deep learning network and constructing a multi-model collaborative prediction framework, the present application has at least the following technical effects: fully exerting the advantages of the Mamba deep learning network in long sequence modeling and computational efficiency; enhancing the sensitivity to key discriminant features by using a machine learning model to improve prediction accuracy; reducing the risk of performance fluctuations of a single model under different data sets or application scenarios; and improving the stability and generalization ability of the algorithm in large-scale transcriptome data analysis.

[0061] S5, input the fusion prediction confidence and the encoding-non-encoding binary classification judgment label of each candidate small peptide encoding region, comprehensively judge the small peptide encoding potential of the corresponding lincRNA according to the judgment result, output no encoding potential conclusion if there is no, otherwise derive the small peptide amino acid sequence from the encoding positive region, and then calculate the comprehensive ranking confidence according to the fusion prediction confidence multi-class evaluation index and the fixed weight, sort the multiple positive candidate small peptides from high to low after weighting, and then put the candidate small peptides with comprehensive ranking confidence higher than the threshold into the high confidence set, and the rest into the low confidence set.

[0062] In step S5, the encoding potential of the lincRNA sequence is comprehensively judged according to the prediction result of step S4, and the corresponding small peptide candidate product information is output.

[0063] Specifically, it includes: encoding potential judgment: judging whether the lincRNA has small peptide encoding potential according to the model output result. Small peptide sequence determination: outputting the corresponding small peptide amino acid sequence for the candidate small peptide encoding region judged to have encoding potential. Candidate result screening and sorting: when there are multiple candidate small peptides corresponding to the same lincRNA, the candidate small peptides can be sorted or screened according to the prediction score, sequence feature or other evaluation index to obtain a high confidence small peptide set. The final output result includes the small peptide encoding potential prediction conclusion of the lincRNA and the corresponding small peptide candidate product information.

[0064] S505, when multiple candidate small peptide encoding regions correspond to the same lincRNA, the present application calculates the comprehensive fusion confidence of each candidate small peptide and outputs them in descending order of score; the comprehensive fusion prediction confidence is obtained by weighting one or more indicators such as fusion prediction confidence, encoding potential related sequence feature, sequence conservation and small peptide length.

[0065] The weight distribution operation criteria for candidate small peptide ranking are as follows: The comprehensive ranking confidence adopts a fixed weight: the fusion prediction confidence weight is 0.60, the coding potential related sequence feature (ORF integrity and triplet preference, etc.) weight is 0.15, the sequence conservation (candidate small peptide coding region conservation score statistics) weight is 0.15, and the small peptide length (mapped to 0-1 score according to the preset preference interval) weight is 0.10. If only part of the indicators are used, the weights of the unused indicators are proportionally distributed to the used indicators, and the total weight is kept at 1.

[0066] As a preferred embodiment, the comprehensive fusion prediction confidence adopts a fixed weight: the fusion prediction confidence weight is 0.60, the coding potential related sequence feature weight is 0.15, the sequence conservation weight is 0.15, and the small peptide length weight is 0.10; wherein the fusion prediction confidence is the coding probability / score output by the fusion model, the sequence feature score is the synthetic score of ORF integrity and triplet preference, etc. features, the conservation is the statistics of the candidate small peptide coding region conservation score, and the small peptide length is mapped to 0-1 score according to the preset preference interval; when only part of the indicators are used, the weights of the unused indicators are proportionally distributed to the used indicators and the total weight is kept at 1, thereby realizing the repeatable ranking of multiple candidate small peptides of the same lincRNA.

[0067] Step S6, construct a training dataset containing positive and negative samples with coding potential annotation, after performing the consistent processing procedures of S1-S3, train the Mamba deep learning network and at least one machine learning model, then train the multi-model fusion strategy to determine the prediction rule, and finally save the trained model parameters and multi-model fusion strategy for prediction analysis of unknown lincRNA sequences.

[0068] S6, when constructing the training dataset containing positive and negative samples and coding potential annotations, the present application uses "candidate coding region" as the basic training sample unit: each lincRNA sequence can generate one or more candidate coding regions through the candidate coding region identification in step S2, and each candidate coding region is independently assigned a binary classification label (coding potential = 1 / non-coding = 0). If necessary, the transcript level label can be further obtained by aggregating the rules such as "any candidate coding region reaching the threshold is judged as positive". The positive sample (coding potential = 1) is derived from at least one traceable coding evidence and can establish a corresponding relationship between the "candidate coding region and evidence", for example, the candidate coding region has translation evidence in ribosome binding / translation initiation related data, or its translation product is detected in mass spectrometry Corresponding peptide segment, or located in the translatable lincRNA / micropeptide coding region of the authoritative set / literature summary (homologous conservation and consistent coding signal can be used as a supplementary weak positive sample and record the evidence level); The negative sample (coding potential = 0) contains at least two types and preferably three types: one is the true background negative sample, that is, from the candidate coding region of the lincRNA which is annotated and lacks translation evidence; The second is the matching control negative sample, that is, for each positive sample, one or more non-coding candidate fragments similar in length and GC content (and optionally dinucleotide / triplet distribution) are constructed to avoid the model from distinguishing only by sequence composition; The third is the hard negative sample, that is, the candidate coding region with a long ORF or a coding style triplet preference in the non-coding background is selected, or a fragment is constructed from the coding sequence by frameshift / damage initiation and termination, etc. But keep the length and composition close, used to inhibit the shortcut learning of "long ORF = coding". After the training set is constructed, redundancy removal and leakage control are performed: divide the training / validation / test set by gene or transcript to ensure that multiple candidate coding regions generated by the same transcript do not appear across sets, and can be clustered and removed according to the sequence similarity threshold; At the same time, in view of the imbalance between positive and negative samples, undersampling / oversampling, class weight or focal loss, and stratified sampling strategies can be used to control the proportion of each type of sample. Each sample at least saves the transcript / gene identifier, candidate coding region sequence and its start and end position / link direction / reading frame, label, and evidence type and level (positive sample optional) and other metadata. Then strictly perform the sequence standardization and length processing, candidate coding region formatting confirmation and feature extraction process on each candidate coding region in the above training dataset consistent with S1-S3, train the Mamba deep learning network and at least one machine learning model respectively, and determine the prediction rule on the validation set. The training completed model parameters and multi-model fusion strategy are finally saved for coding potential prediction analysis of unknown lincRNA sequences.

[0069] The model parameters, fusion weights, threshold values and arbitration rules are stored in a callable form for persistent storage, which is used for the prediction output of unknown lincRNA sequences after performing consistent processing with S1-S3.

[0070] In step S6, the deep learning model and the machine learning model for predicting the small peptide coding potential of lincRNA are trained to obtain model parameters for subsequent prediction analysis.

[0071] (1) Training data set construction In this embodiment, the data used for model training is RNA sequence data or candidate small peptide coding region data with coding potential annotation.

[0072] The training data includes: RNA sequences or their corresponding candidate small peptide coding regions known or highly annotated as having small peptide coding ability; Non-coding RNA sequences or negative control candidate small peptide coding regions annotated as not having small peptide coding ability.

[0073] In one embodiment, the complete lincRNA sequence is constructed into a candidate small peptide coding region according to the method described in step S2, and the candidate small peptide coding region is used as the basic sample unit for model training, so that each lincRNA is converted into multiple training samples that can be used for learning.

[0074] (2) Training data preprocessing and feature encoding Before model training, the RNA sequences or candidate small peptide coding regions in the training data set are preprocessed and feature encoded in the same way as in the prediction stage, including: Sequence quality filtering and standardization processing; Sequence length adaptation and segmentation processing; Candidate small peptide coding region construction; Multi-dimensional sequence feature extraction and encoding.

[0075] By using the same data processing method as in the prediction stage, the data distribution consistency between model training and actual application scenarios is ensured.

[0076] (3) Deep sequence model training based on Mamba deep learning network In this embodiment, the multi-dimensional sequence features obtained in step (two) are input into the Mamba deep learning network based on the state space model, and the model parameters are trained.

[0077] During the training process, the encoding potential label of the candidate small peptide coding region is used as a supervision signal, the difference between the model output result and the true label is measured by a loss function, and the model weight is iteratively optimized by a parameter update mechanism.

[0078] During the training process, the model weight is determined by the training data and the supervision signal: the encoding potential label of the candidate small peptide coding region is used as a supervision signal, the difference between the model output result and the true label is measured by a preset loss function, and the parameter update amount is generated according to the difference to update the model weight in multiple rounds of iteration; during the iteration process, when the evaluation index of the validation set reaches the optimal or meets the preset stopping condition (such as early stopping condition), the model weight of the corresponding round is taken as the final weight and saved. For a machine learning model, the weight is automatically learned by its training algorithm during the training process (determined by multiple rounds of weak learners accumulation or decision rule splitting learning), and the model parameter at the time of optimal validation set or training completion is also saved as the final weight; when the training samples are imbalanced, class weights can be set, which are determined according to the proportion of positive and negative sample numbers and used to adjust the contribution of different class samples in the loss function, so as to obtain stable final model weight.

[0079] Through the above training process, the Mamba model can learn the long-range dependence features and potential translation patterns related to the small peptide coding potential in the candidate small peptide coding region, and output stable deep sequence representation.

[0080] (4) Machine learning model training After or at the same time as completing the Mamba deep learning network training, at least one machine learning model is introduced to model the discrimination of the candidate small peptide coding region. The machine learning model is at least one of random forest, gradient boosting decision tree, support vector machine, and logistic regression.

[0081] The input features of the machine learning model include at least one of the following: statistical features in multi-dimensional sequence features; deep sequence representation output by the Mamba model; combination representation of the above features.

[0082] The machine learning model is trained to learn the discrimination boundary between coding samples and non-coding samples to enhance the classification robustness of the overall model under different sequence structures and lengths.

[0083] (5) Multi-model fusion strategy training In this embodiment, the output results of the Mamba-based deep learning model and the machine learning model are fused to determine the final coding potential prediction strategy.

[0084] The fusion methods include but are not limited to: The prediction scores of different model outputs are combined by weighting; The results are integrated based on a voting or threshold mechanism; The multiple model outputs are input as new features into a higher-level fusion model for secondary training.

[0085] Through the above fusion training process, the influence of single model prediction bias on the final result is reduced, and the stability and generalization ability of the model in different data sets and application scenarios are improved.

[0086] (6) Training completion and model saving After the model training is completed, the Mamba deep learning network parameters, machine learning model parameters and multi-model fusion strategy obtained by training are saved for subsequent prediction analysis of the small peptide coding potential of unknown lincRNA sequences.

[0087] The embodiment of the application provides a lincRNA small peptide coding potential identification method based on a Mamba-machine learning fusion model.

[0088] In one preferred embodiment, the application provides a lincRNA small peptide coding potential identification method based on deep learning and machine learning fusion, which comprises the following steps (such as Figure 1 ).

[0089] S1, from the obtained transcript set data, screening out lincRNA full-length nucleotide sequences, after quality filtering, sequence standardization, length adaptation processing, outputting a standardized lincRNA sequence set with uniform format, length and coding mode; In this embodiment, the lincRNA full-length nucleotide sequences to be analyzed are obtained from public transcriptome databases or transcript sets obtained by high-throughput sequencing.

[0090] First, according to the gene annotation information, the transcript sequences annotated as long-chain intergenic non-coding RNA (lincRNA) are screened out, and protein-coding RNA, pseudogene transcripts and other non-target RNA types are excluded.

[0091] Subsequently, uniform preprocessing operations are performed on the lincRNA sequences, including: 1. Filtering sequences containing uncertain bases or abnormal characters in the sequence; 2. Uniformly processing the sequence direction and converting the base characters to a uniform coding format; 3. For lincRNA sequences with a length exceeding a preset threshold, a sliding window method is used to split them into multiple subsequences; 4. The length of the sequence is adapted by padding.

[0092] After the above processing, all input sequences are consistent in length, format and encoding method to meet the input requirements of subsequent models.

[0093] S2, without presetting the classical protein coding ability, using one or more combinations of multi-reading frame scanning, sequence local feature-based division, and covering window scanning, the potential coding fragment of each standardized lincRNA sequence is identified and scanned, and the candidate small peptide coding region set covering the potential small peptide translation possibility is obtained after integration; In this embodiment, without presetting whether the lincRNA has the classical protein coding ability, a plurality of candidate small peptide coding regions are constructed for the preprocessed sequence.

[0094] As shown in Figure 3 S3, for each candidate small peptide coding region, one or more of the following features are extracted: base level, codon or quasi-codon related, sequence complexity and information, context-related features, the extracted features are encoded in the form of vectors or tensors to form a model-learnable feature representation without manual setting of fixed thresholds, and a multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions is output; For each candidate small peptide coding region, multi-dimensional sequence feature encoding is performed in this embodiment, and the features include: 1. Base composition ratio and local arrangement pattern features in the candidate small peptide coding region; 2. Continuous base combination or quasi-codon usage related features; 3. Statistical features describing sequence complexity and information distribution characteristics; 4. Relative position information of the candidate small peptide coding region in the lincRNA full-length sequence and its upstream and downstream sequence features.

[0095] The above features are represented in the form of vectors or tensors and are used as model inputs, so that the model automatically learns the key feature patterns related to small peptide coding potential during the training process.

[0096] S4, input the multi-dimensional sequence feature representation set into the Mamba deep learning network to obtain a deep sequence representation vector, extract the related features, input at least one machine learning model to complete the coding potential discrimination, fuse the results of the two types of models, and construct an output module to output the coding potential prediction results, and mark the low-confidence candidate small peptide coding region for verification and return to the feature encoding; In this embodiment, the features obtained in step (three) are input into the Mamba deep learning network based on the state space model to model the sequence of the candidate small peptide coding region.

[0097] The Mamba model encodes the sequence features through a state space recursion mechanism to capture long-range dependencies and potential translation signals in the candidate small peptide coding region, and outputs a deep sequence representation vector while maintaining linear time complexity.

[0098] Meanwhile, the deep sequence representation is input into at least one machine learning model for discriminant modeling, and the machine learning model includes but is not limited to a random forest model, a gradient boosting decision tree model, or a support vector machine model.

[0099] Subsequently, the prediction results of the Mamba model and the machine learning model are fused to obtain the small peptide coding potential prediction result of the candidate small peptide coding region.

[0100] S402, after the multi-dimensional sequence features in step S3, the deep sequence representation vector, or the combined features of the two are spliced, the spliced features are input into a machine learning model as input features; S403, the input features are input into at least one machine learning model for coding potential discrimination, and the machine learning single model coding probability of each candidate region is output; the machine learning model is selected from one or more of random forest, gradient boosting decision tree, support vector machine, and logistic regression; S404, one or more of weighted combination prediction score, voting integration judgment result, and inputting the multi-model output as new features into a higher-level model for secondary prediction are used to fuse the results of the Mamba deep learning network and the machine learning model to obtain a fusion result; the fusion method is one or more of weighted combination prediction score, voting integration judgment result, and inputting the multi-model output as new features into a higher-level model for secondary prediction; S405, according to the fusion result, an output module is constructed to output the continuous coding potential score of the candidate small peptide coding region, the coding-non-coding binary classification judgment label, or the coding potential fusion prediction confidence of the combination of the two; and the candidate small peptide coding region with a prediction fusion prediction confidence lower than a threshold value is marked as to be verified and returned for re-feature encoding.

[0101] S5, after inputting the step S4 prediction result with confidence, the small peptide coding potential of the lincRNA is comprehensively judged, and if there is no coding potential, the conclusion is directly output, and if there is coding potential, the small peptide amino acid sequence is derived, after sorting and screening high-confidence small peptides, at least one of the lincRNA coding potential conclusion, the corresponding small peptide amino acid sequence, the corresponding position, and the prediction score is output. In this embodiment, according to the fused prediction result, the small peptide coding potential of each lincRNA is comprehensively judged.

[0102] For the candidate small peptide coding region determined to have coding potential, the corresponding small peptide amino acid sequence is output; when the same lincRNA corresponds to multiple candidate small peptides, they are sorted according to the prediction score to obtain a high-confidence small peptide candidate set.

[0103] The final output result includes: The small peptide coding potential determination result of the lincRNA, and the corresponding small peptide candidate product information.

[0104] As shown in Figure 2 Step S1, obtaining a standardized lincRNA sequence set with uniform format, length and coding mode includes: Step S101, obtaining original transcript set data; Specifically, the original transcript set data is obtained from the transcript set obtained from the public transcriptome database, gene annotation database or high-throughput sequencing data.

[0105] Step S102: obtaining a lincRNA full-length nucleotide sequence set from the original transcript set data according to the gene annotation information; Step S103: quality filtering of the lincRNA full-length nucleotide sequence set to obtain a filtered lincRNA full-length nucleotide sequence set; Step S104: standardizing the filtered lincRNA full-length nucleotide sequence set, unifying the sequence direction and converting the base characters to a uniform coding format to obtain a lincRNA sequence set with uniform format; Step S105, according to the target input length threshold, the lincRNA sequence set with uniform format is classified and processed according to the actual length of the sequence, the long sequence exceeding the threshold is split by sliding window, the short sequence shorter than the threshold is filled according to the fixed alignment rule, the lincRNA sequence is converted into a fixed length subsequence input unit and the start and end position index in the original sequence is retained, and the standardized lincRNA sequence set with uniform format, length and coding mode is output.

[0106] Specifically, first set the target input length threshold (i.e. the fixed input length expected by the model, preferably 512, 1024 or 2048 nt (nucleotides), consistent between training and prediction stages), for all sequences in the lincRNA sequence set with uniform format, length adaptation processing is carried out according to the actual length, long sequences exceeding the threshold are split by sliding window, short sequences shorter than the threshold are filled by fixed rule, effective length markers are generated and subsequence position indexes are retained, and finally the standardized lincRNA sequence set is output to obtain length-adapted lincRNA subsequence units, the specific processing rules are as follows: Long sequence processing beyond threshold: when the sequence length exceeds the target input length threshold, perform sliding window splitting from the 5' end of the sequence with a window length = target input length threshold, continuously intercepting subsequences; the window moving step is a preset value, preferably 1 / 2 of the window length (forming overlapping coverage) or equal to the window length (no overlapping division), ensuring complete coverage of long sequence information without loss. For the residual fragments at the end of the sliding window splitting, at least one end rule processing is adopted: one is end padding, which fills the 3' end of the fragment to the target input length threshold; the other is end alignment cutting, which cuts a window with a length equal to the target input length threshold from the end of the original sequence to the front, or the two rules can be combined (additional end alignment window after regular sliding window), to ensure that the end information of the sequence is complete.

[0107] Short sequence processing shorter than threshold: when the sequence length does not exceed the target input length threshold, keep its original 5'→3' order unchanged, put it into a fixed length container according to the unified alignment rule (preferably 5' alignment or 3' alignment, consistent between training and prediction stages), and fill the part that is not enough for the target length with a preset filler symbol at the other end (such as "N" and other unknown base placeholders, PAD labels and other special filler symbols).

[0108] To avoid interference of the filled area on subsequent feature statistics and deep network modeling, effective length labels (masks) are generated for all filled sequences to accurately identify real base positions and filled positions, so that subsequent base composition calculation, k-mer statistics, and deep model attention / state updating steps can directly ignore the filled area.

[0109] Through the above operations, the lincRNA sequence of any length is uniformly converted into one or more fixed-length subsequence input units, and the start and end position indexes of each subsequence in the original sequence are preserved throughout the process, providing a basis for subsequent candidate small peptide coding region positioning, result backtracking, and multi-subsequence output.

[0110] As Figure 3As shown, the Mamba of step S4 of the present application cooperates with the machine learning prediction process to include: taking a multi-dimensional sequence feature representation set as input, introducing a Mamba deep learning network based on a state space model to model the sequence to capture long-range dependencies in the candidate small peptide coding region and output a deep sequence representation vector; on this basis, the deep sequence representation is fused with statistical features to construct machine learning discriminant features, and input into at least one machine learning model to obtain a preliminary discriminant result of the coding potential; then, the candidate small peptide coding region is screened or re-encoded according to the prediction confidence or threshold condition; finally, the small peptide coding potential judgment result of the candidate small peptide coding region is output, including the binary classification label of the coding potential and the candidate small peptide coding region prediction result with confidence.

[0111] Step S2: The identified potential coding fragments from the standardized lincRNA sequence set are integrated, de-duplicated and merged to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility; Specifically, for each standardized lincRNA sequence in the standardized lincRNA sequence set, without pre-setting its classic protein coding ability, one or more combinations of the following three methods are used to identify potential coding fragments: multi-reading frame scanning, division based on sequence local features, and covering window scanning. The construction method of the candidate small peptide coding region is not limited to a single rule, and the specific implementation rules of each identification method are as follows: Multi-reading frame scanning method: The lincRNA sequence is scanned in different reading frames, and short open reading frames with a length corresponding to ≥20 amino acids are identified to capture potential small peptide coding segments in the sequence; Division based on sequence local features: According to the arrangement features, composition characteristics or statistical feature changes of local bases in the sequence, the lincRNA sequence is divided into multiple sequence fragments with potential coding possibility.

[0112] Covering window scanning method: Fixed window or variable window scanning strategy is used to continuously and uninterruptedly scan the entire lincRNA sequence, so that any position fragment in the sequence is included in the identification range, ensuring that no potential coding fragment is missed.

[0113] By constructing the candidate small peptide coding region through the above-mentioned multiple methods, it can effectively avoid missing small peptide coding events produced under non-classical initiation, non-standard length or special translation mode when relying only on a single open reading frame rule for identification, and ensure comprehensive coverage of the potential small peptide translation possibility of the candidate small peptide coding region. After completing the potential coding fragment identification of all standardized lincRNA sequences, all identified potential coding fragments are integrated, de-duplicated and merged to finally obtain a candidate small peptide coding region set.

[0114] As shown in Figure 4 Step S201, input multi-dimensional sequence feature representation set, input the multi-dimensional sequence features of the candidate small peptide coding region as the model input.

[0115] Step S202, Mamba deep sequence modeling, input the features into the Mamba model to capture long-range dependencies, and output a deep representation vector.

[0116] Step S203, feature fusion, fuse the statistical features and the Mamba model representation to construct machine learning input features.

[0117] Step S204, machine learning discriminant modeling, input the fused features into at least one machine learning model, and output a discriminant result.

[0118] Step S205, prediction confidence judgment, according to the prediction confidence or threshold.

[0119] Step S205, prediction confidence judgment, according to the prediction confidence or threshold. In the embodiment, the prediction confidence is calculated from the prediction score output by the fusion model, specifically: first, obtain the prediction probability score of the same candidate small peptide coding region by the Mamba model and the machine learning model respectively, then fuse according to the weighting combination rule of step S404 to obtain the fused prediction probability as the prediction confidence of the candidate small peptide coding region; when using the voting arbitration rule, if the two models are consistent, the fused prediction probability is taken as the confidence, if the two models are inconsistent, the probability of the model with the greater deviation from the respective fused decision threshold is taken as the confidence, so as to reflect the reliability of the prediction result of the candidate small peptide coding region. The threshold is used to convert the prediction confidence into a binary classification judgment result, and the threshold is set according to the verification set optimization: after training, the threshold is searched on the verification set, and the threshold that makes the preset evaluation index optimal is selected and fixedly saved, wherein the preset evaluation index includes one or more of accuracy, F1 value, sensitivity / specificity balance or Youden index; when there is sample imbalance or stronger emphasis on reducing false positives, the threshold that makes the precision or specificity reach the preset target can be selected preferentially. The reference range of the threshold is 0.40-0.70, and the preferred range is 0.50-0.60; when used for "high confidence candidate" screening, a high confidence threshold can be further set, the reference range is 0.70-0.90 (for example, 0.75 or 0.80), and the candidate small peptide coding region that reaches the high confidence threshold is taken as the small peptide candidate set for priority output.

[0120] Step S206, candidate small peptide coding region labeling and recoding, label low confidence samples as to be verified or re-feature encoding.

[0121] Step S207, output the prediction result, output the binary classification label of the coding potential and the candidate small peptide coding region prediction result with confidence.

[0122] S5, prediction result output and small peptide candidate product determination process.

[0123] As shown in Figure 5 The prediction result output and small peptide candidate product determination process of step S5 of the present application includes: based on the candidate small peptide coding region prediction result with confidence, comprehensively determining the overall small peptide coding potential of lincRNA; when it is determined that there is no small peptide coding potential, directly outputting the conclusion that there is no small peptide coding potential; when it is determined that there is coding potential, performing translation deduction on the candidate small peptide coding region to obtain the corresponding small peptide amino acid sequence, and combining the prediction confidence, sequence conservation and length and other indicators to sort and screen the candidate small peptides, and finally outputting the high-confidence small peptide candidate product and its corresponding coding potential information.

[0124] Step S501, input the candidate small peptide coding region prediction result.

[0125] The fusion confidence of each candidate small peptide coding region and / or the coding-non-coding binary classification determination label obtained from the foregoing steps are input.

[0126] Step S502, comprehensively determine the overall coding potential of lincRNA.

[0127] Based on the prediction result of the candidate small peptide coding region, it is comprehensively determined whether the corresponding lincRNA has small peptide coding potential.

[0128] Step S503, no coding potential determination output.

[0129] When the comprehensive determination result is that there is no high-confidence candidate small peptide coding region, the conclusion of "no small peptide coding potential" is output and the process is ended.

[0130] Step S504, small peptide sequence deduction.

[0131] When it is determined that the lincRNA has small peptide coding potential, the candidate small peptide coding region determined to be coding type is subjected to translation deduction to obtain the corresponding small peptide amino acid sequence.

[0132] Step S505, candidate small peptide sorting.

[0133] According to at least one of the prediction confidence, sequence conservation, and small peptide length, the candidate small peptides are weighted and sorted.

[0134] Step S506, high-confidence candidate small peptide screening.

[0135] Filtering from the sorting results to obtain a high-confidence small peptide candidate set.

[0136] Step S507, result output.

[0137] Output the coding potential conclusion of lincRNA and the high-confidence small peptide candidate product information, including small peptide sequence, corresponding position and prediction score.

[0138] Example 1 Unknown lincRNA candidate small peptide coding region prediction, confidence determination and candidate small peptide output In this embodiment, an unknown lincRNA full-length nucleotide sequence LINC_Test_001 is input, the sequence length is 1800nt, the sequence standardization, candidate small peptide coding region construction and feature extraction are completed according to steps S1-S3, and the candidate small peptide and the sorting result are output according to steps S404, S205 and S5.

[0139] (1) Step S105 (length adaptation, fixed 1024nt input) In this embodiment, the target input length threshold of the model is fixed to 1024nt. When the input sequence length exceeds 1024nt, sliding window splitting is used, the window length is fixed to 1024nt, the step is fixed to 512nt, and the end alignment is used to supplement the end window to cover the tail information of the sequence. Thus, LINC_Test_001 generates the following fixed length subsequence windows: W1 (1-1024nt), W2 (513-1536nt), W3 (777-1800nt).

[0140] (2) Step S2 (candidate small peptide coding region construction) Multiple reading frame scanning and coverage scanning are performed on W1, W2 and W3 respectively to generate a candidate small peptide coding region set, and duplicate candidates are merged and removed (only one of the same or highly overlapping candidates is retained, and its coordinates in the full-length sequence are retained for backtracking). In this embodiment, three candidate small peptide coding regions are obtained as follows (coordinates are full-length coordinates, and the length and small peptide length corresponding relationship satisfies the integer division): C1: full-length 600-824nt, length 225nt, corresponding small peptide length 75aa (amino acid), starting with CTG (non-canonical start); C2: full-length 210-299nt, length 90nt, corresponding small peptide length 30aa (amino acid), starting with ATG (canonical start); C3: full-length 1200-1379nt, length 180nt, corresponding small peptide length 60aa (amino acid), starting with GTG (non-canonical start).

[0141] Wherein, C1, C2 fall within the coverage range of window W1, C3 falls within the coverage range of window W2 / W3; the candidate small peptide coding region set is obtained by the above scanning and deduplication rules, and can cover non-canonical initiation and potential small peptide coding events under different length ranges.

[0142] (3) Step S3 (feature extraction and context construction, matching 1024 input) For each candidate small peptide coding region, sequence features are extracted, and context association features are constructed. The upstream and downstream windows in this embodiment are fixed at: upstream 128 nt, downstream 128 nt; the insufficient part is filled.

[0143] For Mamba model: for each candidate small peptide coding region, a fixed length 1024 nt input fragment is constructed, the construction rule is "upstream 128 nt + candidate small peptide coding region body + downstream 128 nt", and the remaining length part is filled according to the rule and ignored by the effective length marker, so that the input length is constant 1024 nt; if the candidate small peptide coding region body is too long to be accommodated, the candidate small peptide coding region body is first split according to the sliding window rule of S105, then multiple 1024 nt input fragments are constructed and output.

[0144] For machine learning model (RF): the same type of statistical features (including base composition, k-mer / quasi-codon frequency, complexity / information entropy, etc.) are extracted for the candidate small peptide coding region, the upstream window and the downstream window, respectively, and then the start and end relative position features of the candidate small peptide coding region in the full length are spliced to form a unified feature vector as input.

[0145] (4) Step S404 (double model prediction and fusion, fixed weight and threshold) The trained and saved Mamba model and RF model are called to output the coding probability of each candidate small peptide coding region, and a weighted combination prediction score is used for fusion, with the fusion weight fixed as: Mamba=0.60, RF=0.40; the fusion output probability is used as the prediction confidence of the candidate small peptide coding region.

[0146] (5) Step S205 (prediction confidence judgment and threshold) The fusion judgment threshold is fixed at 0.50, when the fusion prediction confidence of the candidate small peptide coding region is ≥0.50, it is judged as coding potential positive, otherwise it is judged as negative; at the same time, the voting arbitration rule is fixed and saved as a backup: when the two model binary classification results are inconsistent, the label corresponding to the one with relatively larger deviation from the threshold is selected as the final determination.

[0147] (6) Step S5 (candidate small peptide output and sorting) Step S505: Small peptide deduction and output of candidate list; when there are multiple positive candidates for the same lincRNA, sort and output according to the fixed weight rule of step S505, and output the Top-N candidate small peptides and their corresponding full-length coordinates and confidence.

[0148] (7) Output field template (example output format) The system output of this embodiment includes at least: lincRNA ID, candidate number, full-length coordinate, small peptide sequence, small peptide length, fusion prediction confidence (coding probability score), ranking position; for example, as shown in the following table (the probability value is an example output, used to illustrate the process and field, not limited to specific values), as shown in Table 1 below: Table 1: Candidate small peptide prediction results of LINC_Test_001 Step S6: Model training and fusion strategy determination 1. Implementation example of voting integration determination result and secondary prediction of upper model Example 1: Voting integration determination result In this embodiment, the voting integration determination result is obtained by fusing the prediction results of the Mamba model and the machine learning model to improve the final prediction accuracy. The specific steps are as follows: Step 1.1: Model output determination For each candidate region, the coding potential probability is calculated by the Mamba model and the machine learning model respectively. If the prediction probability is greater than or equal to the threshold value (threshold value = 0.5), it is determined as "coding potential", otherwise it is determined as "non-coding".

[0149] Step 1.2: Voting integration determination When the prediction results of the two models are consistent, the prediction result is directly output as the final determination. If the prediction results of the two models are inconsistent, arbitration is performed in the following way: Calculate the deviation of the output probability of the Mamba model and the machine learning model from 0.50, and select the label corresponding to the larger deviation as the final result.

[0150] Output the final determination label and use it as the prediction result of the candidate small peptide coding region.

[0151] Example 2: Secondary prediction of upper model (Stacking / Meta-model) In this embodiment, the outputs of multiple base models are input as new features into an upper model for secondary prediction to optimize the prediction boundary.

[0152] Step 2.1: Constructing the upper-layer input features The probability values, binary classification labels, and confidence bias output by the Mamba model and the machine learning model are used as input features for the upper-layer model. The feature vector is formed as follows: z = (pMamba, pML, yMamba, yML, dMamba, dML, stat) where: pMamba, pML are the predicted probabilities; yMamba, yML are the predicted labels; dMamba, dML are the confidence bias; stat is the supplementary enhanced features (e.g. GC content, sequence length of the candidate small peptide region, etc.).

[0153] Step 2.2: Training the upper-layer model The upper-layer model (such as logistic regression, SVM, random forest, etc.) is trained using the training data to minimize the classification loss or maximize the AUC / F1 performance indicators. After training, the model parameters are saved.

[0154] Step 2.3: Inference output The upper-layer model receives new prediction input (such as the output of the Mamba model and the machine learning model) and makes a second prediction, outputting the fused prediction probability pMeta as the final decision result.

[0155] 2. Define the final vector length and tensor dimension arrangement after multi-feature splicing 2.1 Final vector length after multi-feature splicing In this embodiment, multiple features (such as base-level, quasi-codon frequency, complexity features, etc.) are spliced into a unified vector. The dimension of each feature is as follows: Base-level features: A / C / G / T ratio + GC content, 5 dimensions in total.

[0156] Quasi-codon (3-mer) frequency: 64 normalized frequencies of triplets, 64 dimensions in total.

[0157] Complexity and information: low complexity ratio 1 dimension + repetition 1 dimension + information entropy 1 dimension + triplet information entropy 1 dimension, 4 dimensions in total.

[0158] Position features: starting position and ending position of the candidate region, 2 dimensions in total.

[0159] Therefore, the vector dimension of each candidate region is: 5 + 64 + 4 + 2 = 75 dimensions.

[0160] Taking into account the splicing of candidate regions, upstream windows and downstream windows, the final vector length is: 75 (candidate region) + 75 (upstream window) + 75 (downstream window) = 225 dimensions.

[0161] 2.2 Tensor dimension arrangement For the Mamba model, a fixed input length of 1024 nt is adopted. The tensor dimensions of the Mamba model are: batch_size x 1024 x feature dimension.

[0162] Feature dimension: There are 5 channels for each nt position (one-hot encoding of A / C / G / T and effective length marker), so the feature dimension is 5.

[0163] Tensor dimension: Assuming that each batch (batch_size) contains 32 samples, the final tensor dimension is: 32 x 1024 x 5.

[0164] 3. Performance difference between "the invention and Mamba alone" Compared with using only the Mamba model, the invention has the following advantages: Accuracy improvement: Through multi-model fusion, the advantages of each model can be integrated to improve the prediction accuracy of coding potential. On the validation set, the accuracy of the invention is improved by 3-5 percentage points compared with using only the Mamba model.

[0165] Long sequence processing efficiency: Although the Mamba model has advantages in modeling long sequences, the invention makes supplementary predictions at the candidate region level through machine learning models, further improving the accuracy of long sequence processing while maintaining high efficiency.

[0166] Fusion advantage: Multi-model fusion makes the advantages of each model complementary, thereby improving the overall robustness and performance.

[0167] 4. Differences from traditional multi-model fusion schemes The invention has the following significant differences from traditional multi-model fusion schemes (such as no fixed weights or no low-confidence re-encoding): Standardized feature extraction: The invention performs standardized processing on input feature extraction, ensuring that sequences of different lengths and sources can be compared in the same feature space, avoiding the inconsistency of features in traditional schemes.

[0168] Fixed weight ordering: Unlike the dynamic weights in traditional schemes that may change with batches or datasets, the invention determines and fixes the weights through validation set optimization, making the fusion results more stable and reproducible.

[0169] Low-confidence re-encoding: Traditional schemes often handle low-confidence samples simply, while the invention returns low-confidence samples to the feature encoding stage through a low-confidence fallback mechanism to trigger more feature extraction and stricter judgment, thereby reducing the risk of misjudgment.

[0170] 5. Applicable scenarios for "512 / 1024 / 2048 nt" and learning rate adjustment logic 5.1 Applicable scenarios for input length 512 nt: Suitable for short sequences, especially when the candidate region is short. Also suitable for scenarios where fast processing is required when the data volume is large.

[0171] 1024 nt: As a standard length, it is suitable for most data sets and can provide sufficient context information while maintaining computational efficiency.

[0172] 2048 nt: Suitable for long sequences or sequences with strong upstream and downstream dependencies, which can reduce the number of sliding windows and improve processing efficiency.

[0173] 5.2 Learning rate adjustment logic Low learning rate (1×10^-4 ~ 2×10^-4): When the data volume is small or the sample is not balanced, using a lower learning rate can improve the stability of the model and avoid overfitting.

[0174] Medium learning rate (2×10^-4 ~ 3×10^-4): When the data volume is moderate and there is no significant sample imbalance, using a medium learning rate can speed up model convergence.

[0175] High learning rate (3×10^-4 ~ 5×10^-4): When the data volume is large and the data is relatively balanced, using a higher learning rate can speed up the training process and reduce the convergence time.

[0176] 5.3 Applicable range of other parameters batch_size: The recommended range is 16–128. When the video memory or memory is limited, it is suitable to use a smaller batch_size (e.g., 16–32); when resources are sufficient, a larger batch_size (e.g., 64–128) can be used.

[0177] dropout: The recommended range is 0.05–0.30. When the data volume is small or the model is complex, a larger dropout (0.15–0.30) can be used; when the data volume is large, a smaller dropout (0.05–0.15) can be used.

[0178] Step S6 is the model training and saving process. By constructing a labeled training dataset, performing consistent preprocessing and feature encoding in the prediction stage, training Mamba deep learning network and machine learning model respectively, optimizing multi-model fusion strategy, finally saving reusable model parameters and rules for unknown lincRNA sequence coding potential prediction, the specific sub-steps are as follows: S601, Constructing training dataset with coding potential annotation; Taking "candidate small peptide coding region" as the basic training sample unit, after identifying the candidate region of each lincRNA sequence by step S2, one or more candidate samples are generated, each sample is independently assigned a binary classification label (coding potential = 1 / non-coding = 0), and if necessary, the transcript level label can be summarized as "any candidate region meets the standard, the transcript is positive".

[0179] Positive samples (coding potential = 1): need to have traceable coding evidence, including translation evidence of candidate regions in ribosome binding / translation initiation data, detection of corresponding peptide segments in mass spectrometry, or located in the authoritative translatable lincRNA / micropeptide coding region set (homologous conservation and consistent coding signal can be used as weak positive samples and record evidence level); Negative samples (coding potential = 0): at least three types, the first is the true background negative sample annotated as lincRNA without translation evidence; the second is the matching control negative sample with similar length, GC content (optional dinucleotide / triplet distribution) as the positive sample; the third is the difficult negative sample with longer ORF or triplet preference of coding style (or fragments constructed by coding sequence with frame shift / damaged start and stop), used to inhibit the shortcut learning of "long ORF = coding"; Dataset processing: divide the training set (70%), validation set (20%), and test set (10%) by gene / transcript, ensure that the candidate regions of the same transcript do not cross the set, and cluster and de-duplicate by sequence similarity threshold; use class weight, focal loss or stratified sampling to balance the positive and negative sample ratio; save transcript / gene identification, candidate region sequence and start and end position / link direction / reading frame, label, evidence type and level (positive sample) and other metadata for each sample.

[0180] S602, Preprocessing and feature encoding of training data Perform the same processing flow as steps S1-S3 on the training dataset to ensure consistency of data distribution in the training and prediction stages: Sequence preprocessing: perform quality filtering, standardization processing, and length adaptation (long sequence sliding window splitting, short sequence padding and effective length label generation) according to the target input length threshold of 1024nt; Candidate region construction: use a combination of multi-reading frame scanning, local feature-based division, and overlapping window scanning to identify short open reading frames corresponding to ≥20 amino acids, and integrate and de-duplicate to form a candidate small peptide coding region set; Multi-dimensional feature encoding: extract base-level (base composition, k-mer frequency), codon / quasi-codon (64 normalized frequencies of triplets), sequence complexity and information (low complexity proportion, information entropy), context association (upstream 128 nt + candidate region + downstream 128 nt) features, encode into Mamba adapted tensor form and machine learning model adapted vector form, output multi-dimensional sequence feature representation set.

[0181] S603, Mamba deep learning network training; Input the tensor features output by S602 into the Mamba deep learning network, and use the encoded potential of the candidate region as the supervision signal to iteratively optimize the model parameters: Model fixed parameters: embedding dimension 256, network layer number 6 layers, hidden dimension 512, dropout=0.1, output layer is a binary classification fully connected layer (Sigmoid activation, output encoding probability 0-1); Training hyperparameters: optimizer AdamW, learning rate 3x10 -4 , batch_size=64, maximum training rounds 50, loss function is binary cross-entropy with class weights, early stopping strategy is to stop when the validation set AUC does not improve for 10 consecutive epochs; Training process: input features by batch, capture long-range dependencies through state space model hidden state continuous evolution, output deep sequence representation vector (512 dimensions), calculate prediction probability and real label loss and update parameters by back propagation, save the model weight with the best validation set AUC.

[0182] S604, machine learning model training; Take the statistical features of S602 and the deep sequence representation vector of S603 or their combination as input, and train at least one machine learning model: Model selection: choose random forest (RF), parameters fixed as: number of trees in forest n_estimators=100, maximum depth max_depth=10, node splitting criterion criterion="gini", minimum number of samples in leaf node min_samples_leaf=1, maximum number of features considered when splitting nodes max_features="sqrt"; Training process: after standardizing the input features by Z-score, build a set of independent decision trees through random forest to form an ensemble model, train the training set and calculate the F1 value on the validation set, save the model file after training is completed, and output the feature importance ranking result at the same time.

[0183] S605: multi-model fusion strategy training; Optimize the fusion rule of the Mamba model and the machine learning model on the validation set to determine the final prediction strategy: Fusion method: adopt a weighted combination prediction score as the main fusion method, supplemented by a voting arbitration rule; Weight determination: traverse the Mamba weight (0-1) with a step size of 0.05, calculate the validation set AUC of the fusion score under different weights, and select the optimal weight (in this embodiment, Mamba=0.60, RF=0.40); Threshold determination: traverse the threshold (0.40-0.70, step size 0.01), select the threshold with the maximum Youden index (in this embodiment, 0.50) as the coding / non-coding binary classification criterion; Arbitration rule: when the two models make inconsistent judgments, the "arbitration by the one with larger deviation from the threshold" is adopted, that is, the absolute deviation of the single model probability from the threshold 0.50 is compared, and the model with larger deviation is used as the final result.

[0184] S606: Save the model and the fusion strategy; Save the model parameters and the fusion strategy in a callable form, and form a complete model package: Save the content: Mamba model weight file (.pth) and configuration file (.yaml), RF model file (.model) and feature standardization parameters, fusion weight and threshold configuration file, voting arbitration rule script, feature extraction and encoding script, and result output template; Application scenario: after the unknown lincRNA sequence is preprocessed by S1-S3, the model package can be directly called to output the fusion prediction confidence, binary classification judgment label and high-confidence small peptide set, realizing end-to-end coding potential identification.

[0185] In the embodiment of the application, the input sequence is stored in the computer system in the form of electronic data, and the prediction result is output in the form of a data file or a visualized result. The system can be deployed in a general-purpose computer device, a server or a cloud computing environment.

[0186] Figures 5-10 The structure diagram of the lincRNA small peptide coding potential identification system provided in the embodiment of the application.

[0187] The lincRNA small peptide coding potential identification system in the embodiment of the application comprises: A standardized lincRNA sequence acquisition module 1 for acquiring a set of standardized lincRNA sequences with unified format, length and coding mode; The specific implementation process of the standardized lincRNA sequence acquisition module 1 can be referred to the related description of the aforementioned step S1.

[0188] The standardized lincRNA sequence acquisition module 1 comprises: a raw transcript data acquisition module 101, a lincRNA sequence screening module 102, a lincRNA sequence quality filtering module 103, a lincRNA sequence standardization processing module 104, and a lincRNA sequence length adaptation and index reservation module 105. The specific implementation process of each module can be referred to the related description of the foregoing steps S101, S102, S103, S104, and S105.

[0189] The candidate small peptide coding region construction module 2 is used for integrating, deduplicating, and merging all potential coding fragments obtained from the standardized lincRNA sequence set to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility. The multi-dimensional sequence feature encoding module 3 is used for extracting multi-class multi-dimensional sequence features of each candidate small peptide coding region in the candidate small peptide coding region set, encoding the multi-class multi-dimensional sequence features into a vector or tensor form feature representation that can be learned by a model, and outputting a corresponding multi-dimensional sequence feature representation set. The multi-dimensional sequence feature encoding module 3 comprises: a multi-dimensional sequence feature extraction module 301, a sequence feature vector tensor encoding module 302, and a multi-dimensional sequence feature representation set output module 303. The specific implementation process of each module can be referred to the related description of the foregoing steps S301, S302, and S303.

[0190] The coding potential fusion prediction module 4 is used for inputting the multi-dimensional sequence feature representation set of the candidate small peptide coding region into a trained Mamba deep learning network, outputting a deep sequence representation vector and a Mamba single model coding probability, inputting the multi-dimensional sequence features, the deep sequence representation vector, or a combination of the two into at least one trained machine learning model to output a machine learning single model coding probability, subsequently linearly summing the fusion prediction confidence according to a preset weight, outputting the numerical value and / or an encoding-non-encoding binary classification judgment label, and finally marking the candidate region with a fusion prediction confidence lower than a threshold as a to-be-verified sample and returning to step S3 for re-feature encoding. The coding potential fusion prediction module 4 comprises: a Mamba model feature modeling and coding probability output module 401, a machine learning model coding potential discrimination module 402, a multi-model coding probability fusion and confidence calculation module 403, a fusion result numerical value and binary classification label output module 404, and a low-confidence sample marking and feature encoding backoff module 405. The specific implementation process of each module can be referred to the related description of the foregoing steps S401, S402, S403, S404, and S405.

[0191] The small peptide coding potential determination and screening module 5 is used for inputting the fusion prediction confidence of each candidate small peptide coding region and the coding-non-coding binary classification determination label, comprehensively determining the small peptide coding potential of the corresponding lincRNA according to the determination result, outputting a non-coding potential conclusion if there is no coding potential, deriving a small peptide amino acid sequence from the coding positive region if there is coding potential, combining the fusion prediction confidence multi-class evaluation index to calculate a comprehensive ranking confidence according to a fixed weight, ranking the multiple positive candidate small peptides from high to low after weighting, and classifying the candidate small peptides with a comprehensive ranking confidence higher than a threshold into a high confidence set and the rest into a low confidence set.

[0192] The small peptide coding potential determination and screening module 5 includes a fusion prediction result input module 501, a lincRNA coding potential comprehensive determination module 502, a non-coding potential result output module 503, a coding positive region amino acid sequence derivation module 504, a positive candidate small peptide weighted ranking and confidence calculation module 505, a high-low confidence candidate small peptide classification module 506, and a coding potential conclusion and high confidence small peptide information output module 507. The specific implementation process of each module can be referred to the related description of the foregoing steps S501, S502, S503, S504, S505, S506 and S507.

Claims

1. A method for identifying lincRNA small peptide coding potential, characterized in that, The method comprises the following steps: S1, obtaining a standardized lincRNA sequence set with unified format, length and coding mode; S2, integrating, deduplicating and merging all potential coding fragments identified from the standardized lincRNA sequence set to obtain a candidate small peptide coding region set covering the potential small peptide translation possibility; S3, extracting multi-class multi-dimensional sequence features from each candidate small peptide coding region in the candidate small peptide coding region set, encoding the multi-dimensional sequence features into a vector or tensor form feature representation that can be learned by a model, and outputting a corresponding multi-dimensional sequence feature representation set; S4, inputting the multi-dimensional sequence feature representation set of the candidate small peptide coding region into a trained Mamba deep learning network to output a deep sequence representation vector and a Mamba single model coding probability, then inputting the multi-dimensional sequence features, the deep sequence representation vector or a combination of the two into at least one trained machine learning model to output a machine learning single model coding probability, subsequently linearly summing the fusion prediction confidence according to a preset weight, outputting the numerical value and / or an encoding-non-encoding binary classification judgment label, and finally marking the candidate region with a fusion prediction confidence lower than a threshold value as a to-be-verified sample and returning to step S3 for re-feature encoding; S5, inputting the fusion prediction confidence and the encoding-non-encoding binary classification judgment label of each candidate small peptide coding region, comprehensively judging the small peptide coding potential of the corresponding lincRNA according to the judgment result, outputting a no-coding-potential conclusion if there is no small peptide coding potential, deriving a small peptide amino acid sequence from the coding-positive region if there is a small peptide coding potential, calculating a comprehensive ranking confidence according to a fixed weight based on the fusion prediction confidence and a plurality of evaluation indexes, ranking a plurality of positive candidate small peptides from high to low after weighting, and classifying the candidate small peptides with a comprehensive ranking confidence higher than a threshold value into a high-confidence set and the rest into a low-confidence set.

2. The method of claim 1, wherein, The sub-steps of step S1 include: S101, obtaining original transcript set data; S102, screening lincRNA full-length nucleotide sequence set from the original transcript set data according to gene annotation information; S103, performing quality filtering on the lincRNA full-length nucleotide sequence set to obtain a filtered lincRNA full-length nucleotide sequence set; S104, performing standardization processing on the filtered lincRNA full-length nucleotide sequence set, unifying sequence direction and converting base characters into a unified coding format to obtain a lincRNA sequence set with unified format; S105, according to a target input length threshold, classifying and processing the lincRNA sequence set with unified format according to actual sequence length, splitting long sequences exceeding the threshold using a sliding window, filling short sequences shorter than the threshold according to a fixed alignment rule, converting the lincRNA sequence into a fixed-length subsequence input unit and retaining its start and end position index in the original sequence, and outputting a standardized lincRNA sequence set with unified format, length and coding mode.

3. The method of claim 1, wherein, The sub-steps of step S3 include: S301, for each candidate small peptide coding region in the candidate small peptide coding region set, extract multi-dimensional sequence features, and obtain the multi-dimensional sequence features including one or more of the following: base level, codon or quasi-codon, sequence complexity and information, context association features; S302, encode the extracted multi-dimensional sequence features in the form of a vector or a tensor to form a model learnable feature representation without manual setting of a fixed threshold; S303, output the multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions.

4. The method of claim 1, wherein, The sub-steps of step S4 include: S401, input the multi-dimensional sequence feature representation set of the candidate small peptide coding region into the trained Mamba deep learning network, capture the long-range dependence relationship and non-local sequence pattern of each candidate region through the continuous evolution of the hidden state of the state space model, and output the deep sequence representation vector and the Mamba single model coding probability of each candidate region; S402, splice the multi-dimensional sequence features, the deep sequence representation vector, and the combined features of the two as input features, input them into at least one trained machine learning model for coding potential discrimination, and output the machine learning single model coding probability of each candidate region; S403, in a fusion mode of weighted combination prediction score, linearly sum the Mamba single model coding probability and the machine learning single model coding probability according to a preset weight to calculate a fusion prediction confidence; S404, according to the fusion prediction confidence, output the fusion prediction confidence value and / or the coding-non-coding binary classification judgment label of the candidate small peptide coding region; S405, mark the candidate small peptide coding region with a fusion prediction confidence lower than a threshold as a to-be-verified sample and return to step S3 for re-feature coding.

5. The method of claim 1, wherein, The sub-steps of step S5 include: S501, input the fusion prediction confidence value and the coding-non-coding binary classification judgment label of each candidate small peptide coding region; S502, based on the coding-non-coding binary classification judgment result of the candidate small peptide coding region, comprehensively determine whether the corresponding lincRNA has small peptide coding potential; S503, if it is determined that there is no coding potential, output that the lincRNA has no small peptide coding potential; S504, if it is determined that there is coding potential, translate and deduce the amino acid sequence of the candidate small peptide according to the codon translation rule of the corresponding reading frame for the candidate small peptide coding region determined as coding positive; S505, weight and sort the multiple positive candidate small peptides of the same lincRNA, calculate a comprehensive sorting confidence according to a fixed weight rule, and sort according to the score from high to low; S506, classify the candidate small peptides with a comprehensive sorting confidence higher than a comprehensive sorting confidence threshold into a high confidence set, and classify the rest into a low confidence set; S507, output the lincRNA coding potential conclusion and high confidence small peptide information; the high confidence small peptide information includes: the amino acid sequence of the high confidence small peptide, the corresponding full-length coordinate, the fusion prediction confidence, the comprehensive sorting confidence, and the ranking.

6. A lincRNA small peptide coding potential identification system, characterized in that, It includes: The standardized lincRNA sequence acquisition module is configured to acquire a set of standardized lincRNA sequences with unified format, length and encoding mode. The candidate small peptide coding region construction module is configured to integrate, de-duplicate and merge all potential coding fragments identified from the set of standardized lincRNA sequences to obtain a set of candidate small peptide coding regions covering the potential small peptide translation possibility. The multi-dimensional sequence feature encoding module is configured to extract multi-class multi-dimensional sequence features of each candidate small peptide coding region in the set of candidate small peptide coding regions, encode the multi-class multi-dimensional sequence features into a vector or tensor form feature representation learnable by a model, and output a corresponding set of multi-dimensional sequence feature representations. The coding potential fusion prediction module is configured to input the set of multi-dimensional sequence feature representations of the candidate small peptide coding region into a trained Mamba deep learning network, output a deep sequence representation vector and a Mamba single model coding probability, input the multi-dimensional sequence features, the deep sequence representation vector or a combination of the two into at least one trained machine learning model to output a machine learning single model coding probability, then linearly sum the fusion prediction confidence according to a preset weight, output the value and / or an encoding-non-encoding binary classification judgment label, and finally mark the candidate region with a fusion prediction confidence lower than a threshold as a to-be-verified sample and return to step S3 for re-feature encoding. The small peptide coding potential judgment and screening module is configured to input the fusion prediction confidence and the encoding-non-encoding binary classification judgment label of each candidate small peptide coding region, comprehensively judge the small peptide coding potential of the corresponding lincRNA according to the judgment result, output a no coding potential conclusion if there is no small peptide coding potential, derive a small peptide amino acid sequence from the coding positive region if there is a small peptide coding potential, calculate a comprehensive ranking confidence according to a fixed weight based on the fusion prediction confidence and a multi-class evaluation index, sort a plurality of positive candidate small peptides from high to low according to the weighted ranking, and classify the candidate small peptides with a comprehensive ranking confidence higher than a threshold into a high confidence set and the rest into a low confidence set.

7. The system of claim 6, wherein, The standardized lincRNA sequence acquisition module includes: The original transcript data acquisition module is configured to acquire original transcript set data. The lincRNA sequence screening module is configured to screen lincRNA full-length nucleotide sequence sets from the original transcript set data according to gene annotation information. The lincRNA sequence quality filtering module is configured to filter the lincRNA full-length nucleotide sequence sets to obtain filtered lincRNA full-length nucleotide sequence sets. The lincRNA sequence standardization processing module is configured to perform standardization processing on the filtered lincRNA full-length nucleotide sequence sets, unify the sequence direction and convert the base characters into a unified encoding format to obtain a set of lincRNA sequences with unified format. The lincRNA sequence standardization processing module is configured to perform standardization processing on the filtered lincRNA full-length nucleotide sequence sets, unify the sequence direction and convert the base characters into a unified encoding format to obtain a set of lincRNA sequences with unified format. The lincRNA sequence length adaptation and index retention module is configured to classify the lincRNA sequence set in a uniform format according to a target input length threshold, split long sequences exceeding the threshold using a sliding window, and fill short sequences shorter than the threshold according to a fixed alignment rule, convert the lincRNA sequence into a fixed-length subsequence input unit, and retain the start and end position indexes of the lincRNA sequence in the original sequence, and output a standardized lincRNA sequence set with uniform format, length, and encoding mode.

8. The system of claim 6, wherein, The multi-dimensional sequence feature encoding module comprises: The multi-dimensional sequence feature extraction module is configured to extract multi-dimensional sequence features for each candidate small peptide coding region in the candidate small peptide coding region set, and obtain one or more of the following: base level, codon or quasi-codon, sequence complexity and information, and context association features. The sequence feature vector tensor encoding module is configured to encode the extracted multi-dimensional sequence features in the form of a vector or a tensor to form a model-learnable feature representation without manual setting of a fixed threshold. The multi-dimensional sequence feature representation set output module is configured to output a multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions.

9. The system of claim 6, wherein, The coding potential fusion prediction module comprises: The Mamba model feature modeling and encoding probability output module is configured to input the multi-dimensional sequence feature representation set of the candidate small peptide coding region into a trained Mamba deep learning network, capture long-range dependencies and non-local sequence patterns of each candidate region through continuous evolution of the hidden state of the state space model, and output a deep sequence representation vector and a Mamba single-model encoding probability of each candidate region. The machine learning model coding potential discrimination module is configured to input the multi-dimensional sequence features, the deep sequence representation vector, or the combined features of the two as input features into at least one trained machine learning model to perform coding potential discrimination, and output a machine learning single-model encoding probability of each candidate region. The multi-model encoding probability fusion and confidence calculation module is configured to linearly sum the Mamba single-model encoding probability and the machine learning single-model encoding probability according to a preset weight in a fusion manner of weighted combination prediction score, and calculate a fusion prediction confidence. The fusion result numerical value and binary classification label output module is configured to output a fusion prediction confidence numerical value and / or an encoding-non-encoding binary classification judgment label of the candidate small peptide coding region according to the fusion prediction confidence. The low-confidence sample marking and feature encoding fallback module is configured to mark a candidate small peptide coding region with a fusion prediction confidence lower than a threshold as a to-be-verified sample and return to step S3 to perform feature encoding again.

10. The system of claim 6, wherein, The small peptide coding potential judgment and screening module comprises: The fusion prediction result input module is configured to input the fusion prediction confidence numerical value and the encoding-non-encoding binary classification judgment label of each candidate small peptide coding region. The lincRNA coding potential comprehensive judgment module is configured to comprehensively judge whether the corresponding lincRNA has the small peptide coding potential based on the coding-non-coding binary classification judgment result of the candidate small peptide coding region; The non-coding potential result output module is configured to output that the lincRNA has no small peptide coding potential if it is judged to have no coding potential; The coding positive region amino acid sequence derivation module is configured to, if it is judged to have coding potential, perform translation derivation on the candidate small peptide coding region judged to be coding positive according to the codon translation rules of the corresponding reading frame to obtain the amino acid sequence of the candidate small peptide; The positive candidate small peptide weighted sorting and confidence calculation module is configured to perform weighted sorting on multiple positive candidate small peptides of the same lincRNA, calculate the comprehensive sorting confidence according to the fixed weight rules, and sort the candidate small peptides according to the scores from high to low; The high-low confidence candidate small peptide classification module is configured to classify the candidate small peptides with the comprehensive sorting confidence higher than the comprehensive sorting confidence threshold into a high confidence set, and classify the rest into a low confidence set; The coding potential conclusion and high confidence small peptide information output module is configured to output the lincRNA coding potential conclusion and the high confidence small peptide information; the high confidence small peptide information includes the amino acid sequence, the corresponding full-length coordinate, the fusion prediction confidence, the comprehensive sorting confidence and the ranking of the high confidence small peptide.

Images