A method and system for identifying lincrna small peptide coding potential

Abstract

The application discloses a lincRNA small peptide coding potential identification method and system, and belongs to the technical field of computer technology and bioinformatics. The method first processes lincRNA sequences into a standardized sequence set; then identifies and integrates potential coding fragments to obtain a candidate coding region set; after extracting the multi-dimensional feature codes of each region, the fusion prediction confidence and label are obtained through the Mamba model and the machine learning model; finally, the coding potential is determined, and if there is no coding potential, the conclusion is output, and if there is coding potential, the amino acid sequence is deduced, the comprehensive ranking confidence is calculated and ranked by combining multiple indexes, and the high and low confidence small peptide sets are screened according to the threshold. The application introduces an efficient sequence state space modeling structure, does not need to rely on ribosome sequencing or mass spectrometry data, realizes high-precision, low-complexity and scalable prediction of the potential small peptide coding capacity in the lincRNA sequence, and solves the technical defects that the existing methods are difficult to balance between accuracy and computational efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer technology and bioinformatics, and in particular to a method and system for identifying the coding potential of lincRNA small peptides. Background Technology

[0002] With the rapid development of high-throughput sequencing technology and multi-omics research, an increasing number of studies have discovered that long intergenic non-coding RNAs (lincRNAs), traditionally considered non-coding, can be translated into biologically functional small peptides or microproteins under specific conditions. These small peptides derived from non-coding transcripts are considered important components of the "cryptome," playing potential roles in tumorigenesis, cell signal transduction, and immune regulation. Related research indicates that systematically mining small peptides encoded by lincRNAs is crucial for expanding genome functional annotation and discovering new disease-related molecules.

[0003] However, computational prediction of the coding potential of small peptides in lincRNAs still faces many technical challenges, mainly in the following aspects:

[0004] Existing algorithms are mostly derived from the coding potential assessment framework of linear mRNA or circRNA, lacking specific modeling for the structural and translational characteristics of lincRNA. Most commonly used coding potential prediction methods are based on open reading frame (ORF) length, codon usage preference, or simple sequence statistical features. Their design is mainly aimed at classical protein-coding transcripts, making it difficult to accurately characterize the characteristics of short ORFs, multiple start sites, and non-classical translation patterns in lincRNA, resulting in insufficient sensitivity and specificity for small peptide recognition.

[0005] Traditional feature engineering methods have limited ability to express high-dimensional sequence information and struggle to capture the complex patterns of lincRNA coding potential. Existing methods typically rely on a small number of manually designed sequence features or regular thresholds, which are insufficient for modeling potential long-range dependencies, local pattern combinations, and contextual information in RNA sequences, thus limiting the generalization performance of the models in complex transcriptome contexts.

[0006] Some methods heavily rely on ribosome sequencing or mass spectrometry data, resulting in high application costs and limited applicability. Although ribosome binding evidence and mass spectrometry identification are considered important means of verifying translation events, the related experimental techniques are costly, sample acquisition is difficult, and they are not easily applied systematically in large-scale transcriptome data, limiting the high-throughput prediction and screening of lincRNA small peptides.

[0007] Existing deep learning models face significant bottlenecks in terms of long sequence processing and computational efficiency. While recent deep learning methods have improved the accuracy of encoding potential prediction to some extent, most mainstream models are based on recurrent neural networks or self-attention mechanisms. These models suffer from high computational complexity and slow training and inference speeds when processing long lincRNA sequences, making it difficult to meet the efficiency and stability requirements of large-scale data analysis.

[0008] In summary, current technologies lack an algorithmic approach that addresses the small peptide coding characteristics of lincRNAs, balances prediction accuracy and computational efficiency, and is applicable to large-scale transcriptome data. Therefore, it is necessary to propose a novel technical solution that introduces a more efficient deep learning sequence modeling structure to systematically, rapidly, and accurately predict the potential small peptide coding capabilities of lincRNAs. Summary of the Invention

[0009] To overcome the above shortcomings, the present invention aims to provide a method and system for identifying the coding potential of lincRNA small peptides, so as to solve the problems of insufficient feature expression ability, low efficiency of long sequence modeling, and poor adaptability to large-scale transcriptome data in the existing technology for predicting the coding potential of lincRNA small peptides.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] This invention provides a method and system for identifying the coding potential of small lincRNA peptides. The method includes the following steps:

[0012] S1, Obtain a standardized lincRNA sequence set with uniform format, length, and encoding method;

[0013] S2, integrate, remove duplicates and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the potential translation possibilities of small peptides;

[0014] S3, extract multi-class 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set;

[0015] 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 the 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 the machine learning single model coding probability. Subsequently, linearly sum the values ​​according to the preset weights to obtain the fusion prediction confidence. Output the value and / or the binary classification label for encoding and non-encoding. Finally, mark the candidate regions with a fusion prediction confidence lower than the threshold as samples to be verified and return to step S3 to re-encode the features.

[0016] S5. Input the fusion prediction confidence and coding / non-coding binary classification label of each candidate small peptide coding region. Based on the judgment results, comprehensively judge the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, output the conclusion of no coding potential. If there is, deduce the small peptide amino acid sequence for the coding positive region. Then, combine the fusion prediction confidence with multiple evaluation indicators with fixed weights to calculate the comprehensive ranking confidence. After weighting and sorting multiple positive candidate small peptides from high to low, candidate small peptides with comprehensive ranking confidence higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set.

[0017] This invention also provides a system for identifying the coding potential of lincRNA small peptides, comprising:

[0018] The standardized lincRNA sequence acquisition module is used to acquire a set of standardized lincRNA sequences with uniform format, length, and encoding method.

[0019] The candidate small peptide coding region construction module is used to integrate, remove duplicates, and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the translation possibilities of potential small peptides.

[0020] The multi-dimensional sequence feature encoding module is used to extract multiple types of 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set.

[0021] The encoding potential fusion prediction module is used to input the multi-dimensional sequence feature representation set of candidate small peptide encoding regions into the trained Mamba deep learning network, output the deep sequence representation vector and the Mamba single model encoding probability, and 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 the machine learning single model encoding probability. Subsequently, the fusion prediction confidence is obtained by linear summation according to preset weights, and the value and / or the binary classification judgment label of encoding and non-coding are output. Finally, the candidate regions with fusion prediction confidence below the threshold are marked as samples to be verified and the process returns to step S3 for re-encoding features.

[0022] The small peptide coding potential determination and screening module is used to input the fusion prediction confidence level and coding / non-coding binary classification label of each candidate small peptide coding region. Based on the determination results, it comprehensively judges the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, it outputs the conclusion of no coding potential. If there is, it derives the small peptide amino acid sequence from the coding positive region. Then, it calculates the comprehensive ranking confidence level by combining multiple evaluation indicators such as fusion prediction confidence level with fixed weights. After weighting and ranking multiple positive candidate small peptides from high to low, candidate small peptides with comprehensive ranking confidence level higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set.

[0023] This invention uses full-length lincRNA nucleotide sequences as input. First, the sequences are standardized to obtain a standardized lincRNA sequence set with uniform format, length, and encoding method. Potential coding fragments identified within this set are integrated, deduplicated, and merged to construct a set of candidate small peptide coding regions. Further, multi-dimensional sequence features are extracted from each candidate small peptide coding region and encoded into a model-learnable feature representation set. Based on this, the feature representation set is input into a Mamba deep learning network based on a state-space model to obtain a deep sequence representation vector and single-model encoding probabilities. Then, the multi-dimensional sequence features, deep sequence representation vectors, or combinations thereof are input into at least one machine learning model to obtain corresponding single-model encoding probabilities. The fusion prediction confidence is obtained by fusing the results of multiple models, and binary classification labels (encoded and non-encoded) are output. Low-confidence samples are returned to the feature encoding step for iterative optimization. Finally, the small peptide coding potential of lincRNAs is determined based on a combination of fusion prediction confidence and classification labels. The corresponding amino acid sequences of small peptides are derived from positive coding regions, and then weighted and sorted according to a comprehensive index to divide them into high-confidence and low-confidence candidate peptide sets. The lincRNA coding potential conclusions and corresponding high-confidence candidate peptide products are output. While ensuring prediction accuracy, this approach improves the efficiency of long lincRNA sequence modeling and the stability and generalization ability of the algorithm in large-scale transcriptome data analysis. Attached Figure Description

[0024] Figure 1 is a flowchart of the lincRNA small peptide coding potential identification method provided in the embodiment of the present invention;

[0025] Figure 2 is a flowchart of the sub-steps of step S1 provided in an embodiment of the present invention;

[0026] Figure 3 is a flowchart of the sub-steps of step S3 provided in an embodiment of the present invention;

[0027] Figure 4 is a flowchart of the sub-steps of step S4 provided in an embodiment of the present invention;

[0028] Figure 5 is a flowchart of the sub-steps of step S5 provided in an embodiment of the present invention;

[0029] Figure 6 is a structural block diagram of the lincRNA small peptide coding potential identification system provided in an embodiment of the present invention;

[0030] Figure 7 is a block diagram of the submodule structure of the standardized lincRNA sequence acquisition module 1 provided in an embodiment of the present invention;

[0031] Figure 8 is a block diagram of the sub-module structure of the multi-dimensional sequence feature encoding module 3 provided in an embodiment of the present invention;

[0032] Figure 9 is a block diagram of the sub-module structure of the coding potential fusion prediction module 4 provided in an embodiment of the present invention;

[0033] Figure 10 is a block diagram of the submodule structure of the small peptide coding potential determination and screening module 5 provided in an embodiment of the present invention. Detailed Implementation

[0034] The present invention is illustrated by the following examples.

[0035] Gene annotation information refers to structured annotation data used to describe the identity, category, and genomic structural location of genes and their transcripts. This annotation data can originate from public annotation databases or equivalent annotation files (e.g., annotation collections from GENCODE, Ensembl, and RefSeq, GTF / GFF3 format files). The gene annotation information includes at least one or more of the following: gene and transcript identifiers (gene_id, transcript_id, and / or gene_name), transcript type / gene type (used to distinguish lincRNAs from protein-coding genes, etc.), chromosome number, start and end coordinates and chain direction information of the gene or transcript on the reference genome, exon structure information (number of exons and coordinates of each exon), and annotation version and reference genome version information. This gene annotation information is used to screen and determine the lincRNA sequences to be analyzed, establish the correspondence between sequences and genomic coordinates, and support location backtracking and result annotation of predicted candidate small peptide coding regions.

[0036] This embodiment provides a method for training and saving a model for predicting the coding potential of unknown lincRNA sequences. First, the training data is processed according to steps S1–S3: tagged lincRNA sequences and their candidate small peptide coding regions are acquired and organized, sequence standardization, candidate small peptide coding region construction, and feature extraction are completed to form a training sample set that can be used for supervised learning.

[0037] S606 explicitly states that the trained Mamba model parameters, machine learning model parameters, and multi-model fusion strategies are saved for subsequent prediction and analysis of unknown lincRNA sequences. In this embodiment, the input target length threshold of the Mamba model is consistent with S105, fixed at 1024 nt; sequences longer than 1024 nt are split using a sliding window (window length 1024 nt, step size 512 nt), and short sequences are padded to 1024 nt, generating an effective length mask to ignore padded regions. The Mamba model structure parameters are fixed as follows: embedding dimension 256, network layers 6, hidden dimension 512, dropout=0.1, and the output layer is a binary fully connected layer; the training hyperparameters are fixed as follows: optimizer AdamW, learning rate 3×10^-4, batch_size=64, training epochs 50, and the early stopping strategy is to stop if the validation set metrics show no improvement for 10 consecutive epochs, and the trained network weights and configuration files are saved. The machine learning model uses a random forest (RF) with fixed parameters: learning_rate=0.05, n_estimators=1000, max_depth=8, num_leaves=64, subsample=0.8, and colsample_bytree=0.8. The trained model file is saved. The multi-model fusion strategy is consistent with S404, using weighted combined prediction scores as the primary fusion method. The fusion weights are determined and fixed through validation set optimization, resulting in a Mamba weight of 0.60, an RF weight of 0.40, and a fusion decision threshold of 0.50. A fixed voting arbitration rule is also saved as a backup: when the binary classification results of the two models are inconsistent, the model with the larger deviation from its respective threshold (higher confidence) is selected as the final decision. The above model parameters, fusion weights, thresholds, and arbitration rules are persistently stored in a callable format for predicting unknown lincRNA sequences after performing the same processing as S1–S3.

[0038] In step S105, to meet the fixed input length requirement of the deep learning model, length adaptation processing is performed on the lincRNA sequence, with the target input length threshold fixed at 1024 nt (nucleotides). When the sequence length exceeds 1024 nt, a sliding window approach is used for splitting, with a fixed window length of 1024 nt and a fixed step size of 512 nt, ensuring that the long sequence is covered by multiple overlapping subsequences. When the sequence length does not exceed 1024 nt, padding is used to complete the sequence to 1024 nt, and an effective length mask is generated simultaneously to distinguish between the actual bases and the padding positions, thus ignoring the influence of the padding regions during subsequent feature statistics and deep network modeling. The split subsequences retain their start and end position indices in the original sequence to support subsequent candidate small peptide coding region localization and result backtracking.

[0039] In step S6, a training dataset with encoded potential labels containing positive and negative samples is constructed. After performing the same processing flow as S1–S3, a Mamba deep learning network (which can be simply referred to as the Mamba model) and at least one machine learning model are trained respectively. In this embodiment, the Mamba deep learning network structure parameters are fixed as follows: input embedding dimension 256, number of network layers 6, hidden dimension 512, dropout=0.1, and the output layer is a binary classification fully connected layer that outputs the confidence of candidate small peptide coding regions. The training hyperparameters are fixed as follows: AdamW optimizer is used, learning rate 3×10^-4, batch_size=64, maximum number of training epochs 50, and an early stopping policy is set to terminate training if the validation set evaluation index does not improve for 10 consecutive epochs. After training, the network weight file and model configuration file are saved. The machine learning model uses a random forest (RF), with the following key parameters fixed: learning_rate=0.05, n_estimators=1000, max_depth=8, num_leaves=64, subsample=0.8, and colsample_bytree=0.8. After training, the corresponding model file is saved.

[0040] In step S404, the outputs of the Mamba model and the machine learning model are fused. In this embodiment, a weighted combined prediction score is used as the main fusion method. The fusion weights are determined and fixed by optimization on the validation set, resulting in a Mamba weight of 0.60 and an RF weight of 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 encoded / non-encoded. At the same time, the voting arbitration rule is fixed as a backup fusion rule: when the binary classification results of the two models are inconsistent, the deviation of their respective output probabilities from their fusion decision threshold is compared, and the label corresponding to the one with the greater deviation is selected as the final decision.

[0041] In step S606, the trained Mamba deep learning network parameters, RF machine learning model parameters, and multi-model fusion strategies (including fusion weights, fusion decision thresholds, and voting arbitration rules) are persistently 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 and S1–S3. Then, the saved Mamba deep learning network and RF machine learning model are called respectively to obtain the predicted probability / score of the candidate small peptide coding region. The fusion prediction confidence is calculated according to the saved fusion weights and thresholds, and the final coding potential prediction result is output, thereby realizing the coding potential analysis and priority screening of candidate small peptides for unknown lincRNA sequences.

[0042] like Figure 1 As shown, this embodiment of the invention provides a method for identifying the coding potential of lincRNA small peptides, including:

[0043] Step S1: Obtain a standardized lincRNA sequence set with uniform format, length, and encoding method;

[0044] lincRNA sequences can be obtained from public transcriptome databases, gene annotation databases, or transcript collections acquired through high-throughput sequencing. During the acquisition process, transcripts belonging to the long-chain intergenic non-coding RNA type can be screened based on gene annotation information to exclude protein-coding RNAs or other non-target RNA types that may interfere with subsequent analysis.

[0045] After sequence acquisition, the lincRNA sequences undergo uniform preprocessing, including quality filtering, normalization, and length adaptation. Specifically:

[0046] Sequence quality filtering: Remove sequences containing uncertain bases or abnormal characters to ensure the stability of subsequent feature encoding and model calculation.

[0047] Sequence normalization: The sequence orientation is standardized and the base characters are converted into a uniform encoding format so that sequences from different sources are represented in a consistent manner.

[0048] Sequence length adaptation processing: The target input length threshold is set to 1024 nt. For lincRNA sequences with a length exceeding this threshold, a sliding window method with a window length of 1024 nt and a step size of 512 nt is used for splitting. For residual fragments with a length less than a complete window after splitting, end-padding or end-alignment truncation is performed. For short sequences shorter than this threshold, they are padded to 1024 nt according to a fixed alignment rule. At the same time, an effective length marker is generated for all padded sequences to ignore the padded region. The start and end position indices of the split and padded subsequences are retained in the original sequence.

[0049] Through the above preprocessing steps, lincRNA sequences of arbitrary length are converted into fixed-length subsequence input units, ensuring that all input sequences are consistent in format, length, and encoding method, thereby guaranteeing that the subsequent deep learning model can perform computation stably and efficiently.

[0050] Step S2 involves integrating, deduplicating, and merging all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the potential translation possibilities of small peptides.

[0051] For each preprocessed standardized lincRNA sequence, without presupposing its classical protein coding ability, one or more combinations of multiple reading frame scanning, sequence local feature-based segmentation, and covering window scanning are used to identify potential coding fragments (multiple reading frame scanning needs to identify short open reading frames with a length of ≥20 amino acids). After all potential coding fragments are identified, they are integrated, deduplicated, and merged to construct multiple candidate small peptide coding regions, forming a set of candidate small peptide coding regions to comprehensively cover potential small peptide coding events with non-classical initiation, non-standard length, or special translation modes.

[0052] The construction of candidate small peptide coding regions is not limited to a single rule, but can be achieved through one or a combination of the following methods:

[0053] Multiple reading frame scanning method: The lincRNA sequence is scanned under different reading frames to identify short open reading frames with lengths within a preset range in order to capture potential small peptide coding regions.

[0054] Sequence segmentation based on local features: Based on the changes in local base arrangement, composition, or statistical features in the sequence, the sequence is divided into multiple segments with potential coding possibilities.

[0055] Coverage window scanning method: Using a fixed window or variable window strategy, the entire lincRNA sequence is continuously scanned, so that sequence fragments at any position have the opportunity to be included in the candidate small peptide coding region set.

[0056] By constructing multiple candidate small peptide coding regions, we can avoid missing small peptide coding events generated by non-classical initiation, non-standard length, or special translation modes that rely solely on a single ORF rule.

[0057] Step S3: Extract multi-class, 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set.

[0058] like Figure 3 As shown, the extracted multi-level and multi-dimensional sequence features include base level, codons or quasi-codons, sequence complexity and information, and contextual features. Among them, the construction of contextual features requires fixing the window range of 128nt upstream and 128nt downstream, and padding the insufficient part to generate an effective length label. The feature encoding of the Mamba deep learning network is concatenated and padded to 1024nt by "upstream 128nt + candidate small peptide coding region + downstream 128nt" to form a tensor form feature representation. The feature encoding of the machine learning model is to concatenate the candidate region features, upstream and downstream window features and position features to form a vector form feature representation. Finally, the multi-dimensional sequence feature representation set corresponding to all candidate regions is output.

[0059] Multidimensional sequence feature extraction and encoding operation standards:

[0060] For candidate small peptide coding regions, four types of multi-dimensional sequence features are extracted: base level, codon / quasi-codon, sequence complexity and information, and contextual association. These features are then encoded into model-learnable representations in vector or tensor form. All features are not judged by manually fixed thresholds; instead, the model autonomously learns feature patterns related to the small peptide coding potential during training. The specific extraction criteria are as follows:

[0061] Base level characteristics:

[0062] Base composition ratio: The proportion of the four bases A, C, G, and T (U) in the candidate region is calculated to form a 4-dimensional vector, and the GC content is added as a scalar feature;

[0063] Local arrangement patterns: Statistically count the k-mer frequencies of dinucleotides (16 types) or trinucleotides (64 types) to generate corresponding dimensional pattern vectors;

[0064] Distribution characteristics: The candidate region is divided into a fixed number of segments, and the base composition ratio vector or k-mer frequency vector of each segment is calculated and then spliced ​​together.

[0065] Characteristics of codons or quasi-codons:

[0066] The quasi-codon is a continuous combination of bases of length 3 (3-mer), without translation reading frame or start / stop codon restrictions; overlapping triplets are extracted with a step size of 1 or non-overlapping triplets with a step size of 3, and the normalized frequencies of 64 triplets are statistically analyzed to generate a 64-dimensional vector that reflects the potential translation preference of the sequence.

[0067] Sequence complexity and information characteristics:

[0068] Low complexity ratio: The proportion of the total length of fragments with consecutive ≥6 nt identical bases within a candidate region to the total length of the region;

[0069] Repeatability index: The ratio of the number of unique triplet types to the total number of all triplet types;

[0070] Information entropy: Information entropy is calculated based on single base distribution (4 base ratios) and triplet distribution (64 triplet frequencies) to characterize the differences in sequence structure and composition.

[0071] Contextual features:

[0072] Relative position: The proportion of the start and end positions of the candidate region to the full length of the lincRNA is used as a scalar feature;

[0073] Upstream and downstream windows: Fixed window range of 128nt upstream and 128nt downstream of the candidate region. When the window length is insufficient, it is filled and an effective length marker (mask) is generated.

[0074] Feature fusion: The input to the Mamba model is a tensor formed by concatenating the upstream 128nt + candidate small peptide coding region + downstream 128nt and padding it to 1024nt; the input to the machine learning model is a vector formed by concatenating candidate region features + upstream and downstream window features + position features.

[0075] Multidimensional sequence feature encoding of candidate small peptide coding regions:

[0076] Multi-dimensional sequence feature encoding is performed on candidate small peptide coding regions to form an input representation that can be processed by a deep learning model. The core encoded sequence features are divided into the following four categories, and each type of feature is input into the model in vector / tensor form:

[0077] Base-level characteristics: encompassing the base composition ratio, local arrangement pattern, and base distribution characteristics of candidate regions;

[0078] Codon or quasi-codon related features: Encoding the usage characteristics of consecutive base combinations within candidate regions to reflect potential translation preferences of the sequence;

[0079] Sequence complexity and information features: including the proportion of low sequence complexity, repetition index, information entropy, etc., to describe the structural and compositional differences of candidate regions;

[0080] Contextual features: These features are constructed by combining the relative positions of candidate regions within the full length of the lincRNA with sequence information from upstream and downstream fixed windows, reflecting potential translation-related contextual features.

[0081] 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 the Mamba single model encoding 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 the machine learning single model encoding probability. Subsequently, linearly sum the values ​​according to preset weights to obtain the fusion prediction confidence, and output the value and / or the binary classification label for encoding and non-coding. Finally, mark the candidate regions with a fusion prediction confidence lower than the threshold as samples to be verified and return to step S3 for re-encoding features.

[0082] In this embodiment, the Mamba deep learning network uses a pre-trained model with the following structural parameters: embedding dimension 256, network layers 6, hidden dimension 512, dropout=0.1, and output layer is a binary classification fully connected layer. The machine learning model used is a pre-trained Random Forest (RF) with the following parameters: 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 scoring method, with a preset Mamba weight of 0.60 and an RF weight of 0.40. The fusion decision threshold is set to 0.50, while retaining a voting arbitration rule as a backup. That is, when the binary classification results of the two models are inconsistent, the label corresponding to the one with the larger deviation from the threshold is selected as the final decision. Candidate regions with a fusion prediction confidence of less than 0.50 are marked as samples to be verified, and the process returns to step S3 for re-encoding features.

[0083] In step S4, to further improve the accuracy and stability of lincRNA small peptide coding potential prediction and overcome the problem of limited generalization ability of a single model in the context of complex transcriptomes, this invention introduces a state-space model-based Mamba deep learning network structure and further integrates machine learning models to construct a multi-model collaborative prediction framework.

[0084] (1) Deep sequence representation learning based on Mamba deep learning network

[0085] First, the multi-dimensional sequence features of the candidate small peptide coding regions obtained in step S3 are input into the Mamba deep learning network.

[0086] The Mamba deep learning network is based on a state-space modeling mechanism that recursively updates sequence features. It models the internal and contextual information of candidate small peptide coding regions through the continuous evolution of hidden states, thereby achieving the following technical effects: capturing long-range dependencies in candidate small peptide coding regions; effectively expressing non-local sequence patterns; and reducing the computational resource consumption required for long sequence modeling while maintaining linear time complexity.

[0087] In this way, the Mamba deep learning network outputs a deep sequence representation vector to characterize the potential translational properties of candidate small peptide coding regions.

[0088] (2) Discriminative feature modeling based on machine learning model

[0089] While obtaining the deep sequence representation of the output of the Mamba deep learning network, this invention further introduces at least one traditional machine learning model to model the discriminative features of candidate small peptide coding regions, so as to supplement the discriminative ability of the deep learning model under specific data distributions.

[0090] The machine learning models include, but are not limited to: random forest models; gradient boosting decision tree models; support vector machine models; and logistic regression models. The input features of the machine learning models may include:

[0091] The statistical features extracted in step S3; the deep sequence representation output by the Mamba deep learning network; or a combination of the above two types of features. The feature space is partitioned using a tree model or a margin maximization model, which makes the model more robust in discriminating candidate small peptide coding regions of different lengths and structural types.

[0092] (3) Multi-model collaborative prediction and result fusion

[0093] After achieving independent predictions from the Mamba deep learning network and the machine learning model, this invention employs one or more of the following fusion methods to accurately fuse the results of the two models. All fusion rules are determined and fixed during the training phase using a validation set or cross-validation, and are then used in the subsequent inference phase:

[0094] The weight allocation operation standard for multi-model fusion is as follows:

[0095] 1. Weighted combination prediction score weights:

[0096] Determining the optimal validation set: Using a step size of 0.05 or 0.1, traverse the weights within the range of 0-1, select the weights that optimize AUC, F1, accuracy, or PR-AUC, and fix them.

[0097] Performance metrics are normalized and determined: weights are assigned according to the proportion of AUC or F1 of each model on the validation set, with higher weights given to models with better performance.

[0098] 2. Voting Rules:

[0099] Majority voting: When there are 3 or more models, output 1 if more than half of them are positive, otherwise output 0;

[0100] Arbitration based on tie: Decisions are made based on confidence level (comparing the distance between the model output probability and the threshold) or preset model priority.

[0101] Firstly, the weighted combined prediction score is obtained by linearly weighting and summing the predicted probabilities / scores output by each model. The weights are preferably obtained by at least one of the following methods: "determining by validation set optimization" or "determining by performance index normalization". Validation set optimization involves iterating through the weights on the validation set (e.g., enumerating in the range of 0 to 1 with a step size of 0.05 or 0.1) to select the weight that optimizes the preset evaluation index (preferably one or more of AUC, F1, accuracy, or PR-AUC), and fixing this weight as the fusion weight. Performance index normalization involves first calculating the performance index values ​​of each base model on the validation set, and then determining the weights according to their relative performance proportions (e.g., using the proportion of AUC or F1 of each model as the weight), thus giving higher weights to models with better performance.

[0102] Secondly, the voting integration and judgment results are as follows: First, the probabilities / scores output by each model are converted into binary classification labels through a preset threshold, and then the final judgment is obtained according to the voting rules; the voting rules include two types: majority voting and confidence-based arbitration.

[0103] Majority voting: When there are three or more base models, if more than half of the models output a positive label, the final decision is 1; otherwise, the output is 0. In the event of a tie, the decision is made by arbitration based on confidence level or by a pre-set model priority order.

[0104] Confidence-based arbitration: When the labels of the models are inconsistent, compare the distance between the output probability of each model and its judgment threshold (or compare the maximum probability value), and select the label corresponding to the one with higher confidence as the final output; when the binary classification results of the two models are inconsistent, the "probability corresponding to the one that deviates more from the threshold" can be used as the prediction confidence of the candidate small peptide coding region for arbitration.

[0105] Third, secondary prediction of the upper-level model: The outputs of each base model are used as new feature inputs to the upper-level fusion model for secondary learning and prediction. The new features include at least the prediction probability / score of each model output and its binary classification label, and may optionally include the confidence of each model, candidate small peptide coding region-level statistics (such as the proportion of positive windows in the same transcript, maximum / mean scores, etc.) as enhancement features. The upper-level model type 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 its parameters are learned based on the validation set or cross-validation during the training phase to minimize classification loss and maximize preset evaluation indicators. Finally, the upper-level model parameters, fusion weights, voting rules, etc. are saved together for the prediction analysis of unknown lincRNA sequences.

[0106] By employing the aforementioned multi-model fusion strategy, the impact of single-model prediction bias on the final result can be effectively reduced, improving the accuracy and stability of identifying real small peptide coding events and providing a reliable basis for subsequent threshold determination of candidate small peptide coding regions.

[0107] The result fusion methods include, but are not limited to: weighted combination prediction scores of the fusion prediction confidence of different model outputs; integration of coding potential judgment results using a voting mechanism; and inputting the outputs of multiple models as new features into the upper-level judgment model for secondary prediction.

[0108] By using a multi-model collaborative prediction mechanism, the impact of single-model prediction bias on the final result can be effectively reduced, thereby improving the accuracy and stability of identifying real small peptide coding events.

[0109] In this invention, "confidence level" includes two types of evaluation indicators: candidate coding regions and candidate small peptides. The definitions and application scenarios of each type of confidence level are as follows:

[0110] (1) Single model encoding probability: refers to the encoding probability of candidate small peptide coding regions output by the Mamba deep learning network, and the encoding probability of the same candidate coding region output by the machine learning model (such as RF). Both are the original results predicted independently by each model.

[0111] (2) Fusion prediction confidence: refers to the fusion output probability obtained by fusing the above two types of single model encoding probabilities according to preset weights, which is used for threshold determination in step S205; when the encoding / non-encoding determination results of the two models are inconsistent, the "probability corresponding to the one that deviates more from the threshold" can be used as the prediction confidence of the candidate encoding region for arbitration.

[0112] (3) Comprehensive ranking confidence: When the same lincRNA corresponds to multiple candidate peptides, in order to achieve the ranking and priority output of candidate peptides, multi-source information such as sequence feature score, conservation statistics and peptide length mapping score are further introduced on the basis of fusion prediction confidence. The comprehensive score is obtained by weighting according to preset weights. This score is used for the ranking of candidate peptides and the output of Top-N results in step S505.

[0113] Among the confidence metrics mentioned above, the single-model encoding probability and the fusion prediction confidence are used for binary classification of the coding and non-coding regions of candidate small peptides, while the comprehensive ranking confidence is used for priority ranking of candidate small peptides and final result output.

[0114] (4) Construction of the prediction output module

[0115] Based on the fused prediction results, a coding potential prediction output module is constructed to output the confidence level of the fusion prediction of the small peptide coding potential of candidate small peptide coding regions or the binary classification result. In this step, the output format can be selected according to the actual application requirements, including but not limited to: continuous coding potential fusion prediction confidence level; coding and non-coding judgment labels; or a combination of the two.

[0116] (5) Description of technical effects

[0117] By introducing machine learning models and constructing a multi-model collaborative prediction framework based on the Mamba deep learning network, this invention has at least the following technical effects: fully leveraging the advantages of the Mamba deep learning network in long sequence modeling and computational efficiency; enhancing the sensitivity to key discriminative features and improving prediction accuracy by utilizing machine learning models; reducing the risk of performance fluctuations of a single model in different datasets or application scenarios; and improving the stability and generalization ability of the algorithm in large-scale transcriptome data analysis.

[0118] S5. Input the fusion prediction confidence and coding / non-coding binary classification label of each candidate small peptide coding region. Based on the judgment results, comprehensively judge the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, output the conclusion of no coding potential. If there is, deduce the small peptide amino acid sequence for the coding positive region. Then, combine the fusion prediction confidence with multiple evaluation indicators with fixed weights to calculate the comprehensive ranking confidence. After weighting and sorting multiple positive candidate small peptides from high to low, candidate small peptides with comprehensive ranking confidence higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set.

[0119] In step S5, based on the prediction results of step S4, the coding potential of the lincRNA sequence is comprehensively determined, and the corresponding small peptide candidate product information is output.

[0120] Specifically, the process includes: **Coding Potential Determination:** Based on the model output, determining whether the lincRNA possesses small peptide coding potential. **Small Peptide Sequence Determination:** For candidate small peptide coding regions determined to have coding potential, outputting the corresponding small peptide amino acid sequence. **Candidate Result Screening and Ranking:** When the same lincRNA corresponds to multiple candidate small peptides, the candidate small peptides can be ranked or screened based on prediction scores, sequence characteristics, or other evaluation indicators to obtain a high-confidence set of small peptides. The final output includes the lincRNA's small peptide coding potential prediction conclusion and the corresponding small peptide candidate product information.

[0121] S505, when the same lincRNA corresponds to multiple candidate small peptide coding regions, the present invention calculates the comprehensive fusion confidence score for each candidate small peptide and outputs it in descending order of score; the comprehensive fusion prediction confidence score is obtained by weighting one or more indicators among the fusion prediction confidence score, coding potential related sequence features, sequence conservation and small peptide length.

[0122] The weighting criteria for ranking candidate small peptides are as follows:

[0123] The overall ranking confidence score uses a fixed weight: fusion prediction confidence score weight 0.60, coding potential related sequence features (ORF integrity and triplet preference, etc.) weight 0.15, sequence conservation (candidate small peptide coding region conservation score statistic) weight 0.15, and small peptide length (mapped to 0-1 score according to preset preference interval) weight 0.10.

[0124] If only some indicators are used, the weights of the unused indicators are proportionally allocated to the used indicators, and the total weight remains at 1.

[0125] As a preferred embodiment, the comprehensive fusion prediction confidence score adopts a fixed weight: the fusion prediction confidence score weight is 0.60, the coding potential related sequence feature weight is 0.15, the sequence conservation weight is 0.15, and the peptide length weight is 0.10. Specifically, the fusion prediction confidence score is the coding probability / score output by the fusion model; the sequence feature score is the composite score of features such as ORF integrity and triplet preference; the conservation score is the statistical measure of the conservation score of the candidate peptide coding region; and the peptide length is mapped to a 0–1 score according to a preset preference interval. When only some indicators are used, the weights of unused indicators are proportionally allocated to the used indicators while maintaining a total weight of 1, thereby achieving repeatable ranking of multiple candidate peptides for the same lincRNA.

[0126] Step S6: Construct a training dataset with coded potential annotations containing positive and negative samples. After performing the same processing flow as S1-S3, train the Mamba deep learning network and at least one machine learning model respectively. Then train the multi-model fusion strategy to determine the prediction rules. Finally, save the trained model parameters and multi-model fusion strategy for the prediction analysis of unknown lincRNA sequences.

[0127] S6. When constructing a training dataset containing positive and negative samples and labeled with coding potential, this invention uses "candidate coding region" as the basic training sample unit: each lincRNA sequence can generate one or more candidate coding regions after being identified by the candidate coding region in step S2, and each candidate coding region is independently assigned a binary label (coding potential = 1 / non-coding = 0). If necessary, transcript-level labels can be obtained by further summarizing according to the rule that "if any candidate coding region reaches the threshold, the transcript is judged to be positive". Positive samples (coding potential = 1) must originate from at least one traceable coding evidence and establish a "candidate peptide coding region - evidence" correspondence. For example, the candidate peptide coding region must have translational evidence in ribosome binding / translation initiation related data, or its translation product must have a corresponding peptide detected in mass spectrometry, or it must be located in a translatable lincRNA / micropeptide coding region compiled from authoritative collections / literature (homologically conserved regions with consistent coding signals can be used as supplementary weak positive samples and their evidence level recorded). Negative samples (coding potential = 0) must include at least two types, preferably three: first, true background negative samples, i.e., candidate peptide coding regions annotated as lincRNAs and lacking translational evidence; second, matched control negative samples, i.e., constructing one or more non-coding candidate fragments for each positive sample that are similar in length and GC content (and optional dinucleotide / triplex distribution) to the positive sample, to avoid the model making judgments solely based on sequence composition differences; and third, hard negative samples. Negatives are selected by choosing candidate small peptide coding regions with long ORFs or a preference for coding pattern triplet formation in non-coding backgrounds, or by constructing segments from coding sequences using frameshifting / breaking start and end points while maintaining similar length and composition. These are used to suppress shortcut learning that assumes "long ORF = coding". After training set construction, redundancy and leakage control are performed: training / validation / test sets are divided by gene or transcript to ensure that multiple candidate small peptide coding regions generated from the same transcript do not appear across sets, and clustering can be performed to remove duplicates based on sequence similarity thresholds. Simultaneously, to address the imbalance between positive and negative samples, strategies such as undersampling / oversampling, class weights or focal loss, and stratified sampling can be used to control the proportion of each type of sample. Each sample at least stores metadata such as transcript / gene identifier, candidate small peptide coding region sequence and its start / end positions / chain direction / reading frame, label, and evidence type and level (optional for positive samples). Subsequently, the sequence standardization and length processing, formatting confirmation of candidate small peptide coding regions, and feature extraction processes consistent with those in S1–S3 were strictly performed on each candidate small peptide coding region in the above training dataset. Mamba deep learning network and at least one machine learning model were trained respectively, and a multi-model fusion strategy was trained on the validation set to determine the prediction rules. Finally, the trained model parameters and multi-model fusion strategy were saved for the coding potential prediction analysis of unknown lincRNA sequences.

[0128] The aforementioned model parameters, fusion weights, thresholds, and arbitration rules are all persistently stored in a callable format for predicting unknown lincRNA sequences after performing the same processing as S1–S3.

[0129] In step S6, the present invention trains a deep learning model and a machine learning model for predicting the coding potential of lincRNA small peptides to obtain model parameters for subsequent predictive analysis.

[0130] (1) Construction of training dataset

[0131] In this embodiment, the data used for model training is RNA sequence data with coding potential annotations or candidate small peptide coding region data.

[0132] The training data includes:

[0133] Known or high-confidence annotations are RNA sequences capable of coding small peptides or their corresponding candidate small peptide coding regions;

[0134] Non-coding RNA sequences that are labeled as not having small peptide coding ability or negative control candidate small peptide coding regions.

[0135] In one implementation, candidate small peptide coding regions are constructed from the complete lincRNA sequence according to the method described in step S2, and the candidate small peptide coding regions are used as the basic sample units for model training, thereby converting each lincRNA into multiple training samples that can be used for learning.

[0136] (2) Training data preprocessing and feature encoding

[0137] Before model training, the RNA sequences or candidate small peptide coding regions in the training dataset undergo the same preprocessing and feature encoding process as in the prediction stage, including:

[0138] Sequence quality filtering and standardization;

[0139] Sequence length adaptation and segmentation;

[0140] Construction of candidate small peptide coding regions;

[0141] Multidimensional sequence feature extraction and encoding.

[0142] By adopting the same data processing methods as in the prediction phase, the consistency of data distribution between model training and actual application scenarios is ensured.

[0143] (3) Training of deep sequence models based on Mamba deep learning network

[0144] In this embodiment, the multi-dimensional sequence features obtained in step (ii) are input into a Mamba deep learning network based on a state-space model to train the model parameters.

[0145] During training, the coding potential labels of candidate small peptide coding regions are used as supervision signals. The difference between the model output and the true labels is measured by the loss function, and the model weights are iteratively optimized by the parameter update mechanism.

[0146] During training, model weights are determined jointly by training data and supervision signals: The coding potential labels of candidate small peptide coding regions are used as supervision signals; a preset loss function is used to measure the difference between the model output and the true labels; and parameter updates are generated based on these differences, with multiple rounds of iterative updates to the model weights. During iteration, when the validation set evaluation metric reaches its optimum or a preset stopping condition (such as early stopping) is met, the model weights of the corresponding round are taken as the final weights and saved. For machine learning models, the weights are automatically learned by their training algorithm during training (determined through multiple rounds of weak learner accumulation or decision rule splitting learning), and are similarly saved as the final weights based on the optimal validation set or the model parameters at the end of training. When there is class imbalance in the training samples, class weights can be set. These class weights are determined based on the ratio of positive to negative samples and are used to adjust the contribution of different classes of samples in the loss function, thereby obtaining stable final model weights.

[0147] Through the above training process, the Mamba model is able to learn long-range dependent features and potential translation patterns related to the small peptide coding potential in candidate small peptide coding regions, and output stable deep sequence representations.

[0148] (4) Machine learning model training

[0149] During or after training the Mamba deep learning network, at least one machine learning model is introduced to discriminate and model candidate small peptide coding regions. The machine learning model is at least one of random forest, gradient boosting decision tree, support vector machine, and logistic regression.

[0150] The input features of the machine learning model include at least one of the following:

[0151] Statistical features in multidimensional sequence features;

[0152] The depth sequence representation output by the Mamba model;

[0153] The combination of the above features is represented.

[0154] By training a machine learning model to learn the discrimination boundary between encoded and non-coded samples, the classification robustness of the overall model under different sequence structures and lengths can be enhanced.

[0155] (5) Multi-model fusion strategy training

[0156] In this embodiment, the outputs of the Mamba-based deep learning model and the machine learning model are fused to determine the final coding potential prediction strategy.

[0157] Integration methods include, but are not limited to:

[0158] The predicted scores output by different models are weighted and combined.

[0159] Results are integrated based on voting or threshold mechanisms;

[0160] The outputs of multiple models are used as new feature inputs to the upper-layer fusion model for secondary training.

[0161] Through the above-mentioned fusion training process, the impact of prediction bias of a single model on the final result is reduced, and the stability and generalization ability of the model are improved in different datasets and application scenarios.

[0162] (6) Training completion and model saving

[0163] After the model training is completed, the parameters of the trained Mamba deep learning network, the parameters of the machine learning model, and the multi-model fusion strategy are saved for subsequent prediction and analysis of the small peptide coding potential of unknown lincRNA sequences.

[0164] This invention provides a method for identifying the coding potential of lincRNA small peptides based on the Mamba-machine learning fusion model.

[0165] In a preferred embodiment, the present invention provides a method for identifying the coding potential of lincRNA small peptides based on the fusion of deep learning and machine learning, the method comprising the following steps (e.g. Figure 1 ).

[0166] S1. From the obtained transcript set data, full-length nucleotide sequences of lincRNAs are screened out. After quality filtering, sequence standardization, and length adaptation, a standardized lincRNA sequence set with uniform format, length, and coding method is output.

[0167] In this embodiment, the full-length nucleotide sequence of the lincRNA to be analyzed is derived from a public transcriptome database or a collection of transcripts obtained through high-throughput sequencing.

[0168] First, transcript sequences annotated as long intergenic non-coding RNAs (lincRNAs) are screened based on gene annotation information, while protein-coding RNAs, pseudogene transcripts, and other non-target RNA types are excluded.

[0169] Subsequently, a uniform preprocessing operation is performed on the lincRNA sequence, including:

[0170] 1. Filter sequences containing uncertain bases or abnormal characters;

[0171] 2. The sequence orientation is standardized, and the base characters are converted into a unified encoding format;

[0172] 3. For lincRNA sequences whose length exceeds a preset threshold, a sliding window method is used to split them into multiple sub-sequences;

[0173] 4. For sequences that are not long enough, length adaptation is achieved by padding.

[0174] After the above processing, all input sequences are kept consistent in length, format and encoding method to meet the input requirements of subsequent models.

[0175] S2, without presupposing classical protein coding capabilities, uses one or more combinations of multiple reading frame scanning, sequence local feature-based segmentation, and covering window scanning to identify and scan potential coding fragments for each standardized lincRNA sequence, and integrates them to obtain a set of candidate small peptide coding regions covering the potential translation possibilities of small peptides.

[0176] In this embodiment, without presupposing whether the lincRNA has classical protein coding ability, multiple candidate small peptide coding regions are constructed from the preprocessed sequence.

[0177] S3, for each candidate small peptide coding region, extracts one or more features from various categories, including base level, codon or quasi-codon correlation, sequence complexity and information, and contextual association. These extracted features are encoded as vectors or tensors to form a model-learnable feature representation that does not require manually set fixed thresholds. The system then outputs a multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions, such as... Figure 3 As shown;

[0178] For each candidate small peptide coding region, multi-dimensional sequence feature encoding is performed in this embodiment, the features including:

[0179] 1. Base composition ratio and local arrangement pattern characteristics in the coding regions of candidate small peptides;

[0180] 2. Characteristics related to continuous base combinations or the use of quasi-codons;

[0181] 3. Statistical characteristics describing sequence complexity and information distribution;

[0182] 4. The relative position of candidate small peptide coding regions in the full-length lincRNA sequence and the characteristics of their upstream and downstream sequences.

[0183] The aforementioned features are represented in vector or tensor form and used as input to the model, enabling the model to automatically learn key feature patterns related to the small peptide encoding potential during training.

[0184] S4. Input the multi-dimensional sequence feature representation set into the Mamba deep learning network to obtain the deep sequence representation vector, extract relevant features and input them into at least one machine learning model to complete the coding potential discrimination, fuse the results of the two models and construct the output module to output the coding potential prediction result, and mark the low-confidence candidate small peptide coding regions for verification and return them for re-feature encoding.

[0185] In this embodiment, the features obtained in step S3 are input into a Mamba deep learning network based on a state-space model to perform sequence modeling on the candidate small peptide coding regions.

[0186] The Mamba model encodes sequence features through a state-space recursion mechanism to capture long-range dependencies and potential translation signals in candidate small peptide coding regions, and outputs a deep sequence representation vector while maintaining linear time complexity.

[0187] Simultaneously, the deep sequence representation and some statistical features are input into at least one machine learning model for discriminative modeling. The machine learning model includes, but is not limited to, random forest model, gradient boosting decision tree model, or support vector machine model.

[0188] Subsequently, the prediction results of the Mamba model and the machine learning model were fused to obtain the small peptide coding potential prediction results of the candidate small peptide coding regions.

[0189] S402, extract the multi-dimensional sequence features from step S3, the deep sequence representation vector, or a combination of both, and concatenate them as input features for the machine learning model;

[0190] S403, the input features are input into at least one machine learning model for encoding potential discrimination, and the machine learning single-model encoding 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;

[0191] S404 employs one or more of the following methods: weighted combination prediction score, voting integration judgment result, and using the output of multiple models as new features as input to the upper-level model for secondary prediction. The results of the Mamba deep learning network and the machine learning model are then fused to obtain the fusion result.

[0192] S405, construct an output module based on the fusion result, 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 mark the candidate small peptide coding regions whose prediction fusion prediction confidence is lower than the threshold as to be verified and return to re-encode the features.

[0193] S5. After inputting the confidence prediction results from step S4, comprehensively determine the small peptide coding potential of lincRNA. If there is no coding potential, output the conclusion directly. If there is coding potential, deduce the amino acid sequence of the small peptide. After sorting and screening high-confidence small peptides, output the lincRNA coding potential conclusion and at least one of the following: small peptide amino acid sequence, corresponding position, and prediction score.

[0194] In this embodiment, the small peptide coding potential of each lincRNA is comprehensively determined based on the prediction results after fusion.

[0195] For candidate small peptide coding regions identified as having 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 in order to screen and obtain a set of small peptide candidates with high confidence.

[0196] The final output includes:

[0197] Results of small peptide coding potential determination for lincRNAs, and information on corresponding small peptide candidate products.

[0198] like Figure 2 As shown, step S1, obtaining a standardized lincRNA sequence set with uniform format, length, and encoding method, includes:

[0199] Step S101: Obtain the original transcript set data;

[0200] Specifically, raw transcript set data is obtained from transcript sets obtained from public transcriptome databases, gene annotation databases, or high-throughput sequencing data.

[0201] Step S102: Obtain the full-length nucleotide sequence set of lincRNAs from the original transcript set data based on gene annotation information;

[0202] Step S103: Perform quality filtering on the full-length nucleotide sequence set of lincRNA to obtain the filtered full-length nucleotide sequence set of lincRNA;

[0203] Step S104: Standardize the filtered full-length nucleotide sequence set of lincRNA, unify the sequence orientation and convert the base characters into a unified encoding format to obtain a lincRNA sequence set with a unified format;

[0204] Step S105: Based on the target input length threshold, the lincRNA sequence set with uniform format is classified and processed according to the actual sequence length. Long sequences exceeding the threshold are split using a sliding window, and short sequences shorter than the threshold are filled according to fixed alignment rules. The lincRNA sequence is converted into a fixed-length subsequence input unit and its start and end position index in the original sequence is retained. A standardized lincRNA sequence set with uniform format, length and encoding method is output.

[0205] Specifically, a target input length threshold is first set (i.e., the fixed input length expected by the model, preferably 512, 1024, or 2048 nt (nucleotides), consistent with the training and prediction phases). For all sequences in the uniformly formatted lincRNA sequence set, length adaptation processing is performed in two categories based on their actual length. A sliding window is used to split long sequences exceeding the threshold, and short sequences shorter than the threshold are filled using fixed rules. At the same time, effective length markers are generated and subsequence position indices are preserved. Finally, a standardized lincRNA sequence set is output, resulting in length-adapted lincRNA subsequence units. The specific processing rules are as follows:

[0206] Processing of long sequences exceeding the threshold: When the sequence length exceeds the target input length threshold, a sliding window is executed starting from the 5' end of the sequence, with the window length equal to the target input length threshold, continuously truncating subsequences. The window movement step size is a preset value, preferably 1 / 2 of the window length (forming overlapping coverage) or equal to the window length (no overlapping segmentation), ensuring complete coverage and no loss of long sequence information. For residual segments at the end of the sliding window segmentation that are less than a complete window, at least one end rule is used for processing: one is end padding, padding the segment at the 3' end to the target input length threshold; the other is end alignment truncation, truncating a window with a length exactly equal to the target input length threshold from the end of the original sequence. The two rules can also be combined (adding an end alignment window after the regular sliding window) to ensure that the information at the end of the sequence is completely included.

[0207] Short sequence processing (shorter than the threshold): When the sequence length does not exceed the target input length threshold, its original 5'→3' order remains unchanged. It is placed into a fixed-length container according to a unified alignment rule (preferably 5' end alignment or 3' end alignment, consistent during training and prediction). For parts that are shorter than the target length, they are padded at the other end with preset padding symbols (such as "N" placeholders for unknown bases, PAD markers, and other special padding symbols).

[0208] To avoid interference from the padded regions on subsequent feature statistics and deep network modeling, effective length markers (masks) are generated synchronously for all padded sequences to accurately identify the real base positions and padded positions, so that subsequent steps such as base composition calculation, k-mer statistics, and deep model attention / state updates can directly ignore the padded regions.

[0209] Through the above operations, lincRNA sequences of arbitrary length are uniformly converted into one or more fixed-length subsequence input units, and the start and end position indices of each subsequence in the original sequence are preserved throughout the process, providing a foundation for subsequent candidate small peptide coding region localization, result backtracking, and multi-subsequence summary output.

[0210] Step S2 involves integrating, deduplicating, and merging all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the potential translation possibilities of small peptides.

[0211] Specifically, for each standardized lincRNA sequence in the standardized lincRNA sequence set, without presupposing its classical protein coding ability, one or more combinations of three methods are used to identify potential coding fragments: multiple reading frame scanning, segmentation based on local sequence features, and covering window scanning. The construction method of candidate small peptide coding regions is not limited to a single rule. The specific implementation rules of each identification method are as follows:

[0212] Multiple reading frame scanning method: The entire sequence of lincRNA is scanned under different reading frames to identify short open reading frames with a length of ≥20 amino acids, thereby capturing potential small peptide coding regions in the sequence;

[0213] The segmentation method based on local sequence features: Based on the arrangement features, composition characteristics or statistical features of local bases in the sequence, the lincRNA sequence is divided into multiple sequence segments with potential coding possibilities.

[0214] Coverage window scanning method: Using a fixed window or variable window scanning strategy, the entire lincRNA sequence is continuously and uninterruptedly scanned, so that fragments at any position in the sequence are included in the recognition range, ensuring that no potential coding fragments are missed.

[0215] Constructing candidate small peptide coding regions using the aforementioned methods effectively avoids overlooking small peptide coding events arising from non-classical initiations, non-standard lengths, or special translation patterns when relying solely on a single open reading frame rule. This ensures comprehensive coverage of potential small peptide translation possibilities by the candidate small peptide coding regions. After identifying potential coding fragments from all standardized lincRNA sequences, all identified potential coding fragments are integrated, deduplicated, and merged to ultimately obtain a set of candidate small peptide coding regions.

[0216] The Mamba and machine learning collaborative prediction process in step S4 of this invention includes: 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, so as to capture long-range dependencies in candidate small peptide coding regions and output a deep sequence representation vector; on this basis, fusing the deep sequence representation with statistical features to construct machine learning discriminative features, and inputting them into at least one machine learning model to obtain a preliminary discrimination result of coding potential; then filtering or re-encoding candidate small peptide coding regions according to prediction confidence or threshold conditions; finally outputting the small peptide coding potential determination result of candidate small peptide coding regions, including binary classification labels of coding potential and prediction results of candidate small peptide coding regions with confidence.

[0217] like Figure 4 As shown, in step S401, the multi-dimensional sequence feature representation set of the candidate small peptide coding region is input into the trained Mamba deep learning network. Through the continuous evolution of the hidden state of the state space model, the long-range dependency relationship and non-local sequence pattern of each candidate region are captured, and the deep sequence representation vector of each candidate region and the Mamba single model encoding probability are output.

[0218] Step S402: The multi-dimensional sequence features, the deep sequence representation vector, and the combined features of the two are concatenated and used as input features. This input is then fed into at least one pre-trained machine learning model for encoding potential discrimination, and the machine learning single-model encoding probability of each candidate region is output.

[0219] Step S403: Using a weighted combination prediction score fusion method, the Mamba single-model encoding probability and the machine learning single-model encoding probability are linearly summed according to preset weights to calculate the fusion prediction confidence.

[0220] Step S404: Based on the fusion prediction confidence level, output the fusion prediction confidence level value of the candidate small peptide coding region and / or the binary classification label for coding and non-coding regions.

[0221] In this embodiment, the prediction confidence is calculated from the prediction score output by the fusion model. Specifically, the prediction probability scores of the Mamba model and the machine learning model for the same candidate peptide coding region are obtained separately. Then, they are fused according to the weighted combination rule in step S403 to obtain the fused prediction probability as the prediction confidence of the candidate peptide coding region. When the voting arbitration rule is adopted, if the two models make the same judgment, the fused prediction probability is used as the confidence. If the two models make different judgments, the probability of the model whose output probability deviates more from its respective fusion judgment threshold is used as the confidence, thereby reflecting the reliability of the prediction result of the candidate peptide coding region. The threshold is used to convert the prediction confidence into a binary classification judgment result. The threshold is set based on the optimization of the validation set: after training, the threshold is traversed and searched on the validation set to select the threshold that makes the preset evaluation index optimal and save it. The preset evaluation index includes one or more of the following: accuracy, F1 value, sensitivity / specificity balance, or Youden index. When there is sample imbalance or more emphasis is placed on reducing false positives, the threshold that makes the accuracy or specificity reach the preset target can be selected first. The reference range for the threshold is 0.40–0.70, with a preferred range of 0.50–0.60. When used for “high confidence candidate” screening, a higher confidence threshold can be further set, with a reference range of 0.70–0.90 (e.g., 0.75 or 0.80), and candidate peptide coding regions that reach the higher confidence threshold will be used as the preferred set of small peptide candidates for output.

[0222] Step S405: Candidate small peptide coding regions with fusion prediction confidence below the threshold are marked as samples to be verified and returned to step S3 for re-encoding of features.

[0223] S5, Prediction Results Output and Small Peptide Candidate Product Identification Process.

[0224] like Figure 5 As shown, the prediction result output and small peptide candidate product determination process in step S5 of the present invention includes: based on the prediction results of candidate small peptide coding regions with confidence, a comprehensive judgment is made on the overall small peptide coding potential of lincRNA; when it is determined that there is no small peptide coding potential, the conclusion of no small peptide coding potential is directly output; when it is determined that there is coding potential, the candidate small peptide coding region is translated and deduced to obtain the corresponding small peptide amino acid sequence, and the candidate small peptides are sorted and screened in combination with indicators such as prediction confidence, sequence conservation and length, and finally, high-confidence small peptide candidate products and their corresponding coding potential information are output.

[0225] Step S501: Input the prediction results of the candidate small peptide coding region.

[0226] Input the fusion confidence score and / or binary classification label for each candidate peptide coding region obtained from the preceding steps.

[0227] Step S502: Comprehensively determine the overall coding potential of lincRNA.

[0228] Based on the prediction results of the candidate small peptide coding regions, a comprehensive determination is made as to whether the corresponding lincRNA has the potential to encode small peptides.

[0229] Step S503, output the no-coding potential determination.

[0230] When the overall judgment result indicates that there is no high-confidence candidate small peptide coding region, the conclusion "no small peptide coding potential" is output and the process ends.

[0231] Step S504, small peptide sequence derivation.

[0232] When a lincRNA is determined to have the potential to encode a small peptide, the coding region of the candidate small peptide is translated to obtain the corresponding small peptide amino acid sequence.

[0233] Step S505: Sort candidate small peptides.

[0234] Candidate peptides are ranked by weight based on at least one of the following indicators: prediction confidence, sequence conservation, and peptide length.

[0235] Step S506: Screening of high-confidence candidate small peptides.

[0236] A set of small peptide candidates with high confidence was obtained from the sorting results.

[0237] Step S507, output the result.

[0238] Output the coding potential of lincRNA and information on high-confidence small peptide candidate products, including the small peptide sequence, corresponding position, and prediction score. Example

[0239] Prediction of candidate small peptide coding regions, confidence determination, and output of candidate small peptides for unknown lincRNAs

[0240] This embodiment takes an unknown full-length nucleotide sequence LINC_Test_001 as input, with a sequence length of 1800 nt. Sequence standardization, candidate small peptide coding region construction and feature extraction are completed according to steps S1–S3, and candidate small peptides and sorting results are output according to steps S404, S205 and S5.

[0241] (1) Step S105 (Length adaptation, fixed 1024nt input)

[0242] In this embodiment, the target input length threshold is fixed at 1024 nt. When the input sequence length exceeds 1024 nt, a sliding window is used for splitting, with a fixed window length of 1024 nt and a fixed step size of 512 nt. End alignment is used to supplement the end window to cover the information at the end of the sequence. Therefore, LINC_Test_001 generates the following fixed-length subsequence windows:

[0243] W1 (1–1024nt), W2 (513–1536nt), W3 (777–1800nt).

[0244] (2) Step S2 (Construction of candidate small peptide coding regions)

[0245] Multiple reading frame scans and overlay scans are performed on W1, W2, and W3 respectively to generate a set of candidate small peptide coding regions. Duplicate candidates are merged and deduplicated (only one of the same or highly overlapping candidates is retained, and its coordinates in the full-length sequence are preserved for backtracking). This embodiment exemplarily yields three candidate small peptide coding regions as follows (all coordinates are full-length coordinates, and the length corresponds to the peptide length in an integer divisible relationship):

[0246] C1: Full length 600–824nt, length 225nt, corresponding to a small peptide length of 75aa (amino acids), starting with CTG (non-classical initiation).

[0247] C2: Full length 210–299nt, length 90nt, corresponding to a small peptide length of 30aa (amino acids), starting with ATG (classical initiation).

[0248] C3: Full length 1200–1379nt, length 180nt, corresponding to a small peptide length of 60aa (amino acids), starting with GTG (non-classical initiation).

[0249] Among them, C1 and C2 fall within the coverage area of ​​window W1, and C3 falls within the coverage area of ​​window W2 / W3; the candidate small peptide coding region set is obtained by the above scanning and deduplication rules, which can cover small peptide coding events that may be generated under non-classical initiation and different length ranges.

[0250] (3) Step S3 (feature extraction and context construction, matching 1024 inputs)

[0251] Sequence features are extracted for each candidate small peptide coding region, and context-related features are constructed. In this embodiment, the upstream and downstream windows are fixed at 128 nt upstream and 128 nt downstream; padding is used for any insufficient areas.

[0252] For the Mamba model: For each candidate small peptide coding region, a fixed-length 1024nt input fragment is constructed, with the construction rule being "upstream 128nt + candidate small peptide coding region body + downstream 128nt". The remaining length is filled according to the rule and ignored by the effective length marker, so that the input length is constant at 1024nt. If the candidate small peptide coding region body is too long to accommodate, the candidate small peptide coding region body is first split according to the sliding window rule of S105, and then multiple 1024nt input fragments are constructed and summarized for output.

[0253] For the machine learning model (RF): extract similar statistical features (including base composition, k-mer / quasi-codon frequency, complexity / information entropy, etc.) from the candidate small peptide coding region, upstream window and downstream window respectively, and then concatenate them with the start and end relative position features of the candidate small peptide coding region in the whole length to form a unified feature vector as input.

[0254] (4) Step S404 (Dual model prediction and fusion, fixed weights and thresholds)

[0255] The trained and saved Mamba and RF models are called to output the coding probability of each candidate small peptide coding region, and the prediction scores are fused using a weighted combination. The fusion weights are fixed at Mamba=0.60 and RF=0.40. The fused output probability is used as the prediction confidence of the candidate small peptide coding region.

[0256] (5) Step S205 (Prediction confidence judgment and threshold)

[0257] 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 positive coding potential; otherwise, it is judged as negative. At the same time, the voting arbitration rule is fixed and saved as a backup: when the binary classification results of the two models are inconsistent, the label corresponding to the one whose output probability deviates more from the threshold is selected as the final judgment.

[0258] (6) Step S5 (output and sorting of candidate small peptides)

[0259] For candidate peptide coding regions identified as positive, small peptide derivation is performed and a candidate list is output. When multiple positive candidates exist for the same lincRNA, they are sorted and output according to the fixed weight rule in step S505, and the Top-N candidate small peptides and their corresponding full-length coordinates and confidence scores are output.

[0260] (7) Output field template (output format of example)

[0261] The system output in this embodiment includes at least: lincRNA_ID, candidate number, full-length coordinates, small peptide sequence, small peptide length, fusion prediction confidence score (Coding probability score), and ranking; an example is shown in the table below (the probability values ​​are exemplary outputs used to illustrate the process and fields, and are not limited to specific values), as shown in Table 1 below:

[0262] Table 1. Prediction results of candidate small peptides in LINC_Test_001

[0263] lincRNA_ID Candidate Number Total coordinates (nt) Small peptide sequence Small peptide length (aa) Fusion prediction confidence Ranking LINC_Test_001 C1 600–824 NEAPI 75 0.788 1 LINC_Test_001 C2 210–299 FVPPE 30 0.49 — LINC_Test_001 C3 1200–1379 KRVEV 60 0.662 2

[0264] Step S6: Determining Model Training and Fusion Strategies

[0265] 1. Example of integrating voting results with secondary predictions from the upper-level model

[0266] Example 1: Voting Integration Determination Results

[0267] In this embodiment, the voting integration decision result improves the final prediction accuracy by fusing the prediction results of the Mamba model and the machine learning model. The specific steps are as follows:

[0268] Step 1.1: Model Output Determination

[0269] For each candidate region, the encoding potential probability is calculated using both the Mamba model and a machine learning model. If the predicted probability is greater than or equal to a threshold (threshold = 0.5), it is determined to be "encoding potential"; otherwise, it is "non-encoding".

[0270] Step 1.2: Voting Consolidation Determination

[0271] When the predictions of the two models are consistent, that prediction is directly output as the final decision. If the predictions of the two models are inconsistent, arbitration is conducted in the following manner:

[0272] Calculate the deviation of the output probabilities of the Mamba model and the machine learning model from 0.50, and select the label corresponding to the one with the larger deviation as the final result.

[0273] The final determination label is output and used as the prediction result for candidate small peptide coding regions.

[0274] Example 2: Secondary prediction using an upper-layer model (Stacking / Meta-model)

[0275] In this embodiment, the outputs of multiple base models are used as new features to input into the upper-layer model for secondary prediction, thereby optimizing the prediction boundary.

[0276] Step 2.1: Construct upper-layer input features

[0277] The probability values, binary classification labels, and confidence biases output by the Mamba model and the machine learning model are used as input features for the upper-level model, forming a feature vector.

[0278] z = (pMamba, pML, yMamba, yML, dMamba, dML, stat)

[0279] Where: pMamba and pML are predicted probabilities; yMamba and yML are predicted labels; dMamba and dML are confidence biases; stat is supplementary enhancement features (such as GC content of candidate small peptide regions, sequence length, etc.).

[0280] Step 2.2: Train the upper-layer model

[0281] Use the training data to train the upper-level model (such as logistic regression, SVM, random forest, etc.) to minimize the classification loss or maximize performance metrics such as AUC / F1. Save the model parameters after training.

[0282] Step 2.3: Inference Output

[0283] The upper-level model receives new prediction inputs (such as the outputs of the Mamba model and the machine learning model) and performs secondary predictions, outputting the fused prediction probability pMeta as the final judgment result.

[0284] 2. Clarify the final vector length and tensor dimensionality arrangement after multi-feature concatenation.

[0285] 2.1 Length of the final vector after multi-feature concatenation

[0286] In this embodiment, multiple features (such as base level, quasi-codon frequency, complexity features, etc.) are concatenated into a unified vector. The dimensions of each feature are as follows:

[0287] Base level characteristics: A / C / G / T ratio + GC content, a total of 5 dimensions.

[0288] Quasi-codon (3-mer) frequencies: 64 triplet normalized frequencies, totaling 64 dimensions.

[0289] Complexity and Information: Low complexity ratio 1 dimension + repetition 1 dimension + information entropy 1 dimension + triplet information entropy 1 dimension, for a total of 4 dimensions.

[0290] Location features: the starting and ending positions of the candidate region, totaling 2 dimensions.

[0291] Therefore, the vector dimension of each candidate region is: 5 + 64 + 4 + 2 = 75 dimensions.

[0292] Considering the concatenation of candidate regions, upstream windows, and downstream windows, the final vector length is: 75 (candidate regions) + 75 (upstream windows) + 75 (downstream windows) = 225 dimensions.

[0293] 2.2 Tensor Dimensional Arrangement

[0294] For the Mamba model, a fixed input length of 1024nt is used. The tensor dimension of the Mamba model is: batch_size × 1024 × feature dimension.

[0295] Feature dimension: There are 5 channels at each nt position (one-hot encoding of A / C / G / T and effective length marker), so the feature dimension is 5.

[0296] Tensor Dimension: Assuming each batch (batch_size) contains 32 samples, the final tensor dimension is: 32×1024×5.

[0297] 3. Compare the performance differences between "this invention and Mamba used alone".

[0298] This invention has the following advantages compared to using only the Mamba model:

[0299] Improved accuracy: By fusing multiple models, the advantages of each model can be combined to improve the accuracy of predicting coding potential. On the validation set, this invention improves the accuracy by 3-5 percentage points compared to using only the Mamba model.

[0300] Long sequence processing efficiency: Although the Mamba model has advantages in modeling long sequences, this invention further improves the processing accuracy of long sequences by using a machine learning model to perform supplementary predictions at the candidate region level, while maintaining high efficiency.

[0301] Advantages of fusion: Multi-model fusion allows the advantages of each model to complement each other, thereby improving the overall robustness and performance.

[0302] 4. Differences from traditional multi-model fusion solutions

[0303] This invention differs significantly from traditional multi-model fusion schemes (such as those without fixed weights or low-confidence recoding) in the following ways:

[0304] Standardized feature extraction: This 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 feature inconsistency problem in traditional schemes.

[0305] Fixed-weight sorting: Unlike the dynamic weights in traditional schemes that may change with batches or datasets, this invention determines and fixes the weights through validation set optimization, making the fusion results more stable and reproducible.

[0306] Low-confidence recoding: Traditional schemes often handle low-confidence samples in a relatively simple way, while this invention uses a low-confidence backoff mechanism to return low-confidence samples to the feature encoding stage, thereby triggering more feature extraction and more stringent judgment, thus reducing the risk of misjudgment.

[0307] 5. Applicable scenarios and learning rate adjustment logic of “512 / 1024 / 2048nt”

[0308] 5.1 Applicable Scenarios for Input Length

[0309] 512nt: Suitable for short sequences, especially when the candidate region is short. Also suitable for scenarios requiring fast processing when dealing with large amounts of data.

[0310] 1024nt: As a standard length, it is suitable for most datasets, providing sufficient contextual information while maintaining computational efficiency.

[0311] 2048nt: Suitable for long sequences or sequences with strong upstream and downstream dependencies, it can reduce the number of sliding windows, thereby improving processing efficiency.

[0312] 5.2 Learning Rate Adjustment Logic

[0313] Low learning rate (1×10^-4 ~ 2×10^-4): When the amount of data is small or the samples are imbalanced, using a lower learning rate can improve the stability of the model and avoid overfitting.

[0314] Medium learning rate (2×10^-4 ~ 3×10^-4): When the amount of data is moderate and there is no significant imbalance of samples, using a medium learning rate can accelerate model convergence.

[0315] High learning rate (3×10^-4 ~ 5×10^-4): When the amount of data is large and the data is relatively balanced, using a higher learning rate can accelerate the training process and reduce convergence time.

[0316] 5.3 Scope of application of other parameters

[0317] batch_size: The recommended range is 16–128. When video memory or memory is limited, a smaller batch_size (e.g., 16–32) is suitable; when resources are sufficient, a larger batch_size (e.g., 64–128) can be used.

[0318] 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) should be used.

[0319] Step S6 is the model training and saving process. It involves constructing a labeled training dataset, performing preprocessing and feature encoding consistent with the prediction stage, training the Mamba deep learning network and machine learning model separately, optimizing the multi-model fusion strategy, and finally saving reusable model parameters and rules for predicting the coding potential of unknown lincRNA sequences. The specific sub-steps are as follows:

[0320] S601, Construct a training dataset with encoded potential annotations;

[0321] Using "candidate small peptide coding regions" as the basic training sample unit, each lincRNA sequence generates one or more candidate samples after candidate region identification in step S2. Each sample is independently assigned a binary label (coding potential = 1 / non-coding = 0). If necessary, transcript-level labels can be summarized according to "if any candidate region meets the criteria, the transcript is positive".

[0322] Positive samples (coding potential = 1): must have traceable coding evidence, including translation evidence in ribosome binding / translation initiation data, detection of corresponding peptides in mass spectrometry of translation products, or being located in an authoritative set of translatable lincRNA / micropeptide coding regions (homologous conserved and with consistent coding signals can be used as weak positive samples and their evidence level recorded).

[0323] Negative samples (coding potential = 0): These include at least three categories: first, real background negative samples annotated as lincRNAs with no translation evidence; second, matching control negative samples with length and GC content (optional dinucleotide / triple distribution) similar to positive samples; and third, difficult negative samples with longer ORFs or coding-pattern triple preference (or fragments constructed from coding sequences through frameshifting / breaking start and stop structures), used to suppress shortcut learning of "long ORF = coding".

[0324] Dataset processing: The dataset is divided into training set (70%), validation set (20%), and test set (10%) based on genes / transcripts to ensure that candidate regions of the same transcript do not cross sets. Clustering is performed to remove duplicates based on sequence similarity thresholds. The ratio of positive to negative samples is balanced using class weights, focal loss, or stratified sampling. Each sample stores metadata such as transcript / gene identifier, candidate region sequence and start / end position / chain direction / reading frame, label, evidence type and level (positive sample).

[0325] S602, Preprocessing and Feature Encoding of Training Data

[0326] Perform the same processing procedure as steps S1-S3 on the training dataset to ensure consistency in data distribution between the training and prediction phases:

[0327] Sequence preprocessing: Perform quality filtering, standardization, and length adaptation (sliding window splitting for long sequences, padding for short sequences, and generating effective length markers) based on the target input length threshold of 1024nt.

[0328] Candidate region construction: A combination of multiple reading frame scanning, local feature-based segmentation, and covering window scanning is used to identify short open reading frames corresponding to ≥20 amino acids, which are then integrated and deduplicated to form a set of candidate small peptide coding regions;

[0329] Multi-dimensional feature encoding: Extract features at the base level (base composition, k-mer frequency), codon / quasi-codon (64 triplet normalized frequencies), sequence complexity and information (low complexity ratio, information entropy), and contextual association (upstream 128nt + candidate region + downstream 128nt), encode them into tensor form adapted to Mamba and vector form adapted to machine learning models, and output a multi-dimensional sequence feature representation set.

[0330] S603, Mamba deep learning network training;

[0331] The tensor features output by S602 are input into the Mamba deep learning network, and the encoded potential labels of candidate regions are used as supervision signals to iteratively optimize the model parameters:

[0332] The model has the following fixed parameters: embedding dimension 256, network layers 6, hidden dimension 512, dropout=0.1, and the output layer is a binary classification fully connected layer (Sigmoid activation, output encoding probability 0-1).

[0333] Training hyperparameters: Optimizer AdamW, learning rate 3×10 -4 The batch size is 64, the maximum number of training epochs is 50, the loss function is binary cross-entropy with class weights, and the early stopping strategy is to stop if the validation set AUC does not improve for 10 consecutive epochs.

[0334] Training process: Input features in batches, capture long-range dependencies through continuous evolution of hidden states in the state-space model, output a deep sequence representation vector (512 dimensions), calculate the loss between the predicted probability and the true label, backpropagate to update parameters, and save the model weights with the optimal AUC on the validation set.

[0335] S604, Machine Learning Model Training;

[0336] Train at least one machine learning model using the statistical features of S602, the deep sequence representation vector of S603, or a combination of both as input:

[0337] Model selection: Random Forest (RF) was selected with the following parameters: number of trees in the forest n_estimators=100, maximum depth max_depth=10, node splitting criterion="gini", minimum number of samples per leaf node min_samples_leaf=1, and maximum number of features to consider when splitting nodes max_features="sqrt".

[0338] Training process: After the input features are standardized by Z-score, multiple independent decision trees are constructed through random forest to form an ensemble model. The model is trained on the training set and the F1 score is calculated on the validation set. After training is completed, the model file is saved and the feature importance ranking results are output.

[0339] S605: Multi-model fusion strategy training;

[0340] Optimize the fusion rules of the Mamba model and the machine learning model on the validation set to determine the final prediction strategy:

[0341] Integration method: The weighted combination prediction score is used as the main integration method, supplemented by voting arbitration rules;

[0342] Weight determination: Traverse the Mamba weights (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).

[0343] Threshold determination: Iterate through the thresholds (0.40-0.70, step size 0.01), and select the threshold with the largest Youden index (0.50 in this example) as the criterion for binary classification of encoded / non-encoded.

[0344] Arbitration rule: When the two models make inconsistent decisions, the "arbitration of the model with the greater deviation from the threshold" shall be adopted, that is, the absolute deviation of the single model probability from the threshold of 0.50 shall be compared, and the decision of the model with the greater deviation shall be the final result.

[0345] S606: Model and fusion strategy storage;

[0346] The trained model parameters and fusion strategy are persistently stored in a callable format to form a complete model package:

[0347] Saved contents: Mamba model weight file (.pth) and configuration file (.yaml), RF model file (.model) and feature normalization parameters, fusion weight and threshold configuration file, voting arbitration rule script, feature extraction and encoding script, and result output template;

[0348] Application scenario: After preprocessing unknown lincRNA sequences with S1-S3, this model package can be directly called to output fusion prediction confidence, binary classification judgment labels, and a set of high-confidence small peptides, realizing end-to-end coding potential identification.

[0349] In this embodiment of the invention, 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 data file or visualization result. The system can be deployed on general computer equipment, servers or cloud computing environment.

[0350] Figures 5-10 This is a schematic diagram of the lincRNA small peptide coding potential identification system provided in an embodiment of the present invention.

[0351] The lincRNA small peptide coding potential identification system of this invention includes:

[0352] Standardized lincRNA sequence acquisition module 1 is used to acquire a set of standardized lincRNA sequences with uniform format, length, and encoding method;

[0353] For details on the implementation of the standardized lincRNA sequence acquisition module 1, please refer to the description of step S1 above.

[0354] The standardized lincRNA sequence acquisition module 1 includes: 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 preservation module 105. The specific implementation process of each of the above modules can be referred to the relevant descriptions of steps S101, S102, 103, S104, and 105 above.

[0355] Candidate small peptide coding region construction module 2 is used to integrate, remove duplicates and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the translation possibilities of potential small peptides;

[0356] The multi-dimensional sequence feature encoding module 3 is used to extract multiple types of 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set.

[0357] The multi-dimensional sequence feature encoding module 3 includes: 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 processes of each of the above modules can be referred to the relevant descriptions of steps S301, S302, and S303 above.

[0358] The encoding potential fusion prediction module 4 is used to input the multi-dimensional sequence feature representation set of candidate small peptide coding regions into the trained Mamba deep learning network, output the deep sequence representation vector and the Mamba single model encoding 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 the machine learning single model encoding probability, then linearly sum them according to preset weights to obtain the fusion prediction confidence, output the value and / or the binary classification judgment label of encoding and non-coding, and finally mark the candidate regions with fusion prediction confidence below the threshold as samples to be verified and return to step S3 for re-feature encoding;

[0359] The coding potential fusion prediction module 4 includes: 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 and binary classification label output module 404, and a low-confidence sample labeling and feature coding backoff module 405. The specific implementation process of each of the above modules can be referred to the relevant descriptions of steps S401, S402, S403, S404, and S405 above.

[0360] The small peptide coding potential determination and screening module 5 is used to input the fusion prediction confidence and coding-non-coding binary classification label of each candidate small peptide coding region. Based on the determination results, it comprehensively judges the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, it outputs the conclusion of no coding potential. If there is, it derives the small peptide amino acid sequence for the coding positive region. Then, it calculates the comprehensive ranking confidence by combining the fusion prediction confidence and multiple evaluation indicators with fixed weights. After weighting and ranking multiple positive candidate small peptides from high to low, the candidate small peptides with a comprehensive ranking confidence higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set.

[0361] 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 no-coding potential result output module 503, a coding positive region amino acid sequence deduction module 504, a positive candidate peptide weighted sorting and confidence calculation module 505, a high and low confidence candidate peptide classification module 506, and a coding potential conclusion and high confidence peptide information output module 507. The specific implementation process of each module can be referred to the relevant descriptions of steps S501, S502, S503, S504, S505, S506, and S507 above.

Claims

1. A method for identifying the coding potential of lincRNA small peptides, characterized in that, include: S1, Obtain a standardized lincRNA sequence set with uniform format, length, and encoding method; S2, integrate, remove duplicates and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the potential translation possibilities of small peptides; S3, extract multi-class 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set; 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 the Mamba single model coding probability. Then, concatenate the multi-dimensional sequence features, the deep sequence representation vector and the combined features of the two and input them into at least one trained machine learning model to output the machine learning single model coding probability. Then, linearly sum the results according to the preset weights to obtain the fusion prediction confidence. Output the fusion prediction confidence and / or the encoded and non-encoded binary classification judgment label. Finally, mark the candidate regions with fusion prediction confidence below the threshold as samples to be verified and return to step S3 to re-encode the features. S5. Input the fusion prediction confidence and coding / non-coding binary classification label of each candidate small peptide coding region. Based on the judgment results, comprehensively judge the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, output the conclusion that there is no coding potential. If there is, deduce the small peptide amino acid sequence for the coding positive region. Then, combine the fusion prediction confidence with multiple evaluation indicators with fixed weights to calculate the comprehensive ranking confidence. After weighting and sorting multiple positive candidate small peptides from high to low, candidate small peptides with comprehensive ranking confidence higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set. The sub-steps of step S4 include: S401: Input the multi-dimensional sequence feature representation set of candidate small peptide coding regions into the trained Mamba deep learning network. Through the continuous evolution of the hidden state of the state space model, capture the long-range dependency relationship and non-local sequence pattern of each candidate region, and output the deep sequence representation vector of each candidate region and the Mamba single model encoding probability. S402, the multi-dimensional sequence features from step S3, the deep sequence representation vector, and the combined features of the two are concatenated and input into at least one trained machine learning model for encoding potential discrimination, and the machine learning single-model encoding probability of each candidate region is output. S403 uses a weighted combination prediction score fusion method, which linearly sums the Mamba single-model encoding probability and the machine learning single-model encoding probability according to preset weights to calculate the fusion prediction confidence. S404, based on the fusion prediction confidence, output the fusion prediction confidence value of the candidate small peptide coding region and / or the binary classification label for coding and non-coding; S405, candidate peptide coding regions with fusion prediction confidence below the threshold are marked as samples to be verified and returned to step S3 for feature encoding again.

2. The method according to claim 1, characterized in that, The sub-steps of step S1 include: S101, Obtain the original transcript set data; S102, the set of full-length nucleotide sequences of lincRNAs was obtained by screening the original transcript set data based on gene annotation information; S103, quality filtering is performed on the full-length nucleotide sequence set of lincRNA to obtain the filtered full-length nucleotide sequence set of lincRNA; S104. The filtered full-length nucleotide sequence set of lincRNA is standardized to unify the sequence orientation and convert the base characters into a unified encoding format, resulting in a lincRNA sequence set with a unified format. S105. Based on the target input length threshold, the lincRNA sequence set with uniform format is classified and processed according to the actual sequence length. Long sequences exceeding the threshold are split using a sliding window, and short sequences shorter than the threshold are filled according to fixed alignment rules. The lincRNA sequence is converted into a fixed-length subsequence input unit and its start and end position index in the original sequence is retained. The output is a standardized lincRNA sequence set with uniform format, length and encoding method.

3. The method according to claim 1, characterized in that, 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, and contextual association features; S302 encodes the extracted multi-dimensional sequence features in vector or tensor form to form a model-learnable feature representation that does not require manual setting of fixed thresholds; S303 outputs the multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions.

4. The method according to claim 1, characterized in that, The sub-steps of step S5 include: S501, Input the fusion prediction confidence value of each candidate small peptide coding region and the binary classification label for coding and non-coding; S502, based on the binary classification of coding and non-coding regions of candidate small peptide coding regions, comprehensively determines whether the corresponding lincRNA has the potential to code small peptides. S503, if determined to have no coding potential, output lincRNA with no small peptide coding potential; S504. If a candidate peptide is determined to have coding potential, the coding region of the candidate peptide that is determined to be coding positive is translated according to the codon translation rules of the corresponding reading frame to obtain the amino acid sequence of the candidate peptide. S505: Multiple positive candidate peptides of the same lincRNA are weighted and sorted, and the overall ranking confidence is calculated according to a fixed weight rule, and then sorted from high to low score. S506, candidate small peptides with a comprehensive ranking confidence score higher than the comprehensive ranking confidence score threshold are classified into the high confidence set, and the rest are classified into the low confidence set; S507, outputs the lincRNA encoding potential conclusion and high-confidence peptide information; the high-confidence peptide information includes: the amino acid sequence of the high-confidence peptide, the corresponding full-length coordinates, the fusion prediction confidence, the comprehensive ranking confidence, and the ranking rank.

5. A system for identifying the coding potential of lincRNA small peptides, characterized in that, include: The standardized lincRNA sequence acquisition module is used to acquire a set of standardized lincRNA sequences with uniform format, length, and encoding method. The candidate small peptide coding region construction module is used to integrate, remove duplicates, and merge all potential coding fragments identified from the standardized lincRNA sequence set to obtain a set of candidate small peptide coding regions covering the translation possibilities of potential small peptides. The multi-dimensional sequence feature encoding module is used to extract multiple types of 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 representations that the model can learn, and output the corresponding multi-dimensional sequence feature representation set. The encoding potential fusion prediction module is used to input the multi-dimensional sequence feature representation set of candidate small peptide encoding regions into the trained Mamba deep learning network, output the deep sequence representation vector and the Mamba single model encoding probability, then concatenate the multi-dimensional sequence features, the deep sequence representation vector and the combined features of the two and input them into at least one trained machine learning model to output the machine learning single model encoding probability, then linearly sum them according to preset weights to obtain the fusion prediction confidence, output the fusion prediction confidence and / or the encoded and non-encoded binary classification judgment label, and finally mark the candidate regions with fusion prediction confidence below the threshold as samples to be verified and return to step S3 for re-encoding features; The small peptide coding potential determination and screening module is used to input the fusion prediction confidence and coding-non-coding binary classification label of each candidate small peptide coding region. Based on the determination results, it comprehensively judges the small peptide coding potential of the corresponding lincRNA. If there is no coding potential, it outputs the conclusion of no coding potential. If there is, it derives the small peptide amino acid sequence from the coding positive region. Then, it calculates the comprehensive ranking confidence by combining the fusion prediction confidence and multiple evaluation indicators with fixed weights. After weighting and ranking multiple positive candidate small peptides from high to low, the candidate small peptides with a comprehensive ranking confidence higher than the threshold are classified into the high confidence set, and the rest are classified into the low confidence set. The coding potential fusion prediction module includes: The Mamba model feature modeling and encoding probability output module is used to input the multi-dimensional sequence feature representation set of candidate small peptide encoding regions into the trained Mamba deep learning network. Through the continuous evolution of the hidden state of the state space model, it captures the long-range dependencies and non-local sequence patterns of each candidate region and outputs the deep sequence representation vector of each candidate region and the encoding probability of the Mamba single model. The machine learning model encoding potential discrimination module is used to concatenate the multi-dimensional sequence features of the Mamba model feature modeling and encoding probability output module, the deep sequence representation vector, and the combined features of the two into at least one trained machine learning model for encoding potential discrimination, and output the machine learning single model encoding probability of each candidate region. The multi-model encoding probability fusion and confidence calculation module is used to calculate the fusion prediction confidence by linearly summing the Mamba single-model encoding probability and the machine learning single-model encoding probability according to preset weights using a weighted combination prediction score fusion method. The fusion result numerical and binary classification label output module is used to output the fusion prediction confidence value of the candidate small peptide coding region and / or the coding and non-coding binary classification judgment label according to the fusion prediction confidence. The low-confidence sample labeling and feature encoding backoff module is used to label candidate small peptide coding regions with fusion prediction confidence below the threshold as samples to be verified and return to step S3 to re-encode features.

6. The system according to claim 5, characterized in that, The standardized lincRNA sequence acquisition module includes: The raw transcript data acquisition module is used to acquire raw transcript set data; The lincRNA sequence screening module is used to screen and obtain the full-length nucleotide sequence set of lincRNAs from the original transcript set data based on gene annotation information. The lincRNA sequence quality filtering module is used to filter the full-length lincRNA nucleotide sequence set to obtain the filtered full-length lincRNA nucleotide sequence set. The lincRNA sequence normalization module is used to normalize the filtered full-length lincRNA nucleotide sequence set, unify the sequence orientation, and convert the base characters into a uniform encoding format to obtain a lincRNA sequence set with a uniform format. The lincRNA sequence length adaptation and index preservation module is used to classify and process a uniformly formatted lincRNA sequence set according to the actual length of the sequence based on the target input length threshold. Long sequences exceeding the threshold are split using a sliding window, while short sequences shorter than the threshold are padded according to fixed alignment rules. The module converts the lincRNA sequence into a fixed-length subsequence input unit and preserves its start and end position indices in the original sequence. It outputs a standardized lincRNA sequence set with uniform format, length, and encoding method.

7. The system according to claim 5, characterized in that, The multi-dimensional sequence feature encoding module includes: The multi-dimensional sequence feature extraction module is used to extract multi-dimensional sequence features for each candidate small peptide coding region in the candidate small peptide coding region set. The multi-dimensional sequence features include one or more of the following: base level, codon or quasi-codon, sequence complexity and information, and contextual features. The sequence feature vector tensor encoding module is used to encode the extracted multi-dimensional sequence features in vector or tensor form, forming a model-learnable feature representation that does not require manual setting of fixed thresholds; The multi-dimensional sequence feature representation set output module is used to output the multi-dimensional sequence feature representation set corresponding to all candidate small peptide coding regions.

8. The system according to claim 5, characterized in that, The small peptide encoding potential determination and screening module includes: The fusion prediction result input module is used to input the fusion prediction confidence value of each candidate small peptide coding region and the binary classification label for coding and non-coding. The lincRNA coding potential comprehensive determination module is used to comprehensively determine whether the corresponding lincRNA has small peptide coding potential based on the coding and non-coding binary classification results of the candidate small peptide coding region. The no-coding-potential result output module is used to output "lincRNA has no small peptide coding potential" if it is determined to have no coding potential. The amino acid sequence derivation module for positive coding regions is used to deduce the amino acid sequence of candidate small peptides by translating the coding regions of the candidate small peptides according to the codon translation rules of the corresponding reading frame if the region is determined to have coding potential. The positive candidate peptide weighted ranking and confidence calculation module is used to perform weighted ranking of multiple positive candidate peptides of the same lincRNA, calculate the comprehensive ranking confidence according to fixed weight rules, and sort them from high to low scores. The high and low confidence candidate peptide classification module is used to classify candidate peptides with a comprehensive ranking confidence score higher than the comprehensive ranking confidence score threshold into the high confidence set, and the rest into the low confidence set. The module for outputting coding potential conclusions and high-confidence peptide information is used to output lincRNA coding potential conclusions and high-confidence peptide information. The high-confidence peptide information includes: the amino acid sequence of the high-confidence peptide, its corresponding full-length coordinates, fusion prediction confidence, comprehensive ranking confidence, and ranking position.

Images