Method and system for generating and predicting function of mRNA untranslated region sequence conditioned on coding sequence

Abstract

The application provides a method and system for generating and predicting the function of mRNA untranslated region sequence based on coding sequence. The method comprises: constructing a pre-training data set and a downstream task data set; using the pre-training data set to perform autoregressive training on the constructed conditional generation model to obtain a UTR sequence generation model; using the downstream task data set to fine-tune the UTR sequence generation model to obtain a downstream task function prediction model; generating candidate UTR sequences based on the UTR sequence generation model, and evaluating the generation ability of the UTR sequence generation model in the non-coding region sequence generation task; and predicting the function attribute of the UTR sequence based on the downstream task function prediction model, and evaluating the prediction ability of the downstream task function prediction model in multiple UTR related downstream tasks. The application considers the synergistic relationship between UTR and CDS in the UTR sequence generation process, and improves the efficiency and rationality of UTR sequence design.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and genetic engineering technology, specifically relating to a method and system for generating and predicting the function of mRNA untranslated region sequences based on coding sequences. Background Technology

[0002] Messenger RNA (mRNA) typically consists of a 5′ untranslated region (5′ UTR), a coding sequence (CDS), and a 3′ untranslated region (3′ UTR). While the UTR sequence does not participate in protein coding, it plays a crucial role in regulating mRNA stability, translation efficiency, and protein expression levels. Therefore, the rational design of UTR sequences is of great significance in gene expression regulation, mRNA drug design, and protein expression optimization.

[0003] In existing technologies, the design and optimization of UTR sequences mainly include the following methods: The first type of method is the natural sequence screening method, which involves screening natural UTR sequences with high expression efficiency from public databases or literature and applying them directly to the construction of target gene expression, or modifying the sequence to a certain extent on this basis.

[0004] The second category of methods consists of machine learning or deep learning-based approaches. These methods typically utilize existing UTR sequence data and their corresponding translation efficiency or expression level data to train predictive models, thereby predicting or scoring UTR sequence function and providing a reference for UTR sequence optimization. In recent years, with the accumulation of high-throughput sequencing technology and translatomics data, some deep learning models have been used to predict UTR sequence-related functional attributes, such as mean ribosome load (MRL) or protein expression levels, thus providing technical means for data-driven UTR design.

[0005] Although existing technologies can optimize or predict the function of UTR sequences through sequence screening or machine learning models, certain limitations still exist: Translation-related phenotypic metrics (such as protein expression level (EL) and translation efficiency (TE)) typically arise from the synergistic effect of the UTR and CDS. Existing experimental studies have shown that the same UTR sequence can produce significantly different expression outputs in different CDS contexts, while different CDS may exhibit different translation efficiencies in the same UTR environment. This context-dependent relationship indicates that the regulatory role of the UTR is not entirely independent but is influenced by the coding sequence background.

[0006] However, most existing UTR design or functional prediction models typically treat the UTR as a relatively independent regulatory module, failing to explicitly consider the potential statistical dependencies and functional associations between it and the CDS. For example, UTR-LM primarily predicts translational function based on the UTR sequence itself, without incorporating the corresponding CDS sequence as a conditional variable into the modeling process, thus neglecting the potential interaction between the UTR and the coding sequence during model construction. Ignoring the interaction between the UTR and CDS during UTR sequence modeling or design may lead to models that fail to accurately characterize translational regulatory mechanisms in real biological systems, thereby affecting the reliability of prediction or design results and limiting their generalization ability and application effectiveness in practical mRNA sequence design and optimization.

[0007] In addition, in the existing technology, there are still relatively few methods that can simultaneously perform generative design for UTR sequences and conditional modeling in combination with encoded sequence information, which to some extent limits the efficiency and flexibility of UTR sequence design in practical applications. Summary of the Invention

[0008] To address the shortcomings of existing technologies in mRNA untranslated region (UTR) sequence design and functional modeling, such as insufficient utilization of coding and untranslated region (UTR) sequence features, inadequate mining of UTR-CDS association information, lack of further application of sequence features learned by pre-trained models for downstream functional prediction, and the lack of functional evaluation and screening mechanisms for UTR sequence generation, this invention provides a method and system for mRNA UTR sequence generation and functional prediction based on coding sequences. This method constructs a machine learning or deep learning-based UTR sequence generation model, combined with different processing strategies for CDS and UTR sequences, to achieve automated generation and design of UTR sequences. Simultaneously, it utilizes the sequence features learned by the pre-trained model to construct a downstream task functional prediction model to predict UTR-related functional attributes. Furthermore, by functionally evaluating and screening multiple generated candidate UTR sequences, a UTR sequence that better matches the target CDS and has superior functional attributes is obtained, thus forming a technical solution integrating UTR generation, functional prediction, and candidate screening.

[0009] To achieve the above objectives, the present invention provides the following solution: A method for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, the method comprising: Collect and preprocess mRNA transcript data to obtain a pre-training dataset; Collect downstream functional prediction task data of UTR to obtain downstream task dataset; A conditional generation model is constructed based on the Text-To-Text Transfer Transformer architecture, and the conditional generation model is trained autoregressively using the pre-trained dataset to obtain a UTR sequence generation model. The UTR sequence generation model is fine-tuned and trained using the downstream task dataset to obtain a downstream task function prediction model. Based on the UTR sequence generation model, candidate UTR sequences are generated under the condition of a given coding sequence CDS, and the generation capability of the UTR sequence generation model in the non-coding region sequence generation task is evaluated. The downstream task function prediction model is used to predict the functional attributes of UTR sequences, and the predictive ability of the downstream task function prediction model in multiple UTR-related downstream tasks is evaluated.

[0010] Preferably, the conditional generation model includes: an encoder, a decoder, and a hybrid granularity word segmenter; At the encoder end, the hybrid granularity word segmenter represents the CDS sequence according to 3-mer granularity, that is, it segments the words with codons as the basic modeling unit to obtain the semantic representation of the CDS. At the decoder end, the hybrid granularity word segmenter represents the UTR sequence at a 1-mer granularity, that is, it segments the sequence using single nucleotides as the basic modeling unit, preserving the local features of the UTR sequence at base resolution.

[0011] Preferably, the method for constructing a conditional generation model and using the pre-training dataset to perform autoregressive training on the conditional generation model to obtain a UTR sequence generation model includes: For the 5′ UTR generation task, the sequence formed by 5′ UTR and the corresponding CDS is concatenated and then reversed as a whole. The original sequence of “5′ UTR + CDS” is concatenated and then the direction is adjusted so that it presents a structure of “CDS + 5′ UTR” when input into the model.

[0012] Preferably, the method for fine-tuning and training the UTR sequence generation model using the downstream task dataset to obtain the downstream task function prediction model includes: The hidden representations output by the UTR sequence generation model are used as input features for downstream tasks, and a regression prediction head is added thereafter to output the predicted values ​​of the target functional attributes. The regression prediction head uses a ResNet network.

[0013] Preferably, the method for evaluating the generation capability of the UTR sequence generation model in non-coding region sequence generation tasks includes: The generative model's ability to generate sequences in non-coding regions is evaluated by sequence similarity and related metrics, including edit distance, 4-mer distance, and GC content.

[0014] Preferably, the method further includes: The downstream task function prediction model is used to evaluate and score candidate UTR sequences to obtain the UTR sequence with the best functional attributes.

[0015] The present invention also provides a system for generating and predicting the function of mRNA untranslated region sequences based on coding sequences. The system is used to implement the aforementioned method and includes: a first data collection module, a second data collection module, a first model acquisition module, a second model acquisition module, a sequence generation and evaluation module, and a function prediction and evaluation module. The first data collection module is used to collect mRNA transcript data and preprocess it to obtain a pre-training dataset; The second data collection module is used to collect downstream functional prediction task data of UTR to obtain downstream task dataset; The first model acquisition module is used to construct a conditional generation model based on the Text-To-Text Transfer Transformer architecture, and to perform autoregressive training on the conditional generation model using the pre-trained dataset to obtain a UTR sequence generation model; The second model acquisition module is used to fine-tune and train the UTR sequence generation model using the downstream task dataset to obtain a downstream task function prediction model. The sequence generation and evaluation module is used to generate candidate UTR sequences based on the UTR sequence generation model, given a coding sequence CDS, and to evaluate the generation capability of the UTR sequence generation model in non-coding region sequence generation tasks. The function prediction and evaluation module is used to predict the functional attributes of the UTR sequence based on the downstream task function prediction model, and to evaluate the prediction capability of the downstream task function prediction model in multiple UTR-related downstream tasks.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes a UTR sequence generation model based on machine learning or deep learning, using the coding sequence CDS as a conditional variable. By training on large-scale transcriptome data containing UTR sequences and their corresponding CDS, the model learns the association between UTRs and CDS. Based on this, it achieves conditional generation and functional evaluation of UTR sequences, thereby generating and selecting matching 5′ or 3′ UTR sequences given a CDS. Through this technical solution, this invention can directly consider the synergistic effect between UTRs and CDS during the UTR generation process, obtaining UTR sequences with superior functional properties, improving the rationality and efficiency of UTR sequence design, and thus solving the technical problem in existing technologies where UTRs are designed as independent modules, making it difficult to obtain UTR sequences adapted to the target CDS. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the pre-trained UTR sequence generation model cUTRGen obtained in an embodiment of the present invention; Figure 3 This is a schematic diagram of the downstream function prediction task in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the analysis results generated in an embodiment of the present invention; Figure 5 This is a schematic diagram of the TE / EL prediction task results according to an embodiment of the present invention, wherein (a) is a schematic diagram of the TE prediction task results; and (b) is a schematic diagram of the EL prediction task results. Figure 6 This is a schematic diagram of the results of the mRNA stability prediction task in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 This invention provides a method for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, comprising: Collect and preprocess mRNA transcript data to obtain a pre-training dataset; Collect downstream functional prediction task data of UTR to obtain downstream task dataset; A conditional generation model is constructed, and the conditional generation model is trained autoregressively using the pre-trained dataset to obtain a UTR sequence generation model; The UTR sequence generation model is fine-tuned and trained using the downstream task dataset to obtain a downstream task function prediction model. Based on the UTR sequence generation model, candidate UTR sequences are generated under the condition of a given coding sequence CDS, and the generation capability of the UTR sequence generation model in the non-coding region sequence generation task is evaluated. The downstream task function prediction model is used to predict the functional attributes of UTR sequences, and the predictive ability of the downstream task function prediction model in multiple UTR-related downstream tasks is evaluated.

[0022] like Figure 1 As shown, the specific implementation process of the present invention is as follows: S1. Collect pre-trained large-scale mRNA transcript data and perform preprocessing operations on the data to obtain a pre-trained dataset.

[0023] In this embodiment, S1 first collects large-scale mRNA transcript data to construct a pre-training dataset. Specifically, it integrates approximately 3 million high-quality mammalian mRNA transcript data from public databases such as NCBI, Ensembl, and UTRdb. The data mainly comes from 91 species, including humans, chickens, house mice, and rats.

[0024] Subsequently, the transcript data underwent data cleaning, redundancy removal, and quality control to construct a standardized pre-training corpus. Specifically, completely repetitive transcript sequences were first filtered out, and then clustering and redundant sequence removal were performed using the mmseqs easy-linclust tool with a 90% sequence identity threshold. Through these processing steps, approximately 2 million high-quality mammalian mRNA transcript sequences were finally obtained.

[0025] Based on this, according to the corresponding gene annotation files (GTF / GFF), the mRNA transcript sequences were divided into 5′ UTR, CDS, and 3′ UTR. Further, based on the structural division of functional regions within the same transcript, the corresponding UTR sequences were combined with their CDS to construct training samples containing UTR-CDS. During sample construction, the data underwent further quality filtering, removing samples with UTR lengths less than 5 bases, and truncating the UTR length to 256 and the CDS length to 512 to ensure the validity and efficiency of the training data. Ultimately, approximately 1.3 million 3′ UTR–CDS paired samples and approximately 1.4 million 5′ UTR–CDS paired samples were obtained, and divided into training, validation, and independent test sets in a 98:1:1 ratio for subsequent model training and performance evaluation.

[0026] S2. Collect downstream function prediction task data related to UTR, construct a downstream task dataset, and use it for subsequent model fine-tuning training.

[0027] In this embodiment, S2 collects downstream functional prediction task data for model fine-tuning and performance evaluation to obtain a downstream task dataset. The tasks include translation efficiency (TE) prediction task, expression level (EL) prediction task, and mRNA stability prediction task.

[0028] For TE and EL prediction tasks, endogenous gene datasets from three human cell types were collected from existing literature: human muscle tissue, human prostate cancer cell line PC3, and human embryonic kidney cell line HEK293T. Since the original datasets only contained 5′ UTR sequences without corresponding coding sequence information, the corresponding transcript sequences and their annotation files were retrieved from the Ensembl database based on the Ensembl transcript IDs provided in the datasets. This allowed for the matching of corresponding CDS sequences to each 5′ UTR sequence, constructing a data sample containing 5′ UTR and CDS pairing information.

[0029] For the mRNA stability prediction task, two datasets of different sizes were collected from relevant existing literature. The smaller dataset contains approximately 1967 3′ UTR sequences, while the larger dataset contains approximately 90,000 3′ UTR sequences. These datasets were used to evaluate the model's performance in the 3′ UTR function prediction task. After processing, these datasets were used for subsequent model fine-tuning training and downstream task performance evaluation.

[0030] S3. Construct a conditional generation model based on the Text-To-Text Transfer Transformer (T5) architecture, and use the pre-trained dataset to perform autoregressive training on the conditional generation model to obtain the UTR sequence generation model.

[0031] In this embodiment, S3 constructs and pre-trains a generative language model to achieve UTR conditional generation based on encoded sequence CDS. Specifically, this includes the following steps: Model Architecture Construction: A conditional generative model for UTR sequence generation and functional representation learning is constructed. This model uses paired samples of the encoded sequence CDS and the UTR sequence as training data, simultaneously achieving UTR sequence generation and sequence functional representation learning within a unified framework.

[0032] Specifically, the conditional generation model adopts a T5 structure, including an encoder and a decoder. The encoder is used to learn the representation of the input CDS sequence, and the decoder is used to generate the UTR sequence under the conditional constraints of the CDS representation of the encoded sequence. During the forward computation, the T5 model naturally outputs the hidden states of each layer. In this embodiment, the last hidden state is extracted as the sequence feature representation (i.e., hidden representation) for subsequent function prediction tasks. Through this structure, the conditional generation model can obtain a sequence feature representation reflecting the CDS-UTR correlation while learning the UTR generation rules.

[0033] Compared with existing methods that only input CDS sequences into the generation model to obtain UTR sequences, this embodiment not only uses the model to generate UTR sequences, but also further utilizes the hidden representations obtained in the pre-training stage to construct a downstream functional prediction model, thereby enabling the same model framework to have both generation and functional representation capabilities.

[0034] Tokenizer Construction: To characterize the structural features of coding and non-translated regions at different biological levels, a hybrid granularity tokenizer is constructed. Specifically, the hybrid granularity tokenizer adopts a hybrid granularity segmentation approach. At the encoder end, the CDS sequence is represented with 3-mer granularity, that is, the coding sequence is segmented using codons as the basic modeling unit to obtain the CDS semantic representation, thereby enhancing the model's ability to represent the semantic information of the encoded sequence. At the decoder end, the UTR sequence is represented with 1-mer granularity, that is, the sequence is segmented using single nucleotides as the basic modeling unit, in order to preserve the local features of the UTR sequence at base resolution.

[0035] By using the above hybrid granularity word segmentation method, the model can learn representations for different biological attributes of CDS and UTR respectively. The 3-mer representation on the CDS side helps to capture high-level semantic information at the codon level, while the 1-mer representation on the UTR side helps to achieve fine generation at the base level, thereby improving the model's ability to model transcript structural features.

[0036] Furthermore, for the 5′UTR generation task, since the 5′UTR is located upstream of the CDS in the natural transcript structure, in order to enable the 5′UTR generation task and the 3′UTR generation task to adopt a unified conditional generation form, a sequence flipping operation is performed on the sequence formed by the 5′UTR and the corresponding CDS, so that the model input satisfies the modeling form of "conditional sequence first, target sequence second", thereby facilitating the generation of 5′UTR and 3′UTR under a unified framework.

[0037] Pre-training: The UTR-CDS paired data obtained from the preprocessing in step S1 is sequence encoded by a hybrid granularity word segmenter and then input into the T5-based conditional generation model for pre-training.

[0038] During the pre-training phase, the conditional generation model is trained autoregressively using the Next Token Prediction (NTP) strategy. This is given by the encoded sequence X and the target UTR sequence. The conditional generation model predicts the true base at each position t. The conditional probability, where The values ​​are derived from the base set {A, T, C, G}. Within this framework, the generation probability of the UTR sequence can be expressed as: Where X represents the input CDS sequence, This represents the actual base at position t. This indicates the UTR base sequence preceding position t.

[0039] The model training objective is to minimize the negative log-likelihood loss of the true sequence, and its loss function is defined as: Where T represents the length of the UTR sequence, This represents the probability that the model predicts the true base at position t.

[0040] By minimizing the loss function described above, the model can learn the contextual dependencies of the UTR sequence and the association between the UTR and the CDS, thereby gaining the ability to generate UTR sequences given the CDS.

[0041] Model Examples: In this embodiment, the above model is trained using 3′ UTR-CDS paired data and 5′ UTR-CDS paired data respectively, resulting in two instance models (UTR sequence generation models, cUTRGen). The model trained on 3′ UTR data is named c3UTRGen, and the model trained on 5′ UTR data is named c5UTRGen. Both models maintain the same model architecture and training strategy, differing only in the training data. Figure 2 As shown.

[0042] This invention further constructs a downstream task function prediction model based on the hidden representation of the pre-trained UTR sequence generation model. By fine-tuning the pre-trained UTR sequence generation model, it can predict UTR-related functional attributes and can be used to evaluate and screen the generated candidate UTR sequences, thus forming an integrated design framework of "generation-evaluation-screening".

[0043] S4. Fine-tune the pre-trained conditional generation model using the downstream task dataset to obtain the downstream task function prediction model.

[0044] In this embodiment, S4 uses the UTR sequence generation model obtained from the pre-training in step S3 and fine-tunes it in conjunction with the downstream task dataset to obtain a downstream task function prediction model for UTR function prediction.

[0045] Specifically, the UTR sequence generation model obtained in step S3 has learned the correlation between the UTR sequence and the CDS. Therefore, in step S4, the model is no longer retrained from random parameters, but directly inherits the network parameters of the UTR sequence generation model, and continues to train on this basis in combination with downstream task data, so that the model can adapt to the specific functional prediction task.

[0046] For different downstream tasks, corresponding UTR sequence generation models were selected as the parameter base. Specifically, for the 5′UTR-related translation efficiency (TE) prediction task and expression level (EL) prediction task, the model c5UTRGen, which was pre-trained on 5′UTR–CDS paired data, was used as the parameter base; for the 3′UTR-related stability prediction task, the model c3UTRGen, which was pre-trained on 3′UTR–CDS paired data, was used as the parameter base.

[0047] like Figure 3As shown, during fine-tuning training, downstream task data is first input into the UTR sequence generation model, and the hidden representations output from the last layer of the model are extracted. Then, these hidden representations are used as input to the regression prediction head, which outputs the predicted values ​​of the target functional attributes. The regression prediction head employs a ResNet network structure.

[0048] During fine-tuning training, downstream task data is input into the initialized model, the last hidden layer representation is extracted, and the prediction result for the corresponding task is output via the ResNet regression head. For continuous functional attribute prediction tasks, Huber Loss is used as the optimization objective to reduce the impact of outliers on the training process and improve model training stability. Specifically, let the true label of the sample be... The model predicts the value. Then Huber Loss is defined as: in, For segmented threshold parameters, in this embodiment, the parameter is... .

[0049] In the above manner, the UTR sequence and CDS association information already learned in the UTR sequence generation model can be further used for downstream task prediction, thereby obtaining a downstream task functional prediction model for translation efficiency, expression level or stability prediction.

[0050] S5. Using the UTR sequence generation model, candidate UTR sequences are generated under the given coding sequence CDS. The generation capability of the generation model in the non-coding region sequence generation task is evaluated by sequence similarity and related indicators.

[0051] In this embodiment, step S5 evaluates the generation capability of the UTR sequence generation model obtained in step S3 in the non-coding region sequence generation task using sequence similarity and related indicators. All evaluation experiments are conducted on the independent test set obtained in step S1.

[0052] Specifically, for each intrinsic coding sequence (CDS) in the test set, multiple candidate UTR sequences are generated using the UTR sequence generation model. In this embodiment, 10 candidate UTR sequences are generated for each CDS. Simultaneously, to construct a control group, a UTR sequence is randomly generated for each CDS, wherein the length of the random sequence is consistent with the length of the corresponding wild-type UTR sequence to ensure the fairness of the comparative experiment.

[0053] Subsequently, the generated UTR sequences, the wild-type UTR sequences corresponding to the CDS, and the randomly generated UTR sequences were compared and analyzed. Multiple sequence-level evaluation metrics were used to assess the sequence generation capability of the generative model. The evaluation metrics used included: (1) Edit distance, used to measure the degree of sequence difference between the generated UTR sequence and the corresponding wild-type UTR sequence; (2) 4-mer distance, which measures the similarity of sequence composition by statistically analyzing the differences in the frequency distribution of k-mers of length 4 in the sequence; (3) GC content is used to assess the degree of similarity between the generated sequence and the wild-type sequence in terms of base composition.

[0054] During the evaluation process, the average and minimum values ​​of the above indicators were calculated between the generated UTR sequence and the wild-type UTR sequence, and compared with the corresponding results between the random sequence and the wild-type sequence. Experimental results are as follows: Figure 4 As shown in the figure. The results indicate that, in terms of edit distance, 4-mer distance, and GC content difference, the difference between the generated UTR sequence and the wild-type UTR sequence is smaller than the difference between the random sequence and the wild-type sequence, indicating that the generative model can generate candidate sequences that are closer to the real UTR sequence in terms of sequence features.

[0055] Furthermore, to further evaluate the similarity between the generated sequences and the training data, BLAT sequence alignment analysis was performed on the generated UTR sequences and the UTR sequences in the training set, and metrics such as identity, maximal match length, and maximal identity were calculated. Experimental results are as follows: Figure 4 As shown in the figure, the matching rate between the generated sequence and the sequences in the training set is significantly lower than that between the training sequences. Furthermore, the maximum matching length and maximum similarity are also lower than in the self-matching case. This indicates that the generated sequence is not a simple copy of the training data, but rather a UTR sequence with a certain degree of novelty generated based on the learned sequence statistical patterns.

[0056] The above evaluation experiments verify that the UTR sequence generation model can generate new UTR sequences while maintaining the statistical characteristics of biological sequences, thus proving its effectiveness in non-coding region sequence generation tasks.

[0057] S6. Use the downstream task function prediction model to predict the functional attributes related to the UTR sequence, and evaluate the prediction ability of the downstream task function prediction model in multiple UTR-related downstream tasks.

[0058] In this embodiment, step S6 is used to evaluate the predictive performance of the downstream task function prediction model obtained in step S4 in the non-coding region function prediction task. The evaluation includes a 5′ UTR-related translation efficiency (TE) prediction task and an expression level (EL) prediction task, as well as a 3′ UTR-related mRNA stability prediction task. The experimental results are as follows: Figure 5 and Figure 6 As shown.

[0059] For the TE and EL prediction tasks, the human muscle tissue, prostate cancer cell line PC3, and human embryonic kidney cell line HEK datasets described in S2 were used for evaluation. During fine-tuning training, CDS and UTR sequences were input simultaneously, enabling the downstream task functional prediction model to utilize the UTR–CDS correlation information learned during the pre-training phase for prediction. During model evaluation, 10-fold cross-validation was used to assess the performance of the downstream task functional prediction model to improve the stability and reliability of the evaluation results.

[0060] To verify the model's performance, the downstream task function prediction model was compared with several existing UTR modeling methods. In the comparative experiments, considering that existing methods typically only use UTR sequences as input, this embodiment uniformly configured the model, where conditions permit, to simultaneously receive CDS and UTR information, thus ensuring the fairness of the comparison.

[0061] Experimental results show that, on multiple datasets, the downstream task function prediction model in this embodiment exhibits superior prediction performance in both TE and EL prediction tasks, outperforming the comparative methods in terms of correlation metrics. This indicates that jointly modeling CDS and UTR sequences helps improve the predictive ability of translation-related phenotypes. Furthermore, experiments on a model variant that only inputs UTR sequences revealed a decrease in predictive performance compared to the complete downstream task function prediction model, suggesting that CDS information plays a crucial role in the model's capture of regulatory signals.

[0062] For the 3′ UTR stability prediction task, the datasets described in S2 were used for evaluation. The small dataset contains approximately 1967 3′ UTR sequences, while the large dataset contains approximately 90,000 3′ UTR sequences. The stability prediction task was trained and tested using pre-divided datasets without cross-validation. Experimental results show that the downstream task function prediction model in this embodiment achieves good prediction performance on datasets of different sizes. Its performance in mean squared error and correlation metrics is superior to or no less than existing methods, indicating that the downstream task function prediction model can effectively capture stability regulation signals in 3′ UTR sequences and has good generalization ability.

[0063] In summary, the evaluation results on multiple downstream tasks demonstrate that the downstream task function prediction model can not only be used for UTR sequence generation, but also learn the functional features related to UTR sequences and coding sequences, thus showing good performance in various non-coding region function prediction tasks.

[0064] Furthermore, in this embodiment, based on the generation of multiple candidate UTR sequences, the downstream task function prediction model obtained in step S4 is used to evaluate and score the candidate UTR sequences to obtain the UTR sequence with the best functional attributes.

[0065] Specifically, for the same CDS, the UTR sequence generation model can generate multiple UTRs as candidate sequences. Each candidate UTR sequence and its corresponding CDS are input into the downstream task functional prediction model to obtain the prediction results of each candidate UTR sequence on the target functional attribute. Based on the prediction results, the candidate UTR sequences are sorted and filtered to select the UTR sequence with the best score or that meets preset conditions. The target functional attribute includes, but is not limited to, translation efficiency, expression level, and mRNA stability.

[0066] Through the above methods, the present invention can not only generate multiple candidate UTR sequences under a given CDS condition, but also further combine downstream task function prediction models to quantitatively evaluate and screen the candidate sequences, thereby obtaining UTR sequences that are more compatible with the target CDS and have better functional attributes.

[0067] In summary, this invention provides a method for generating and predicting the function of mRNA untranslated region (UTR) sequences based on coding sequences. By constructing a deep learning-based UTR sequence generation model and introducing coding sequence CDS as a conditional variable, the method achieves automated generation and design of UTR sequences. The synergistic relationship between UTR and CDS is considered during UTR sequence generation, improving the efficiency and rationality of UTR sequence design. Furthermore, this invention utilizes a downstream task function prediction model to evaluate and screen the generated candidate UTR sequences, thereby obtaining UTR sequences with superior functional attributes.

[0068] Example 2 Based on the same inventive concept, the present invention also provides a system for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, for implementing the method described in the foregoing embodiments. The system includes: a first data collection module, a second data collection module, a first model acquisition module, a second model acquisition module, a sequence generation and evaluation module, and a function prediction and evaluation module. The first data collection module is used to collect mRNA transcript data and preprocess it to obtain a pre-training dataset; The second data collection module is used to collect downstream functional prediction task data of UTR to obtain downstream task dataset; The first model acquisition module is used to construct a conditional generation model based on the Text-To-Text Transfer Transformer architecture, and to perform autoregressive training on the conditional generation model using the pre-trained dataset to obtain a UTR sequence generation model; The second model acquisition module is used to fine-tune and train the UTR sequence generation model using the downstream task dataset to obtain a downstream task function prediction model. The sequence generation and evaluation module is used to generate candidate UTR sequences based on the UTR sequence generation model, given a coding sequence CDS, and to evaluate the generation capability of the UTR sequence generation model in non-coding region sequence generation tasks. The function prediction and evaluation module is used to predict the functional attributes of the UTR sequence based on the downstream task function prediction model, and to evaluate the prediction capability of the downstream task function prediction model in multiple UTR-related downstream tasks.

[0069] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, characterized in that, The method includes: Collect and preprocess mRNA transcript data to obtain a pre-training dataset; Collect downstream functional prediction task data of UTR to obtain downstream task dataset; A conditional generation model is constructed based on the Text-To-Text Transfer Transformer architecture, and the conditional generation model is trained autoregressively using the pre-trained dataset to obtain a UTR sequence generation model. The UTR sequence generation model is fine-tuned and trained using the downstream task dataset to obtain a downstream task function prediction model. Based on the UTR sequence generation model, candidate UTR sequences are generated under the condition of a given coding sequence CDS, and the generation capability of the UTR sequence generation model in the non-coding region sequence generation task is evaluated. The downstream task function prediction model is used to predict the functional attributes of UTR sequences, and the predictive ability of the downstream task function prediction model in multiple UTR-related downstream tasks is evaluated.

2. The method according to claim 1, characterized in that, The conditional generation model includes: an encoder, a decoder, and a hybrid granularity word segmenter; At the encoder end, the hybrid granularity word segmenter represents the CDS sequence according to 3-mer granularity, that is, it segments the words with codons as the basic modeling unit to obtain the semantic representation of the CDS. At the decoder end, the hybrid granularity word segmenter represents the UTR sequence at a 1-mer granularity, that is, it segments the sequence using single nucleotides as the basic modeling unit, preserving the local features of the UTR sequence at base resolution.

3. The method according to claim 1, characterized in that, The method for obtaining a UTR sequence generation model by performing autoregressive training on the conditional generation model using the pre-trained dataset includes: For the 5′ UTR generation task, the sequence formed by 5′ UTR and the corresponding CDS is concatenated and then reversed as a whole. The original sequence of "5′ UTR + CDS" is concatenated and then the direction is adjusted so that it presents the structure of "CDS + 5′ UTR" when input into the model.

4. The method according to claim 1, characterized in that, The method for fine-tuning and training the UTR sequence generation model using the downstream task dataset to obtain the downstream task function prediction model includes: The hidden representations output by the UTR sequence generation model are used as input features for downstream tasks, and a regression prediction head is added thereafter to output the predicted values ​​of the target functional attributes. The regression prediction head uses a ResNet network.

5. The method according to claim 1, characterized in that, Methods for evaluating the generation capability of the UTR sequence generation model in non-coding region sequence generation tasks include: The generative model's ability to generate sequences in non-coding regions is evaluated by sequence similarity and related metrics, including edit distance, 4-mer distance, and GC content.

6. The method according to claim 1, characterized in that, The method further includes: The downstream task function prediction model is used to evaluate and score candidate UTR sequences to obtain the UTR sequence with the best functional attributes.

7. A system for generating and predicting the function of mRNA untranslated region sequences based on coding sequences, said system being used to implement the method described in any one of claims 1-6, characterized in that, The system includes: a first data collection module, a second data collection module, a first model acquisition module, a second model acquisition module, a sequence generation and evaluation module, and a functional prediction and evaluation module; The first data collection module is used to collect mRNA transcript data and preprocess it to obtain a pre-training dataset; The second data collection module is used to collect downstream functional prediction task data of UTR to obtain downstream task dataset; The first model acquisition module is used to construct a conditional generation model based on the Text-To-Text Transfer Transformer architecture, and to perform autoregressive training on the conditional generation model using the pre-trained dataset to obtain a UTR sequence generation model; The second model acquisition module is used to fine-tune and train the UTR sequence generation model using the downstream task dataset to obtain a downstream task function prediction model. The sequence generation and evaluation module is used to generate candidate UTR sequences based on the UTR sequence generation model, given a coding sequence CDS, and to evaluate the generation capability of the UTR sequence generation model in non-coding region sequence generation tasks. The function prediction and evaluation module is used to predict the functional attributes of the UTR sequence based on the downstream task function prediction model, and to evaluate the prediction capability of the downstream task function prediction model in multiple UTR-related downstream tasks.

Images