Method, computer program product and system for optimizing protein expression

Abstract

Methods of optimizing protein expression are provided. The method comprises: obtaining, by a processor (102), a plurality of amino acid sequences and corresponding known efficiency values, each known efficiency value indicating an efficiency of expressing a protein having a corresponding one of the plurality of amino acid sequences; for each of a plurality of prediction algorithms, obtaining, by the processor (102), a prediction function from one of the plurality of prediction algorithms, where the prediction function outputs a prediction efficiency value expressing a protein having an amino acid sequence corresponding to an input numerical vector; evaluating, by the processor (102), the obtained prediction function by comparing a predicted efficiency value output by the obtained prediction function with the known efficiency value; selecting, by the processor (102), at least one prediction algorithm from the plurality of prediction algorithms based on the evaluation; predicting, by the processor (102), one or more efficiency values expressing one or more proteins having one or more specified amino acid sequences, respectively, using the at least one prediction algorithm and a prediction function obtained with the at least one prediction algorithm; and outputting, by the processor (102), the one or more specified amino acid sequences and the one or more efficiency values predicted for the one or more specified amino acid sequences.

Description

[0001] The present application relates to computer-implemented methods, computer program products, and systems for optimizing protein expression.

[0002] background

[0003] Protein expression analysis is typically performed using wet lab experiments in a controlled environment. For example, with the influx of high-throughput sequencing and genetic engineering technologies, experimental approaches are available to analyze recombinant antibody sequence parameters and their effects on protein expression. Major bottlenecks to the widespread adoption of these experimental approaches may include high cost, time, and / or the need for a controlled environment. Summary of the invention

[0004] According to an aspect, the problem relates to optimizing protein expression with increased efficiency. Said problem is solved by the features disclosed in the independent claims. Further exemplary embodiments are defined by the dependent claims.

[0005] According to aspects, a method for optimizing protein expression is provided. In the method, each of a plurality of prediction algorithms comprises a combination of an encoding algorithm, a dimensionality reduction algorithm, and a regression algorithm, and the combination is different from the combination included in any other of the plurality of prediction algorithms. Further, the method comprises:

[0006] obtaining, by a processor, a plurality of amino acid sequences and corresponding known efficiency values, each known efficiency value indicating an efficiency of expressing a protein having a corresponding one of the plurality of amino acid sequences;

[0007] For each of the plurality of prediction algorithms,

[0008] generating, by the processor, a first numerical vector corresponding to one of the plurality of amino acid sequences for each of the plurality of amino acid sequences by encoding at least a portion of one of the plurality of amino acid sequences according to an encoding algorithm included in one of the plurality of prediction algorithms,

[0009] generating, by the processor, a second numerical vector corresponding to the plurality of amino acid sequences by applying a dimensionality reduction algorithm included in one of the plurality of prediction algorithms to a first numerical vector corresponding to the plurality of amino acid sequences, wherein the dimension of the second numerical vector is smaller than the dimension of the first numerical vector,

[0010] obtaining, by the processor, a prediction function by processing at least a portion of the second numerical vector and known efficiency values ​​of an amino acid sequence corresponding to at least a portion of the second numerical vector according to a regression algorithm included in one of the plurality of prediction algorithms, wherein the prediction function outputs a predicted efficiency value for expressing a protein having an amino acid sequence corresponding to an input numerical vector, and

[0011] evaluating, by the processor, the obtained prediction function by comparing a predicted efficiency value output by the obtained prediction function for at least a portion of the second numerical vector with a known efficiency value for an amino acid sequence corresponding to at least a portion of the second numerical vector;

[0012] selecting, by the processor, at least one prediction algorithm from the plurality of prediction algorithms based on the evaluation;

[0013] Predicting, by the processor, one or more efficiency values ​​of expressing one or more proteins having one or more specified amino acid sequences respectively, using the at least one prediction algorithm and a prediction function obtained using the at least one prediction algorithm; and

[0014] The processor outputs the one or more specified amino acid sequences and one or more efficiency values ​​predicted for the one or more specified amino acid sequences.

[0015] In various aspects and embodiments described herein, the "amino acid sequence" preferably comprises a series of amino acids in single letter notation. However, in some exemplary embodiments, the "amino acid sequence" may comprise a series of amino acids in three letter notation.

[0016] In the present disclosure, the "efficiency" of expressing a protein may indicate the degree of expression of a protein having a specific amino acid sequence. Further, in the present disclosure, the "efficiency" of expressing a protein may also be referred to as the "performance" of protein expression.

[0017] In various aspects and embodiments described herein, a "known efficiency value" corresponding to "one of the multiple amino acid sequences" can be determined by expressing a protein having one of the multiple amino acid sequences using a specific type of host cell and a specific type of expression vector in a wet lab experiment. For example, the "known efficiency value" can be the average efficiency of protein expression of proteins having the corresponding amino acid sequence.

[0018] In various aspects and embodiments described herein, an "encoding algorithm" may be an algorithm for encoding an amino acid sequence into a series of numerical values. A series of numerical values ​​may be considered as a numerical vector (e.g., a "first numerical vector"). The dimension of the numerical vector may be reduced according to a "dimensionality reduction algorithm" (e.g., to produce a "second numerical vector"). A "regression algorithm" may be an algorithm for estimating the relationship between the "second numerical vector" (as an independent variable) and a protein expression efficiency value (as a dependent variable).

[0019] In various aspects and embodiments described herein, different prediction algorithms include different combinations of the encoding algorithms, the dimensionality reduction algorithms, and the regression algorithms. For example, multiple encoding algorithms, multiple dimensionality reduction algorithms, and multiple regression algorithms may be available, and the combination of algorithms included in the prediction algorithm can be determined by selecting one of the multiple encoding algorithms, one of the multiple dimensionality reduction algorithms, and one of the multiple regression algorithms. In some exemplary embodiments, the multiple prediction algorithms can include all possible combinations of algorithms obtained from the available encoding, dimensionality reduction, and regression algorithms.

[0020] Further, in various aspects and embodiments described herein, at least some of the "one or more specified amino acid sequences" can be obtained from a user via an input device connected to the processor. Additionally or alternatively, the "one or more specified amino acid sequences" can be obtained from a computer connected (e.g., via a network) to the processor.

[0021] In various aspects and embodiments described herein, one or more predicted efficiency values ​​predicted for expressing one or more proteins having one or more specified amino acid sequences, respectively, may indicate a predicted efficiency value for expressing the respective protein using a specific type of host cell and a specific type of expression vector for obtaining the known efficiency value.

[0022] Further, in various aspects and embodiments described herein, the “output” of “one or more specified amino acid sequences” and “one or more efficiency values ​​predicted for one or more specified amino acid sequences” can cause a display device connected to the processor to display information indicating the “one or more specified amino acid sequences” and the “one or more efficiency values ​​predicted for one or more specified amino acid sequences”. Additionally or alternatively, the “output” can include transmitting the information indicating the “one or more specified amino acid sequences” and the “one or more efficiency values ​​predicted for one or more specified amino acid sequences” to a computer connected (e.g., via a network) to the processor.

[0023] For example, such a computer that receives the information from the processor may be included in an automated cell culture system configured to perform protein expression. The computer included in the automated cell culture system may, for example, display the received information so that proteins having the one or more specified amino acid sequences, respectively, may be expressed in the automated cell culture system. Further, in some cases, the computer included in the automated cell culture system may control the automated cell culture system to provide a controlled environment (e.g., temperature, humidity, stirring, etc.) suitable for expressing proteins having the one or more specified amino acid sequences, respectively.

[0024] The method according to any of the above aspects or various embodiments thereof may promote protein expression with improved efficiency.

[0025] The method according to the above aspect may further include:

[0026] identifying, by the processor, one of the one or more designated amino acid sequences with the highest predicted efficiency value; and

[0027] By the processor, information is output indicating that the identified one of the one or more specified amino acid sequences has the highest predicted efficiency value.

[0028] In various aspects and embodiments described herein, the plurality of amino acid sequences may be antibody sequences or recombinant antibody sequences.

[0029] For example, synthetic genes can be used to produce recombinant antibodies. Different biotechnology methods can be used to express antibody sequences in protein form, adjust different fragments to enhance affinity, spatial orientation, stability of antigen binding sites, change immunogenicity, analyze sequence parameters, expression and inhibition mechanisms, and many other parameters. Considering the therapeutic potential of recombinant antibodies, the pharmaceutical industry is investing trillions to generate accurate recombinant antibody sequences and to conduct in-depth, defect-free analysis of recombinant antibodies. Therefore, any one of the embodiments in which the multiple amino acid sequences are recombinant antibody sequences can reduce the huge recombinant antibody sequence analysis costs suffered by the pharmaceutical industry to a large extent, without losing the hope of accurate discovery. Further, any one of the embodiments in which the multiple amino acid sequences are recombinant antibody sequences can be adopted by numerous pharmaceutical industries to accurately infer the expression of, for example, new recombinant antibodies.

[0030] In the embodiment where the plurality of amino acid sequences are antibody sequences or recombinant antibody sequences, the plurality of amino acid sequences may be light chain sequences, heavy chain sequences, or light-heavy chain sequences of the antibody or recombinant antibody.

[0031] For example, in some cases, using different types of sequences as the multiple amino acid sequences may result in different prediction functions with different evaluation results, even when the prediction functions are obtained according to the same prediction algorithm. Therefore, the use of the light chain sequence, the heavy chain sequence and / or the light-heavy chain sequence can make it possible to quantify the performance potential of the full-length sequence of each type of sequence and / or the main bottleneck at the amino acid distribution level, which may be preventing the prediction function from providing accurate efficiency predictions.

[0032] Further, in various aspects and embodiments described herein, at least a portion of one of the plurality of amino acid sequences may include:

[0033] The first x% of one of the plurality of amino acid sequences; and / or

[0034] The last y% of one of the multiple amino acid sequences

[0035] where each of x and y is a value greater than 0 and less than or equal to 50,

[0036] wherein x is preferably 25 or 50 and y is preferably 25 or 50.

[0037] Using the subsequence to be encoded (e.g., the first x% and / or the last y% of one of the multiple amino acid sequences) for generating the first numerical vector can allow the most discriminative amino acid distribution to be identified by retaining variable and / or informative regions and discarding constant regions that remain similar in different sequences.

[0038] Further, for example, using different subsequences (e.g., first x% and / or last y%) of different types of sequences to be encoded (e.g., light chains, heavy chains and / or light-heavy chains) for generating the first numerical vector can help to investigate which sequence types and subsequence types can produce prediction functions that provide more accurate efficiency predictions.

[0039] In various aspects and embodiments described herein, each known efficiency value can indicate the titer of expressing a protein having a corresponding one of the multiple amino acid sequences, and the predicted efficiency value output by the prediction function can indicate the predicted titer. Alternatively or additionally, a cell specific productivity (e.g., expressed in pg / cell / day) can be used as a known efficiency value, and the prediction function can output the predicted cell specific productivity as a predicted efficiency value. Further, in some exemplary embodiments, known efficiency values ​​and predicted efficiency values ​​can be related to product quality features such as aggregation, folding, glycosylation, etc.

[0040] In various aspects and embodiments described herein, the generation of the second numerical vector may include:

[0041] generating, by the processor, a plurality of sets of the second numerical vectors using the dimensionality reduction algorithm,

[0042] wherein the second numerical vectors included in the same set of the plurality of sets of the second numerical vectors have the same dimension,

[0043] wherein the second numerical vectors included in different sets of the plurality of sets of the second numerical vectors have different dimensions, and

[0044] Wherein, for each of the plurality of sets of the second numerical vectors, the obtaining of the prediction function and the evaluating of the prediction function are performed.

[0045] Using multiple sets of the second numerical vector as described above can help determine the dimensionality of the second numerical vector, resulting in a more accurate prediction function for predicting protein expression efficiency.

[0046] In various aspects and embodiments described herein, the evaluation of the obtained prediction function may include:

[0047] Determine, by the processor, one or more of the following performance indicators of the obtained prediction function:

[0048] Accuracy,

[0049] Accuracy,

[0050] Recall rate,

[0051] F1 score.

[0052] One or more performance indicators as described above may indicate how efficiently the prediction function can predict the input amino acid sequence.

[0053] Further, in various aspects and embodiments described herein, the combination of the encoding algorithm, the dimensionality reduction algorithm and the regression algorithm may be a combination of the following:

[0054] A k-mer-based encoding algorithm, a count-based encoding algorithm, a K-gap-based encoding algorithm, a window-based encoding algorithm, a group-based encoding algorithm, a physicochemical property-based encoding algorithm, or a word embedding-based encoding algorithm, as the encoding algorithm;

[0055] Principal component analysis PCA, K-means, t-distributed stochastic neighbor embedding TSNE, kernel principal component analysis, local linear embedding LLE, tensor singular value decomposition T-SVD, non-negative matrix factorization NMF, multidimensional scaling MDS, factor analysis, agglomerate feature, Gaussian random projection, sparse random projection or fast independent component analysis fast-ICA, as the dimensionality reduction algorithm; and

[0056] Linear regression, nonlinear regression, penalized linear regression, penalized nonlinear regression, naive Bayes, bagging regression, random forest regressor, boosting regression, partial least squares regression or support vector machine are used as the regression algorithm.

[0057] In the present disclosure, "k-mer" may be understood as a substring of length k contained in a biological sequence such as an amino acid sequence.

[0058] In various aspects and embodiments described herein, according to the encoding algorithm based on physicochemical properties as the encoding algorithm, encoding of at least a portion of one of the plurality of amino acid sequences may include:

[0059] By the processor, assigning a weight value to each amino acid based on one or more physicochemical properties of the protein; and

[0060] generating, by the processor, the first numerical vector including weights corresponding to amino acids included in at least a portion of one of the plurality of amino acid sequences in an order of at least a portion of one of the plurality of amino acid sequences,

[0061] The one or more physicochemical properties include one or more of the following:

[0062] Dissociation,

[0063] Solubility,

[0064] Hydration,

[0065] polarity,

[0066] Charge,

[0067] Hydrophobicity,

[0068] Molecular weight,

[0069] size.

[0070] In various aspects and embodiments described herein, according to the k-mer based encoding algorithm as the encoding algorithm, encoding of at least a portion of one of the plurality of amino acid sequences may include:

[0071] determining, by the processor, a k-mer of at least a portion of one of the plurality of amino acid sequences;

[0072] Determining, by the processor, a weight for each of the k-mers based at least on:

[0073] the number of times one of the k-mers occurs in at least a portion of one of the plurality of amino acid sequences, and

[0074] a number of times one of the k-mers occurs in at least a portion of the plurality of amino acid sequences;

[0075] The first numerical vector is generated by the processor, and the first numerical vector includes weights of the k-mers according to the order in which the k-mers appear in at least a portion of one of a plurality of amino acid sequences.

[0076] A specific but non-limiting example of the k-mer based encoding algorithm involves Okapi-BM25, which is a ranking function used in information retrieval and can provide a score indicating the relevance of a document to a query based on the frequency with which terms contained in the query appear in the document.

[0077] Further, the method according to any one of the above aspects and various embodiments thereof may further include:

[0078] In an automated cell culture system, one or more nucleic acids respectively encoding the one or more designated amino acid sequences are used to express one or more proteins respectively comprising the one or more designated amino acid sequences.

[0079] When one or more nucleic acids encoding the one or more specified amino acid sequences, respectively, are used in the automated cell culture system, a specific type of host cell and a specific type of expression vector for obtaining a known efficiency value may be used.

[0080] In the present disclosure, the "automated cell culture system" may be understood as a system configured and / or capable of expressing one or more proteins.

[0081] According to another aspect, a computer program product is provided. The computer program product comprises computer readable instructions which, when loaded and executed on a computer, cause the computer to perform the method according to any one of the above aspects and various embodiments thereof.

[0082] According to another aspect, a prediction system for predicting protein expression efficiency is provided. The prediction system comprises:

[0083] A processor configured to execute the method according to any one of the above aspects and various embodiments thereof; and

[0084] A storage medium is in communication with the processor and is configured to store a plurality of amino acid sequences and corresponding known efficiency values.

[0085] According to yet another aspect, a system for optimizing protein expression is provided. The system comprises:

[0086] A prediction system according to the above aspect; and

[0087] An automated cell culture system capable of expressing one or more proteins.

[0088] According to any of the various aspects and embodiments described herein, protein expression can be inferred by developing a computational framework that evaluates the validity of a wide range of intrinsic relationships of sequence residues and machine learning regression without the need for extensive wet laboratory experiments. In other words, for protein expression analysis, wet laboratory experiments can be replaced by calculations using artificial intelligence.

[0089] According to any one of the various aspects and embodiments described herein, developing a computational framework that can accurately infer protein expression can promote the expression of target proteins. For example, different amino acid sequences of the same protein class (e.g., antibodies) may show different efficiencies in the final expression. Therefore, as in any one of the various aspects and embodiments described herein, obtaining a prediction function that can provide accurate efficiency prediction based on the amino acid sequence of input can produce a better understanding of how to adjust the sequence to improve the efficiency of target protein expression (which can promote protein expression).

[0090] Further, any of the various aspects and embodiments described herein may provide a computational framework that is less expensive, more adaptable, scalable, and / or more robust for analyzing amino acid sequences and inferring their expression compared to wet lab experimental methods.

[0091] The subject matter described in this application may be implemented as a method or system, possibly in the form of one or more computer program products. The subject matter described in this application may be implemented in a data signal or on a machine-readable medium, wherein the medium is embodied in one or more information carriers, such as a CD-ROM, a DVD-ROM, a semiconductor memory, or a hard disk. Such a computer program product may cause a data processing device to perform one or more operations described in this application.

[0092] In addition, the subject matter described in this application can also be implemented as a system, the system including a processor and a memory coupled to the processor. The memory can encode one or more programs to enable the processor to perform one or more methods described in this application. In some examples, the system can be a general-purpose computer system. In other examples, the system can be a special-purpose computer system including an embedded system. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] The details of one or more embodiments are set forth in the following exemplary drawings and description. Other features will be apparent from the description, drawings, and from the claims. However, it should be understood that even if embodiments are described separately, individual features of different embodiments may be combined into further embodiments.

[0095] Figure 1 An exemplary development process for generating a protein expressing cell line is shown.

[0096] Figure 2 Examples of product titers obtained by exemplary protein expression of the same class of proteins (eg, antibodies) having different amino acid sequences are shown.

[0097] Figure 3 A functional block diagram of an exemplary system according to the present disclosure is shown.

[0098] Figure 4 A flow chart of an exemplary process performed by an exemplary forecasting system according to the present disclosure is shown.

[0099] Figure 5 An exemplary workflow for estimating and evaluating prediction functions according to various prediction algorithms is shown.

[0100] Figure 6 Examples of different types of subsequences of antibodies or recombinant antibodies are shown.

[0101] Fig. 7A An exemplary encoding algorithm that may be used with the present disclosure is shown.

[0102] Figure 7B Another exemplary encoding algorithm that may be used with the present disclosure is shown.

[0103] Figure 8 A specific example of encoding an amino acid sequence to generate a first numerical vector is shown.

[0104] Fig. 9 An example of a graph generated in an encoding algorithm based on an exemplary word embedding is shown.

[0105] Fig.10 A specific example of reducing the dimension of a first numerical vector to generate a second numerical vector is shown.

[0106] Fig.11 A specific example of estimating a prediction function based on a regression algorithm is shown.

[0107] Fig.12 A specific example of classifying the output of the prediction function to evaluate the prediction function is shown.

[0108] Fig.13 An exemplary hardware configuration of a computer that can be used to implement at least a portion of a system according to the present disclosure is shown.

[0109] Detailed description of implementation plan

[0110] Hereinafter, a detailed description of the examples will be given with reference to the accompanying drawings. It should be understood that various modifications may be made to the examples. In particular, one or more elements of an example may be combined and used in other embodiments to form new embodiments.

[0111] Protein expression

[0112] Figure 1 An exemplary development process for generating a protein-expressing cell line is shown. Figure 1 As shown, the target protein can be represented by sequence data including the amino acid sequence of the protein. Protein sequence data, such as amino acid sequence, can be reverse translated into DNA sequence data. Once the DNA sequence data is obtained, the DNA sequence can be transfected into a host cell line. After transfection, the cell can be subjected to a selection process, in which a cell pool carrying the DNA sequence can be obtained. The pool can then be used to implement clonal generation, and amplification of the clones can be performed. A further selection process can be performed on the clones of the cells. The titer of the target protein can be measured, for example, at the end of the protein expression process. Additionally or alternatively, the titer of the target protein can be measured when the cell pool is obtained, during amplification and selection and / or after amplification and selection before a further selection process. Titer, also referred to as "product titer", can be understood as indicating the efficiency or performance of protein expression.

[0113] Figure 1 Details of at least a portion of the exemplary process shown can be found in Lai et al., “Advances in Mammalian Cell Line Development Technologies for Recombinant Protein Production”, Pharmaceuticals 2013, 6(5), p.579-603, doi: 10.3390 / ph6050579.

[0114] Figure 2 Examples of product titers obtained as a result of exemplary protein expression of antibodies of the same species having different amino acid sequences are shown. Figure 2 Antibodies A, B, and C mentioned in the Figure 1 The exemplary automated cell culture system shown is expressed. Further, Figure 2The antibodies A, B and C mentioned in are antibodies of the same type but have different amino acid sequences. Figure 2 The average titer of antibody B was shown to be 40% lower than that of antibody A, and the average titer of antibody C was shown to be 66% lower than that of antibody A.

[0115] from Figure 2 As can be seen in Figure 1, in a class of proteins (e.g., antibodies), different amino acid sequences can and mostly will result in different efficiencies (e.g., product titers) when the protein is expressed. It is often difficult to identify and / or predict the correlation between the amino acid sequence of a protein and the expression efficiency of the protein. The present disclosure is directed to optimizing protein expression with improved efficiency by making it possible to predict the efficiency from the amino acid sequence.

[0116] System Configuration

[0117] Figure 3 A functional block diagram of an exemplary system according to the present disclosure is shown. Figure 3 The exemplary system shown includes a prediction system 10 and an automated cell culture system 20 .

[0118] like Figure 3 As shown, the prediction system 10 may include a processor 102 and a storage medium 104 .

[0119] Processor 102 can be configured to predict the efficiency of protein expression from the amino acid sequence. For the prediction, a variety of prediction algorithms can be used. Each prediction algorithm includes a combination of an encoding algorithm, a dimensionality reduction algorithm, and a regression algorithm. Different prediction algorithms can include different combinations of these algorithms.

[0120] For example, the processor 102 may be configured to obtain a plurality of amino acid sequences and corresponding known efficiency values, each of which indicates the efficiency of expressing a protein having a corresponding one of the plurality of amino acid sequences. The processor 102 may be configured to estimate a prediction function using the plurality of amino acid sequences and the known efficiency values ​​according to each of the prediction algorithms, the prediction function taking a numerical vector corresponding to the amino acid sequence as input and outputting a predicted efficiency value of expressing a protein having the amino acid sequence. The processor 102 may be further configured to evaluate the prediction function estimated according to a plurality of prediction algorithms, and select at least one prediction algorithm based on the evaluation. The processor 102 may then use at least one prediction algorithm and the prediction function obtained using the at least one prediction algorithm to predict one or more efficiency values ​​of one or more proteins expressing one or more specified amino acid sequences, respectively. The processor 102 may be further configured to output one or more specified amino acid sequences and one or more efficiency values ​​predicted for one or more specified amino acid sequences. Further details of the prediction process performed by the processor 102 will be described later below.

[0121] The storage medium 104 may be configured to store information necessary for the processor 102 to predict protein expression efficiency. For example, the storage medium 104 may be configured to store a plurality of amino acid sequences and corresponding known efficiency values ​​for prediction as described above. Further, for example, the storage medium 104 may be configured to store information about a plurality of prediction algorithms. Further, for example, the storage medium 104 may be configured to store program code that enables the processor 102 to perform the prediction process as described above. In addition, the storage medium 104 may be configured to store information that is obtained during and / or as a result of the prediction process of the processor 102 as described above.

[0122] The prediction system 10 may be connected to the automated cell culture system 20 via one or more wired and / or wireless communication networks.

[0123] The automated cell culture system 20 may be a system capable of controlling the expression of one or more proteins. As a specific example, the automated cell culture system 20 may be configured to perform the above-described Figure 1The exemplary development process described. For example, the automated cell culture system 20 may include one or more holders of one or more receptacles, in which cells can be cloned, amplified and / or selected. Further, the automated cell culture system 20 may include at least one container, which is configured to provide a closed space for accommodating the receptacles and their respective holders. The closed space can provide a controlled environment (e.g., temperature, humidity, stirring, etc.) for cloning, amplification and / or selection of cells. In addition, the automated cell culture system 20 may include a device for stirring cells, such as a vibrator, a stirrer, etc. The system used as the automated cell culture system 20 may include a combination of Ambr15 (Sartorius) and ALS CellCelector (Sartorios) and / or Octet (Sartoris) and / or a liquid handler and / or a robotic arm and / or a cell culture incubator.

[0124] In some exemplary embodiments, the automated cell culture system 20 may further include a computer (not shown) configured to control the operation of the automated cell culture system 20. For example, the computer of the automated cell culture system 20 may control one or more parameters of protein expression (e.g., temperature, humidity, stirring, etc.). The computer included in the automated cell culture system 20 may be further configured to receive one or more specified amino acid sequences and respective predicted efficiency values ​​from the processor 102 of the prediction system 10, so that one or more proteins having the one or more specified amino acid sequences, respectively, may be expressed in the automated cell culture system 20.

[0125] In some exemplary embodiments, the system does not necessarily include an automated cell culture system 20. For example, in these exemplary embodiments, the output from the prediction system 10 can be used in a manual cell culture workflow.

[0126] Efficiency prediction

[0127] Figure 4 A flow chart of an exemplary process performed by an exemplary prediction system according to the present disclosure is shown. For example, Figure 3 The processor 102 of the prediction system 10 shown may execute Figure 4 As shown in the exemplary process above Figure 3 As stated, Figure 4 The exemplary process shown may be considered an example of a prediction process performed by processor 102 .

[0128] For example, when user input indicating initiation of the exemplary process is received by the processor 102 of the prediction system 10 via an input device (not shown), Figure 4 The exemplary process shown may be initiated.

[0129] In step S10, processor 102 may obtain a plurality of amino acid sequences and corresponding known efficiency values. In some exemplary embodiments, each amino acid sequence may include a series of amino acids in single-letter notation. For example, a plurality of amino acid sequences and corresponding known efficiency values ​​may be received from storage medium 104 and / or from a computer (e.g., a database server) outside prediction system 10. The known efficiency values ​​may be determined by expressing a protein having a corresponding amino acid sequence using a specific type of host cell and a specific type of expression vector for protein expression (see Figure 1 ). Information indicating a specific type of host cell and a specific type of expression vector may also be obtained in step S10 and stored in the storage medium 104. In some exemplary embodiments, the known efficiency value may be the titer of expressing proteins having the corresponding amino acid sequences.

[0130] In some exemplary embodiments, the multiple amino acid sequences obtained in step S10 may be of an antibody or a recombinant antibody. Further, in some cases where the multiple amino acid sequences are of an antibody or a recombinant antibody, the multiple amino acid sequences may include a light chain sequence of the antibody or the recombinant antibody, a heavy chain sequence of the antibody or the recombinant antibody, or a light-heavy chain sequence of the antibody or the recombinant antibody.

[0131] Figure 4 The exemplary process may proceed to step S20 after step S10.

[0132] In step S20, processor 102 can use the multiple amino acid sequences obtained in step S10 and the corresponding known efficiency values ​​to estimate and evaluate the prediction function according to the available prediction algorithm. Each of the available prediction algorithms can include a combination of an encoding algorithm, a dimensionality reduction algorithm, and a regression algorithm. Using the encoding algorithm, the amino acid sequence can be encoded as a numerical vector. The dimension of the numerical vector can be reduced by a dimensionality algorithm. Further, the prediction function can be estimated according to the regression algorithm. Also as described above, the prediction function takes the numerical vector corresponding to the amino acid sequence as input, and outputs the predicted efficiency value of expressing the protein with the amino acid sequence. In the exemplary embodiment in which the known efficiency value is the titer as described above, the predicted efficiency value can also be the predicted titer of expressing the protein with the amino acid sequence.

[0133] Figure 5 An exemplary workflow for estimating and evaluating prediction functions according to various prediction algorithms is shown. Figure 5 The exemplary workflow shown can be implemented Figure 4 The estimation and evaluation of step S20 of the exemplary process shown. Figure 5 As shown, the exemplary workflow includes inputting a sequence (step S200), encoding the sequence (step S202), reducing dimension (step S204), machine learning regression (step S206), and evaluating (step S208). The processor 102 may implement process modules corresponding to steps S200 to S208, respectively, to perform Figure 4 Step S20 of the exemplary process is shown.

[0134] exist Figure 5 In step S200 of the exemplary workflow shown in FIG. 1 , the processor 102 may pre-process the Figure 4 The processor 102 may extract at least a portion of each of the plurality of amino acid sequences. For example, the processor 102 may extract at least a portion of each of the plurality of amino acid sequences. For example, at least a portion of each amino acid sequence may include the first x% of the amino acid sequence and / or the last y% of the amino acid sequence, wherein each of x and y may be a value greater than 0 and less than or equal to 50. At least a portion of each amino acid sequence may also be referred to as a "subsequence".

[0135] By extracting subsequences of multiple amino acid sequences, sequence regions with the most informative and discriminative patterns can be retained for further processing. In other words, the idea of ​​generating subsequences can be understood as skipping highly similar sequence regions to increase the uniqueness level.

[0136] The regression algorithm used to estimate the prediction function in step S206 can then learn the discriminant pattern in a data-driven manner, which helps the regression algorithm predict the appropriate amino acid sequence group unknown to the regression algorithm. In some exemplary embodiments, in order to explore which specific sequence type has more potential to support the regression algorithm in terms of the efficiency of predicting protein expression, different types of amino acid sequences can be input into the exemplary workflow. For example, in the case where multiple amino acid sequences are antibodies or recombinant antibodies, the following three types of sequences can be used as input: light chain sequence, heavy chain sequence or light-heavy chain sequence. The light-heavy chain sequence can be generated by combining light chain sequence and heavy chain sequence.

[0137] Figure 6 Specific examples of different types of subsequences of antibodies or recombinant antibodies are shown. Figure 6 In the present invention, "full HC", "full HCLC" and "full LC" represent the full-length heavy chain sequence, light-heavy chain sequence and heavy chain sequence, respectively. Figure 6 In , “S-25” and “S-50” represent the first 25% and first 50% of the corresponding sequence, respectively. Figure 6 In the table, “E-25” indicates the last 25% of the corresponding sequence. Figure 6The exemplary subsequence (a) shown includes the first 25% and last 25% of the heavy chain sequence. Figure 6 The exemplary subsequence (b) shown includes the first 50% of the light chain sequence. Figure 6 The exemplary subsequence (c) shown includes a combination of subsequences (a) and (b), in other words, the first 25% and the last 25% of the heavy chain sequence and the first 50% of the light chain sequence. After extensive experiments on three types of amino acid sequences, heavy chain, light chain and light and heavy chain, it has been found that Figure 6 The exemplary subsequences (a), (b) and (c) shown can represent the most discriminative parts of the light chain, heavy chain and light-heavy chain sequence types, respectively. Figure 5 The exemplary workflow shown is performed using the following types of amino acid sequences as input:

[0138] - The top 25% of each sequence type (heavy chain, light chain, light-heavy chain);

[0139] - The top 50% of each sequence type (heavy chain, light chain, light-heavy chain);

[0140] - the bottom 25% of each sequence type (heavy chain, light chain, light-heavy chain);

[0141] - the bottom 50% of each sequence type (heavy chain, light chain, light-heavy chain);

[0142] - The top 25% and bottom 25% of each sequence type (heavy chain, light chain, light-heavy chain); and

[0143] - The first 50% and the last 50% of each sequence type (heavy chain, light chain, light-heavy chain), in other words, the full length of each sequence type.

[0144] In the experiment, Figure 6 The exemplary subsequence (c) shown shows the best performance.

[0145] Refer again Figure 5 In some cases, for example, when processing multiple full-length amino acid sequences, the preprocessing of step S200 can be skipped. For a specific example, when the light chain sequence, heavy chain sequence or light-heavy chain sequence of an antibody or a recombinant antibody is Figure 4 Step S10 is obtained in the case, and if the full length of these sequences are to be processed, it can be skipped Figure 5 Step S200.

[0146] In step S202, the processor 102 may encode the multiple amino acid sequences preprocessed in step S200 according to a plurality of encoding algorithms to generate a numerical vector corresponding to the multiple amino acid sequences. The numerical vector generated in step S202 may be referred to as a "first numerical vector".

[0147] As also described above, each amino acid sequence can include a series of amino acids in single letter notation. Thus, each amino acid sequence can be a series of alphabetical characters. On the other hand, Figure 5 The machine learning algorithm used in step S206 of the exemplary workflow shown may require a statistical representation of the amino acid sequence, such as a numerical representation, for predictive modeling. Therefore, the plurality of amino acid sequences may be encoded in step S202 to generate a first numerical vector corresponding to the plurality of amino acid sequences. In some exemplary embodiments, a plurality of encoding algorithms may be used to perform Figure 5 The encoding of step S202.

[0148] Therefore, in step S202, for example, for each of the plurality of amino acid sequences, the processor 102 may encode at least a portion of one of the plurality of amino acid sequences (e.g., the subsequence obtained in step S200) according to one of the plurality of encoding algorithms to generate a first numerical vector. This encoding step may be repeated using different algorithms in the plurality of encoding algorithms to generate multiple sets of first numerical vectors.

[0149] Figure 5 Examples of the multiple encoding algorithms used in step S202 may include, but are not limited to, three types of algorithms: statistical, physical-chemical, and word embedding. A statistical encoding algorithm (also referred to as a "statistical encoder") may capture the distribution information of amino acids. An example of a statistical encoder may be based on okapi-BM25, which will be described in detail later below.

[0150] Physicochemical property-based encoding algorithms (also called "physicochemical property-based encoders") can make use of pre-calculated values ​​for amino acids. In this category, some algorithms, such as AAINDEX, can replace amino acids with their respective values ​​(see Fig. 7A ), and some algorithms, such as CTDC (composition / transition / distribution), can use physicochemical properties together with some statistical procedures (for detailed examples, see Figure 7B ).related Figure 7B Specific examples of CTDC are shown, and Table 1 below summarizes exemplary properties and groups that can be used for the specific examples.

[0151] Table 1

[0152]

[0153] Furthermore, the following equation can be used for Figure 7B The specific example of CTDC is shown.

[0154]

[0155]

[0156] The word embedding-based encoding algorithm (also called "word embedding-based encoder") can utilize a deep learning model and learn the distribution of amino acids in an unsupervised manner.

[0157] Also as mentioned above, in some exemplary embodiments, multiple coding algorithms can include one or more algorithms based on the physicochemical properties of protein. In the coding algorithm based on physicochemical properties, according to the physical and chemical characteristics of protein, such as dissociation, solubility, hydration, polarity, charge, hydrophobicity, molecular weight, size, etc., different weight values ​​can be assigned to each amino acid. For example, the weight value can be calculated by the abundant wet laboratory experiment carried out by field experts. As a specific example, the following table 2 shows the weight values ​​(referred to as "characteristic value" in Table 2) assigned to 20 kinds of amino acids based on physicochemical properties.

[0158] Table 2

[0159]

[0160]

[0161] Take the following amino acid sequence as an exemplary hypothetical sequence for a protein:

[0162] (SEQ ID NO:38)MLKKRFLGALA...,

[0163] The numerical vector (eg, the first numerical vector) generated using the weights based on the physicochemical properties shown in Table 2 may include the following:

[0164] [-0.239 -0.267 0.243 0.243 0.171 -0.329 -0.267 0.218 0.008 -0.2670.008…].

[0165] In an exemplary embodiment where the plurality of encoding algorithms includes one or more encoding algorithms based on physicochemical properties, weight values ​​such as those shown in Table 2 may be stored in Figure 3 In the storage medium 104 of the prediction system 10 shown in FIG. Figure 5 When encoding the amino acid sequence to generate a first numerical vector in step S202 , the processor 102 may refer to the weight values ​​stored in the storage medium 104 .

[0166] Further, in some exemplary embodiments, the plurality of encoding algorithms may include one or more algorithms based on k-mers of amino acid sequences. A "k-mer" of an amino acid sequence may be understood as a substring of length k contained in the amino acid sequence. In a k-mer-based encoding algorithm, a k-mer of at least a portion of a plurality of amino acid sequences may first be generated by rotating a window of size k with a step size of S. The frequency of each generated k-mer is used to encode an amino acid sequence comprising a k-mer.

[0167] As a specific example of a k-mer based encoding algorithm, Okapi-BM25 may be used. Okapi-BM25 is a ranking function used in information retrieval and may provide a score indicating the relevance of a document to a query based on the frequency with which terms contained in the query appear in the document.

[0168] In the encoding algorithm involving Okapi-BM25, all k-mers can be considered as queries, and all sequences to be encoded can be considered as documents. To generate an encoding of an amino acid sequence, the first step can be to generate k-mers of the amino acid sequence, the k-mers being generated by rotating a window of size k with a step size of S. For better illustration, consider an exemplary hypothetical corpus containing 2 sequences, where the 2-mers are generated by rotating a window of size 2 with a step size of 1 on the sequence.

[0169] Sequence 1: ACGTTGACTT (SEQ ID NO: 39; 2-mer=AC CG GT TT TG GA AC CT TT)

[0170] Sequence 2: AGCTGTACTT (SEQ ID NO: 40; 2-mer=AG GC CT TG GT TA AC CT TT)

[0171] After generating the k-mer, the second step may be to find a unique k-mer vocabulary (e.g., V = {AC, CG, GT, TT, TG, GA, AC, CT, TT, AG, GC, TA}). The processor 102 may then calculate the elements of the numeric vector (e.g., the first numeric vector) for each k-mer present in the sequence. Specifically, for example, using the k-mer vocabulary, the processor 102 may calculate the k-mer frequency (KF) in the amino acid sequence Seq, the number of amino acid sequences in which the k-mer appears as a sequence frequency (SF), and the k-mer anti-sequence frequency (ISF) to calculate the k-mer BM25 score and / or the OkapiBM25 score using the following equation:

[0172]

[0173] OkapiBM25 score (q i , Seq) = BM25 score (q i ,Seq)×ISF(q i ) (3)

[0174] In the above equation, q i can represent a specific k-mer from a unique k-mer vocabulary V, |Seq| can represent the number of k-mers present in the amino acid sequence Seq, and S avg It can represent the average number of k-mers present in the amino acid sequence corpus. Further, k1 and b can be hyperparameter values, for example, they can fall within the range of k1=[1.2, 2.0] and b=0.75.

[0175] Using the above equation, assuming k1 = 1.2, b = 0.75, |Seq| = 9, S avg =9, for kmer (AC) that appears twice in sequence 1, KF will be 2, ISF (AC) will be -0.6989, BM25 score (AC, sequence 1) will be 1.375, and OkapiBM25 score (AC, sequence 1) will be -0.96. The value of BM25 score (AC, sequence 1) or Okapi BM25 score (AC, sequence 1) can be regarded as the weight of kmer (AC) in sequence 1. Similarly, the weights of other kmers (e.g., the values ​​of individual BM25 scores or individual Okapi BM25 scores) can be calculated to generate a numerical vector for sequence 1. The numerical vectors for sequence 2 and all other amino acid sequences present in the corpus can also be calculated in a similar manner.

[0176] As described above, the value of the BM25 score (Equation (2)) or the value of the Okapi BM25 score (Equation (3)) can be used as a weight assigned to the kmer. In the case of using the value of the Okapi BM25 score as the weight, all of the above equations (1) to (3) can be used for calculation. In the case of using the value of the BM25 score as the weight, the above equation (2) can be used for calculation, and equations (1) and (3) are not required for calculation.

[0177] Figure 8 A specific example of encoding an amino acid sequence using Okapi-BM25 according to a k-mer based encoding algorithm to generate a first numerical vector is shown. Figure 8 In the specific example of , the full-length light-heavy chain sequences are encoded as a numeric vector.

[0178] As also described above, in some exemplary embodiments, the plurality of encoding algorithms may include one or more encoding algorithms based on word embedding. A practical example of an encoding algorithm based on word embedding generated based on a statistical representation represented by a graph will be described below. To make the example simple, only four different amino acids are used in the following examples, rather than 20 different amino acids.

[0179] Graph Representation Embedding

[0180] Sequence data:

[0181] GCCGAGACTA (SEQ ID NO:41)

[0182] GAGTCACATC (SEQ ID NO:42)

[0183] AGTCCTGACA (SEQ ID NO:43)

[0184] AGTCCTGACA (SEQ ID NO:43)

[0185] AGTCCTGACA (SEQ ID NO:43)

[0186] Step 1: Generate bimers for each sample

[0187] GCCGAGACTA:['GC','CC','CG','GA','AG','GA','AC','CT','TA']

[0188] GAGTCACATC:['GA','AG','GT','TC','CA','AC','CA','AT','TC']

[0189] AGTCCTGACA:['AG','GT','TC','CC','CT','TG','GA','AC','CA']

[0190] AGTCCTGACA:['AG','GT','TC','CC','CT','TG','GA','AC','CA']

[0191] AGTCCTGACA:['AG','GT','TC','CC','CT','TG','GA','AC','CA']

[0192] Step 2: Make a list of bimers by concatenating all samples

[0193] ['GC','CC','CG','GA','AG','GA','AC','CT','TA','GA','AG','GT','TC','CA','AC','CA','AT','TC','AG','GT','TC','CC','C T','TG','GA','AC','CA','AG','GT','TC','CC','CT','TG','GA','AC','CA','AG','GT','TC','CC','CT','TG','GA','AC','CA']

[0194] Step 3: Make bimer pairs

[0195] [('GC','CC'),('CC','CG'),('CG','GA'),('GA','AG'),('AG','GA'),('GA','AC'),('AC','CT'),('CT','TA'),('TA','GA'),('GA','AG'),('AG','GT') ,('GT','TC'),('TC','CA'),('CA','AC'),('AC','CA'),('CA','AT'),('AT','TC'),('TC','AG'),('AG','GT'),('GT','TC'),('TC','CC'),('CC','CT') ,('CT','TG'),('TG','GA'),('GA','AC'),('AC','CA'),('CA','AG'),('AG','GT'),('GT','TC'),('TC','CC'),('CC','CT'),('CT','TG'),('TG','GA') ,('GA','AC'),('AC','CA'),('CA','AG'),('AG','GT'),('GT','TC'),('TC','CC'),('CC','CT'),('CT','TG'),('TG','GA'),('GA','AC'),('AC','CA')]

[0196] Step 4: Generate a graph with unique bimers as nodes and bimer pair relationships as edges

[0197] ['GC','CC','CG','GA','AG','AC','CT','TA','GT','TC','CA','AT','TG']

[0198] The above bimers are unique bimers, and each bimer can be represented by an integer. Fig. 9 An example of the resulting graph is shown in .

[0199] Step 5: Create an adjacency matrix from the graph:

[0200] 1. Read the graph and get all nodes

[0201] 2. Create an adjacency matrix by taking the edges from the graph.

[0202] 3. And the matrix as an array can be: [[0.1.0.0.0.0.0.0.0.0.0.0.] [1.0.1.0.0.0.1.0.0.1.0.0.0.] [0.1.0.1.0.0.0.0.0.0.0.0.] [0.0.1.0.1.1.0.1.0.0.0.0.1.] [0.0.0.1.0.0.0.0.1.1.1.0.0.] [0.0.0.1.0.0.0.0.1.0.0.] [0.1.0.0.0.1.0.1.0.0.0.0.1.] [0.0.0.1.0.0.1.0.0.0.0.0.0.] [0.0.0.0.1.0.0.0.0.1.0.0.0.] [0.1.0.0.1.0.0.0.1.0.1.0.] [0.0.0.0.1.1.0.0.0.1.0.1.0.] [0.0.0.0.0.0.0.0.0.1.1.0.0.]

[0215] [0.0.0.1.0.0.1.0.0.0.0.0.0.]].

[0216] Step 6: At this point, the algorithm can read the matrix and the sum of each array and create a matrix of size (1,13) which can be:

[0217] [[1.4.2.5.4.3.4.2.2.5.4.2.2.]].

[0218] Step 7: The algorithm can then repeat the above matrix with node size 13 and create a matrix such as: [[1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.] [1.4.2.5.4.3.4.2.2.5.4.2.2.]

[0231] [1.4.2.5.4.3.4.2.2.5.4.2.2.]].

[0232] Step 8: The algorithm can then take the log of the adjacency matrix / repetition matrix and subtract the log of 1 / node_size.

[0233] node_size = maximum number of nodes

[0234] Matrix = log(adjacency matrix / repetition matrix)-log(1 / node_size)

[0235] And the matrix can be:

[0236]

[0237] Step 9: At this point, the above matrix with dimension size can be passed to SVD (Singular Value Decomposition) function, which returns 3 values, for example:

[0238] U: The eigenvector matrix of (matrix transpose and matrix dot product) and returns a matrix of dimension (13,2) which can be:

[0239]

[0240] Sigma: The singular values ​​of the eigenvector matrix (matrix transpose and matrix dot product), which can be:

[0241] [4.06050863 5.00618494].

[0242] VT: The transpose of U of dimension (2,13), which can be:

[0243]

[0244] The sigma vector can be transformed into (2,2), which can be:

[0245] [[4.06050863 0.],[0.5.00618494]].

[0246] Step 10: Subsequently, U may be multiplied by the alpha value of sigma raised to the power of 0.5, and a vector of shape (13,2) may be returned.

[0247] Step 11: Further, the transpose of V may be multiplied by the power of the product of sigma and an alpha value of 0.5, and a vector of shape (13, 2) may be returned.

[0248] Step 12: Then, the two vectors of steps 10 and 11 above can be added, and a vector of dimension (13,2) can be returned.

[0249] Step 13: The vector from step 12 above can be normalized and a vector of shape (13,2) can be returned:

[0250]

[0251]

[0252] Step 14: The vector can be transformed to (13,4), and the embedding can be returned, for example, by getting the embedding of each node from a lookup table:

[0253] GC -0.28827766755584794 0.9575468585863358 -0.5711549132568590.8208422900061557

[0254] CC -0.3007563592459236 0.9537010078494922 0.53104872070666680.847341286752754

[0255] CG 0.14399019303043473 0.9895791147306304 0.091403213740203370.9958139648136908

[0256] GA 0.18047274984044 0.9835799848334806 -0.81870689491591970.5742116510635545

[0257] AT -0.7696477736771328 0.6384687184771335 0.47194900550313930.8816258481944583

[0258] AC -0.08907567614296603 0.9960248610951804 0.73801568032198120.6747835620396242

[0259] CT 0.27387810042293925 0.9617644129976544 -0.777470625041060.6289192533213332

[0260] TA 0.3980041721080683 0.9173836051426748 0.74747667855901970.6642880512325778

[0261] GT-0.8884333530344746 0.45900563963411234 -0.487815063994093860.8729469991588482

[0262] TC -0.8469286500794678 0.531706555982311 -0.41493593988485310.9098506282856947

[0263] CA -0.7847655072557946 0.6197927868421471 -0.687459565754060.7262226555631816

[0264] AT-0.8596587968915401 0.5108686258980776 0.246334790104278250.9691847972313026

[0265] TG 0.39800417210806815 0.9173836051426748 0.74747667855901960.6642880512325778.

[0266] The various examples of encoding algorithms described above are merely examples and do not limit the encoding algorithms. In other words, the various encoding algorithms may include algorithms other than the exemplary algorithms described above. Further non-limiting examples of encoding algorithms are summarized as follows:

[0267] Word embedding based encoder

[0268] -Word2vec[1]

[0269] -FastText[4][5]

[0270] -DeepWalk

[12]

[0271] - Graph Auto Encoder (GAE)

[13]

[0272] -Graph Factorization (GF)

[14]

[0273] - Graph Representation (GraRep)

[18]

[0274] -Large scale Information Network Embedding

[0275] Network Embedding (LINE)

[19]

[0276] -High-Order Proximity Preserving Embedding

[0277] preserved Embedding(HOPE))

[21]

[0278] -Laplacian

[23]

[0279] -Structural Deep Network Embedding

[0280] Embedding (SDNE)

[25]

[0281] -Singular Value Decomposition (SVD)

[27]

[0282] -Node2Vec

[32]

[0283] Statistical encoder

[0284] Composition of k-spaced amino acid pairs

[0285] Amino Acid Pairs (CKSAAP))[2]

[0286] -MonoMonoKGap[6]

[0287] -MonoDiKGap[6]

[0288] -MonoTriKGap[6]

[0289] -DiMonoKGap[6]

[0290] -DiDiKGap[6]

[0291] -DiTriKGap[6]

[0292] -TriMonoKGap[6]

[0293] -TriDiKGap[6]

[0294] -Accumulated Nucleotide Frequency (ANF)

[45]

[0295] -K-mer

[28]

[29]

[0296] -EAAC

[33] [2]

[0297] -TFIDF

[35]

[36]

[0298] -One-Hot Vector Encoding

[35]

[0299] -Cross-Covariance

[38]

[39]

[40]

[0300] -Okapi-BM25

[41]

[0301] -GAAC

[33] [2]

[0302] -WSRC-Local

[42]

[0303] -WSRC-Global

[42]

[0304] -EGAAC

[33] [2]

[0305] -GTPC

[33] [2]

[0306] -Ctriad

[43]

[0307] -KSCTriad

[33] [2]

[0308] -CKSAAGP

[33] [2]

[0309] -PseKRAAC (Type 1 to Type 16)

[44]

[0310] -KSCTriad

[33] [2]

[0311] Encoders based on physicochemical properties

[0312] -AAIndex[3]

[0313] -CTDC[7][8][9]

[10]

[11]

[0314] -CTDT[7][8][9]

[10]

[11]

[0315] -CTDD[7][8][9]

[10]

[11]

[0316] -Soc-Number

[15]

[16]

[17]

[0317] -QSOrder

[15]

[16]

[17]

[0318] -PAAC

[20]

[0319] -APAAC

[22]

[0320] -BLOSUM62

[24]

[0321] -Z-Scale

[26]

[0322] -Moran

[30]

[31]

[0323] -Geary

[34]

[0324] -NMBroto

[37]

[0325] -Auto-covariance

[38]

[39]

[40]

[0326] Auto-Cross-Covariance

[38]

[39]

[40]

[0327] Thanks for watching[waiting] for a snowflake in a snowstorm:

[0328] [1]Kenneth Ward Church.Word2vec.Natural Language Engineering,23(1):155–162,2017.

[0329] [2]Zhen Chen,Pei Zhao,Fuyi Li,Tatiana T Marquez-Lago,Andre′Leier,Jerico Revote,Yan Zhu,David R Powell,Tatsuya Akutsu,Geoffrey I Webb,etal sequencedata.Briefings in Bioinformatics,21(3):1047–1057,2020.

[0330] [3]Chun-Wei Tung and Shinn-Ying Ho.Computational Identification ofUbiquitylation Sites from Protein Sequences.BMC Bioinformatics,9(1):1–15,2008.

[0331] [4]Michal Busta,Lukas Neumann,and Jiri Matas.Fastext:An efficientunconstrained scene text detector.In Proceedings of the IEEE InternationalConference on Computer Vision,pages 1206–1214,2015.

[0332] [5]Alexander Filonenko,Konstantin Gudkov,Aleksei Lebedev,IvanZagaynov,and Nikita Orlov.Fastext:Fast and small text extractor.In2019International Conference on Document Analysis and Recognition Workshops(ICDARW),volume 4,pages 49–54.IEEE,2019.

[0333] [6]Rafsanjani Muhammod,Sajid Ahmed,Dewan Md Farid,Swakkhar Shatabda,Alok Sharma,and Abdollah Dehzangi.Pyfeat:a python-based effective featuregeneration tool for dna,rna and protein sequences.Bioinformatics,35(19):3831-3833,2019.

[0334] [7]CZ Cai,LY Han,Zhi Liang Ji,X Chen,and Yu Zong Chen.Svm-prot:web-based support vector machine software for functional classification of aprotein from its primary sequence.Nucleic acids research,31(13):3692–3697,2003.

[0335] [8]CZ Cai,LY Han,ZL Ji,and YZ Chen.Enzyme family classification bysupport vector machines.Proteins:Structure,Function,and Bioinformatics,55(1):66–76,2004.

[0336] [9]Inna Dubchak,Ilya Muchnik,Stephen R Holbrook,and Sung-HouKim.Prediction of protein folding class using global description of aminoacid sequence.Proceedings of the National Academy of Sciences,92(19):8700–8704,1995.

[0337]

[10] Inna Dubchak,Ilya Muchnik,Christopher Mayor,Igor Dralyuk,andSung-Hou Kim.Recognition of a protein fold in the context of the scopclassification.Proteins:Structure,Function,and Bioinformatics,35(4):401–407,1999.

[0338]

[11] Lian Yi Han,Cong Zhong Cai,Siew Lin Lo,Maxey CM Chung,and Yu ZongChen.Prediction of rna-binding proteins from primary sequence by a supportvector machine approach.Rna,10(3):355–368,2004.

[0339]

[12] Bryan Perozzi,Rami Al-Rfou,and Steven Skiena.Deepwalk:Onlinelearning of social representations.In Proceedings of the 20th ACM SIGKDDinternational conference on Knowledge discovery and data mining,pages 701–710,2014.

[0340]

[13] Thomas N Kipf and Max Welling.Variational graph auto-encoders.arXiv preprint arXiv:1611.07308,2016.

[0341]

[14] Palash Goyal and Emilio Ferrara.Graph embedding techniques,applications,and performance:A survey.Knowledge-Based Systems,151:78–94,2018.

[0342]

[15] Kuo-Chen Chou.Prediction of protein subcellular locations byincorporating quasi-sequence-order effect.Biochemical and biophysicalresearch communications,278(2):477–483,2000.

[0343]

[16] Kuo-Chen Chou and Yu-Dong Cai.Prediction of protein subcellularlocations by go–fund–pseaa predictor.Biochemical and Biophysical ResearchCommunications,320(4):1236–1239,2004.

[0344]

[17] Gisbert Schneider and Paul Wrede.The rational design of aminoacid sequences by artificial neural networks and simulated molecularevolution:de novo design of an idealized leader peptidase cleavagesite.Biophysical Journal,66(2):335–344,1994.

[0345]

[18] Shaosheng Cao,Wei Lu,and Qiongkai Xu.Grarep:Learning graphrepresentations with global structural information.In Proceedings of the 24thACM international on conference on information and knowledge management,pages891–900,2015.

[0346]

[19] Jian Tang,Meng Qu,Mingzhe Wang,Ming Zhang,Jun Yan,and QiaozhuMei.Line:Large-scale information network embedding.In Proceedings of the 24 th international conference on world wide web,pages1067–1077,2015.

[0347]

[20] Kuo-Chen Chou.Prediction of protein cellular attributes usingpseudo-amino acid composition.Proteins:Structure,Function,and Bioinformatics,43(3):246–255,2001.

[0348]

[21] Dingyuan Zhu,Peng Cui,Ziwei Zhang,Jian Pei,and Wenwu Zhu.High-order proximity preserved embedding for dynamic networks.IEEE Transactions onKnowledge and Data Engineering,30(11):2134–2144,2018.

[0349]

[22] Kuo-Chen Chou.Using amphiphilic pseudo amino acid composition topredict enzyme subfamily classes.Bioinformatics,21(1):10–19,2005.

[0350]

[23] Russell Merris.Laplacian matrices of graphs:a survey.Linearalgebra and its applications,197:143–176,1994.

[0351]

[24] Tzong-Yi Lee,Shu-An Chen,Hsin-Yi Hung,and Yu-Yen Ou.Incorporatingdistant sequence features and radial basis function networks to identifyubiquitin conjugation sites.PloS one,6(3):e17331,2011.

[0352]

[25] Daixin Wang,Peng Cui,and Wenwu Zhu.Structural deep networkembedding.In Proceedings of the 22nd ACM SIGKDD international conference onKnowledge discovery and data mining,pages 1225–1234,2016.

[0353]

[26] Yong-Zi Chen,Zhen Chen,Yu-Ai Gong,and Guoguang Ying.Sumohydro:anovel method for the prediction of sumoylation sites based on hydrophobicproperties.PloS one,7(6):e39195,2012.

[0354]

[27] HervéAbdi.Singular value decomposition(svd)and generalizedsingular value decomposition.Encyclopedia of measurement and statistics,pages907–912,2007.

[0355]

[28] William Stafford Noble,Scott Kuehn,Robert Thurman,Man Yu,and JohnStamatoyannopoulos.Predicting the in vivo signature of human gene regulatorysequences.Bioinformatics,21(suppl 1):i338–i343,2005.

[0356]

[29] Shobhit Gupta,Jonathan Dennis,Robert E Thurman,Robert Kingston,John A Stamatoyannopoulos,and William Stafford Noble.Predicting humannucleosome occupancy from primary sequence.PLoS computational biology,4(8):e1000134,2008.

[0357]

[30] Zhi-Ping Feng and Chun-Ting Zhang.Prediction of membrane proteintypes based on the hydrophobic index of amino acids.Journal of proteinchemistry,19(4):269–275,2000.

[0358]

[31] Zong Lin and Xian-Ming Pan.Accurate prediction of proteinsecondary structural content.Journal of Protein Chemistry,20(3):217–220,2001.

[0359]

[32] Aditya Grover and Jure Leskovec.node2vec:Scalable featurelearning for networks.In Proceedings of the 22nd ACM SIGKDD internationalconference on Knowledge discovery and data mining,pages855–864,2016.

[0360]

[33] Chao Zhou,Changshi Wang,Hongbo Liu,Qiangwei Zhou,Qian Liu,YanGuo,Ting Peng,Jiaming Song,Jianwei Zhang,Lingling Chen,et al.Identificationand analysis of adenine n6-methylation sites in the rice genome.Natureplants,4(8):554–563,2018.

[0361]

[34] Robert R Sokal and Barbara A Thomson.Population structureinferred by local spatial autocorrelation:an example from an amerindiantribal population.American Journal of Physical Anthropology:The OfficialPublication of the American Association of Physical Anthropologists,129(1):121–131,2006.

[0362]

[35] Juan Ramos et al.Using tf-idf to determine word relevance indocument queries.In Proceedings of the first instructional conference onmachine learning,volume 242,pages 29–48.Citeseer,2003.

[0363]

[36] Ari Aulia Hakim,Alva Erwin,Kho I Eng,Maulahikmah Galinium,andWahyu Muliady.Automated document classification for news article in bahasaIndonesia based on term frequency inverse document frequency(tf-idf)approach.In 2014 6th international conference on information technology andelectrical engineering(ICITEE),pages 1–4.IEEE,2014.

[0364]

[37] David S Horne.Prediction of protein helix content from anautocorrelation analysis of sequence hydrophobicities.Biopolymers:OriginalResearch on Biomolecules,27(3):451–477,1988.

[0365]

[38] Bin Liu,Fule Liu,Longyun Fang,Xiaolong Wang,and Kuo-ChenChou.repdna:a python package to generate various modes of feature vectors fordna sequences by incorporating user-defined physicochemical properties andsequence-order effects.Bioinformatics,31(8):1307–1309,2015.

[0366]

[39] Qiwen Dong,Shuigeng Zhou,and Jihong Guan.A new taxonomy-basedprotein fold recognition approach based on autocross-covariancetransformation.Bioinformatics,25(20):2655–2662,2009.

[0367]

[40] Yanzhi Guo,Lezheng Yu,Zhining Wen,and Menglong Li.Using supportvector machine combined with auto covariance to predict protein–proteininteractions from protein sequences.Nucleic acids research,36(9):3025–3030,2008.

[0368]

[41] John S Whissell and Charles LA Clarke.Improving documentclustering using okapi bm25 feature weighting.Information retrieval,14(5):466–487,2011.

[0369]

[42] Meng Kong, Yusen Zhang, Da Xu, Wei Chen, and Matthias Dehmer. Fctp-wsrc: protein–protein interactions prediction via weighted sparse representation based classification. Frontiers in genetics, 11:18, 2020.

[0370]

[43] Juwen Shen, Jian Zhang, Xiaomin Luo, Weiliang Zhu, Kunqian Yu, Kaixian Chen, Yixue Li, and Hualiang Jiang. Predicting protein–protein interactions based only on sequences information. Proceedings of the National Academy of Sciences, 104(11):4337–4341, 2007.

[0371]

[44] Yongchun Zuo, Yuan Li, Yingli Chen, Guangpeng Li, Zhenhe Yan, and Lei Yang. Psekraac: a flexible web server for generating pseudo k-tuple reduced amino acids composition. Bioinformatics, 33(1):122–124, 2017.

[0372]

[45] Wei Chen, Hong Tran, Zhiyong Liang, Hao Lin, and Liqing Zhang. Identification and analysis of the n6-methyladenosine in the saccharomyces cerevisiae transcriptome. Scientific reports, 5(1):1–8, 2015.

[0373] Refer again Figure 5After encoding the amino acid sequence in step S202, the processor 102 may perform dimensionality reduction in step S204. Dimensionality reduction may be performed on each set of the encoded sequence (e.g., each set of first numerical vectors generated according to one of the plurality of encoding algorithms) according to a plurality of dimensionality reduction algorithms. As a result of the dimensionality reduction, a further set of numerical vectors may be generated, wherein the further set of numerical vectors has a dimension that is less than the dimension of the encoded sequence. The further set of numerical vectors generated in step S204 may be referred to as a "second numerical vector".

[0374] Dimensionality reduction can be considered as investigating whether the first numerical vector contains redundant and repeated patterns. In some exemplary embodiments, the multiple dimensionality reduction algorithms used in step S204 may include at least some of the following:

[0375] - Principal Component Analysis (PCA), in which the principal components of a data set are calculated and used to perform a change of basis on the data set;

[0376] - K-means, where n observations in a dataset are divided into k clusters, where each observation belongs to the cluster with the closest mean (e.g., cluster center), and the closest means of the k clusters are used for dimensionality reduction;

[0377] -t-Distributed Stochastic Nearest Neighbor Embedding (TSNE), a statistical method for visualizing high-dimensional data by giving the location of each data point in a two-dimensional or three-dimensional map;

[0378] -Kernel principal component analysis is an extension of PCA using kernel method techniques;

[0379] - Locally Linear Embedding (LLE), which provides a low-dimensional projection of the dataset that preserves distances within a local neighborhood;

[0380] - Tensor Singular Value Decomposition (T-SVD), which computes the orthonormal subspaces associated with the different factors that appear in the tensor product of the vector space in which the tensor lies;

[0381] - Non-negative matrix factorization (NMF), which is a set of algorithms in multivariate analysis and linear algebra where a matrix V can be decomposed into two or more matrices W and H, ..., with the property that all matrices V, W, H, ... have no negative elements;

[0382] - Multidimensional Scaling (MDS), which is a means of visualizing the similarity levels of instances in a dataset;

[0383] - Factor analysis, which is a statistical method used to describe the variability between observed related variables in terms of an underlying smaller number of unobserved variables called factors;

[0384] - Aggregate features, which use hierarchical clustering to group features with similar behavior together;

[0385] - Gaussian random projection, which reduces the dimensionality by projecting the original input space onto a random matrix generated using a Gaussian distribution;

[0386] - Sparse random projection, which projects the original input space using a sparse random matrix to reduce the dimensionality;

[0387] - fast independent component analysis (fast-ICA), which is a type of independent component analysis, which is a computational method for separating multivariate signals into additive subcomponents;

[0388] -t-Distributed Stochastic Nearest Neighbor Embedding (TSNE), a statistical method for visualizing high-dimensional data by giving the location of each data point in a two-dimensional or three-dimensional map;

[0389] - Kernel principal component analysis is an extension of principal component analysis PCA using kernel method techniques, which is a class of pattern analysis algorithms.

[0390] Fig.10 A specific example of reducing the dimension of a first numerical vector to generate a second numerical vector is shown. Fig.10 In the specific example shown, 70 first numerical vectors with a dimension of 275 are processed according to first-ICA, and 70 second numerical vectors with a dimension of 4 are generated.

[0391] In dimensionality reduction, it should be noted that there is usually a strong possibility of losing important information when converting a high-dimensional space to a low-dimensional space. Therefore, it may not be simple to determine the minimum number of components that can capture the most information of the high-dimensional space.

[0392] Therefore, in some exemplary embodiments, an incremental method can be used to determine the optimal (reduced) dimensionality. In the incremental method, the high-dimensional space can be mapped to K possible low-dimensional spaces before the optimal dimensionality is determined through empirical evaluation.

[0393] For example, in the incremental method, in order to reduce the dimension of the original sequence vector, the L dimension (for example, in Fig.10 In the specific example shown, L = 275) feature space is mapped to multiple low-dimensional feature spaces, and the feature space with the lowest dimension that captures the essence of the high-dimensional space can be selected. The incremental method can start by selecting N components of the dimensionality reduction algorithm that are equal to x0% (e.g., x0 = 5) of the original feature vector (e.g., the first numerical vector), mapping the L-dimensional feature space to a low-dimensional space, and evaluating its efficacy in protein expression prediction. The following will be based on Figure 5Step S208 of describes the details of this evaluation. Subsequently, N components equal to (x0+n)% (e.g., n=5) of the original feature vector can be selected, and the efficacy of their protein expression can be evaluated. The percentage can be increased at intervals of n% until the threshold percentage, and the effectiveness of the obtained N components for the actual task is evaluated. For example, when x0=5, n=5, and the threshold percentage is 95%, the percentages of the N components to be evaluated will be 5%, 10%, 15%, 20%, 25%, 30%, ..., 95%. In this way, different target dimensions can be found for different encoding algorithms through empirical evaluation. For example, some encoding algorithms can perform better when their high-dimensional feature vectors are reduced to 5% of the dimension, some have 10% of the reduced dimension, and so on. For various encoding algorithms, the optimal reduced dimension of the original feature vector can be regarded as a hyperparameter. To calculate this hyperparameter for each of the plurality of encoding algorithms, different reduction dimensions may be iteratively evaluated, the different reduction dimensions falling within a range of x0% (e.g., 5%) to a threshold percentage (e.g., 95%), with a step size of n% (e.g., 5%). Thus, in an incremental approach, a single dimensionality reduction algorithm may be used to generate multiple sets of second numerical vectors having different reduction dimensions.

[0394] Reference again Figure 5 , the processor 102 may use the second numerical vector generated by the dimensionality reduction in step S204 to perform the machine learning regression in step S206. For example, in step S206, the processor 102 may process at least a portion of the second numerical vector and a known efficiency value (e.g., Figure 4 In some exemplary embodiments, a prediction function according to one of a plurality of regression algorithms may be obtained for each set of second numerical vectors generated using one of a plurality of encoding algorithms and one of a plurality of dimensionality reduction algorithms having one of a reduced dimension. In other words, a prediction function may be obtained for each prediction algorithm, the prediction algorithm being a combination of an encoding algorithm, a regression algorithm, and a regression algorithm. The obtained prediction function may be used to infer the efficiency value of expressing proteins having new amino acid sequences, respectively.

[0395] Algorithms that can handle complex nonlinear data distributions can be used as available multiple regression algorithms. Examples of multiple regression algorithms may include, but are not limited to, linear regression, nonlinear regression, penalized linear regression, penalized nonlinear regression, naive Bayes, bagging regression, random forest regressor, boosting regression, partial least squares regression, support vector machine, etc.

[0396] Linear regression is a linear method used to model the relationship between a scalar response (also known as the dependent variable) and one or more explanatory variables (also known as independent variables). On the other hand, in nonlinear regression, the relationship between the dependent and independent variables can be modeled as a nonlinear function.

[0397] Penalized linear regression / penalized nonlinear regression can be understood as linear / nonlinear regression that penalizes certain variables to avoid overfitting.

[0398] Naive Bayes can be understood as a conditional probability model based on applying Bayes' theorem under the assumption of strong (in other words, "naive") independence between features.

[0399] Bagging regression is a regression algorithm that involves a machine learning ensemble meta-algorithm designed to improve the stability and accuracy of the regression algorithm.

[0400] The random forest regressor can be understood as an ensemble learning method for regression tasks, which works by building multiple decision trees at training time, where the mean or average prediction of the individual trees is returned for regression.

[0401] Boosting regression can be understood as a machine learning regression algorithm that gives a prediction model in the form of a collection of weak prediction models, which are usually decision trees.

[0402] In partial least squares regression, a linear regression model is found by projecting the predicted and observed variables into a new space.

[0403] A support vector machine can be a supervised learning model that can assign new examples to one category or another, and it can be used to analyze regression data.

[0404] Fig.11 A specific example of estimating a prediction function based on a regression algorithm is shown. Fig.11 In the specific example shown, a prediction function f is estimated according to a random forest regression algorithm, which returns the titer using an input comprising a second numerical vector of scores.

[0405] Reference again Figure 5 , in step S208, the processor 102 may evaluate the prediction function estimated in step S206. For example, the processor 102 may evaluate each prediction function by comparing the prediction efficiency value output by the prediction function of at least a portion of the second numerical vector with a known efficiency value of the amino acid sequence corresponding to at least a portion of the second numerical vector. More specifically, for example, the processor 102 may determine one or more of the following performance indicators for each prediction function: accuracy, precision, recall, F1 score.

[0406] For example, in order to determine the performance indicators of accuracy, precision, recall and / or F1 score, more than one second numerical vector can be input into the prediction function, and the prediction efficiency value output by the prediction function can be divided into two categories: high efficiency and low efficiency. For example, the prediction efficiency value below the specified threshold can be classified as low efficiency, and the prediction efficiency value equal to or higher than the specified threshold can be classified as high efficiency. The specified threshold can be changed according to the background (e.g., production target) of the desired analysis. Then, the predicted efficiency value can be compared with the corresponding known efficiency value, and each efficiency value can be marked as true positive, false positive, true negative or false negative. True positive can be understood as the prediction efficiency value and the corresponding known efficiency value are equal to or higher than the specified threshold. False positive can be understood as the prediction efficiency value is equal to or higher than the specified threshold, and the corresponding known efficiency value is lower than the specified threshold. True negative can be understood as the prediction efficiency value and the corresponding known efficiency value are both lower than the specified threshold. False negative can be understood as the prediction efficiency value is lower than the specified threshold, and the corresponding known efficiency value is equal to or higher than the specified threshold.

[0407] Fig.12 A specific example of classifying the output of a prediction function to evaluate the prediction function is shown. Fig.12 In the specific example of , 70 second numerical vectors (corresponding to amino acid sequences) as input to the prediction function are classified as high producers (e.g., having a value equal to or higher than the threshold value (in Fig.12 In the specific example of 2.0), a high efficiency) and a low producer (e.g., having a low efficiency below a threshold). Fig.12 The graph on the left shows the known efficiency values ​​for 70 inputs (in Fig.12 The actual performance) is compared with the predicted efficiency value ( Fig.12 The number of inputs labeled as true positives, false positives, true negatives, and false negatives is shown in Fig.12 In the confusion matrix on the right. Fig.12 The confusion matrix in Figure 3 shows that in this specific example, there are 21 true positives, 4 false positives, 43 true negatives, and 9 false negatives.

[0408] The number of inputs labeled as true positives, false positives, true negatives, and false negatives can be used to determine the above performance metrics: accuracy, precision, recall, and F1 score.

[0409] For example, accuracy can be calculated as follows:

[0410]

[0411] exist Fig.12In the specific example shown, the accuracy would be 0.91 (=(21+43) / (21+43+4+9)).

[0412] Furthermore, the accuracy can be calculated as follows:

[0413]

[0414] exist Fig.12 In the case of the specific example shown, the accuracy would be 0.84 (=21 / (21+4)).

[0415] Furthermore, the recall can be calculated as follows:

[0416]

[0417] exist Fig.12 In the case of the specific example shown, the recall would be 0.7 (=21 / (21+9)).

[0418] Furthermore, the F1 score can be calculated as follows:

[0419]

[0420] exist Fig.12 In the case of the specific example shown, the F1 score would be 0.76 (= 0.84*0.7 / (0.84+0.7)).

[0421] Reference again Figure 5 , Figure 5 The exemplary workflow may end after the evaluation of step S208.

[0422] In the above reference Figures 5 to 11 In the exemplary embodiment description, before entering the subsequent steps, the processor 102 performs each of steps S202, S204 and S206 according to multiple encoding algorithms, multiple dimensionality reduction algorithms and multiple regression algorithms. However, in some exemplary embodiments, the processor 102 may first select a combination of a single encoding algorithm, a single dimensionality reduction algorithm and a single regression algorithm, and perform steps S202, S204 and S206 to estimate the prediction function of the algorithm combination, evaluate the prediction function in step S208, and then repeat the same steps S202 to S208 for another algorithm combination until all possible combinations are processed.

[0423] Reference again Figure 4 After evaluating the prediction function, step S20 may end. The exemplary process may then proceed to step S30.

[0424] In step S30, the processor 102 may, based on the Figure 5 Step S208) performs the evaluation of the prediction function, and selects at least one prediction algorithm from the available prediction algorithms. For example, the processor 102 may select at least one prediction algorithm, which has Figure 5 The highest value of one or more performance indicators determined in step S208. After step S30, the exemplary process may proceed to step S40.

[0425] In step S40, the processor 102 may use at least one prediction algorithm and corresponding prediction function determined in step S30 to predict one or more efficiency values ​​of expressing one or more proteins having one or more specified amino acid sequences, respectively. The efficiency values ​​predicted in step S40 may indicate the efficiency of expressing the respective proteins using a specific type of host cell and a specific type of expression vector for determining the known efficiency values ​​obtained in step S10. In some exemplary embodiments, the one or more specified amino acid sequences may be input by a user via an input device connected to the processor 102. Additionally or alternatively, the processor 102 may receive one or more specified amino acid sequences from a computer (e.g., a server computer, a client computer, etc.) outside the prediction system 10. In some exemplary embodiments, the one or more specified amino acid sequences may include one or more amino acid sequences not included in the multiple amino acid sequences obtained in step S10. Additionally or alternatively, the one or more specified amino acid sequences may include one or more of the multiple amino acid sequences obtained in step S10. The exemplary process may proceed to step S50 after step S40.

[0426] In step S50, the processor 102 may output one or more designated amino acid sequences and one or more predicted efficiency values. In some exemplary embodiments, the one or more designated amino acid sequences and one or more predicted efficiency values ​​may be displayed on a display device (not shown) connected to the processor 102. Additionally or alternatively, the processor 102 may transmit information indicating the one or more designated amino acid sequences and one or more predicted efficiency values ​​to the automated cell culture system 20. In some exemplary embodiments, a specific type of host cell and a specific type of expression vector used to determine the known efficiency value obtained in step S10 may also be output in step S50.

[0427] In some exemplary embodiments, processor 102 may further identify one of the one or more specified amino acid sequences predicted to have the highest efficiency value, and further output information indicating that the identified one of the one or more specified amino acid sequences has the highest predicted efficiency value at step S50.

[0428] Figure 4 The exemplary process may end after step S50.

[0429] The computational framework provided by any one of the above exemplary embodiments can deeply explore which antibody sequence type and which sequence region contain rich residue relationships for protein expression analysis, and which statistical representation learning scheme effectively captures the biological characteristics of residues. Further, the computational framework provided by any one of the above exemplary embodiments can investigate whether the generated statistical representation needs to be reduced to alleviate the curse of dimensionality, and which machine learning regressor performs better for the actual task. In a complete end-to-end (end-to-end) predictive modeling workflow, the computational framework provided by any one of the above exemplary embodiments can find the best algorithm combination on the prediction pipeline at different levels, for example, which sequence encoding scheme performs better with which dimensionality reduction algorithm, and which encoder-dimensionality reduction pair works better with which machine learning regressor. The robustness, transparency and effectiveness of the computational framework provided by any one of the above exemplary embodiments can be used as a good alternative to the time-consuming and expensive wet laboratory experiments for protein expression analysis (e.g., recombinant antibody sequence expression analysis).

[0430] Application of efficiency prediction in protein expression

[0431] Referenced above Figures 4 to 12 The described efficiency prediction results can be used for protein expression performed in the automated cell culture system 20.

[0432] For example, the automated cell culture system 20 may be used in Figure 4 The one or more designated amino acid sequences outputted in step S50 can be used to express one or more target proteins respectively comprising the one or more designated amino acid sequences. Specifically, for example, in the automated cell culture system 20, one or more nucleic acids encoding the one or more designated amino acid sequences can be used to express the one or more target proteins, wherein the one or more designated amino acid sequences can be used to determine the one or more target proteins. Figure 4 Step S10 obtains known efficiency values ​​for a specific type of host cell and a specific type of expression vector.

[0433] Further, according to Figure 4 The exemplary process shown, for example, by verifying the data (see also e.g. Figure 5 ) to find the best prediction pipeline based on sequence encoding, dimensionality reduction, and machine learning regressors (e.g., Figure 4 Whenever it is necessary to obtain a target expression value for a test amino acid sequence (eg, a recombinant antibody sequence), the pipeline with the best performance (eg, the selected prediction algorithm) can be saved and used.

[0434] Further, for example, Figure 4 The exemplary process shown can be used to generate the accurate amino acid sequence of the target protein (e.g., recombinant antibody), and the target protein is subjected to in-depth defect-free analysis. For example, the selected prediction algorithm can be used to predict the efficiency value of expressing the target protein containing different amino acid sequences respectively. Then one or more amino acid sequences that may show high efficiency can be identified for protein expression in wet laboratory experiments.

[0435] Hardware Configuration

[0436] Fig.13 An exemplary hardware configuration of a computer that can be used to implement at least a portion of the system described above is shown. For example, Figure 3 At least a portion of the prediction system 10 and / or automated cell culture system 20 shown may be used Fig.13 The computer 7 shown is implemented. Fig.13 The computer 7 shown includes a central processing unit (CPU) 70, a system memory 72, a network interface 74, a hard disk drive (HDD) interface 76, an external disk drive interface 78, and an input / output (I / O) interface 80. These components of the computer are coupled to each other by a system bus 82. The CPU 70 can perform arithmetic, logic, and / or control operations by accessing the system memory 72. The system memory 72 can store information and / or instructions used in conjunction with the CPU 70. The system memory 72 can include volatile and non-volatile memory, such as a random access memory (RAM) 720 and a read-only memory (ROM) 722. A basic input / output system (BIOS) containing routines that help transfer information between elements within the computer 7, such as during startup, can be stored in the ROM 722. The system bus 82 can be any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures.

[0437] The computer may include a network interface 74 for communicating with other computers and / or devices via a network.

[0438] Further, the computer may include a hard disk drive (HDD) 84 for reading from and writing to a hard disk (not shown), and an external disk drive 86 for reading from or writing to a removable disk (not shown). The removable disk may be a disk for a disk drive or an optical disk such as a CD ROM for an optical disk drive. The HDD 84 and the external disk drive 86 are connected to the system bus 82 via the HDD interface 76 and the external disk drive interface 78, respectively. The drive and its associated computer-readable medium provide non-volatile storage of computer-readable instructions, data structures, program modules, and other data for a general-purpose computer. The data structure may include relevant data for implementing the exemplary methods described herein and variations thereof. The relevant data may be organized in a database, such as a relational database or an object database.

[0439] Although the exemplary environment described herein employs a hard disk (not shown) and an external disk (not shown), those skilled in the art will appreciate that other types of computer-readable media that can store computer-accessible data, such as cassette tapes, flash memory cards, digital video disks, random access memory, read-only memory, etc., may also be used in the exemplary operating environment.

[0440] A number of program modules may be stored on the hard disk, external disk, ROM 722 or RAM 720, including an operating system (not shown), one or more application programs 7202, other program modules (not shown), and program data 7204. The application program may include at least a portion of the functionality described above.

[0441] The computer 7 may be connected to an input device 92 such as a mouse and / or a keyboard and a display device 94 such as a liquid crystal display via corresponding I / O interfaces 80a and 80b and a system bus 82. In the case where the computer 7 is implemented as a tablet computer, for example, a touch panel that displays information and receives input may be connected to the computer 7 via corresponding I / O interfaces and the system bus 82. Further, in some examples, although Fig.13 Not shown, the computer 7 may be further connected to a printer and / or an imaging device such as a camera via corresponding I / O interfaces and the system bus 82 .

[0442] In addition to using Fig.13 In addition to or as an alternative to the embodiment of the computer 7 shown, some or all of the functions of the exemplary embodiments described herein may be implemented as one or more hardware circuits. Examples of such hardware circuits may include, but are not limited to, large-scale integrated circuits (LSI), reduced instruction set circuits (RISC), application-specific integrated circuits (ASICs), and field programmable gate arrays (FPGAs).

Claims

1. Methods for optimizing protein expression, Each of the plurality of prediction algorithms comprises a combination of an encoding algorithm, a dimensionality reduction algorithm, and a regression algorithm, wherein the combination is different from the combination included in any other of the plurality of prediction algorithms, The method comprises: obtaining, by a processor (102), a plurality of amino acid sequences and corresponding known efficiency values, each known efficiency value indicating an efficiency of expressing a protein having a corresponding one of the plurality of amino acid sequences; For each of the plurality of prediction algorithms, generating, by the processor (102), a first numerical vector corresponding to one of the plurality of amino acid sequences for each of the plurality of amino acid sequences by encoding at least a portion of one of the plurality of amino acid sequences according to an encoding algorithm included in one of the plurality of prediction algorithms, By means of the processor (102), a second numerical vector corresponding to the plurality of amino acid sequences is generated by applying a dimensionality reduction algorithm included in one of the plurality of prediction algorithms to a first numerical vector corresponding to the plurality of amino acid sequences, wherein the dimension of the second numerical vector is smaller than the dimension of the first numerical vector, obtaining, by the processor (102), a prediction function by processing at least a portion of the second numerical vector and known efficiency values ​​of an amino acid sequence corresponding to at least a portion of the second numerical vector according to a regression algorithm included in one of the plurality of prediction algorithms, wherein the prediction function outputs a predicted efficiency value for expressing a protein having an amino acid sequence corresponding to the input numerical vector, and evaluating, by the processor (102), the obtained prediction function by comparing a predicted efficiency value output by the obtained prediction function for at least a portion of the second numerical vector with a known efficiency value for an amino acid sequence corresponding to at least a portion of the second numerical vector; selecting, by the processor (102), at least one prediction algorithm from the plurality of prediction algorithms based on the evaluation; Predicting, by the processor (102), one or more efficiency values ​​of expressing one or more proteins having one or more specified amino acid sequences respectively, using the at least one prediction algorithm and a prediction function obtained by the at least one prediction algorithm; as well as The processor (102) outputs the one or more designated amino acid sequences and one or more efficiency values ​​predicted for the one or more designated amino acid sequences.

2. The method according to claim 1, further comprising: Identifying, by the processor (102), one of the one or more designated amino acid sequences having the highest predicted efficiency value; as well as Information is outputted by the processor (102) indicating which of the one or more specified amino acid sequences is identified as having the highest predicted efficiency value.

3. The method according to claim 1 or 2, wherein the plurality of amino acid sequences are antibody sequences or recombinant antibody sequences. 4 . The method according to claim 3 , wherein the plurality of amino acid sequences are light chain sequences, heavy chain sequences, or light-heavy chain sequences of an antibody or a recombinant antibody.

5. The method according to any one of claims 1 to 4, wherein at least a portion of one of the plurality of amino acid sequences comprises: The first x% of one of the plurality of amino acid sequences; and / or The last y% of one of the multiple amino acid sequences where each of x and y is a value greater than 0 and less than or equal to 50, wherein x is preferably 25 or 50 and y is preferably 25 or 50.

6. The method according to any one of claims 1 to 5, wherein each known efficiency value indicates a titer of expressing a protein having a corresponding one of the plurality of amino acid sequences; and The predicted efficiency value output by the prediction function indicates the predicted titer.

7. The method according to any one of claims 1 to 6, wherein the generation of the second numerical vector comprises: generating, by the processor (102), a plurality of sets of the second numerical vectors using the dimensionality reduction algorithm, wherein the second numerical vectors included in the same set of the plurality of sets of the second numerical vectors have the same dimension, wherein the second numerical vectors included in different sets of the plurality of sets of the second numerical vectors have different dimensions, and Wherein, for each of the plurality of sets of the second numerical vectors, the obtaining of the prediction function and the evaluating of the prediction function are performed.

8. The method according to any one of claims 1 to 7, wherein the evaluation of the obtained prediction function comprises: Determine, by the processor (102), one or more of the following performance indicators of the obtained prediction function: Accuracy, Accuracy, Recall rate, F1 score.

9. The method according to any one of claims 1 to 8, wherein the combination of the encoding algorithm, the dimensionality reduction algorithm and the regression algorithm is a combination of: A k-mer-based encoding algorithm, a counting-based encoding algorithm, a K-gap-based encoding algorithm, a window-based encoding algorithm, a group-based encoding algorithm, a physicochemical property-based encoding algorithm, or a word embedding-based encoding algorithm, as the encoding algorithm; Principal component analysis PCA, K-means, t-distributed stochastic neighbor embedding TSNE, kernel principal component analysis (kernel-PCA), local linear embedding LLE, tensor singular value decomposition T-SVD, non-negative matrix factorization NMF, multidimensional scaling MDS, factor analysis, aggregate features, Gaussian random projections, sparse random projections or fast independent component analysis fast-ICA, as the dimensionality reduction algorithm; and Linear regression, nonlinear regression, penalized linear regression, penalized nonlinear regression, naive Bayes, bagging regression, random forest regressor, boosted regression, partial least squares regression or support vector machine as the regression algorithm.

10. The method according to claim 9, wherein the encoding of at least a portion of one of the plurality of amino acid sequences according to the encoding algorithm based on physicochemical properties as the encoding algorithm comprises: By means of the processor (102), a weight value is assigned to each amino acid based on one or more physicochemical properties of the protein; as well as generating, by the processor (102), the first numerical vector including weights corresponding to the amino acids included in at least a portion of one of the plurality of amino acid sequences in the order of at least a portion of one of the plurality of amino acid sequences, The one or more physicochemical properties include one or more of the following: Dissociation, Solubility, Hydration, polarity, Charge, Hydrophobicity, Molecular weight, size.

11. The method of claim 9, wherein the encoding of at least a portion of one of the plurality of amino acid sequences according to the k-mer based encoding algorithm as the encoding algorithm comprises: determining, by the processor (102), a k-mer of at least a portion of one of the plurality of amino acid sequences; The processor (102) determines a weight for each of the k-mers based at least on: the number of times one of the k-mers occurs in at least a portion of one of the plurality of amino acid sequences, and a number of times one of the k-mers occurs in at least a portion of the plurality of amino acid sequences; The first numerical vector is generated by the processor (102), and the first numerical vector includes weights of the k-mers according to the order in which the k-mers appear in at least a portion of one of the multiple amino acid sequences.

12. The method according to any one of claims 1 to 11, further comprising: In the automated cell culture system (20), one or more nucleic acids encoding the one or more designated amino acid sequences are used to express one or more proteins comprising the one or more designated amino acid sequences.

13. A computer program product comprising computer readable instructions which, when loaded and run on a computer, cause the computer to perform the method according to any one of claims 1 to 11.

14. A prediction system (10) for predicting protein expression efficiency, the prediction system comprising: A processor (102) configured to execute the method according to any one of claims 1 to 11; A storage medium (104) is in communication with the processor and is configured to store the plurality of amino acid sequences and the corresponding known efficiency values.

15. A system for optimizing protein expression, the system comprising: The prediction system (10) according to claim 14; as well as An automated cell culture system (20) capable of expressing said one or more proteins.

Images