Systems and methods for improving the stability of synthetic proteins

Abstract

A computer-implemented method for training a neural network to improve a property of a protein, comprising collecting a set of amino acid sequences from a database, compiling each amino acid sequence into a three-dimensional crystal structure of a folded protein, training a neural network with a subset of the three-dimensional crystal structures, using the neural network to identify candidate residues to be mutated in a target protein, and using the neural network to identify predicted amino acid residues to replace the candidate residues to generate mutant proteins, wherein the mutant proteins exhibit an improvement in property over the target protein. A system for improving a property of a protein is also described. An improved blue fluorescent protein generated using the system is also described.

Description

[0001] Related Applications

[0002] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 62 / 841,906, filed on May 2, 2019, entitled “SYSTEMS AND METHODS FOR IMPROVING THE STABILITY OF SYNTHETIC PROTEINS,” the contents of which are incorporated herein by reference in their entirety.

[0003] STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0004] This invention was made with government support under Grant No. R43 NS105463 awarded by the National Institutes of Health and Grant No. FA9550-14-1-0089 awarded by the Air Force Office of Scientific Research. The Government has certain rights in this invention. Background Art

[0005] Protein engineering is a transformative approach in biotechnology and biomedicine that aims to endow existing proteins with new functions or to make them more durable in non-native environments. A design consideration that influences both engineering approaches is the overall stability of the protein. In the former case, the introduction of gain-of-function mutations to expand the role of a protein, either by rational design or directed evolution, often has a thermodynamic cost. Since most natural proteins are only slightly stable, functional mutations that destabilize a protein to the point of unfolding may be missed, and increasing stability before selection has been shown to improve the evolvability of a protein.

[0006] A significant barrier to converting useful naturally occurring biocatalysts to industrial use is the adaptation of proteins to radically different environmental conditions, temperatures, and solvents. Increasing protein stability can alleviate many of these stresses, allowing for large-scale expression with higher yields and lower costs. Thus, stability is critical to the success of many protein engineering efforts.

[0007] There are a large number of methods for designing proteins, and all generally represent a compromise between how to quickly and accurately measure protein variants and how to efficiently sample the panorama of protein variants. Techniques such as mutagenic polymerase chain reaction (PCR) require minimal knowledge of the relationship between sequence and function, but rely on high-throughput screening or selection to isolate large libraries of protein variants. Structural data and computational methods can be used to narrow the search space while reducing the amount of downstream characterization. These tools are becoming increasingly important for proteins where desired properties are difficult to measure, especially at scale. However, due to our incomplete understanding of protein sequence / structure / function relationships, different computational tools for protein engineering often provide radically different or even conflicting solutions. This is especially true for properties such as stability and folding that are often the result of many small interactions distributed throughout the protein sequence.

[0008] Typically, in silico approaches will identify residues that destabilize a protein by performing computationally intensive folding simulations. The level of detail involved in these simulations varies—some even invoke quantum mechanics (MOE) to account for molecular interactions, while others use more coarse-grained approaches (Rosetta). As a first approximation, coarse-grained approaches identify problematic residues by looking for gaps in the protein structure (RosettaVIP), performing fast local free energy calculations (foldX), or looking for residues that are evolutionarily aberrant (PROS). More suitable residues are then suggested, either by hydrophobic packing or by reversion to evolutionary consensus. The effects of these substitutions on protein stability are then estimated through energy simulations of mutants. Overall, this process (residue identification, substitution suggestions, renaturation, and free energy calculations) can take anywhere from a few hours to a few days.

[0009] Machine learning is an attractive option because it does not require a priori knowledge of specific protein features or time-consuming manual inspection and assignment of individual structural features. Recently, Torng and Altman (Torng et al., “3D deep convolutional neural networks for amino acid environment similarity analysis,” BMC Bioinformatics, 18:302, 2017, incorporated herein by reference) described a general framework that applies 3D convolutional neural networks (3DCNNs) to protein structure analysis by predicting the properties of amino acids given information about the surrounding protein microenvironment. The neural network achieved 42% prediction accuracy in assigning amino acids relative to the wild-type sequence and outperformed other computer methods that rely on identifying pre-assigned structure-based features. Furthermore, given structural data for the model protein T4 lysozyme, the 3D CNN often predicted wild-type residues at positions known to be destabilizing by mutations and showed a strong preference for wild-type residues when given the structures of these known destabilizing mutants. Summary of the invention

[0010] Given that the proteome must simultaneously exhibit several unrelated and even conflicting phenotypes, such as folding geometry, stability, catalysis, and binding specificity, it seems plausible that amino acids that are structural outliers at positions far from the active site may affect folding and stability but not function. Therefore, there is a need in the art for improved protein engineering techniques that utilize artificial intelligence to learn the shared microenvironment of different amino acids and scan the entire structure to identify residues that deviate from structural consensus. These residues are considered to have a low wild-type probability and are considered to be unstable sites, and are therefore good candidates for mutagenesis and stability engineering. Implementation of the systems and methods discussed herein provides such improved protein engineering techniques.

[0011] In one aspect, a computer-implemented method for training a neural network to improve a property of a protein comprises collecting a set of amino acid sequences from a database, compiling a set of three-dimensional crystal structures with chemical environments for the set of amino acids, converting the chemical environments into voxelized matrices, training a neural network with a subset of the voxelized matrices, identifying candidate residues to be mutated in a target protein with the neural network, and identifying predicted amino acid residues to replace the candidate residues with the neural network to generate mutant proteins, wherein the mutant proteins exhibit an improvement in properties over the target protein. In one embodiment, the method further comprises the step of adding a spatial arrangement of features selected from the group consisting of hydrogen positions, partial charges, beta factors, secondary structure, aromaticity, electron density, polarity, and combinations thereof to at least one three-dimensional crystal structure.

[0012] In one embodiment, the method further comprises adjusting the set of amino acid sequences to reflect their natural frequencies. In one embodiment, the method further comprises sampling at least 50% of the amino acids in the set of amino acid sequences from random positions in the sequences. In one embodiment, the method further comprises training a second independent neural network using a second subset of the three-dimensional crystal structure or voxelized matrix, and identifying candidate and predicted residues based on the results of the two neural networks. In one embodiment, the characteristic is stability, maturity, folding, or a combination thereof.

[0013] On the other hand, a system for improving protein properties includes a processor and a non-transitory computer-readable medium having instructions stored thereon, which when executed by the processor performs the following steps, including providing a target protein comprising a residue sequence, providing a set of three-dimensional models surrounding amino acids and providing a set of protein feature values ​​for each three-dimensional model, estimating a set of parameters for each point in each three-dimensional model, training a neural network using the three-dimensional models, the parameters and the protein feature values, identifying candidate residues to be mutated in the target protein using the neural network, and identifying predicted amino acid residues that replace the candidate residues using the neural network, generating mutant proteins, wherein the mutant proteins show improvements in properties that are superior to the target protein.

[0014] In one embodiment, the protein characteristic is stability. In one embodiment, the steps include recompiling at least one amino acid sequence of the folded amino acid sequence to generate an updated three-dimensional model. In one embodiment, the steps include adding a spatial arrangement of the characteristic to at least one amino acid sequence of the folded amino acid sequence prior to recompiling.

[0015] On the other hand, the present invention relates to a protein comprising a secBFP2 variant, which has one or more mutations at one or more residues selected from T18, S28, Y96, S114, V124, T127, D151, N173 and R198 relative to full-length wild-type secBFP2. In one embodiment, the protein comprises a secBFP2 variant comprising an amino acid sequence of one of SEQ ID NO:2 to SEQ ID NO:28. In one embodiment, the secBFP2 variant comprises a variant of an amino acid sequence of one of SEQ ID NO:2 to SEQ ID NO:28. In one embodiment, the secBFP2 variant comprises a fusion protein comprising an amino acid sequence of one of SEQ ID NO:2 to SEQ ID NO:28. In one embodiment, BFP comprises a fragment of an amino acid sequence of one of SEQ ID NO:2 to SEQ ID NO:28.

[0016] In another aspect, the present invention relates to a nucleic acid molecule comprising a nucleotide sequence encoding a protein comprising a secBFP2 variant. In one embodiment, the nucleotide sequence encodes an amino acid sequence as described in SEQ ID NO:2 to SEQ ID NO:28, a variant thereof, a fusion protein thereof, or a fragment thereof. In one embodiment, the molecule is a plasmid. In one embodiment, the molecule is an expression vector. In one embodiment, the nucleic acid molecule further comprises a multiple cloning site for inserting a heterologous protein coding sequence. In another aspect, the present invention includes: a composition comprising the above-mentioned protein, a composition comprising the above-mentioned nucleic acid molecule, a kit comprising the above-mentioned protein or the above-mentioned nucleic acid molecule. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The patent or application file contains at least one drawing in color. Copies of this patent or patent application publication with color drawings will be provided by the Office upon request and payment of the necessary fee.

[0019] The above objects and features and other objects and features will become apparent with reference to the following description and accompanying drawings, which are included to provide an understanding of the present invention and constitute a part of the specification, wherein like numerals represent like elements, and wherein:

[0020] Figure 1A is a schematic diagram of a computer-implemented neural network used to increase the properties of a synthetic protein;

[0021] Figure 1B is a flow chart of the implementation of the method for determining the central amino acid residue of the microenvironment;

[0022] Figure 1C is a flow chart for implementing a method for increasing the properties of a synthetic protein during testing;

[0023] Figure 1D is a block diagram of an implementation of a neural network for increasing the properties of a synthetic protein during training;

[0024] Figure 1E is a block diagram of the implementation of a convolutional neural network for increasing the properties of synthetic proteins;

[0025] Figure 2A is a graph of experimental results for implementation of methods and systems for increasing properties of synthetic proteins;

[0026] Figure 2B is another graph of experimental results of an embodiment of the method and system for increasing the properties of a synthetic protein;

[0027] Figure 3A is another graph of experimental results of an embodiment of the method and system for increasing the properties of a synthetic protein;

[0028] Figure 3B is a photograph of a protein synthesized using a modification proposed by an embodiment of a system for increasing properties of a synthesized protein;

[0029] Figure 4A is another graph of experimental results of an embodiment of the method and system for increasing the properties of a synthetic protein;

[0030] Figure 4B is a schematic diagram of a proposed protein modification by an embodiment of a system for increasing the properties of a synthetic protein;

[0031] Figure 5 is a set of photographs of experimental results of an embodiment of a system for increasing the properties of synthetic proteins;

[0032] Figure 6 and Figure 7 is a graph of experimental results of an embodiment of a system for increasing the properties of a synthetic protein;

[0033] Figure 8 is a graph showing the fluorescence fold change of 17 blue fluorescent protein variants relative to the wild-type protein;

[0034] Fig. 9is a graph showing the fold change in fluorescence of blue fluorescent protein variants relative to the wild-type protein;

[0035] Fig.10 Provides exemplary images of fluorescence of a blue fluorescent protein variant "bluebonnet" comprising S28A, S114T, N173H, and T127L mutations compared to the parent protein and other blue fluorescent proteins; and

[0036] Fig.11A and 11B is a block diagram depicting an embodiment of a system for increasing properties of a synthetic protein.

[0037] Detailed description

[0038] It should be understood that the drawings and description of the present invention have been simplified to illustrate the elements relevant to a clear understanding of the present invention, while for clarity, many other elements found in the related systems and methods have been eliminated. It will be appreciated by those of ordinary skill in the art that other elements and / or steps are desirable and / or required when implementing the present invention. However, since such elements and steps are well known in the art, and since they are not conducive to a better understanding of the present invention, discussion of such elements and steps is not provided herein. The present invention relates to all such changes and modifications to such elements and methods known to those skilled in the art.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, only exemplary methods and materials are described herein.

[0040] As used herein, the following terms have the meanings associated with them in this section.

[0041] As used herein, the articles "a" and "an" refer to one or more (ie, to at least one) of the grammatical object of the article. For example, "an element" refers to one or more than one element.

[0042] As used herein, the term "about" when referring to a measurable value (eg, an amount or duration, etc.) is meant to encompass variations of 20%, 10%, 5%, 1%, 0.5% or even 0.1% of the stated amount if such variations are appropriate.

[0043] The term "nucleic acid molecule" or "polynucleotide" refers to a deoxyribonucleic acid or ribonucleotide polymer in single-stranded or double-stranded form, unless otherwise specifically indicated, including polynucleotides containing known natural nucleotide analogs, which function similarly to natural nucleotides. It should be understood that when a nucleic acid molecule is represented by a DNA sequence, it also includes RNA molecules having the corresponding RNA sequence, in which "U" (uridine) replaces "T" (thymidine).

[0044] The term "recombinant nucleic acid molecule" refers to a non-natural nucleic acid molecule containing two or more linked polynucleotide sequences. Recombinant nucleic acid molecules can be produced by recombinant methods, particularly genetic engineering techniques, or can be produced by chemical synthesis methods. Recombinant nucleic acid molecules can encode fusion proteins, such as fluorescent protein variants linked to a polypeptide of interest as suggested by the systems and methods discussed herein. The term "recombinant host cell" refers to a cell containing a recombinant nucleic acid molecule. Therefore, a recombinant host cell can express a polypeptide of a "gene" that is not present in the natural (non-recombinant) form of the cell.

[0045] Reference to a polynucleotide "encoding" a polypeptide means that, following transcription of the polynucleotide and translation of the mRNA produced thereby, the polypeptide is produced. An encoding polynucleotide is considered to include both a coding strand having a nucleotide sequence identical to that of the mRNA and its complementary strand. It should be appreciated that such an encoding polynucleotide is considered to include a degenerate nucleotide sequence encoding the same amino acid residue. A nucleotide sequence encoding a polypeptide may include a polynucleotide containing introns as well as encoding exons.

[0046] The term "expression control sequence" refers to a nucleotide sequence that regulates the transcription or translation of a polynucleotide or the localization of a polypeptide to which it is operatively linked. An expression control sequence is "operably linked" when it controls or regulates the transcription and (as the case may be) translation of a nucleotide sequence (i.e., a transcriptional or translational regulatory element, respectively) or the encoded polypeptide is localized to a specific compartment of the cell. Thus, an expression control sequence may be a promoter, an enhancer, a transcription terminator, an initiation codon (ATG), a splice signal for intron excision and maintenance of the correct reading frame, a stop codon, a ribosome binding site, or a sequence that targets a polypeptide to a specific location, such as a cell compartmentalization signal that targets a polypeptide to the cytoplasm, the nucleus, the plasma membrane, the endoplasmic reticulum, the mitochondrial membrane or matrix, the chloroplast membrane or lumen, the medial trans-Golgi cisternae or the lysosome or the nucleolus. Cellular compartmentalization domains include, for example, a peptide containing amino acid residues 1 to 81 of human type II membrane-anchored protein galactosyltransferase, or the leading amino acid residues 1 to 12 of subunit IV of cytochrome c oxidase (see also Hancock et al., EMBO J. 10:4033-4039, 1991; Buss et al., Mol. Cell. Biol. 8:3960-3963, 1988; US Pat. No. 5,776,689, each of which is incorporated herein by reference).

[0047] The terms "operatively linked" or "operably linked" or "operatively joined", etc., when used to describe chimeric proteins, refer to polypeptide sequences that are in a physical and functional relationship with each other. In the most preferred embodiments, the function of the polypeptide component of the chimeric molecule is unchanged compared to the functional activity of the isolated parts. For example, the fluorescent protein suggested by the systems and methods discussed herein can be fused to the polypeptide of interest. In this case, the preferred fusion molecule retains its fluorescence, and the polypeptide of interest retains its original biological activity. In some embodiments of the systems and methods discussed herein, the activity of the fluorescent protein or the protein of interest can be reduced relative to its isolated activity. Such fusions can also be used in the systems and methods discussed herein.

[0048] The term "label" refers to a composition that can be detected with or without instrumentation, such as by visual inspection, spectroscopy or photochemistry, biochemistry, immunochemistry, or chemical reaction. Useful labels include, for example, phosphor-32, fluorescent dyes, fluorescent proteins, electron-dense reagents, enzymes (e.g., commonly used in ELISA), small molecules (e.g., biotin, digoxin or other haptens or peptides, the antisera or antibodies thereof can be monoclonal antibodies). It should be recognized that the fluorescent protein variants suggested by the embodiments of the systems and methods discussed herein, which are detectable proteins themselves, can be labeled so as to be detected by means other than their own fluorescence, such as by incorporating a radionuclide label or a peptide label into the protein, so as to facilitate the identification of the protein during, for example, its expression and separation of the expressed protein, respectively. Labels for the purposes of implementing the systems and methods discussed herein typically produce a measurable signal, such as a radioactive signal, fluorescence, enzyme activity, etc., any of which can be used, for example, to quantify the amount of fluorescent protein variants in a sample.

[0049] The term "polypeptide" or "protein" refers to a polymer of two or more amino acid residues. The term applies to amino acid polymers in which one or more amino acid residues is an artificial chemical analog of the corresponding natural amino acid, as well as to natural amino acid polymers. The term "recombinant protein" refers to a protein produced by expressing a nucleotide sequence encoding the amino acid sequence of the protein from a recombinant DNA molecule.

[0050] The terms "isolated" or "purified" refer to a material that is substantially or essentially free from components that normally accompany the material in its native state. Purity or homogeneity is usually determined using analytical chemistry techniques (e.g., polyacrylamide gel electrophoresis, high performance liquid chromatography, and the like). A polynucleotide or polypeptide is considered isolated when it is the predominant species present in a preparation. An isolated protein or nucleic acid molecule generally represents more than 80% of the macromolecular species in a preparation, often more than 90% of all macromolecular species, and usually more than 95% of the macromolecular species, and, in particular, is a polypeptide or polynucleotide that is purified to substantial homogeneity so that it is the only species detected when the purity of such molecules is determined using conventional methods.

[0051] The term "naturally occurring" refers to a protein, nucleic acid molecule, cell, or other material that occurs in nature. For example, a polypeptide or polynucleotide sequence that occurs in an organism (including a virus). A naturally occurring material can exist in the form in which it occurs in nature and can be modified by man, for example, in an isolated form.

[0052] The term "antibody" refers to a polypeptide substantially encoded by one or more immunoglobulin genes or antigen-binding fragments thereof, which specifically binds to and recognizes an analyte (antigen). Recognized immunoglobulin genes include kappa, lambda, alpha, gamma, delta, epsilon and mu constant region genes, as well as numerous immunoglobulin variable region genes. Antibodies exist as complete immunoglobulins and characteristic antigen-binding fragments of antibodies, which can be produced by peptidase digestion or using recombinant DNA methods. Antigen-binding fragments of such antibodies include, for example, Fv, Fab' and F(ab)'2 fragments. The term "antibody" as used herein includes antibody fragments produced by modifying complete antibodies or antibody fragments synthesized de novo using recombinant DNA methods. The term "immunoassay" refers to an assay that utilizes antibodies to specifically bind to an analyte. The characteristic of an immunoassay is to use the specific binding properties of a particular antibody to separate, target and / or quantify an analyte.

[0053] When used for two or more polynucleotide sequences or two or more polypeptide sequences, the term "identical" refers to the residues in the sequence that are identical when aligned for maximum correspondence. When the percentage of sequence identity is used for a reference polypeptide, it is recognized that one or more residue positions that are otherwise different may differ by conservative amino acid substitutions, in which the first amino acid residue is replaced by another amino acid residue with similar chemical properties, such as similar charge or hydrophobic or hydrophilic properties, and thus does not change the functional properties of the polypeptide. When polypeptide sequences differ in conservative substitutions, the percentage of sequence identity can be adjusted upward to correct for the conservative nature of the substitution. For example, such an adjustment can be made by counting conservative substitutions as partial mismatches rather than complete mismatches, thereby increasing the percentage of sequence identity. Thus, for example, if the score for the same amino acid is 1 and the score for a non-conservative substitution is 0, the score for a conservative substitution is between 0 and 1. The score for conservative substitutions can be calculated using the following formula: for example, Meyers and Miller, Comp. Appl. Biol. Sci. 4:11-17, 1988; Smith and Waterman, Adv. Appl. Math. 2:482, 1981; Needleman and Wunsch, J. Mol. Biol. 48:443, 1970; Pearson and Lipman, Proc. Natl. Acad. Sci., USA 85:2444 (1988); Higgins and Sharp, Gene 73:237-244, 1988; Higgins and Sharp, CABIOS 5:151-153; 1989; Corpet et al., Nucl. Acids Res. 16:10881-10890, 1988; Huang, et al., Comp. Appl. Biol. Sci. 8:155-165, 1992; Pearson et al., Meth. Mol. Biol., 24:307-331, 1994, each of which is incorporated herein by reference. Alignment can also be performed by simple visual inspection and manual sequence alignment.

[0054] The term "conservatively modified variation" when used for a particular polynucleotide sequence refers to different polynucleotide sequences that encode the same or substantially the same amino acid sequence, or in the case where the polynucleotide does not encode an amino acid sequence, refers to substantially the same sequence. Due to the degeneracy of the genetic code, a large number of functionally identical polynucleotides encode any given polypeptide. For example, the codons CGU, CGC, CGA, CGG, AGA, and AGG all encode the amino acid arginine. Therefore, at each position where arginine is specified by a codon, the codon can be changed to any of the corresponding codons described without changing the encoded polypeptide. This nucleotide sequence variation is a "silent variation" and can be considered a type of "conservatively modified variation". Therefore, it should be recognized that each polynucleotide sequence encoding a fluorescent protein variant disclosed herein also describes each possible silent variation. It should also be recognized that, except for AUG, which is usually the only codon for methionine, and UUG, which is usually the only codon for tryptophan, each codon in the polynucleotide can be modified by standard techniques to produce functionally identical molecules. Therefore, each silent variation of a polynucleotide that does not change the sequence of the encoded polypeptide is implicitly described herein. In addition, it should be recognized that single substitutions, deletions or additions of a single amino acid or a small percentage of amino acids (typically less than 5%, usually less than 1%) in the coding sequence can be considered conservatively modified variants, provided that the changes result in the replacement of the amino acid with an amino acid with similar chemical properties. Conservative amino acid substitutes that provide functionally similar amino acids may include the following six groups, each group containing amino acids that are considered conservative substitutes for each other:

[0055] 1) Alanine (Ala, A), serine (Ser, S), threonine (Thr, T);

[0056] 2) Aspartic acid (Asp, D), glutamic acid (Glu, E);

[0057] 3) Asparagine (Asn, N), glutamine (Gln, Q);

[0058] 4) Arginine (Arg, R), Lysine (Lys, K);

[0059] 5) Isoleucine (Ile, I), Leucine (Leu, L), Methionine (Met, M), Valine (Val, V); and

[0060] 6) Phenylalanine (Phe, F), tyrosine (Tyr, Y), tryptophan (Trp, W).

[0061] If two or more amino acid sequences or two or more nucleotide sequences share at least 80% sequence identity with each other or with a reference sequence over a given comparison window, then two or more amino acid sequences or two or more nucleotide sequences are considered to be "substantially identical" or "substantially similar." Thus, substantially similar sequences include, for example, sequences having at least 85% sequence identity, at least 90% sequence identity, at least 95% sequence identity, or at least 99% sequence identity.

[0062] A subject nucleotide sequence is considered to be "substantially complementary" to a reference nucleotide sequence if the complement of the subject nucleotide sequence is substantially identical to the reference nucleotide sequence.

[0063] Fluorescent molecules are very useful in fluorescence resonance energy transfer (FRET), which involves one donor molecule and one acceptor molecule. To optimize the efficiency and detectability of FRET between donor and acceptor molecules, several factors need to be balanced. The emission spectrum of the donor should overlap as much as possible with the excitation spectrum of the acceptor to maximize the overlap integral. In addition, the quantum yield of the donor moiety and the extinction coefficient of the acceptor should be as high as possible to maximize the RO, which represents the distance at which the energy transfer efficiency is 50%. However, the excitation spectra of the donor and acceptor should overlap as little as possible so that a wavelength region can be found in which the donor can be efficiently excited without directly exciting the acceptor, because the fluorescence generated by direct excitation of the acceptor is difficult to distinguish from that generated by FRET. Similarly, the emission spectra of the donor and acceptor should overlap as little as possible so that the two emissions can be clearly distinguished. If the emission from the acceptor is measured as the only readout or as part of the emission ratio, a high fluorescence quantum yield of the acceptor moiety is desirable. One factor to consider when selecting a donor and acceptor pair is the efficiency of fluorescence resonance energy transfer between them. Preferably, the efficiency of FRET between the donor and the acceptor is at least 10%, more preferably at least 50%, even more preferably at least 80%.

[0064] The term "fluorescence property" refers to the molar extinction coefficient at the appropriate excitation wavelength, the fluorescence quantum efficiency, the shape of the excitation or emission spectrum, the excitation wavelength maximum and the emission wavelength maximum, the ratio of the excitation amplitudes at two different wavelengths, the ratio of the emission amplitudes at two different wavelengths, the excited state lifetime, or the fluorescence anisotropy. A measurable difference in any of these properties between the wild-type or parent fluorescent protein and a spectral variant or mutant thereof is useful. The measurable difference can be determined by determining the amount of any quantitative fluorescent property, for example, the amount of fluorescence at a particular wavelength, or the integral of fluorescence over the emission spectrum. Determining the ratio of the excitation amplitudes or emission amplitudes at two different wavelengths ("excitation amplitude ratio" and "emission amplitude ratio", respectively) is particularly advantageous because the ratiometric procedure provides an internal reference and cancels out variations in the absolute brightness of the excitation source, the sensitivity of the detector, and light scattering or quenching by the sample.

[0065] As used herein, the term "fluorescent protein" refers to any protein that is capable of fluorescing when excited with appropriate electromagnetic radiation, with the exception of chemically labeled proteins, where the fluorescence is due to the chemical label, and polypeptides that emit light only due to the presence of certain amino acids (e.g., tryptophan or tyrosine), whose emission peaks are at ultraviolet wavelengths (i.e., less than about 400 nm) are not considered fluorescent proteins for use in embodiments of the systems and methods discussed herein. In general, fluorescent proteins used to prepare compositions of embodiments of the systems described herein or for embodiments of the methods described herein are proteins that derive their fluorescence from the autocatalytic formation of a chromophore. Fluorescent proteins may comprise amino acid sequences that occur naturally or have been engineered (i.e., variants or mutants). When used with reference to a fluorescent protein, the term "mutant" or "variant" refers to a protein that is different from the reference protein.

[0066] The term "blue fluorescent protein" is used broadly herein to refer to a protein that emits blue light fluorescence. The term "blue fluorescent protein" or "BFP" is used in the broadest sense, specifically including mTagBFP, secBFP2 and blue fluorescent proteins from any species, and variants thereof, as long as they retain the ability to emit blue light.

[0067] The term "mutant" or "variant" is used herein to refer to a fluorescent protein that contains a mutation with respect to a corresponding wild-type or parent fluorescent protein. In addition, reference is made herein to a "spectral variant" or "spectral mutant" of a fluorescent protein to refer to a mutant fluorescent protein that has different fluorescence properties relative to the corresponding wild-type fluorescent protein.

[0068] In the present disclosure, various aspects of the implementation of the systems and methods discussed herein can be presented in a range format. It should be understood that the description of the range format is only for convenience and brevity and should not be interpreted as a rigid limitation on the scope of the invention. Therefore, it should be considered that the description of the range has specifically disclosed all possible sub-ranges and each numerical value within the range. For example, the description of a range such as 1 to 6 should be considered to have specifically disclosed sub-ranges, such as 1 to 3, 1 to 4, 1 to 5, 2 to 4, 2 to 6, 3 to 6, etc., and single numbers within the range, for example, 1, 2, 2.7, 3, 4, 5, 5.3, 6 and all and part increments therebetween. This applies to any range width.

[0069] In some aspects of the systems and methods discussed herein, software executing the instructions provided herein may be stored on a non-transitory computer-readable medium, wherein when executed on a processor, the software performs some or all steps of an implementation of the methods discussed herein.

[0070] Aspects of the systems and methods discussed herein relate to algorithms executed in computer software. Although certain embodiments may be described as being written in a specific programming language, or executed on a specific operating system or computing platform, it should be understood that the implementation of the systems and methods discussed herein is not limited to any specific computing language, platform, or combination thereof. The software executing the algorithms described herein may be written, compiled, or interpreted in any programming language, including but not limited to C, C++, C#, Objective-C, java, JavaScript, Python, PHP, Perl, Ruby, or Visual Basic. It should also be understood that the elements of the systems and methods discussed herein may be executed on any acceptable computing platform, including but not limited to servers, cloud instances, workstations, thin clients, mobile devices, embedded microcontrollers, televisions, or any other suitable computing devices.

[0071] Partial implementations of the systems discussed herein are described as software running on a computing device. Although the software described herein may be disclosed as operating on one particular computing device (e.g., a dedicated server or workstation), for purposes of implementation of the systems and methods discussed herein, the software may be portable in nature, and software running on a dedicated server may also run on a variety of devices, including desktop or mobile devices, laptops, tablets, smartphones, watches, wearable electronic devices or other wireless digital / mobile phones, televisions, cloud instances, embedded microcontrollers, thin client devices, or any other suitable computing device.

[0072] Similarly, some implementations of the systems discussed herein are described as communicating over various wireless or wired computer networks. For purposes of implementing the systems and methods discussed herein, the terms "network," "networked," and "networking" are understood to include wired Ethernet, fiber optic connections, wireless connections including any of the various 802.11 standards, cellular WAN infrastructure such as 3G, 4G / LTE or 5G networks, Bluetooth ® 、Bluetooth ® LowEnergy (BLE) or Zigbee ® A communication link, or any other method by which one electronic device can communicate with another electronic device. In some embodiments, elements of the networking portion of the implementation of the systems discussed herein may be implemented via a virtual private network (VPN).

[0073] Aspects of the implementation of the systems and methods discussed herein relate to machine learning algorithms, machine learning engines, or neural networks. The neural network can be trained based on various properties of the protein (e.g., the atomic environment of the amino acids in the known protein), and can output suggestions for changes to one or more amino acids in the protein based on the properties. In some embodiments, the properties may include atom type, electrostatics, beta factor, solvent accessibility, secondary structure, aromaticity, or polarity. The resulting amino acids can then be judged based on one or more quality metrics, and the weights of the properties can be optimized to maximize the quality metrics. In this way, the neural network can be trained to predict and optimize any quality metric that can be measured experimentally. Examples of quality metrics that can be trained on the neural network include wild-type amino acid accuracy, known stable / unstable positions, amino acid group accuracy, and any other suitable type of quality metric that can be measured. In some embodiments, the neural network can have multi-tasking capabilities and allow for simultaneous prediction and optimization of multiple quality metrics.

[0074] In embodiments implementing such a neural network, queries can be performed in a variety of ways. A query can request that the neural network identify amino acids within a given protein to increase a desired parameter, such as protein stability, which can be thermodynamically represented by a melting curve or chemically represented by guanidine or urea denaturation. A neural network implemented by the systems and methods discussed herein can identify one or more amino acid residues of a protein whose predicted properties (assessed by the neural network) differ from its native properties, thereby indicating that an improved protein can be produced by mutating the native amino acid residue to the predicted amino acid residue. As contemplated herein, the predicted amino acid residue can be any natural or non-natural (e.g., artificial or synthetic) amino acid.

[0075] In some embodiments, the neural network can be trained by using the values ​​of the expected parameters associated with the input amino acid sequence or residue to update the neural network. Updating the neural network in this manner can improve the ability of the neural network to propose the best amino acid residue. In some embodiments, training the neural network may include using the values ​​of the expected parameters associated with the protein mutated at the predicted amino acid residue. For example, in some embodiments, training the neural network may include predicting the value of the expected parameter of the proposed amino acid, comparing the predicted value and the corresponding value of the parameter associated with the known amino acid, and training the neural network based on the comparison. If the predicted value is the same or substantially similar to the known value, the neural network may be minimally updated or not updated at all. If the predicted value is different from the predicted value of the known amino acid, the neural network may be substantially updated to better correct for the difference. Regardless of how the neural network is retrained, the retrained neural network can be used to propose additional amino acids.

[0076] Although the technology of the present application is carried out in the context of improving protein stability, it should be understood that this is a non-limiting application of these technologies because they can be applied to other types of protein parameters or properties, such as half-life, activity, resistance to degradation, solubility, thermal stability, post-translational modification, extended pH tolerance, shortened maturation time, nucleic acid binding, protein-protein interaction, hydrophobicity or a combination thereof. Depending on the type of data used to train the neural network, the neural network can be optimized for different types of proteins, protein-protein interactions and / or protein properties. In this way, the neural network can be trained to improve the recognition of the amino acid sequence (also referred to as peptide) of the protein. The query neural network may include the initial amino acid sequence of the input protein. The neural network may have been trained using different amino acid sequences before. The query to the neural network may be a proposed amino acid sequence for a protein with higher stability than the initial amino acid sequence. The proposed amino acid sequence can be received from the neural network, which specifies a specific amino acid for each residue of the proposed amino acid sequence.

[0077] The techniques described herein involve iteratively querying a neural network whose input is a sequence having a discrete representation, receiving an output from a neural network having a continuous representation, and discretizing the output before providing it as input to the neural network, which can be applied to other machine learning applications. Such techniques may be particularly useful in applications that require a final output having a discrete representation. Such techniques can be generalized to recognize a series of discrete attributes by applying a model generated by a neural network trained with data that correlates the discrete attribute with features of a series of discrete attributes. In the context of identifying amino acids in a sequence, the discrete attributes may include different amino acids.

[0078] In some embodiments, the model may receive as input an initial sequence having discrete attributes at each position of the sequence, including but not limited to data generated by molecular simulation. Each discrete attribute in the initial sequence is one of a plurality of discrete attributes. Querying the neural network may include inputting an initial sequence of discrete attributes and generating an output sequence of discrete attributes having a feature level different from the feature level of the initial sequence. In response to querying the neural network, an output sequence and values ​​associated with different discrete attributes at each position of the output sequence may be received from the neural network. For each position of the sequence, if a discrete attribute is selected for the position and a continuous value data set is formed, the value of each discrete attribute may correspond to the prediction of the neural network about the feature level. These values ​​may span the discrete attributes of the position and may be used to identify a discrete version of the output sequence. In some embodiments, identifying a discrete version of the output sequence may include, for each position of the sequence, selecting the discrete attribute with the highest value from the values ​​of the different discrete attributes at the position. A proposed series of discrete attributes may be received as an output for identifying a discrete version.

[0079] In some embodiments, an iterative process is formed by querying the output sequence of the neural network, receiving the output sequence, and identifying discrete versions of the output sequence. Additional iterations of the iterative process may include inputting discrete versions of the output sequence from an immediately previous iteration. The iterative process may stop when the current output sequence matches the previous output sequence of the previous iteration.

[0080] In some embodiments, identifying a proposed amino acid sequence having expected values ​​of multiple quality metrics (e.g., higher than the value of another sequence), rather than the expected value of a single quality metric, includes training a neural network to identify an amino acid sequence having multiple quality metrics. Such techniques may be particularly useful in applications where it is necessary to identify proposed amino acid sequences of proteins having different features. In implementations of these techniques, the training data may include data associated with different features of each amino acid sequence used to train the neural network. The model generated by training the neural network may have one or more parameters corresponding to different combinations of features. In some embodiments, the parameter may represent a weight between a first feature and a second feature, which may be used to balance the likelihood that the proposed amino acid sequence has the first feature compared to the second feature. In some embodiments, training the neural network includes assigning scores to different features, and the scores may be used to estimate the values ​​of model parameters for predicting the proposed amino acid sequence. The training data in some such embodiments may include an amino acid sequence associated with an atomic microenvironment, which, when used to train the neural network, generates a model for predicting the proposed amino acid sequence. Training the neural network may involve assigning scores, and the scores may be used to estimate the values ​​of the parameters.

[0081] Convolutional neural networks have relatively few biological applications. Instead of being analyzed as amino acid sequences, proteins are increasingly being evaluated in their crystallized form to resolve their three-dimensional structures. One aspect of an embodiment of the method discussed herein involves training a 3D convolutional neural network that characterizes the chemical environment unique to each of the 20 amino acids. The same neural network can predict the amino acid that best fits a given environment. The neural network described herein has been trained on 1.6 million amino acid environments, spanning 19,000 phylogenetically distant protein structures. After training, the network had an in-sample accuracy of 80.0% and an out-of-sample accuracy of 72.5%, which is about 20% to 30% improvement over the state of the art, which has an out-of-sample accuracy of about 40%.

[0082] Sites with large differences between expected and observed amino acids provide targets for designing protein properties such as stability and folding maturation. The systems and methods described herein experimentally characterize three biological cases—a beta-lactamase antibiotic marker, a coral-derived blue fluorescent protein, and a phosphomannose isomerase from Candida albicans—where predictions from the neural network demonstrated enhanced protein function and stability in vivo. These results predict new biological tools at the intersection of artificial intelligence and molecular biology.

[0083] In one embodiment, implementations of the methods discussed herein utilize neural networks, such as implementations of neural networks disclosed by Torng and Altman as described above. As shown in the experimental results discussed below, implementations of the systems and methods discussed herein greatly improve upon the disclosed neural network designs. The original Torng and Altman set contained 3696 training and 194 test protein families, resulting in 32760 training and 1601 test structures.

[0084] The implementation of the systems and methods discussed in this article is based on the Torng and Altman framework to solve the protein stabilization problem. In a basic example, the crystal structure of a protein is viewed as a 3D image. In any given image, there are many observations of individual amino acids and their atomic environments. Some methods build a consistent reference frame centered on one of these amino acids. From this vantage point, the oxygen, nitrogen, sulfur, and carbon atoms are separated within a 20 x 20 x 20 angstrom box, and all atoms associated with the central amino acid are removed. This set of environments and the amino acid fits within the environments can be used as an annotated training set for a 3D convolutional neural network. With this trained neural network, experimentally introduced destabilizing mutations can be detected.

[0085] Implementations of the systems and methods discussed herein improve upon the basic model to identify novel stabilizing mutations. The improvements described herein result in predictions of sufficient quality to not only demonstrate known destabilizing mutations, but also to identify unknown destabilizing residues and propose stabilizing mutations.

[0086] In some embodiments, the systems and methods discussed herein allow for the identification of wild-type amino acids located in favorable environments on an input protein. Such an implementation may narrow the sequence space of residues with very low wild-type probabilities. The improvements in the implementation of the systems and methods discussed herein relative to the prior art can be described as several discrete improvements that, when combined, form a significantly improved model for identifying candidate protein residues to achieve an overall improved utility.

[0087] Figure 1Ais a diagram of a computer-implemented neural network implementation for increasing the properties of a synthetic protein. An engineer may wish to alter some of the properties of a protein, such as maturation kinetics, thermal stability, Km, Kcat, dependence on cations or anions for proper folding, and pH tolerance. At 101, a protein may be transformed into a microenvironment for each residue in the protein, and a three-dimensional model of the protein and its microenvironment may be generated. Some methods for generating a three-dimensional model include fragment assembly (when an unknown protein model is constructed from a library of candidate fragments extracted from known protein structures); fragment matching (when known protein fragments are matched to amino acid sequences); or comparative protein modeling based on spatial constraint satisfaction (when a known protein model ("template") is selected, residues of the amino acid sequence are mapped to residues in the template sequence ("alignment"), and various distance, angle, dihedral constraints in the sequence are derived from the alignment with the template structure, and the violation of the constraints is minimized) and other methods. When a three-dimensional model of a protein crystal structure is generated, a corresponding microenvironment associated with the structure is generated.

[0088] In some embodiments, the three-dimensional model may only illustrate or represent proteins without microenvironment. In some implementations, the three-dimensional model may be mapped to a three-dimensional array. In one example, the coordinates of the three-dimensional model are stored in the three-dimensional array. In some embodiments, a three-dimensional image can be generated from the three-dimensional model, and the three-dimensional image can be mapped to the three-dimensional array. The image data in the array can be referred to as a voxelized matrix. Just as a pixel can represent an addressable element of an image in a two-dimensional space, a voxel represents an addressable element in a three-dimensional space.

[0089] In some implementations, features of the image can be extracted through 3D convolution and max pooling layers. The 3D filter in the 3D convolution layer searches for periodic spatial patterns that best capture local biochemical features to separate the 20 amino acid microenvironments. The max pooling layer performs downsampling on the input to increase the translation invariance of the network. The convolutional neural network structure will be discussed further below.

[0090] The first convolutional layer 121 detects low-level features via filters. Convolutional neural networks use convolutions to highlight features in a data set. In the convolutional layer of a convolutional neural network, filters are applied to a three-dimensional array to generate a feature map. In the convolutional layer, the filter slides over the element-by-element dot product of the input and the filter, and the input is stored as a feature map. In some embodiments, a 3x3x3 filter can be applied to a three-dimensional image.

[0091] The feature map and image from the convolution filter are shown in 102. In some embodiments, a reference frame can be created around the central amino acid in the image, and features can be extracted around the central amino acid. The feature map created by the convolution of the image and the filter summarizes the presence of filter-specific features in the image. Increasing the number of filters applied to the image will increase the number of features that can be tracked. In 102, 100 filters are applied to create an 18x18x18 feature map. In other implementations, other numbers of filters may be used. The generated feature map can then be passed through an activation function to account for nonlinear patterns in the features.

[0092] In some embodiments, a rectifier linear function with the formula f(x) = max(0, x) may be applied as an activation function to a feature map. The rectifier linear activation function behaves linearly for positive values, making the function easy to optimize and subsequently allowing the neural network to achieve high prediction accuracy. The rectifier linear activation function also outputs zero for any negative input, meaning it is not a true linear function. Therefore, the output of a convolutional layer in a convolutional neural network is a feature map, where the values ​​in the feature map can be passed through the rectifier linear activation function.

[0093] A second convolutional layer is shown at 122. Increasing the number of convolutional layers can increase the complexity of the features that can be tracked. The convolutional layer at 122 incorporates an additional 100 filters to track the features. In some embodiments, the filters are the same as the filters in the first convolutional layer to ensure accuracy of the tracked features. In alternative embodiments, different filters can be incorporated into the second convolutional layer. In some embodiments, atoms associated with the central amino acid can be removed by the filters.

[0094] In some embodiments, a smaller dataset of size 16×16×16 is represented by 103 (although in other implementations, other sizes may be used, or a greater or lesser number of filters may be applied). The dot product in the second convolutional layer convolution reduces the size of the dataset. The dataset 103 includes feature maps that have tracked complex features from the original protein image 101.

[0095] In some embodiments, a first pooling layer of dimension 2x2x2 may be implemented at 123. A pooling layer may be implemented to downsample the data. A pooling window may be applied to a feature map. In some embodiments, the pooling layer outputs the maximum value of the data in the window, downsampling the data in the window. Maximum pooling highlights the most prominent features in the pooling window. In other embodiments, the pooling layer outputs the average value of the data in the window.

[0096] The downsampled data at 104 represents 200 separate 8 x 8 x 8 arrays. The downsampled data allows the neural network to retain relevant information. Although having a large amount of data can be advantageous because it allows the network to fine-tune the accuracy of its weights, as described below, a large amount of data may cause the neural network to take a large amount of time to process. Downsampled data can be important in a neural network to reduce the necessary computations in the network. Although a pooling layer 123 of size 2×2×2 and downsampled data of size 8×8×8 are shown, in other embodiments, other sizes of pooling windows and downsampled data may be utilized.

[0097] In some embodiments, the subsequent convolutional layer 124 uses 200 independent 2x2x2 filters, then reprocesses the downsampled data and highlights the features in the new feature map. A smaller filter 2x2x2 is implemented in the convolutional layer at 124, instead of 3x3x3, to account for the downsampled data. The depth of the convolutional filter should be the same as the depth of the data to successfully perform the dot product matrix multiplication. In other embodiments, as described above, other sizes or dimensions of the filter can be utilized.

[0098] The feature map from the convolutional layer 124 and the image is shown as 105. The feature map created by the convolution of the downsampled data and the filter summarizes the presence of filter-specific features in the image. In the embodiment shown in 105, there are 200 independent 7x 7x 7 arrays. The dot product of the convolution further reduces the size of the data.

[0099] As shown, convolutional layer 125 may use additional filters to extract more complex features, such as by using 400 separate 2x2x2 filters from lower resolution dataset 105. Increasing the number of filters applied to the image increases the number of features that can be tracked. Since the data has been downsampled from pooling layer 123 and significantly reduced in size, more filters may be applied in this convolutional layer to extract and highlight features of the image of protein 101 without requiring excessive processing or memory.

[0100] The feature map from the convolutional layer 125 is shown as 106. The feature map created by the convolution of the downsampled data and the filter summarizes the presence of filter-specific features in the image. In the embodiment shown at 106, there are 400 separate 6x6x6 arrays, although other numbers or sizes of arrays may be used in various embodiments. The dot product of the convolution further reduces the size of the data.

[0101] In some embodiments, a second pooling layer of dimension 2x2x2 (or any other appropriate dimension size) is implemented at 126 to further downsample the data. In some embodiments, the same type of pooling layer as implemented in the first pooling layer may be implemented in the second pooling layer. The type of pooling layer determines the pooling window used to downsample the data. For example, a maximum pooling layer may be implemented at 123 and 126. In other embodiments, different pooling layers may be implemented in the convolutional neural network. For example, a maximum pooling layer may be implemented at 123, while an average pooling layer may be implemented at 126. The maximum pooling layer highlights the most prominent features in the pooling window, while the average pooling layer outputs the average of the data in the window.

[0102] In the illustrated embodiment, the downsampled data at 107 represents 400 separate 3x3x3 arrays, although other numbers or sizes of arrays may be used. While having a large amount of data may be advantageous because it allows the network to fine-tune the accuracy of its weights, as described below, a large amount of data may cause the neural network to take a large amount of time to process. Downsampled data may be useful in a neural network to reduce the necessary computations in the network.

[0103] When reducing the size of the data, in some embodiments, the data can be further flattened, meaning that the data can be arranged into a one-dimensional vector. The data is flattened in order to perform the matrix multiplication flattening in the fully connected layer. The fully connected layer 127 can accordingly receive a flattened one-dimensional vector of length 10800 (e.g., a 400×3×3 array from step 107, although the vector can have a different length in other implementations). In the fully connected convolutional neural network layer, each number in the one-dimensional vector is applied to a neuron. The neuron sums the input and applies an activation function. In some embodiments, the activation function is a rectifier linear function. In alternative embodiments, the activation function can be a hyperbolic tangent function or a sigmoid function.

[0104] In the illustrated embodiment, the first fully connected layer 127 outputs a one-dimensional vector at 108 of length 10800 (although other lengths may be used, as described above). The vector output by the fully connected layer represents a vector of real numbers. In some embodiments, the real numbers may be output and classified. In other embodiments, the real numbers may be further input to subsequent fully connected layers to improve the accuracy of the convolutional neural network.

[0105] In this embodiment, the output of the first fully connected layer 108 is input to the second fully connected layer, as shown at 128. The output of the first fully connected layer 108 is already a one-dimensional vector, so it does not need to be flattened before it is input to the subsequent fully connected layer. In some embodiments, additional fully connected layers are implemented to increase the accuracy of the neural network. The number of additional fully connected layers may be limited by the processing power of the computer running the neural network. Alternatively, the increase in fully connected layers may be limited by the small increase in accuracy compared to the increase in computing time to process the additional fully connected layers.

[0106] In the illustrated embodiment, the second fully connected layer 128 outputs a one-dimensional vector of length 1000 at 109 (although other lengths may be used). The fully connected layer outputs a vector of real numbers. In some embodiments, the real numbers may be output and classified. In other embodiments, the real numbers may be further input to subsequent fully connected layers to improve the accuracy of the convolutional neural network.

[0107] At 129, in some embodiments, the output of the fully connected layer 109 is input into a softmax classifier. The softmax classifier uses a softmax function or a normalized exponential function to convert a real input into a normalized probability distribution over the predicted output class. In an alternative embodiment, the output of the convolutional neural network can be classified using a sigmoid function. If there is one class, a sigmoid function can be used. The softmax function is a multi-class sigmoid function.

[0108] At 110, the output of the softmax layer is the probability that each of the 20 identified amino acids improves the properties of the target protein (although a greater or lesser number of amino acids may be used in other embodiments.) This output may be input into an additional convolutional neural network so that the additional convolutional neural network can perform a different query given a predicted amino acid sequence, or the output 110 may be used directly as a predicted amino acid that improves the properties of the target protein.

[0109] Figure 1Bis a flow chart of an implementation of a method for determining amino acid residues at the center of a microenvironment. A neural network can be trained on known input / output pairs so that the neural network can learn how to classify the output given a specific input. Once the neural network has learned how to classify known input / output pairs, the neural network can operate on unknown inputs to predict what the classified output should be. In this embodiment, the neural network is trained to predict amino acids at the center of a microenvironment. During testing, the neural network can be provided with an amino acid sequence, analyze the microenvironment surrounding the amino acid, and predict amino acid residues that are different from the natural amino acid residues. The amino acids predicted by the neural network indicate that an improved protein can be generated by mutating the natural amino acid residues to the predicted amino acid residues.

[0110] In step 130, in some embodiments, a diverse protein sample set to be used to train the neural network may be compiled or constructed. The more diverse the sample set, the more robust the neural network's classification. For example, the neural network will attempt to classify an input / output pair during the first iteration of learning. If during the next learning iteration, the input / output pair is similar to the learned input / output pair of the first iteration, the neural network may artificially perform higher than it should simply because the data is similar, not because the neural network is robust. If a different input / output pair is then input into the network in a third iteration, the classification error may be much higher than if the first two input / output pairs were different. The similarity of the first two input / output pairs may cause the neural network to fine-tune itself to learn the similar input / output pairs of the first two iterations. This may be referred to as "overtraining" the network.

[0111] Alternatively, if the second iteration of training uses different input / output pairs than the first iteration, the neural network will be forced to be able to classify a wider range of input / output pairs. During testing, the output is unknown, so ideally the network is able to classify a wide range of input / output pairs.

[0112] Thus, in some embodiments of step 130, the training data set for the neural network is constructed from a plurality of proteins, all of which are phylogenetically divergent beyond a certain threshold. In various embodiments, the data set is constructed from proteins that are at least 20%, 30%, 40%, or 50% phylogenetically divergent. This filtering improves efficiency by removing very similar / duplicate proteins that may appear multiple times in the training set. This improvement can reduce the bias towards oversampled proteins that exists in the current state of the art.

[0113] In some embodiments, individual proteins in the training dataset are modified by adding hydrogen atoms to those Protein DataBase (PDB) structures that lack annotations. In one embodiment, the addition of hydrogen atoms is accomplished using a software converter (e.g., pdb2pqr). In another embodiment, atoms are further separated by the bonding ability of each atom and the inclusion of other atoms (e.g., phosphorus in the DNA backbone).

[0114] In some embodiments, individual proteins in the training set are modified by adding biophysical channels to the protein models taking into account additional properties of the protein including, but not limited to, partial charge, beta factor, secondary structure, aromaticity, and polarity.

[0115] In some embodiments, training data can be removed, where high-resolution and low-resolution models of the same protein can coexist in a protein database. According to some embodiments of the methods discussed herein, all genes with related structures having a resolution below a threshold can be grouped in groups with sequence similarity above a certain percentage threshold. As used herein, "resolution" refers to the resolution of the electron density map of a molecule, typically measured in angstroms (Å). Molecular models with "lower" resolution have higher quality than molecular models with "higher" resolution because the electron density map can be decomposed into lower point spacings, which means that more molecular structural features can be seen. In one example, all genes with related structures and a resolution below 2.5Å and a sequence similarity of at least 50% are grouped together, and the available structure with the lowest resolution is selected for training the model, removing molecular models with higher resolution (lower quality).

[0116] In some embodiments, amino acid sampling is normalized to the abundance of cysteine ​​in the PDB, rather than equal representation of all 20 amino acids. In one embodiment, amino acid sampling may be normalized to natural occurrence. In one embodiment, amino acid sampling may be normalized to natural occurrence within a given species. Cysteine ​​amino acids were modified in the data samples because cysteine ​​may be artificially assigned at any given position. Cysteine ​​is the rarest amino acid observed in the PDB, and therefore, it is possible that more abundant amino acids are undersampled and the diversity of protein microenvironments that may be occupied is not fully represented. Modifying the cysteine ​​amino acids in the data samples significantly improved the accuracy of the wild type. On a per amino acid basis, the accuracy ranged from 96.7% to 32.8% (see Figure 2A ).

[0117] In step 131, amino acids in the protein can be randomly sampled from the amino acid sequence. In one embodiment, up to 50% of the amino acids in the protein are sampled unless the protein is large, in which case no more than 100 amino acids are sampled from a single protein. In another embodiment, the upper limit is 200 amino acids per protein. The disclosed sampling method eliminates the bias in the data set towards the outer residues of the protein.

[0118] In step 132, a three-dimensional model of the protein crystal structure and the microenvironment associated with each amino acid comprising the structure may be created. For example, some methods of generating a three-dimensional model include fragment assembly (when an unknown protein model is constructed from a library of candidate fragments extracted from known protein structures); fragment matching (when a known protein fragment is matched to an amino acid sequence); or comparative protein modeling based on spatial constraint satisfaction (when a known protein model ("template") is selected, the residues of the amino acid sequence are mapped to the residues in the template sequence ("alignment"), and various distances, angles, and dihedral constraints in the sequence are derived from the alignment with the template structure, and constraint violations are minimized) and other methods. When a three-dimensional model of a protein crystal structure is generated, a microenvironment associated with each amino acid that constitutes the structure is also generated. One difficulty with existing protein structure databases is that different methods are used to create crystal structures as new proteins are added. Different methods for creating three-dimensional structures may add different biases or artifacts, thereby affecting the accuracy of the model. Rebuilding structures using the latest version of the same method ensures that the training structures differ in chemical composition, rather than artifacts or errors that occurred in older versions.

[0119] In step 133, the three-dimensional model generated from step 132 may be mapped into a three-dimensional array. In one example, the coordinates of the three-dimensional model are stored in a three-dimensional array. In some embodiments, a three-dimensional image may be generated from the three-dimensional model, and the three-dimensional image may be mapped into the three-dimensional array. The image data in the array may be referred to as a voxelized matrix. A pixel represents an addressable element of an image in a two-dimensional space, and a voxel represents an addressable element in a three-dimensional space.

[0120] In step 134, the image is input into a convolutional layer in a convolutional neural network. The convolutional layer detects features in the image through filters. Filters are used to detect the presence of certain features in an image. In a simplified example, a high pass filter detects the presence of a high frequency signal. The output of the high pass filter is the high frequency portion of the signal. Similarly, image filters can be designed to track certain features in an image. The more filters applied to an image, the more features can be tracked.

[0121] In step 135, the image is convolved with the filters in the convolutional layer to extract filter-specific features in the image. In the convolutional layer, the filter slides over the element-wise dot product of the input and the filter, and the input is stored as a feature map.

[0122] The decision at 136 depends on whether there are more filters. As discussed above, more filters implemented may mean more features that can be tracked in the image. Each filter is convolved with the image separately to create an independent feature map. If more filters are to be convolved with the image, steps 134 and 135 may be repeated. If all filters have been convolved with the image, processing proceeds to step 137. In some embodiments, feature maps may be concatenated together to create a feature map as deep as the number of filters applied to the image. In other embodiments, feature maps may be processed one at a time.

[0123] In step 137, an activation function is applied to the feature map in the convolutional layer of the convolutional neural network. The activation function allows the neural network to detect non-linear patterns in the extracted feature map. A rectifier linear function with the formula f(x) = max(0, x) may be applied to the feature map. The rectifier linear activation function behaves linearly when it comes to positive values, making the function easy to optimize, allowing the neural network to achieve higher accuracy. The rectifier linear activation function also outputs zero for any negative input, meaning it is not a true linear function. Therefore, the output of the convolutional layer in the convolutional neural network is a feature map where the values ​​in the feature map have been passed through the rectifier linear activation function.

[0124] The decision at 138 depends on whether there are more convolutional layers. Increasing the number of convolutional layers may increase the complexity of the trackable features. If there are additional convolutional layers, new filters may be applied to the image and the process may repeat steps 134-138. In some embodiments, the filters may be the same as the filters in the first convolutional layer to ensure accuracy of the tracked features. In alternative embodiments, different filters may be incorporated into the second convolutional layer. If there are no more convolutional layers, processing proceeds to step 139.

[0125] In step 139, the pooling layer downsamples the data. The pooling window can be applied to the feature map. In some embodiments, the pooling layer outputs the maximum value of the data in the window and downsamples the data in the window. Max pooling highlights the most prominent features in the pooling window. In other embodiments, the pooling layer outputs the average value of the data in the window.

[0126] The decision in 140 depends on whether there are more convolutional layers. Increasing the number of convolutional layers may increase the complexity of the trackable features. If there are additional convolutional layers, new filters may be applied to the image and the process may repeat steps 134-140. In some embodiments, the filters are the same as the filters in the first convolutional layer to ensure accuracy of the tracked features. In alternative embodiments, different filters may be incorporated into the second convolutional layer. Thus, repeated iterations of steps 134-136, 134-138, and 134-140 provide flexibility and increased complexity of the tracked features. If there are no more convolutional layers, processing proceeds to step 141.

[0127] In step 141, in some embodiments, the downsampled data is flattened. This means that the data is arranged into a one-dimensional vector. The data is flattened in order to perform matrix multiplication in the fully connected layer.

[0128] In step 142, in some embodiments, the flattened one-dimensional vector is input into a fully connected layer of a neural network. In a fully connected convolutional neural network layer, each number in the one-dimensional vector is applied to a neuron as input. The neuron sums the inputs and applies an activation function. In some embodiments, the activation function is a rectifier linear function. In alternative embodiments, the activation function can be a hyperbolic tangent function or a sigmoid function.

[0129] In some embodiments, the outputs of the first group of neurons in a fully connected layer may be input to another group of neurons via weights. Each subsequent group of neurons may be referred to as a "hidden layer" of neurons. The number of hidden layers in a fully connected network may be reduced. In other words, the number of hidden layers in a neural network may change adaptively as the neural network learns how to classify the outputs.

[0130] In step 143, in some embodiments, the neurons that make up the fully connected network are connected to other neurons through weights. The weights are adjusted to strengthen the role of some neurons and weaken the role of other neurons. Adjusting the strength of each neuron can enable the neural network to better classify the output. As the neural network learns how to classify inputs or "trains", the weights connecting the neurons are adjusted. In some embodiments, the number of neurons in the neural network can be pruned. In other words, as the neural network learns how to classify outputs, the number of active neurons in the neural network changes adaptively.

[0131] The decision in 144 depends on whether there are additional fully connected layers. In some embodiments, the output of one fully connected layer can become the input of a second fully connected layer. In some embodiments, additional fully connected layers are implemented to increase the accuracy of the neural network. The number of additional fully connected layers may be limited by the processing power of the computer running the neural network. Alternatively, the addition of fully connected layers may be limited by the small increase in accuracy compared to the increase in computing time to process the additional fully connected layers. In alternative embodiments, the output of one fully connected layer is sufficient to classify the image. If there are additional fully connected layers, steps 142 and 143 are repeated so that inputs are provided to neurons connected to each other by weights. If no additional fully connected layers are added, processing proceeds to step 145.

[0132] In step 145, in some implementations, the fully connected layer outputs a real vector. In some embodiments, the real numbers can be output and classified. In an alternative embodiment, the output of the fully connected layer is input into a softmax classifier. The softmax classifier uses a softmax function or a normalized exponential function to convert the real input into a normalized probability distribution on the predicted output class. In other embodiments, the output of the convolutional neural network can be classified using a sigmoid function. If it is a class, a sigmoid function can be used. The softmax function is a multi-class sigmoid function. In some embodiments, the output of the neural network represents the predicted amino acid residues at the center of the chemical microenvironment.

[0133] For example, the neural network can output a vector of length 20, including 20 real numbers. Since there are 20 possible amino acids in the center of the microenvironment, the length of the vector is 20. The real numbers in the vector are passed through the softmax classifier, so the values ​​in the vector represent the likelihood of the amino acid being present in the center of the microenvironment.

[0134] In step 146, in some embodiments, the predicted amino acid residue is compared to the natural amino acid at the center of the chemical environment. For example, the real amino acid vector can be a vector of length 20, where a single "1" represents the natural amino acid at the center of the chemical environment, and the other values ​​in the vector remain "0".

[0135] This type of learning in neural networks, where known input / output pairs are compared during training, is called supervised learning. The difference between the predicted and known values ​​can be determined and the information propagated back through the neural network. The weights can then be modified by the error signal. This method of training a neural network is called backpropagation.

[0136] In step 147, in some implementations, the weights are updated by steepest descent. Equation 1 below illustrates how the weights are adjusted at each iteration n.

[0137] (1)

[0138] In equation 1 above, represents the weight connecting neuron i and neuron j.

[0139] Steepest descent is an optimization technique that minimizes an objective function. In other words, steepest descent adjusts unknown parameters in the direction of steepest descent. During training, the values ​​of the weights that optimize the classification accuracy of a neural network are unknown. Therefore, the weights are the unknown parameters that are adjusted in the direction of steepest descent.

[0140] In some embodiments, the objective function may be a cross entropy error function. Minimizing the cross entropy error function represents minimizing the difference between the probability distribution of the predicted amino acid vector and the probability distribution of the natural amino acid vector. In other embodiments, the objective function may be a square error function. Minimizing the square error objective function represents minimizing the instantaneous error of each neuron.

[0141] In each training iteration, the weights are adjusted to get closer to their optimal values. Depending on the position of the neuron in the network, different formulas are used to determine how the weights are adjusted based on the objective function. Equation 2 below illustrates how the weights between neuron i and neuron j are adjusted based on the cross entropy error function.

[0142] (2)

[0143] If the weight is too small, meaning that the output of the neuron may not have a significant impact on the classification, a positive change in the weight occurs because when the small weight is compared to the optimal weight, the weight has a negative slope and a negative sign appears in the equation. If the weight is too large, a negative change in the weight occurs because when the larger weight is compared to the optimal weight, the weight has a positive slope and a negative sign appears in the equation. Therefore, the weight trains itself to approach the optimal value. As shown in step 147, the modification of the weight can be temporarily stored.

[0144] In some embodiments, the weights may be adjusted each time a modification of the weights is determined. This type of training may be referred to as online training or incremental training. One advantage of incremental training is that the neural network is able to track small changes in the input. In some embodiments, the weights may be modified after the neural network receives a batch of input / output pairs. This type of training may be referred to as batch training. One advantage of batch training includes that the neural network converges to the optimized weight values ​​more quickly. In this embodiment, the neural network was trained on 1.6 million amino acid and microenvironment pairs. In this embodiment, the batch size used was 20. In step 148, a counter is incremented. When the counter reaches 20, the neural network completes a round of batch training. In other words, when the neural network performs self-evaluation based on 20 input / output pairs, a round of training is completed.

[0145] The decision in 149 depends on whether the training samples for the current batch have been completed. If the required number of training samples for a batch has been reached, the network continues to step 150. As mentioned above, a batch of training requires 20 input / output pairs. If the required number of samples for a batch has not been reached, the neural network repeats steps 134-149.

[0146] In step 150, the weight modifications temporarily stored in step 147 are summed. The weight values ​​are modified according to the summed modifications so that a new batch of 20 input / output pairs are evaluated using the newly modified weight values.

[0147] The decision in 151 depends on whether the maximum number of training iterations has been reached. When a round of batch training is completed, a training iteration is completed. In some cases, the weights may never reach their optimal values ​​because the weights will oscillate around their optimal values. Therefore, in some embodiments, a maximum number of iterations can be set to prevent the neural network from training the network indefinitely.

[0148] If the maximum number of iterations has not been reached, the neural network may be allowed to train the network again using another input / output pair from the data sample created in step 130. After the neural network completes a batch of training, the iteration counter is incremented in step 153.

[0149] If the maximum number of iterations has been reached, the neural network may store the weight values. Step 152 illustrates storing the values ​​of the weights. These weights will be stored in memory because they are the weights that have been trained by the network and will be used later when testing the neural network.

[0150] If the number of iterations is not reached, the error between the predicted amino acid residue and the known natural amino acid residue can be evaluated. This evaluation is performed at step 154. In some cases, the error between the predicted value and the known natural value is very small, so it can be considered that the error is acceptable and the neural network does not need to continue training. In these cases, the weight values ​​that produce such a small error rate can be stored and then used for testing. In some embodiments, the neural network must maintain a small error rate over several iterations to ensure that the neural network does not learn well how to predict an output or accidentally predict an output well. Requiring the network to maintain a small error over several iterations increases the possibility that the network correctly classifies different input ranges. If the error between the predicted value and the known value is still too large, the neural network can continue to train itself and repeat steps 131-154. In many implementations, during the repeated iterations of steps 131-154, the neural network will use a new data set to train the neural network.

[0151] Figure 1C is a flow chart of an implementation of a method for increasing synthetic protein properties during testing. In step 160, the weights stored from the training scene are set as the weights in the fully connected layer in step 172. When it is necessary to classify unknown inputs, these weights are used because the weights have been trained with a wide variety of input sets, so the weights are likely to be able to accurately classify unknown inputs.

[0152] In step 161, in some implementations, the unknown protein is randomly sampled. In one embodiment, up to 50% of the amino acids in the protein are sampled unless the protein is large, in which case no more than 100 amino acids are sampled from a single protein. In another embodiment, the upper limit is 200 amino acids per protein. The disclosed sampling method eliminates the bias in the data set towards the outer residues of the protein.

[0153] In step 162, a three-dimensional model of the protein crystal structure and the microenvironment associated with each amino acid comprising the structure may be created. Some methods for generating a three-dimensional model include fragment assembly (when an unknown protein model is constructed from a library of candidate fragments extracted from known protein structures); fragment matching (when known protein fragments are matched to amino acid sequences); or comparative protein modeling based on spatial constraint satisfaction (when a known protein model ("template") is selected, the residues of the amino acid sequence are mapped to the residues in the template sequence ("alignment"), the angles of various distances in the sequence, and the dihedral constraints are derived from the alignment with the template structure, and the violation of the constraints is minimized) and other methods. When a three-dimensional model of the protein crystal structure is generated, a microenvironment associated with each amino acid comprising the structure is also generated. One difficulty with existing protein structure databases is that different methods are used to create crystal structures as new proteins are added. Different methods for creating three-dimensional structures may add different biases or artifacts that affect the accuracy of the model. Rebuilding the structure using the latest version of the same method ensures that the training structure is different in chemical composition, rather than artifacts or errors that appeared in older versions.

[0154] In step 163, the three-dimensional model generated from step 162 can be mapped into a three-dimensional array. In one example, the coordinates of the three-dimensional model are stored in the three-dimensional array. In some embodiments, a three-dimensional image can be generated from the three-dimensional model, and the three-dimensional image can be mapped into the three-dimensional array. The image data in the array can be referred to as a voxelized matrix. A pixel represents an addressable element of an image in two-dimensional space, and a voxel represents an addressable element in three-dimensional space.

[0155] In step 164, the image may be input into a convolutional layer in a convolutional neural network. The convolutional layer detects features in the image through filters. Filters are used to detect the presence of certain features in an image. In a simplified example, a high pass filter detects the presence of a high frequency signal. The output of the high pass filter is the high frequency portion of the signal. Similarly, image filters may be designed to track certain features in an image. The more filters applied to an image, the more features that can be tracked.

[0156] In step 165, the image is convolved with the filters in the convolutional layer to extract filter-specific features in the image. In the convolutional layer, the filter slides over the element-wise dot product of the input and the filter, and the input is stored as a feature map.

[0157] The decision at 166 depends on whether there are more filters. As discussed above, more filters implemented means more features can be tracked in the image. Each filter is convolved with the image separately to create an independent feature map. If more filters are to be convolved with the image, steps 164 and 165 can be repeated. If all filters have been convolved with the image, the process proceeds to step 167. In some embodiments, feature maps can be concatenated together to create a feature map as deep as the number of filters applied to the image. In other embodiments, one feature map can be processed at a time.

[0158] In step 167, in some implementations, an activation function is applied to the feature map in the convolution layer of the convolutional neural network. The activation function allows the neural network to detect non-linear patterns in the extracted feature map. A rectifier linear function with the formula f(x) = max(0, x) may be applied to the feature map as an activation function. The rectifier linear activation function behaves linearly when the values ​​are positive, making the function easy to optimize and subsequently allowing the neural network to achieve high prediction accuracy. The rectifier linear activation function also outputs zero for any negative input, meaning that it is not a true linear function. Therefore, the output of the convolution layer in the convolutional neural network is a feature map, where the values ​​in the feature map can be passed through the rectifier linear activation function.

[0159] The decision at 168 depends on whether there are more convolutional layers. Increasing the number of convolutional layers may increase the complexity of the trackable features. If there are additional convolutional layers, new filters are applied to the image and steps 164-168 may be repeated. In some embodiments, the filters are the same as those in the first convolutional layer to ensure accuracy of the tracked features. In alternative embodiments, different filters may be incorporated into the second convolutional layer. If there are no more convolutional layers, processing proceeds to step 169.

[0160] In step 169, the pooling layer downsamples the data. A pooling window can be applied to a feature map. In some embodiments, the pooling layer outputs the maximum value of the data in the window and downsamples the data in the window. Max pooling highlights the most prominent features in the pooling window. In other embodiments, the pooling layer outputs the average value of the data in the window.

[0161] The decision at 170 depends on whether there are more convolutional layers. Increasing the number of convolutional layers may increase the complexity of the trackable features. If there are additional convolutional layers, new filters are applied to the image and steps 164-170 may be repeated. In some embodiments, the filters are the same as those in the first convolutional layer to ensure accuracy of the tracked features. In alternative embodiments, different filters may be incorporated into the second convolutional layer. If there are no more convolutional layers, processing proceeds to step 171.

[0162] In step 171, in some implementations, the downsampled data is flattened. This means that the data is arranged into a one-dimensional vector. The data is flattened in order to perform matrix multiplication in the fully connected layer.

[0163] In step 172, in some implementations, the flattened one-dimensional vector is input into a fully connected layer of a neural network. In a fully connected convolutional neural network layer, each number in the one-dimensional vector is applied to a neuron. The neuron sums the inputs and applies an activation function. In some embodiments, the activation function is a rectifier linear function. In alternative embodiments, the activation function can be a hyperbolic tangent function or a sigmoid function.

[0164] In step 173, in some embodiments, the neurons that make up the fully connected network are multiplied by weights. The weights in the fully connected network are the weights initialized in step 160. These weights are used when evaluating unknown inputs because the weights have been trained with a wide variety of input sets, so the weights are likely to be able to accurately classify unknown inputs.

[0165] The decision in 174 depends on whether there are additional fully connected layers. In some embodiments, the output of one fully connected layer can become the input of a second fully connected layer. In some embodiments, additional fully connected layers are implemented to increase the accuracy of the neural network. The number of additional fully connected layers may be limited by the processing power of the computer running the neural network. Alternatively, the addition of fully connected layers may be limited by a small increase in accuracy compared to the increase in computational time to process the additional fully connected layers. In alternative embodiments, the output of one fully connected layer is sufficient to classify the image. If there are additional fully connected layers, steps 172 and 173 are repeated so that the input vectors are provided to neurons connected to each other by weights. If there are no additional fully connected layers, the process proceeds to step 175.

[0166] In step 175, the fully connected layer outputs a real vector. In some embodiments, the real numbers can be output and classified. In an alternative embodiment, the output of the fully connected layer is input into a softmax classifier. The softmax classifier uses a softmax function or a normalized exponential function to convert the real input into a normalized probability distribution over the predicted output class. In other embodiments, the output of the convolutional neural network can be classified using a sigmoid function. If there is one class, a sigmoid function can be used. The softmax function is a multi-class sigmoid function. In some embodiments, the output of the neural network represents predicted candidate residues and amino acid residues to improve the quality metric of the protein.

[0167] In step 176, a synthetic protein may be generated based on the output of the neural network. The synthetic protein may be generated by a computing device executing the neural network, another computing device in communication with the computing device executing the neural network, a third party manufacturer or laboratory, or another entity that makes substitutions based on candidate amino acid residues identified by the neural network and predicted amino acid residues. For example, in some embodiments, a synthetic protein may be obtained by an entity making one or more substitutions based on predicted amino acid residues and candidate residues identified by the neural network and / or at the direction of the neural network or the computing device executing the neural network. In some embodiments, the neural network may predict an amino acid residue that is the same as a natural amino acid residue. In other embodiments, the neural network may predict an amino acid residue that is different from a natural amino acid residue. The amino acid predicted by the neural network indicates that an improved protein may be generated by mutating the natural amino acid residue to the predicted amino acid residue. Thus, a synthetic protein may be generated based on the output of the neural network.

[0168] Figure 1D is a block diagram of a neural network during training according to some implementations. At 180, input is provided to the neural network. As described above, the neural network can accept various inputs. In some embodiments, the neural network accepts an amino acid sequence or residue. In other embodiments, the neural network may receive a series of amino acids with discrete attributes at each position in the series.

[0169] In the block diagram, 181 represents a neural network that changes over time. As described above, during training, the neural network adaptively updates the new input / output of each iteration. The neural network adaptively updates because the weights are updated based on an error signal calculated based on the difference between the predicted output and the known output.

[0170] In the block diagram, 182 represents the output that the neural network predicts will satisfy the query. For example, the neural network can be queried and trained to identify specific amino acid residues that can be modified. In these cases, the output of the neural network can be amino acid residues, where the amino acid residues can be used to synthesize new proteins with improved properties. In other embodiments, the output of the neural network can be amino acid residues that can be used as substitutes, where the substitutes can be used to synthesize new proteins with improved properties. In other embodiments, the neural network can be queried to obtain proposed amino acid sequences for proteins with different parameters from the initial amino acid sequence. In these cases, the output of the neural network can be an amino acid sequence indicating the specific amino acids for each residue of the amino acid sequence.

[0171] In the block diagram, 186 represents the expected value. This type of training is called supervised training because to train the neural network, the input corresponding to the output must be known. During training, the neural network is required to output results as close to the expected value as possible.

[0172] The expected value 186 is compared to the output value from the neural network 182 at 185. The difference between the output value and the expected value is determined and turned into an error signal 183 which is passed back through the neural network so that the neural network can learn from the error. The weights are updated based on the error signal as shown in equations 1 and 2 above.

[0173] Figure 1E is a block diagram of a convolutional neural network according to some implementations. In the block diagram, 190 represents a convolutional layer. The convolutional layer can detect features in an image through filters. Filters are used to detect the presence of certain features in an image. In a simplified example, a high-pass filter detects the presence of a high-frequency signal. The output of the high-pass filter is the high-frequency portion of the signal. Similarly, image filters can be designed to track certain features in an image. The more filters applied to an image, the more features can be tracked.

[0174] In some implementations, an image is convolved with filters in a convolutional layer to extract filter-specific features in the image. In the convolutional layer, the filter slides over the element-wise dot product of the input and the filter, and the input is stored as a feature map. An activation function is applied to the feature map of the convolutional layer of the convolutional neural network. The activation function allows the neural network to detect nonlinear patterns in the extracted feature map. A rectified linear function with the formula f(x) = max(0, x) may be applied to the feature map. The rectifier linear activation function behaves linearly for positive values, making the function easy to optimize, thereby allowing the neural network to achieve high prediction accuracy. The rectifier linear activation function also outputs zero for any negative input, which means that it is not a true linear function. Therefore, the output of the convolutional layer in the convolutional neural network is a feature map, where the values ​​in the feature map have been passed through the rectifier linear activation function. In other embodiments, a sigmoid function or a hyperbolic tangent function may be applied as an activation function.

[0175] As shown in 191, the extracted feature map that has been acted upon by the activation function can then be input to a pooling layer. The pooling layer downsamples the data. A pooling window can be applied to the feature map. In some embodiments, the pooling layer outputs the maximum value of the data in the window, downsampling the data in the window. Max pooling highlights the most prominent features in the pooling window.

[0176] In some embodiments, the downsampled pooled data may then be flattened before being input into the fully connected layer 192 of the convolutional neural network.

[0177] In some embodiments, a fully connected layer may have only one set of neurons. In an alternative embodiment, a fully connected layer may have one set of neurons in the first layer 193 and one set of neurons in a subsequent hidden layer 194. The number of hidden layers in a fully connected network may be pruned. In other words, the number of hidden layers in a neural network may change adaptively as the neural network learns how to classify the output.

[0178] In the fully connected layers, the neurons in each layer of layers 193 and 194 are connected to each other. Neurons are connected by weights. During the training process, the weights are adjusted to strengthen the role of some neurons and weaken the role of other neurons. Adjusting the strength of each neuron can make the neural network better classify the output. In some embodiments, the number of neurons in the neural network can be pruned. In other words, as the neural network learns how to classify the output, the number of active neurons in the neural network changes adaptively.

[0179] After training, the error between the predicted value and the known value may be so small that the error may be deemed acceptable and the neural network need not continue to be trained. In these cases, the weight values ​​that produced such a small error rate may be stored and used for testing later. In some embodiments, the neural network must meet a small error rate for several iterations to ensure that the neural network has not learned how to predict an output very well or accidentally predicted an output very well. Requiring the network to maintain a small error for several iterations increases the likelihood that the network will correctly classify different input ranges.

[0180] In the block diagram, 195 represents the output of the neural network. The output of the fully connected layer is a real vector. In some embodiments, the real numbers can be output and classified. In an alternative embodiment, the output of the fully connected layer is input into a softmax classifier.

[0181] In the block diagram, 196 represents a softmax classifier layer. The softmax classifier uses a softmax function or a normalized exponential function to convert a real input into a normalized probability distribution on a predicted output class. In other embodiments, the output of the convolutional neural network can be classified using a sigmoid function. If it is a class, a sigmoid function can be used. The softmax function is a multi-class sigmoid function. In some embodiments, the output of the neural network represents predicted candidate residues and amino acid residues to improve the quality metric of the protein. In other embodiments, the output of the neural network can be an amino acid sequence indicating the specific amino acids of each residue of the amino acid sequence.

[0182] In some embodiments, problem residues are identified and suggestions for new residues are made by combining predictions from multiple independently trained neural networks. By identifying residues based on independently trained neural networks, biases that arise from idiosyncrasies unique to any single neural network that occur when the neural networks are trained are eliminated. The average of many independent neural networks eliminates anomalies associated with any single neural network.

[0183] Various improvements to existing algorithms have gradually improved accuracy. Figure 2B As shown, in one embodiment, the various improvements together improve the model accuracy of wild-type amino acid predictions from about 40% to over 70% for all amino acids.

[0184] Engineered Protein

[0185] Embodiments of the systems and methods discussed herein further provide or identify compositions comprising engineered proteins comprising one or more mutations that modify the desired characteristics or properties of the protein compared to the characteristics or properties of the native or parent protein. In one embodiment, the modified protein generated or identified by the embodiments of the systems and methods discussed herein includes one or more mutations at one or more amino acid residues predicted by the 3D convolutional neural network (3DCNN) prediction pathway implemented by the systems and methods discussed herein to confer the desired characteristics or properties on the protein. The engineered protein generated or identified by the implementation of the systems and methods discussed herein, which is generated to include mutations at residues predicted by 3DCNN prediction pathway analysis, is referred to herein as a 3DCNN engineered protein.

[0186] Exemplary features or properties that may be modified in 3DCNN engineered proteins generated or identified by implementation of the systems and methods discussed herein include, but are not limited to, stability, affinity, activity, half-life, fluorescence properties, and sensitivity to photobleaching.

[0187] 3DCNN engineered proteins produced or identified by the implementation of the systems and methods discussed herein can be manufactured using chemical methods. For example, 3DCNN engineered proteins can be synthesized by solid phase techniques (Roberge JY et al (1995) Science 269: 202-204), excised from resin, and purified by preparative high performance liquid chromatography. For example, automatic synthesis can be achieved using an ABI 431 A peptide synthesizer (Perkin Elmer) according to the instructions provided by the manufacturer.

[0188] 3DCNN engineered proteins can also be prepared by translating coding nucleic acid sequences, by recombinant means, or by cutting from longer protein sequences. The composition of 3DCNN engineered proteins can be confirmed by amino acid analysis or sequencing.

[0189] Variants of 3DCNN engineered proteins generated or identified by the implementation of the systems and methods discussed herein may be (i) variants in which one or more amino acid residues are replaced by conservative or non-conservative amino acid residues (preferably conservative amino acid residues) and the replaced amino acid residues may or may not be residues encoded by the genetic code; (ii) residues in which there are one or more modified amino acid residues, for example, residues modified by connecting a substituent; (iii) fragments of 3DCNN engineered proteins and / or (iv) proteins in which 3DCNN engineered proteins are fused to another protein or polypeptide. These fragments include polypeptides produced by proteolytic cleavage (including multi-site proteolysis) of the original 3DCNN engineered protein sequence. Variants may be post-translationally modified or chemically modified. Such variants are considered to be within the scope of those skilled in the art based on the teachings herein.

[0190] As is known in the art, "similarity" between two polypeptides is determined by comparing the amino acid sequence of one polypeptide and its conservative amino acid substitutes to the sequence of a second polypeptide. Variants are defined as comprising polypeptide sequences that differ from the original sequence by less than 40% of the residues per fragment of interest, by less than 25% of the residues per fragment of interest, by less than 10% of the residues per fragment of interest, or by only a few residues per fragment of interest from the original protein sequence, while having sufficient homology to the original sequence to retain the functionality of the original sequence and / or the ability to bind to ubiquitin or ubiquitinated proteins. Embodiments of the systems and methods discussed herein can be used to generate or identify amino acid sequences that are at least 60%, 65%, 70%, 72%, 74%, 76%, 78%, 80%, 90%, 91%, 92%, 93%, 94%, 95%, 96%, 97%, 98%, or 99% similar or identical to the original amino acid sequence. The identity between two amino acid sequences is preferably determined using the BLASTP algorithm [BLAST Manual, Altschul, S., et al., NCBI NLM NIHBethesda, Md. 20894, Altschul, S., et al., J. Mol. Biol. 215: 403-410 (1990)].

[0191] The 3DCNN engineered proteins produced or identified by the implementation of the systems and methods discussed herein can be post-translationally modified. For example, post-translational modifications within the scope of the implementation of the systems and methods discussed herein include signal peptide cleavage, glycosylation, acetylation, isoprenylation, proteolysis, myristoylation, protein folding, and proteolytic processing, among others. Some modification or processing events require the introduction of additional biological mechanisms. For example, processing events such as signal peptide cleavage and core glycosylation are examined by adding canine microsomal membranes or African clawed frog egg extracts (US Pat. No. 6,103,489) to standard translation reactions.

[0192] 3DCNN engineered proteins produced or identified by the implementation of the systems and methods discussed herein may include unnatural amino acids formed by post-translational modifications or by the introduction of unnatural amino acids during translation. There are a variety of methods for introducing unnatural amino acids during protein translation. For example, special tRNAs, such as tRNAs with inhibitory properties, are used in the process of site-directed unnatural amino acid replacement (SNAAR). In SNAAR, a unique codon is required on the mRNA and the inhibitory tRNA to target the unnatural amino acid to a unique site during protein synthesis (as described in WO90 / 05785). However, the aminoacyl-tRNA synthetases present in the protein translation system cannot recognize the inhibitory tRNA. In some cases, unnatural amino acids can be formed after the tRNA molecule is aminoacylated by a chemical reaction that specifically modifies natural amino acids and does not significantly change the functional activity of the aminoacylated tRNA. These reactions are called post-aminoacylation modifications. For example, with a homologous tRNA (tRNA LYS ) can be modified with an amine-specific photoaffinity tag.

[0193] The 3DCNN engineered proteins produced or identified by the implementation of the systems and methods discussed herein can be combined with other molecules (e.g., proteins) to prepare fusion proteins. This can be achieved, for example, by synthesizing N-terminal or C-terminal fusion proteins, provided that the resulting fusion protein retains the functionality of the 3DCNN engineered protein.

[0194] 3DCNN engineered protein mimics

[0195] In some embodiments, the target component is a peptidomimetic of a 3DCNN engineered protein. Peptidomimetics are compounds based on or derived from peptides and proteins. Peptidomimetics produced or identified by the implementation of the systems and methods discussed herein can generally be obtained by structurally modifying known 3DCNN engineered protein sequences using unnatural amino acids, conformational restrictions, isotopic substitutions, etc. The peptidomimetics constitute a continuous structural space between peptides and non-peptide synthetic structures; therefore, peptidomimetics can be used to define pharmacophores and help convert peptides into non-peptide compounds with the activity of parent 3DCNN engineered proteins.

[0196] In addition, it is apparent from the present invention that mimic epitopes of the subject 3DCNN engineered proteins can be provided. Such peptidomimetics may have the following properties: non-hydrolyzability (e.g., increased stability to proteases or other physiological conditions that degrade the corresponding peptide), increased specificity and / or potency, and increased cell permeability for intracellular localization of the peptidomimetics. For illustrative purposes, peptide analogs generated or identified by embodiments of the systems and methods discussed herein can be generated using, for example, benzodiazepines (e.g., see Freidinger et al. in Peptides: Chemistry and Biology, GR Marshall ed., ESCOM Publisher: Leiden, Netherlands, 1988), substituted gammalactam rings (Garvey et al. in Peptides: Chemistry and Biology, GR Marshall ed., ESCOM Publisher: Leiden, Netherlands, 1988, p123), C-7 mimetics (Huffman et al. in Peptides: Chemistry and Biology, GR Marshall ed., ESCOM Publisher: Leiden, Netherlands, 1988, p.105), keto-methylene pseudopeptides (Ewenson et al. (1986) J Med Chem 29:295; and Ewenson et al.in Peptides: Structure and Function (Proceedings of the 9th American PeptideSymposium) Pierce Chemical Co. Rockland, Ill., 1985), β-turndipeptide cores (Nagai et al. (1985) Tetrahedron Lett 26:647; and Satoet al. (1986) J Chem Soc Perkin Trans 1:1231), β-aminoalcohols (Gordon etal. (1985) Biochem Biophys Res Commun 126:419; and Dann et al.(1986) Biochem Biophys Res Commun 134:71), diaminoketones (Natarajan et al. (1984) Biochem Biophys Res Commun 124:141), and methyleneamino-modified (Roark et al. in Peptides: Chemistry and Biology, GR Marshall ed., ESCOM Publisher: Leiden, Netherlands, 1988, p134). See also Session III: Analytic and synthetic methods, in Peptides: Chemistry and Biology, GR Marshall ed., ESCOM Publisher: Leiden, Netherlands, 1988).

[0197] In addition to the various side chain substitutions that can be made to generate 3DCNN engineered protein peptidomimetics, implementation of the systems and methods discussed herein specifically contemplates the use of conformationally constrained mimics of peptide secondary structures. Many substitutes have been developed for the amide bond of peptides. Frequently used amide bond substitutes include the following groups (i) trans-olefins, (ii) fluoroolefins, (iii) methyleneamino, (iv) phosphoamides, and (v) sulfonamides.

[0198] Nucleic Acids

[0199] In one embodiment, embodiments of the systems and methods discussed herein can be used to generate or identify an isolated nucleic acid comprising a nucleotide sequence encoding a 3DCNN engineered protein.

[0200] The nucleotide sequence encoding the 3DCNN engineered protein may alternatively include sequence changes relative to the original nucleotide sequence, for example, substitution, insertion and / or deletion of one or more nucleotides, provided that the resulting polynucleotide encodes a polypeptide according to the embodiments of the systems and methods discussed herein. Therefore, the embodiments of the systems and methods discussed herein can be used to generate or identify nucleotide sequences that are substantially identical to the nucleotide sequences described herein and encode 3DCNN engineered proteins.

[0201] In the sense used in this specification, a nucleotide sequence is "substantially identical" to any nucleotide sequence described herein when it has at least 60%, at least 70%, at least 85%, at least 95%, at least 96%, at least 97%, at least 98% or at least 99% identity with a nucleotide sequence described herein. Nucleotide sequences substantially homologous to nucleotide sequences encoding 3DCNN engineered proteins can generally be isolated from a producer organism of a polypeptide produced or identified by the implementation of the systems and methods described herein based on the information contained in the nucleotide sequence (e.g., by introducing conservative or non-conservative substitutions). Other examples of possible modifications include inserting one or more nucleotides into the sequence, adding one or more nucleotides at any end of the sequence, or deleting one or more nucleotides at any end or within the sequence. The identity between two nucleotide sequences is preferably determined using the BLASTN algorithm [BLAST Manual, Altschul, S., et al., NCBI NLM NIH Bethesda, Md. 20894, Altschul, S., et al., J. Mol. Biol. 215: 403-410 (1990)].

[0202] On the other hand, embodiments of the systems and methods discussed herein can be used to generate or identify constructs comprising nucleotide sequences encoding 3DCNN engineered proteins or derivatives thereof. In specific embodiments, the construct is operatively bound to transcriptional control elements and optionally operatively bound to translational control elements. The construct can be incorporated into an operatively bound regulatory sequence for expression of a nucleotide sequence generated or identified by the implementation of the systems and methods discussed herein, thereby forming an expression cassette.

[0203] 3DCNN engineered proteins or chimeric 3DCNN engineered proteins can be prepared using recombinant DNA methods. Therefore, nucleic acid molecules encoding 3DCNN engineered proteins or chimeric 3DCNN engineered proteins can be incorporated into appropriate expression vectors that ensure good expression of 3DCNN engineered proteins or chimeric 3DCNN engineered proteins.

[0204] Therefore, on the other hand, the implementation of the system and method discussed herein can be used to generate or identify a vector comprising a nucleotide sequence or construct generated or identified by the implementation of the system and method discussed herein. The selection of the vector will depend on the host cell into which it is subsequently introduced. In a specific embodiment, the vector generated or identified by the implementation of the system and method discussed herein is an expression vector. Suitable host cells include a variety of prokaryotic and eukaryotic host cells. In a specific embodiment, the expression vector is selected from the group consisting of viral vectors, bacterial vectors, and mammalian cell vectors. Systems based on prokaryotes and / or eukaryotic vectors can be used for the implementation of the system and method discussed herein to produce polynucleotides or their homologous polypeptides. Many such systems are widely available commercially.

[0205] In addition, the expression vector can be provided to the cell in the form of a viral vector. Viruses that can be used as vectors include, but are not limited to, retroviruses, adenoviruses, adeno-associated viruses, herpes viruses, and lentiviruses. Typically, a suitable vector comprises a replication origin that works in at least one organism, a promoter sequence, a convenient restriction endonuclease site, and one or more selectable markers. (See, e.g., WO 01 / 96584; WO 01 / 29058; and US Pat. No. 6,326,193).

[0206] Vectors suitable for inserting the polynucleotide are vectors derived from expression vectors in prokaryotes, such as pUC18, pUC19, Bluescript and its derivatives, mp18, mp19, pBR322, pMB9, ColE1, pCR1, RP4, bacteriophage and "shuttle" vectors, such as pSA3 and pAT28, expression vectors in yeast, such as vectors of the 2 micron plasmid, integrating plasmids, YEP vectors, centromeric plasmids and the like, expression vectors in insect cells, such as vectors of the pAC series and pVL, expression vectors in plants, such as pIBI, pEarleyGate, pAVA, pCAMBIA, pGSA, pGWB, pMDC, pMY, the pORE series and the like, as well as vectors based on viruses (adenoviruses, viruses related to adenoviruses, such as retroviruses, in particular lentiviruses) and non-viral vectors (such as pSilencer 4.1-CMV (Ambion), pcDNA3, pcDNA3.1 / hyg, pHMCV / Zeo, pCR3.1, pEFI / His, pIND / GS, pRc / HCMV2, pSV40 / Zeo2, pTRACER HCMV, pUB6 / V5 His, pVAX1, pZeoSV2, pCI, pSVL and PKSV-10, pBPV-1, pML2d and pTDT1.

[0207] As an illustration, the vector into which the nucleic acid sequence is introduced can be a plasmid, which may or may not be integrated into the host cell genome when the plasmid is introduced into the host cell. Illustrative, non-limiting examples of vectors into which a nucleotide sequence or gene construct generated or identified by embodiments of the systems and methods discussed herein can be inserted include tet-on inducible vectors for expression in eukaryotic cells.

[0208] In certain embodiments, the vector is a vector for transforming animal cells.

[0209] The recombinant expression vector may also include a nucleic acid molecule encoding a portion that provides the following properties: increase the expression of 3DCNN engineered protein or chimeric 3DCNN engineered protein; increase the solubility of 3DCNN engineered protein or chimeric 3DCNN engineered protein; and / or assist in purifying 3DCNN engineered protein or chimeric 3DCNN engineered protein by acting as a ligand in affinity purification. For example, a proteolytic cleavage site may be inserted into the 3DCNN engineered protein to allow separation of the 3DCNN engineered protein or chimeric 3DCNN engineered protein from the fusion portion after purification of the fusion protein. Examples of fusion expression vectors include pGEX (Amrad Corp., Melbourne, Australia), pMAL (New England Biolabs, Beverly, Mass.), and pRIT5 (Pharmacia, Piscataway, NJ), which fuse glutathione S-transferase (GST), maltose E binding protein, or protein A to recombinant proteins, respectively.

[0210] Additional promoter elements, called enhancers, regulate the frequency of transcription initiation. Typically, these are located in the region 30–110 bp upstream of the start site, although recently it has been demonstrated that many promoters also contain functional elements downstream of the start site. The spacing between promoter elements is often flexible, so that promoter function is preserved when elements are flipped or moved relative to each other. In the thymidine kinase (tk) promoter, the spacing between promoter elements can be increased to 50 bp before activity begins to decline. Depending on the promoter, it appears that individual elements can activate transcription either cooperatively or independently.

[0211] The promoter can be a promoter naturally associated with a gene or polynucleotide sequence, such as a promoter that can be obtained by separating 5 non-coding sequences located upstream of the coding segment and / or exon. Such a promoter can be referred to as "endogenous". Similarly, an enhancer can be an enhancer naturally associated with a polynucleotide sequence located downstream or upstream of the sequence. Alternatively, certain advantages may be obtained by positioning the coding polynucleotide fragment under the control of a recombinant or heterologous promoter, which refers to a promoter that is usually not associated with a polynucleotide sequence in its natural environment. A recombinant or heterologous enhancer also refers to an enhancer that is usually not associated with a polynucleotide sequence in its natural environment. Such promoters or enhancers may include promoters or enhancers of other genes, as well as promoters or enhancers isolated from any other prokaryotic cells, viral cells or eukaryotic cells, and promoters or enhancers that are not "naturally produced", i.e., different elements comprising different transcriptional regulatory regions, and / or mutations that change expression. In addition to synthetically producing nucleic acid sequences for promoters and enhancers, sequences linked to the components disclosed herein can also be generated using recombinant cloning and / or nucleic acid amplification techniques (including PCR™) (US Pat. No. 4,683,202, US Pat. No. 5,928,906). In addition, it is contemplated that control sequences can also be used that direct the transcription and / or expression of sequences in non-nuclear organelles (e.g., mitochondria, chloroplasts, etc.).

[0212] It is naturally important to use promoters and / or enhancers that effectively direct the expression of the DNA fragment in the selected cell type, organelle and organism. The promoter used can be constitutive, tissue-specific, inducible and / or used to direct high-level expression of the introduced DNA fragment under appropriate conditions, for example to facilitate large-scale production of recombinant proteins and / or peptides. The promoter can be heterologous or endogenous.

[0213] The promoter sequence illustrated in the experimental examples described herein is an immediate early cytomegalovirus (CMV) promoter sequence. This promoter sequence is a powerful constitutive promoter sequence that can drive high-level expression of any polynucleotide sequence operably linked to it. However, other constitutive promoter sequences may also be used, including but not limited to simian virus 40 (SV40) early promoter, mouse mammary tumor virus (MMTV), human immunodeficiency virus (HIV) long terminal repeat (LTR) promoter, Moloney virus promoter, avian leukosis virus promoter, Epstein-Barr virus immediate early promoter, ROS sarcoma virus promoter, and human gene promoters, such as but not limited to actin promoter, myosin promoter, hemoglobin promoter, and muscle creatine promoter. In addition, the implementation of the systems and methods discussed herein is not limited to the use of constitutive promoters. Inducible promoters may also be generated or identified by the implementation of the systems and methods discussed herein. The use of inducible promoters generated or identified by such systems or methods provides a molecular switch that can turn on the expression of its operably linked polynucleotide sequence when expression is desired, or turn off expression when expression is not desired. Examples of inducible promoters include, but are not limited to, metallothionein promoters, glucocorticoid promoters, progesterone promoters, and tetracycline promoters. In addition, implementation of the systems and methods discussed herein may allow the use of tissue-specific promoters that are effective only in the desired tissue. Tissue-specific promoters include, but are not limited to, HER-2 promoters and PSA-related promoter sequences.

[0214] In one embodiment, expression of the nucleic acid is controlled externally. For example, in one embodiment, expression is externally controlled using the doxycycline Tet-On system or other inducible or repressible expression system.

[0215] The recombinant expression vector may also include a selectable marker gene that helps select the host cell for transformation or transfection. Suitable selectable marker genes are genes encoding proteins such as G418 and hygromycin that confer resistance to certain drugs, beta-galactosidase, chloramphenicol acetyltransferase, firefly luciferase, or immunoglobulin or its portion (e.g., the Fc portion of an immunoglobulin (preferably IgG)). The selectable marker may be introduced into a carrier separated from the nucleic acid of interest.

[0216] Reporter genes are used to identify potentially transfected cells and to assess the functionality of regulatory sequences. In general, a reporter gene is a gene that is not present in or expressed by the recipient organism or tissue and that encodes a protein whose expression manifests itself in some easily detectable property, such as enzymatic activity. Expression of the reporter gene is analyzed at an appropriate time after the DNA is introduced into the recipient cells.

[0217] Exemplary reporter genes may include genes encoding luciferase, β-galactosidase, chloramphenicol acetyltransferase, secreted alkaline phosphatase, or fluorescent proteins (including but not limited to the green fluorescent protein gene) (see, e.g., Ui-Tei et al., 2000 FEBS Lett. 479:79-82).

[0218] In one embodiment, the 3DCNN engineered protein produced or identified by the implementation of the systems and methods discussed herein is a reporter gene and is included in a suitable expression system. For example, in one embodiment, the 3DCNN engineered protein generated or identified by such a system or method is a blue fluorescent protein with increased fluorescence activity. In such an embodiment, the nucleotide sequence encoding the 3DCNN engineered protein produced or identified by the embodiments of the systems and methods discussed herein can be incorporated into an expression system to allow detection of heterologous protein sequences.

[0219] Recombinant expression vectors can be introduced into host cells to produce recombinant cells. These cells can be prokaryotic cells or eukaryotic cells. The carrier generated or identified by the embodiments of the systems and methods discussed herein can be used to transform eukaryotic cells, such as yeast cells, saccharomyces cerevisiae or mammalian cells, such as epithelial kidney 293 cells or U2OS cells, or prokaryotic cells, such as bacteria, Escherichia coli or Bacillus subtilis, for example, conventional techniques can be used to introduce nucleic acids into cells, such as calcium phosphate or calcium chloride coprecipitation, DEAE-dextran-mediated transfection, liposomes, electroporation or microinjection. Suitable methods for transformation and transfection of host cells can be found in Sambrook et al. (Molecular Cloning: A Laboratory Manual, 2nd Edition, Cold Spring Harbor Laboratory press (1989)) and other laboratory textbooks.

[0220] For example, 3DCNN engineered proteins or chimeric 3DCNN engineered proteins produced or identified by the implementation of the systems and methods discussed herein can be expressed in bacterial cells, such as E. coli, insect cells (using baculovirus), yeast cells, or mammalian cells. Other suitable host cells can be found in Goeddel, Gene Expression Technology: Methods in Enzymology 185, Academic Press, San Diego, Calif. (1991).

[0221] Modified blue fluorescent protein

[0222] In one embodiment, the realization of the system and method discussed herein can be used for identifying or generating secBFP2 variant protein.In some aspects, this composition relates to the secBFP2 variant protein comprising one or more mutations that enhance stability.In some aspects, compared with wild-type secBFP2, the secBFP2 variant protein shows one or more of enhanced stability, enhanced fluorescence, enhanced half-life and slower photobleaching.

[0223] In some embodiments, the secBFP2 variant protein comprises a secBFP2 having one or more mutations. For example, in some embodiments, the secBFP2 variant protein comprises a secBFP2 having one or more mutations at one or more residues selected from T18, S28, Y96, S114, V124, T127, D151, N173 and R198 relative to the full-length wild-type secBFP2. In one embodiment, the full-length wild-type secBFP2 comprises the following amino acid sequence:

[0224] SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIY NVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN* (SEQ ID NO: 1)

[0225] In certain embodiments, mutations within secBFP2 variant proteins described herein are labeled relative to SEQ ID NO: 1. For example, a secBFP2 variant protein comprising a mutation at T18 refers to secBFP2 but with a mutation at a residue relative to the threonine at position 18 of full-length wild-type secBFP2 (SEQ ID NO: 1).

[0226] In some embodiments, secBFP2 variant proteins include secBFP2 comprising one or more mutations selected from: T18X, S28X, Y96X, S114X, V124X, T127X, D151X, N173X and R198X relative to full-length wild-type secBFP2, wherein X is any amino acid. In some embodiments, secBFP2 variant proteins include secBFP2 comprising one or more mutations selected from: T18W, T18V, T18E, S28A, Y96F, S114V, S114T, V124T, V124Y, V124W, T127P, T127L, T127R, T127D, D151G, N173T, N173H, N173R, N173S, R198V and R198L relative to full-length wild-type secBFP2.

[0227] In one embodiment, the secBFP2 variant protein includes secBFP2 with a T18X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein includes secBFP2 with a T18W mutation, a T18V mutation, or a T18E mutation relative to full-length wild-type secBFP2.

[0228] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEG W VDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVD FTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 2), or a variant or fragment thereof.

[0229] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEG VVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVD FTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 3), or a variant or fragment thereof.

[0230] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEG E VDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVD FTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 4), or a variant or fragment thereof.

[0231] In one embodiment, the secBFP2 variant protein comprises secBFP2 having an S28X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising an S28A mutation relative to full-length wild-type secBFP2.

[0232] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT AEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 5), or a variant or fragment thereof.

[0233] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a Y96X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising a Y96F mutation relative to full-length wild-type secBFP2.

[0234] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTT F EDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 6), or a variant or fragment thereof.

[0235] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a S114X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising a S114V mutation or a S114T mutation relative to full-length wild-type secBFP2.

[0236] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG VLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 7), or a variant or fragment thereof.

[0237] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG T LIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 8), or a variant or fragment thereof.

[0238] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a V124X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising a V124T mutation, a V124Y mutation, or a V124W mutation relative to full-length wild-type secBFP2.

[0239] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRG T DFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 9), or a variant or fragment thereof.

[0240] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRG Y DFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 10), or a variant or fragment thereof.

[0241] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRG W DFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 11), or a variant or fragment thereof.

[0242] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a T127X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising a T127P mutation, a T127L mutation, a T127R mutation or a T127D mutation relative to full-length wild-type secBFP2.

[0243] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDF PSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 12), or a variant or fragment thereof.

[0244] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDF L SNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 13), or a variant or fragment thereof.

[0245] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDF R SNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 14), or a variant or fragment thereof.

[0246] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDF DSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 15), or a variant or fragment thereof.

[0247] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a D151X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 having a D151G mutation relative to full-length wild-type secBFP2.

[0248] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPA G GGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 16), or a variant or fragment thereof.

[0249] In one embodiment, the secBFP2 variant protein comprises secBFP2 having a N173X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 comprising a N173T mutation, a N173H mutation, a N173R mutation or a N173S mutation relative to full-length wild-type secBFP2.

[0250] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA TAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 17), or a variant or fragment thereof.

[0251] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA H AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 18), or a variant or fragment thereof.

[0252] For example, in one embodiment, the secBFP2 variant protein comprises:

[0253] SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA R AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 19), or a variant or fragment thereof.

[0254] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA SAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 20), or a variant or fragment thereof.

[0255] In one embodiment, the secBFP2 variant protein comprises secBFP2 having an R198X mutation, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein comprises secBFP2 having an R198V mutation or an R198L mutation relative to full-length wild-type secBFP2.

[0256] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDY V LERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 21), or a variant or fragment thereof.

[0257] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCTSEGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDGSLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDY L LERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 22), or a variant or fragment thereof.

[0258] In one embodiment, the secBFP2 variant protein includes a secBFP2 having one or more, two or more, three or more, four or more, five or more, six or more, seven or more, eight or more, or all nine mutations of T18X, S28X, Y96X, S114X, V124X, T127X, D151X, N173X, and R198X relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein includes secBFP2 having one or more, two or more, three or more, four or more, five or more, six or more, seven or more, eight or more, or nine or more of T18W, T18V, T18E, S28A, Y96F, S114V, S114T, V124T, V124Y, V124W, T127P, T127L, T127R, T127D, D151G, N173T, N173H, N173R, N173S, R198V, and R198L relative to full-length wild-type secBFP2.

[0259] In one embodiment, secBFP2 variant proteins include secBFP2 with T18X, S28X, S114X, V124X, T127X, D151X, N173X and R198X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, secBFP2 variant proteins include secBFP2 with T18W, S28A, S114V, V124T, T127P, D151G, N173T and R198L mutations relative to full-length wild-type secBFP2.

[0260] For example, in one embodiment, the secBFP2 variant protein comprises:

[0261] SEELIKENMHMKLYMEG W VDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG V LIYNVKIRG T DF P SNGPVMQKKTLGWEAFTETLYPA G GGLEGRNDMALKLVGGSHLIA T AKTTYRSKKPAKNLKMPGVYYVDY LLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 23), or a variant or fragment thereof.

[0262] In one embodiment, secBFP2 variant proteins include secBFP2 with S28X, S114X, T127X and N173X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, secBFP2 variant proteins include secBFP2 with S28A, S114T, T127L and N173H mutations relative to full-length wild-type secBFP2.

[0263] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG T LIYNVKIRGVDF L SNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA H AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 24), or a variant or fragment thereof.

[0264] In one embodiment, the secBFP2 variant protein includes secBFP2 with S28X and S114X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, the secBFP2 variant protein includes secBFP2 with S28A and S114T mutations relative to full-length wild-type secBFP2.

[0265] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG TLIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIANAKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 25), or a variant or fragment thereof.

[0266] In one embodiment, secBFP2 variant proteins include secBFP2 with S28X, S114X and N173X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, secBFP2 variant proteins include secBFP2 with S28A, S114T and N173H mutations relative to full-length wild-type secBFP2.

[0267] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTTYEDGGVLTATQDTSLQDG T LIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA H AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 26), or a variant or fragment thereof.

[0268] In one embodiment, secBFP2 variant proteins include secBFP2 with S28X, Y96X, S114X and N173X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, secBFP2 variant proteins include secBFP2 with S28A, Y96F, S114T and N173H mutations relative to full-length wild-type secBFP2.

[0269] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTT FEDGGVLTATQDTSLQDG T LIYNVKIRGVDFTSNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA H AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 27), or a variant or fragment thereof.

[0270] In one embodiment, secBFP2 variant proteins include secBFP2 with S28X, Y96X, S114X, T127X and N173X mutations relative to full-length wild-type secBFP2, wherein X is any amino acid. In one embodiment, secBFP2 variant proteins include secBFP2 with S28A, Y96F, S114T, T127L and N173H mutations relative to full-length wild-type secBFP2.

[0271] For example, in one embodiment, the secBFP2 variant protein comprises: SEELIKENMHMKLYMEGTVDNHHFKCT A EGEGKPYEGTQTMRIKVVEGGPLPFAFDILATSFLYGSKTFIDHTQGIPDFFKQSFPEGFTWERVTT F EDGGVLTATQDTSLQDG T LIYNVKIRGVDF L SNGPVMQKKTLGWEAFTETLYPADGGLEGRNDMALKLVGGSHLIA H AKTTYRSKKPAKNLKMPGVYYVDYRLERIKEANDETYVEQHEVAVARYSDLPSKLGHKLN (SEQ ID NO: 28), or a variant or fragment thereof.

[0272] In one embodiment, the components generated or identified by embodiments of the systems and methods discussed herein include an isolated nucleic acid molecule comprising a nucleotide sequence encoding a secBFP2 variant protein. In various embodiments, the nucleic acid molecule comprises a sequence encoding at least one amino acid sequence described in SEQ ID NO: 2 to SEQ ID NO: 28, or a variant or fragment thereof.

[0273] The present invention also provides a fusion protein comprising a fluorescent protein variant operably linked to one or more polypeptides of interest. The polypeptides of the fusion protein can be linked by a peptide bond, or the fluorescent protein variant can be linked to the polypeptide of interest by a linker molecule. In one embodiment, the fusion protein is expressed by a recombinant nucleic acid molecule comprising a polynucleotide encoding a fluorescent protein variant, which is operably linked to one or more polynucleotides encoding one or more polypeptides of interest.

[0274] The polypeptide of interest can be any polypeptide, including, for example, a peptide tag (e.g., a polyhistidine peptide) or a cellular polypeptide (e.g., an enzyme, a G protein, a growth factor receptor, or a transcription factor); it can be one of two or more proteins that can be combined to form a complex. In one embodiment, the fusion protein is a tandem fluorescent protein variant construct, which includes a donor fluorescent protein variant, an acceptor fluorescent protein variant, and a peptide linker portion that couples the donor and the acceptor, wherein the cyclized amino acids of the donor emit a light characteristic of the donor, wherein when the donor is excited, the donor and the acceptor exhibit fluorescence resonance energy transfer, and the linker portion does not substantially emit light to excite the donor. Therefore, the fusion protein generated or identified by the embodiments of the systems and methods discussed herein may include two or more operably linked fluorescent protein variants, which may be directly or indirectly linked, and may further include one or more polypeptides of interest.

[0275] Reagent test kit

[0276] In some embodiments, kits may be provided to facilitate and / or standardize the use of components provided or identified by embodiments of the systems and methods discussed herein, as well as to facilitate the performance of the methods discussed herein. Materials and reagents for performing these various methods may be provided in a kit to facilitate the performance of the methods. As used herein, the term "kit" refers to a combination of items that facilitate a process, test, analysis, or procedure.

[0277] The kit may include chemical reagents (e.g., polypeptides or polynucleotides) and other components. In addition, the kits discussed herein may also include, for example, but not limited to, equipment and reagents for sample collection and / or purification, equipment and reagents for product collection and / or purification, reagents for bacterial cell transformation, reagents for eukaryotic cell transfection, previously transformed or transfected host cells, sample tubes, racks, trays, racks, trays, plates, kit user instructions, solutions, buffers or other chemical reagents, suitable samples for standardization, normalization and / or control samples. The kit may be packaged for storage and safe transportation, for example, in a box with a lid.

[0278] In some embodiments, for example, the kits discussed herein can provide fluorescent proteins produced or identified by embodiments of the systems and methods discussed herein, polynucleotide vectors (e.g., plasmids) encoding fluorescent proteins produced or identified by embodiments of the systems and methods discussed herein, bacterial cell strains suitable for propagating the vectors, and reagents for purifying the expressed fusion proteins. In some embodiments, the kits as described herein can provide the reagents necessary to perform mutations of the coral polyp fluorescent protein to produce protein variants with reduced oligomerization propensity.

[0279] The kit may include one or more components generated or identified by embodiments of the systems and methods discussed herein, for example, one or more fluorescent protein variants, which may be part of a fusion protein, or one or more polynucleotides encoding a polypeptide. The fluorescent protein variant may be a mutant fluorescent protein with a reduced tendency to oligomerize, such as a non-oligomer, or may be a tandem dimer fluorescent protein, and in the case where the kit includes multiple fluorescent protein variants, the multiple may be multiple mutant fluorescent protein variants, or tandem dimer fluorescent proteins, or a combination thereof.

[0280] Kits as described herein may also include one or more recombinant nucleic acid molecules that partially encode fluorescent protein variants (which may be the same or different) and, for example, may further include an operably linked second polynucleotide that contains or encodes a restriction endonuclease recognition site or a recombinase recognition site or any polypeptide of interest. In addition, the kit may include instructions for using the kit components (particularly the components contained in the kit that are produced or recognized by the implementation of the systems and methods discussed herein).

[0281] When these kits provide a variety of different fluorescent protein variants, they can be particularly useful because those skilled in the art can easily select one or more proteins with the desired fluorescent properties for a particular application. Similarly, kits containing multiple polynucleotides encoding different fluorescent protein variants provide many advantages. For example, the polynucleotides can be designed to contain convenient restriction endonuclease or recombinase recognition sites to facilitate the operative connection of the polynucleotides to regulatory elements or polynucleotides encoding the polypeptide, or, if desired, to operatively connect two or more polynucleotides encoding fluorescent protein variants to each other.

[0282] Uses of fluorescent protein variants

[0283] Fluorescent protein variants generated or identified by the implementation of the systems and methods discussed herein can be used in any method using fluorescent proteins. Therefore, fluorescent protein variants, including monomeric, dimer, and tandem dimer fluorescent proteins, can be used as fluorescent markers in many ways that fluorescent markers have been used, including, for example, coupling fluorescent protein variants to antibodies, polynucleotides or other receptors for detection analysis, such as immunoassays or hybridization assays, or for tracking the movement of proteins in cells. For intracellular tracking studies, a first (or other) polynucleotide encoding a fluorescent protein variant is fused to a second (or other) polynucleotide encoding a protein of interest, and the construct can be inserted into an expression vector if desired. After intracellular expression, the protein of interest can be located based on fluorescence without worrying about artifacts caused by oligomerization of the fluorescent protein component of the fusion protein. In one embodiment of the method, two proteins of interest are independently fused to two fluorescent protein variants with different fluorescent properties.

[0284] Fluorescent protein variants produced or identified by implementing the systems and methods discussed herein are useful in systems for detecting transcriptional induction. For example, a nucleotide sequence encoding a non-oligomeric monomer, dimer, or tandem dimer fluorescent protein can be fused to a promoter or other expression control sequence of interest that can be included in an expression vector, the construct can be transfected into a cell, and the induction of the promoter (or other regulatory element) can be measured by detecting the presence or amount of fluorescence, thereby allowing observation of the responsiveness of the signaling pathway from the receptor to the promoter.

[0285] Fluorescent protein variants produced or identified by implementation of the systems and methods discussed herein are useful in applications involving FRET, which can detect events as a function of movement of a fluorescent donor and acceptor toward or away from each other. One or both of the donor / acceptor pair can be a fluorescent acceptor. Such donor / acceptor pairs provide a wide separation between the excitation and emission peaks of the donor and provide good overlap between the donor emission spectrum and the acceptor excitation spectrum.

[0286] FRET can be used to detect the cutting of substrate, and donor and acceptor are coupled to substrate on opposite sides of cutting site.When substrate is cut, donor / acceptor is physically separated, and FRET is eliminated.For example, the test can be performed by contacting substrate with sample and measuring qualitative or quantitative change of FRET (for example, see US Pat. No. 5,741,657, which is incorporated herein by reference). Fluorescent protein variant donor / acceptor pair can also be a part of fusion protein coupled with peptide with proteolytic cleavage site (for example, see US Pat. No. 5,981,200, which is incorporated herein by reference). FRET can also be used to detect the change of transmembrane potential.For example, donor and acceptor can be placed on opposite sides of membrane, so that donor and acceptor are translated on membrane in response to voltage change, thereby producing measurable FRET (for example, see US Pat. No. 5,661,035, which is incorporated herein by reference).

[0287] In other embodiments, fluorescent proteins generated or identified by embodiments of the systems and methods described herein can be used to make fluorescent sensors for protein kinase and phosphatase activity, or indicators for small ions and molecules, such as Ca 2+ 、Zn 2 + , cyclic 3′, 5′-adenosine monophosphate and cyclic 3′, 5′-guanosine monophosphate.

[0288] Fluorescence in a sample is typically measured using a fluorometer, in which excitation radiation from an excitation source having a first wavelength passes through an excitation optical device, which causes the excitation radiation to excite the sample. In response, the fluorescent protein variant in the sample emits radiation having a wavelength different from the excitation wavelength. The collection optical element then collects the emission of the sample. The device may include a temperature controller to maintain the sample at a specific temperature while scanning, and may have a multi-axis translation stage that moves a microtiter plate containing multiple samples to position different wells to be exposed. The multi-axis translation stage, temperature controller, autofocus function, and electronics associated with imaging and data acquisition can be managed by a suitably programmed digital computer, which can also convert the data collected during the analysis process into another presentation format. This process can be miniaturized and automated to screen thousands of compounds in a high-throughput format. Some methods for analyzing fluorescent materials include Lakowicz, “Principles of Fluorescence Spectroscopy” (Plenum Press 1983); Herman, “Resonance energy transfer microscopy” In “Fluorescence Microscopy of Living Cells in Culture” Part B, Meth. Cell Biol. 30:219-243 (ed. Taylor and Wang; Academic Press 1989); Turro, “Modern Molecular Photochemistry” (Benjamin / Cummings Publ. Co., Inc. 1978), pp. 296-361, all of which are incorporated herein by reference.

[0289] Therefore, the present invention also provides an implementation of a method for identifying the presence of a molecule in a sample. For example, the method can be performed by connecting a fluorescent protein variant produced or identified by the implementation of the systems and methods discussed herein to a molecule and detecting the fluorescence generated by the fluorescent protein variant in a sample suspected of containing the molecule. The molecule to be detected can be a polypeptide, a polynucleotide, or any other molecule, including, for example, an antibody, an enzyme, or a receptor, and the fluorescent protein variant can be a tandem dimer fluorescent protein.

[0290] The sample to be examined can be any sample, including a biological sample, an environmental sample, or any other sample, in which it is necessary to determine whether a specific molecule is present. Preferably, the sample includes a cell or an extract thereof. The cell can be obtained from a vertebrate (including a mammal, such as a human) or an invertebrate, and can be a cell from a plant or an animal. The cell can be obtained from a culture of such cells (such as a cell line), or can be isolated from an organism. Therefore, the cell can be contained in a tissue sample, which can be obtained from an organism by any means commonly used to obtain a tissue sample, for example, by biopsy of a human body. If the method is performed using intact living cells or freshly isolated tissue or organ samples, the presence of a molecule of interest in a living cell can be identified, thereby providing a method for determining, for example, the intracellular compartment of a molecule. The use of fluorescent protein variants produced or identified by the embodiments of the systems and methods discussed herein provides a great advantage for such purposes because the possibility of abnormal recognition or localization due to oligomerization of fluorescent proteins is greatly reduced.

[0291] Fluorescent protein variants can be directly or indirectly connected to the molecule using any connection that is stable under the conditions to which the protein-molecule complex is exposed. Therefore, the fluorescent protein and the molecule can be connected by a chemical reaction between the reactive groups present on the protein and the molecule, or the connection can be mediated by a linker portion containing the fluorescent protein and the molecule-specific reactive groups. It should be recognized that, for example, appropriate conditions for connecting fluorescent protein variants and molecules are selected according to the chemical properties of the molecule and the desired type of connection. When the molecule of interest is a polypeptide, a convenient method for connecting the fluorescent protein variant and the molecule is by expressing it as a fusion protein from a recombinant nucleic acid molecule, the recombinant nucleic acid molecule comprising a polynucleotide encoding, for example, a tandem dimer fluorescent protein operably connected to the polynucleotide encoding the polypeptide molecule.

[0292] A method for identifying an agent or condition that modulates the activity of an expression control sequence is also provided. For example, the method can be performed by exposing a recombinant nucleic acid molecule to an agent or condition suspected of being able to modulate the expression of a polynucleotide from an expression control sequence, and detecting the fluorescence of a fluorescent protein variant resulting from such exposure, wherein the recombinant nucleic acid molecule includes a polynucleotide encoding a fluorescent protein variant operably linked to an expression control sequence. For example, this method can be used to identify chemical or biological agents (including cellular proteins) whose expression can be regulated by an expression control sequence, including cytokines that participate in tissue-specific expression from a regulatory element. Thus, an expression control sequence can be a transcriptional regulatory element, such as a promoter, an enhancer, a silencer, an intronic splicing recognition site, a polyadenylation site, etc.; or a translational regulatory element, such as a ribosome binding site.

[0293] Fluorescent protein variants generated or identified by embodiments of the systems and methods discussed herein are also useful in methods for identifying specific interactions between a first molecule and a second molecule. For example, the method can be performed by contacting a first molecule connected to a donor first fluorescent protein variant and a second molecule connected to an acceptor second fluorescent protein variant under conditions that allow a specific interaction between the first molecule and the second molecule; Excite the donor; And detect fluorescence or luminescence resonance energy transfer from the donor to the acceptor, thereby identifying a specific interaction between the first molecule and the second molecule. The conditions for this interaction can be any conditions under which it is expected or suspected that these molecules can specifically interact. In particular, if the molecule to be examined is a cellular molecule, the conditions are typically physiological conditions. Therefore, the method can be performed in vitro using buffers, pH, ionic strength, and the like that simulate physiological conditions, or the method can be performed in cells or using cell extracts.

[0294] Luminescence resonance energy transfer involves the transfer of energy from a chemiluminescent, bioluminescent, lanthanide, or transition metal donor to a red fluorescent protein moiety. The longer excitation wavelength of red fluorescent protein allows energy transfer from a wider variety of donors and over greater distances than green fluorescent protein variants. In addition, solid-state photodetectors can detect longer wavelength emissions more efficiently and are particularly valuable for in vivo applications, where red light penetrates tissues much better than shorter wavelengths. Chemiluminescent donors include, but are not limited to, luminol derivatives and peroxyoxalate systems. Bioluminescent donors include, but are not limited to, aequorin, obelin, firefly luciferase, renilla luciferase, bacterial luciferase, and variants thereof. Lanthanide donors include, but are not limited to, terbium chelates containing a UV-absorbing sensitizer chromophore attached to multiple ligand groups to protect the metal ion from the solvent water. Transition metal donors include, but are not limited to, ruthenium and osmium chelates of oligopyridine ligands. Chemiluminescent and bioluminescent donors do not require excitation light but are excited by the addition of substrate, whereas metal-based systems require excitation light but offer longer excited-state lifetimes, facilitating time-gated detection to distinguish unwanted background fluorescence and scattering.

[0295] The first and second molecules can be cellular proteins being studied to determine whether the proteins specifically interact with each other, or to confirm such an interaction. These first and second cellular proteins can be the same (for example, when their oligomerization ability is detected), or they can be different (for example, when the protein is detected as a specific binding partner involved in an intracellular pathway). The first and second molecules can also be polynucleotides and polypeptides, for example, polynucleotides of known or to be tested transcriptional regulatory element activity and polypeptides of known or to be tested transcription factor activity. For example, the first molecule can contain a plurality of nucleotide sequences of transcriptional regulatory element activity to be tested, which can be random or variants of known sequences, and the second molecule can be a transcription factor. This method can be used to identify new transcriptional regulatory elements with desired activity.

[0296] The present invention also provides an embodiment of a method for determining whether a sample contains an enzyme. For example, the method can be performed by the following steps: contacting the sample with a tandem fluorescent protein variant produced or identified by an embodiment of the system and method discussed herein; exciting the donor, and determining the fluorescent properties in the sample, wherein the presence of the enzyme in the sample results in a change in the degree of fluorescence resonance energy transfer. Similarly, the present invention provides an embodiment of a method for determining enzyme activity in a cell. For example, the method can be performed by the following steps: providing a cell expressing a tandem fluorescent protein variant construct, wherein the peptide linker portion includes a cleavage recognition amino acid sequence specific to an enzyme that couples the donor and the acceptor; exciting the donor, and determining the degree of fluorescence resonance energy transfer in the cell, wherein the presence of enzyme activity in the cell results in a change in the degree of fluorescence resonance energy transfer.

[0297] Experimental example

[0298] The implementation of the systems and methods discussed herein is further described in detail by reference to the following experimental examples. These examples are provided for illustration only and are not intended to be limiting unless otherwise specified. Therefore, the systems and methods discussed herein should by no means be interpreted as being limited to the following examples, but should be interpreted as including any and all variations that become apparent as a result of the teachings provided herein.

[0299] Without further description, it is believed that one of ordinary skill in the art can use the preceding description and the following illustrative examples to make and utilize implementations of the systems and methods discussed herein. Therefore, the following working examples specifically point out exemplary embodiments of the systems and methods discussed herein, and should not be construed as limiting the remainder of the invention in any way.

[0300] Example 1: Protein Engineering Using Neural Networks

[0301] To empirically validate the neural network, three different model proteins were chosen, each representing a different protein engineering challenge. The first validation model protein was tem-1β-lactamase, in large part because 1) sensitivity to antibiotics is directly correlated with the overall stability of the protein and 2) the protein has been well characterized in terms of elucidating stabilizing and destabilizing mutations. Next, to improve stability, the metalloprotein phosphomannose isomerase was modified as a reporter protein to incorporate the non-standard amino acid L-DOPA. However, the poor stability of this enzyme precluded its use as a reporter protein. The final protein case was a modification of the blue fluorescent protein variant secBFP2. Although blue fluorescent proteins have been well characterized, rapid photobleaching, slow maturation and folding, and relatively dim fluorescence have prevented their wider application.

[0302] First, the neural network's true negative rate was evaluated by isolating the analysis of residues where the wild-type amino acid had been experimentally validated as the optimal residue at that position. This was tested using a previously published mutation scan of TEM-1 β-lactamase, where the impact of each individual amino acid change was quantified using organismal fitness. Of the 263 sites tested in TEM-1, 136 had relative fitness values ​​less than zero (i.e., sites that could not tolerate mutations away from the wild-type residue without sacrificing organismal fitness). This set of 136 sites constituted the complete set of true negatives for TEM-1 β-lactamase, and for each discrete change in the neural network, true negative sensitivity was benchmarked. The final version accurately identified 92.6% of the 136 true negatives, an increase of nearly 30% over the initial model. Thus, the developed model has a higher ability to identify sites in the protein that are less prone to mutation.

[0303] The experimental results are as follows Figure 3A and Figure 3B shown. Figure 3A Bar graphs showing locations predicted by the neural network to improve BFP fluorescence and the degree of improvement. The rightmost bar graph 301 shows the fluorescence improvement observed by making a specific combination of amino acid substitutions to the wild-type protein, each substitution being individually suggested by the neural network. A visual representation of the improvement is shown in FIG. Figure 3B As shown, the modified blue fluorescent protein 302 emits brighter than the wild-type blue fluorescent protein 303.

[0304] Other results such as Figure 4A and Figure 4B shown. Figure 4BThe bar graph in Figure 4 shows the improvements proposed by the neural network for phosphomannose isomerase (PMI). Individual stabilizing mutations all yielded increases of 15% to 50% over wild type, but when used in combination (bar 401), the improvements were additive, resulting in a significant increase in stability of nearly 600%.

[0305] Figure 4B The Venn diagrams in 411 (blue fluorescent protein, PDB: 3M24) and 412 (phosphomannose isomerase, PDB: 1PMI) show that the neural network predicts unique candidate residues that were not identified by other computational protein stabilization techniques (FoldX Position Scan and RosettaPmut Scan).

[0306] Figure 5 The neural network identifies variants of TEM-1 β-lactamase that enable E. coli to grow at higher ampicillin concentrations compared to the ancestral protein. E. coli expressing single-mutation β-lactamase mutants N52K, F60Y, M182T, E197D, or A249V can grow at ampicillin concentrations equal to or greater than 125 ug / mL, a concentration at which E. coli expressing the ancestral enzyme labeled "WT" cannot grow. E. coli expressing a single enzyme variant containing all five mutations (N52K, F60Y, M182T, E197D, and A249V, labeled "all") can grow at 3000 ug / mL ampicillin concentrations. In other words, the neural network improves a catalysis-related phenotype, which in this example enables E. coli to exhibit greater resistance to the antibiotic ampicillin.

[0307] Figure 6 The neural network was shown to improve the thermal stability of blue fluorescent protein. In one example, after 10 minutes of thermal excitation, the ancestral protein SecBFP2.1 had less residual fluorescence than the derived protein Bluebonnet. Purified blue fluorescent protein was diluted to 0.01 mg / mL in PBS, pH 7.4, and 100uL aliquots were heat treated for 10 minutes based on a thermal gradient in a PCR strip using a thermal cycler. The fluorescence of the thermally excited variants and controls incubated at room temperature was analyzed using excitation and emission wavelengths of 402nm and 457nm, respectively. The fluorescence readings were normalized to the average of the incubation solutions at room temperature (e.g., a measurement of 0.8 indicates that the heat-treated protein retains 80% of its untreated fluorescence). As Figure 6 As shown, Bluebonnet exhibited higher thermal stability than SecBFP2.1 over the entire temperature range (from about 84°C to about 100°C) - for example, when the ancestral protein retained no fluorescence after 10 min thermal excitation at 100°C, Bluebonnet retained more than 20% of the untreated fluorescence.

[0308] Figure 7 Neural networks were shown to improve the chemical stability of blue fluorescent proteins. In another example, the fluorescence half-life of the ancestral protein SecBFP2.1 was less than that of the derived protein Bluebonnet in guanidinium melt. Purified blue fluorescent protein was diluted to 0.01 mg / mL in 6 M guanidine hydrochloride. 100uL aliquots (in triplicate) were added to the wells of a 96-well clear bottom black wall plate and incubated at 25°C for 23 hours. These purified fluorescent proteins were analyzed at 30 minute intervals using excitation and emission wavelengths of 402nm and 457nm, respectively. The plate was shaken before each measurement. The fluorescence value measured at time zero was used to normalize the fluorescence during the rest of the analysis (e.g., a measurement of 0.8 indicates that the protein retains 80% of its initial fluorescence). As Figure 7 As shown, Bluebonnet exhibited higher chemical stability than SecBFP2.1 at all time points from greater than time = 0 to approximately time = 24 hours.

[0309] Example 2: Bluebonnet, a brighter blue fluorescent protein

[0310] When studying how and where proteins move in cells, scientists need specialized genetic tools. One of these tools is a family of proteins that fluoresce under UV light, known as fluorescent proteins. Blue fluorescent protein (BFP, pdb: 3M24) is a derivative of the more commonly used red fluorescent protein, but has poorer in vivo activity. A 3D convolutional neural network approach was used to predict BFP variants that would result in increased fluorescence when expressed inside E. coli cells. Figure 8 Data presented show that 17 neural network predictions were tested for their ability to enhance fluorescence (shown normalized to wild type). Fig. 9 The data presented show that when beneficial mutations are combined, a >8-fold increase in fluorescence over wild type is observed. Fig.10 An increase in fluorescence of the Bluebonnet blue fluorescent protein (including a combination of S28A, S114T, T127L, and N173H mutations) was visible compared to the parental strain as well as other blue fluorescent proteins.

[0311] Computer system diagram

[0312] Fig.11A and 11B is a block diagram depicting an embodiment of a useful computer associated with implementation of the systems and methods discussed herein. Fig.11A and 11B A block diagram of a typical computer 1100 is depicted. Fig.11A and11B As shown, computer 1100 includes a central processing unit 1102 and a main memory unit 1104. Computer 1100 may also include other optional elements, such as one or more input / output devices 130a-130n (generally using reference numeral 1130), a coprocessor 1106, and a cache memory 1140 in communication with central processing unit 1102 and coprocessor 1106.

[0313] The central processing unit 1102 is any logic circuit that responds to and processes instructions retrieved from the main memory unit 1104. In many embodiments, the central processing unit is provided by a microprocessor unit, such as those manufactured by Intel Corporation (Mountain View, California); by Motorola Corporation (Schaumburg, Illinois); by International Business Machines (White Plains, New York); or by Advanced Micro Devices (Sunnyvale, California).

[0314] Similarly, coprocessor 1106 is any logic circuit that responds to and processes instructions fetched from main memory unit 1104. In some embodiments, coprocessor 1106 may include a tensor processing unit ("TPU"), which is an artificial intelligence application-specific integrated circuit, such as those manufactured by Google (Mountain View, California).

[0315] The main memory unit 1104 can be one or more memory chips that are capable of storing data and allowing the main processor 1102 or the microprocessor in the coprocessor 1106 to directly access any storage location, such as static random access memory (SRAM), burst SRAM or SynchBurst SRAM (BSRAM), dynamic random access memory (DRAM), fast page mode DRAM (FPM DRAM), enhanced DRAM (EDRAM), extended data output RAM (EDO RAM), extended data output DRAM (EDO DRAM), burst extended data output DRAM (BEDO DRAM), enhanced DRAM (EDRAM), synchronous DRAM (SDRAM), JEDEC SRAM, PC100 SDRAM, double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct Rambus DRAM (DRAM), or ferroelectric RAM (FRAM).

[0316] exist Fig.11A In the illustrated embodiment, processor 1102 communicates with main memory 1104 via system bus 1120 (described in greater detail below). Similarly, coprocessor 1106 communicates with main memory 1104 via system bus 1120. Fig. 11B An embodiment of a computer system 1100 is depicted in which a processor 1102 communicates directly with a main memory 1104 through a memory port. Fig. 11B In some embodiments, the main memory 1104 may be a DRDRAM. In some embodiments, the neural network engine may reside in the main memory because the main memory may be responsible for storing the values ​​of the training weights.

[0317] Fig.11A and 11BAn embodiment is depicted in which the main processor 1102 communicates directly with the cache memory 1140 via an auxiliary bus (sometimes referred to as a "backend" bus). In some embodiments, the coprocessor 1106 may communicate directly with the cache memory 1140 via the auxiliary bus. In other embodiments, the main processor 1102 communicates with the cache memory 1140 using the system bus 1120. In other embodiments, the coprocessor 1106 may communicate with the cache 1140 using the system bus 1120. The cache memory 1140 typically has a faster response time than the main memory 1104 and is typically provided by SRAM, BSRAM, or EDRAM. In some embodiments, the coprocessor may include a tensor processing unit (TPU) or other coprocessor, such as an application-specific integrated circuit (ASIC) for performing computations related to neural networks (which may be faster or more efficient than performing such computations on the main processor 1102).

[0318] exist Fig.11A In the illustrated embodiment, the processor 1102 and the coprocessor 1106 communicate with various I / O devices 1130 via a local system bus 1120. Various buses may be used to connect the central processing unit 1102 and the coprocessor 1106 to the I / O devices 1130, including a VESA VL bus, an ISA bus, an EISA bus, a Micro Channel Architecture (MCA) bus, a PCI bus, a PCI-X bus, a PCI-Express bus, or a NuBus. For embodiments in which the I / O device is a video display, the processor 1102 and / or the coprocessor 1106 may communicate with the display using an Advanced Graphics Port (AGP). Fig. 11B An embodiment of a computer system 1100 is depicted in which the main processor 1102 communicates directly with the I / O device 1130b via HyperTransport, Rapid I / O, or InfiniBand. Fig. 11B An embodiment in which local bus and direct communication are mixed is also depicted: processor 1102 communicates with I / O device 1130a using a local interconnect bus while communicating directly with I / O device 1130b.

[0319] A variety of I / O devices 1130 may be present in the computer system 1100. Input devices include keyboards, mice, trackpads, trackballs, microphones, and drawing tablets. Output devices include video displays, speakers, inkjet printers, laser printers, and dye sublimation printers. I / O devices may also provide mass storage for the computer system 1100, such as hard disk drives, floppy disk drives for receiving floppy disks (e.g., 3.5-inch, 5.25-inch disks, or ZIP disks), CD-ROM drives, CD-R / RW drives, DVD-ROM drives, tape drives of various formats, and USB storage devices, such as the USB flash drive series of devices manufactured by Twintech Industry, Inc. (Los Alamitos, California), and the iPod Shuffle series of devices manufactured by Apple Computer, Inc., (Cupertino, California).

[0320] In a further embodiment, the I / O device 1130 can be a bridge between the system bus 1120 and an external communication bus (such as a USB bus, an Apple desktop bus, an RS-232 serial connection, a SCSI bus, a FireWire bus, a FireWire 800 bus, an Ethernet bus, an AppleTalk bus, a Gigabit Ethernet bus, an asynchronous transfer mode bus, a HIPPI bus), a Super HIPPI bus, a SerialPlus bus, an SCI / LAMP bus, a Fibre Channel bus, or a serially connected Small Computer System Interface bus.

[0321] Fig.11A and 11B A general-purpose desktop computer of the type shown typically runs under the control of an operating system, which controls the scheduling of tasks and access to system resources. Typical operating systems include: MICROSOFT WINDOWS, manufactured by Microsoft Corp. (Redmond, Washington); MacOS, manufactured by Apple Computer (Cupertino, California); OS / 2, manufactured by International Business Machines (Armonk, New York); and Linux, a free operating system distributed by CalderaCorp. (Salt Lake City, Utah).

[0322] The disclosure of each patent, patent application and publication cited herein is incorporated herein by reference in its entirety. Although the present invention has been disclosed with reference to specific embodiments, it is apparent that other persons skilled in the art may design other embodiments and variations of the present invention without departing from the true spirit and scope of the present invention. The appended claims are intended to be interpreted as including all such embodiments and equivalent variations. Sequence Listing <110> The University of Texas Board of Regents Andrew Ellington Austin Cole Raghav Shroff Rose Tell <120> Systems and methods for improving the stability of synthetic proteins <130> 191740-21F-CNP <150> 62 / 841,906 <151> 2019-05-02 <160> 28 <170> PatentIn version 3.5 <210> 1 <211> 233 <212> PRT <213> Artificial sequence <220> <223> Wild-type secBFP2 <400> 1 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 2 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the T18W mutation <400> 2 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Trp Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 3 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the T18V mutation <400> 3 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Val Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 4 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the T18E mutation <400> 4 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Glu Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 5 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the S28A mutation <400> 5 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 6 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the Y96F mutation <400> 6 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Phe 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 7 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the S114V mutation <400> 7 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Val Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 8 <211> 233 <212> PRT <213> Artificial Sequence <220> secBFP2 containing the S114T mutation <400> 8 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 9 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the V124T mutation <400> 9 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Thr Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 10 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the V124Y mutation <400> 10 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Tyr Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 11 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the V124W mutation <400> 11 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Trp Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 12 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the T127P mutation <400> 12 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Pro Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 13 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 with T127L Mutation <400> 13 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Leu Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 14 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the T127R mutation <400> 14 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Arg Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 15 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the T127D mutation <400> 15 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Asp Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 16 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the D151G mutation <400> 16 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Gly Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 17 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the N173T mutation <400> 17 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Thr Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 18 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the N173H mutation <400> 18 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala His Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 19 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the N173R mutation <400> 19 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Arg Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 20 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 Containing the N173S Mutation <400> 20 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Ser Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 21 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing the R198V mutation <400> 21 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Val Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 22 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the R198L mutation <400> 22 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ser Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Ser Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Leu Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 23 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing T18W, S28A, S114V, V124T, T127P, D151G, N173T and R198L mutations <400> 23 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Trp Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Val Leu Ile Tyr Asn Val Lys Ile Arg Gly Thr Asp Phe Pro Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Gly Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Thr Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Leu Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 24 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing S28A, S114T, T127L and N173H mutations <400> 24 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Leu Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala His Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 25 <211> 233 <212> PRT <213> Artificial sequence <220> <223> Comprising secBFP2 containing S28A and S114T mutations <400> 25 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala Asn Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 26 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing S28A, S114T and N173H mutations <400> 26 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Tyr 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala His Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 27 <211> 233 <212> PRT <213> Artificial Sequence <220> <223> secBFP2 containing S28A, Y96F, S114T and N173H mutations <400> 27 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Phe 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Thr Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala His Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230 <210> 28 <211> 233 <212> PRT <213> Artificial sequence <220> <223> secBFP2 containing the S28A, Y96F, S114T, T127L and N173H mutations <400> 28 Ser Glu Glu Leu Ile Lys Glu Asn Met His Met Lys Leu Tyr Met Glu 1 5 10 15 Gly Thr Val Asp Asn His His Phe Lys Cys Thr Ala Glu Gly Glu Gly 20 25 30 Lys Pro Tyr Glu Gly Thr Gln Thr Met Arg Ile Lys Val Val Glu Gly 35 40 45 Gly Pro Leu Pro Phe Ala Phe Asp Ile Leu Ala Thr Ser Phe Leu Tyr 50 55 60 Gly Ser Lys Thr Phe Ile Asp His Thr Gln Gly Ile Pro Asp Phe Phe 65 70 75 80 Lys Gln Ser Phe Pro Glu Gly Phe Thr Trp Glu Arg Val Thr Thr Phe 85 90 95 Glu Asp Gly Gly Val Leu Thr Ala Thr Gln Asp Thr Ser Leu Gln Asp 100 105 110 Gly Thr Leu Ile Tyr Asn Val Lys Ile Arg Gly Val Asp Phe Leu Ser 115 120 125 Asn Gly Pro Val Met Gln Lys Lys Thr Leu Gly Trp Glu Ala Phe Thr 130 135 140 Glu Thr Leu Tyr Pro Ala Asp Gly Gly Leu Glu Gly Arg Asn Asp Met 145 150 155 160 Ala Leu Lys Leu Val Gly Gly Ser His Leu Ile Ala His Ala Lys Thr 165 170 175 Thr Tyr Arg Ser Lys Lys Pro Ala Lys Asn Leu Lys Met Pro Gly Val 180 185 190 Tyr Tyr Val Asp Tyr Arg Leu Glu Arg Ile Lys Glu Ala Asn Asp Glu 195 200 205 Thr Tyr Val Glu Gln His Glu Val Ala Val Ala Arg Tyr Ser Asp Leu 210 215 220 Pro Ser Lys Leu Gly His Lys Leu Asn 225 230

Claims

1. A computer-implemented method for training a neural network to improve protein properties, comprising: Collect a set of amino acid sequences from a database; compiling a set of three-dimensional structures with chemical environments for the set of amino acids; converting the chemical environment into a voxelized matrix; the step of adding to the at least one three-dimensional structure a spatial arrangement of features selected from the group consisting of hydrogen positions, partial charges, beta factors, secondary structure, aromaticity, electron density, and polarity; training a neural network using a subset of the voxelized matrix; using the neural network to identify candidate residues to be mutated in the target protein; and using the neural network to identify predicted amino acid residues to replace the candidate residues, thereby generating a mutant protein; The mutant protein exhibits improved properties relative to the target protein.

2. The computer-implemented method of claim 1, further comprising adjusting the set of amino acid sequences to reflect their natural frequencies.

3. The computer-implemented method of claim 1, further comprising sampling at least 50% of the amino acids in the set of amino acid sequences from random positions in the sequence.

4. The computer-implemented method of claim 1 , further comprising training a second independent neural network using a second subset of the three-dimensional structures, and identifying candidate residues and predicted residues based on the results of the two neural networks.

5. The computer-implemented method of claim 1, wherein the characteristic is stability, maturity, or folding.

6. The computer-implemented method of claim 1, wherein at least a portion of the set of three-dimensional structures are three-dimensional crystal structures.

7. A system for improving protein properties, comprising a processor and a non-transitory computer readable medium storing instructions, which, when executed by the processor, perform the steps of: providing a target protein comprising an amino acid sequence; providing a set of three-dimensional models around amino acids and providing a set of protein property values ​​for each three-dimensional model; adding a characteristic spatial arrangement to at least one of the amino acid sequences; estimating a set of parameters at different points in each three-dimensional model; Training a neural network using the three-dimensional model, the parameters and the protein characteristic values; identifying candidate residues to be mutated in the target protein using the neural network; and using the neural network to identify predicted amino acid residues to replace the candidate residues to generate mutant proteins; The mutant protein exhibits improved properties relative to the target protein.

8. The system of claim 7, wherein the protein property is stability.

9. The system of claim 7, wherein the step comprises recompiling the at least one amino acid sequence to generate an updated three-dimensional model.

10. A method for improving one or more properties of a target protein, the method comprising: Analyzing the amino acid sequence of the target protein using the trained neural network to identify one or more amino acid residues at specific positions of the amino acid sequence as candidate residues for mutation; and identifying one or more predicted amino acid residues for use as substitutes for at least one of the candidate residues using the neural network, wherein the neural network is trained by: (a) generating a multidimensional array representing a folded protein having a given amino acid residue sequence, wherein the folded protein exhibits one or more properties associated with the microenvironment of each amino acid residue; (b) preprocessing the multidimensional array into a vector; (c) calculating, via the neural network, predicted amino acid residues located at the center of a microenvironment associated with a folded protein based on the preprocessed vectors; (d) determining the differences between the predicted amino acid residues and the amino acid residues associated with the microenvironment; and (e) in response to determining that the difference exceeds the threshold, iteratively repeating steps (a)-(d) for different folded proteins; The preprocessing of the multidimensional array comprises: extracting features from a subset of the multidimensional array and downsampling the extracted features to generate a feature-specific map; and Combine feature-specific maps into a one-dimensional vector.

11. The method of claim 10, further comprising using the neural network to identify one or more predicted amino acid residues for use as substitutes for at least one other of the candidate residues.

12. The method of claim 10, further comprising using the neural network to identify one or more predicted amino acid residues to serve as substitutes for each candidate residue.

13. The method of claim 10, further comprising synthesizing a mutant protein by making one or more substitutions, wherein the mutant protein exhibits one or more improved properties compared to the target protein.

14. The method of claim 10, further comprising generating the multidimensional array from a sample of one or more amino acids from an amino acid sequence of the target protein.

15. The method of claim 10, wherein generating a multidimensional array further comprises mapping the three-dimensional model of the folded protein to a voxelized matrix.

16. The method of claim 10, wherein step (e) further comprises modifying one or more neuron weights of the neural network in response to differences between the predicted candidate residues and amino acid residues and the measured residues and amino acid residues.

Images