Method for enzyme fitness data collection

Abstract

The present invention relates to a method for estimating or determining fitness of enzyme variants and / or optimizing a nucleotide sequence encoding an enzyme variant comprising generating a DNA variant library and linking a unique DNA barcode to each DNA variant, wherein the DNA variants are subjected to long-read and short-read sequencing. In particular, the present invention relates to a method for predicting or determining fitness of enzyme variants and / or optimizing a nucleotide sequence encoding an enzyme variant by combining DNA barcodes, selection pressure, long-read sequencing, and short-read sequencing. Furthermore, the present invention relates to application of machine learning for predicting enzyme fitness and / or optimizing a nucleotide sequence encoding an enzyme variant based on data obtained from said method.

Description

[0001]82843PC01 1Method for enzyme fitness data collectionTechnical field of the inventionThe present invention relates to a method for predicting or determining fitness ofenzyme variants comprising generating a DNA variant library of a given enzyme template and linking a unique DNA barcode to each DNA variant, wherein the DNAvariants are subjected to long-read and short-read sequencing. In particular, thepresent invention relates to a method for estimating fitness of enzyme variantsand / or optimizing a nucleotide sequence encoding an enzyme variant bycombining DNA barcodes, selection pressure based on enzyme catalytic activity,long-read sequencing, and short-read sequencing. Background of the inventionWhile machine learning and deep learning have been applied successfully towardmany protein-related problems, for example structure prediction and enzymeclassification, machine learning has not been particularly powerful for predictingenzyme activities. This is due to a lack of data. Enzyme data collection is typicallytime- and resource-intensive. Although many advanced machine learningalgorithms have been developed over the recent years, machine learning cannotyield good performance without sufficient data. Data for application in machinelearning for predicting enzyme fitness should be of high quality, such ascomprising full-length enzyme sequence, and in large amounts.Yang et al. (2021) (Defining protein variant functions using high-complexitymutagenesis libraries and enhanced mutant detection software ASMv1.0.) describes a method for screening of a variant expression library for a gene of interest for characterizing the functions of thousands of single-amino-acidsubstitutions in a single experiment. The method comprises characterizing proteinfitness of several variants by attaching barcodes to the gene of interest,amplification and expression in cells, growth selection, and deep sequencing. Further, while Yang et al. mentioned that long read sequencing is possible forassociating a set of unique external barcodes to each of the variants in asaturation variant library, so that the screen deconvolution work can be achievedby direct next-generation sequencing of the barcodes, they are doubtful about the 82843PC01 2 feasibility of this approach, and are silent about applying this approach to obtaining enzyme activity data.A review paper by Wei H, Li X. (2023) (Deep mutational scanning: A versatile toolin systematically mapping genotypes to phenotypes. Front Genet. 2023 Jan 12)proposed the possibility of constructing a barcoded mutant variant library, cellgrowth selection, associating each barcode with corresponding variant using long- read sequencing, and afterwards performing deep sequencing of the barcodes.Wei H, Li X. are silent in respect of coupling cell growth to the catalytic activity ofenzymes, and are silent about using such an approach for enzyme activity datacollection. Wei H, Li X. are also silent about using data obtained in such a mannerto predict enzyme sequences or sequence libraries with potentially improved activity.Fowler DM et al. (2014) (Measuring the activity of protein variants on a largescale using deep mutational scanning. Nat Protoc. 2014 Sep) describes analysis ofbig dataset relating to protein activity within a short time period. The techniquecomprises attachment of a unique DNA barcode on each variant and combinesselection with high-throughput DNA sequencing, for a library of tens to hundredsof thousands of variants of some protein and imposes a selection for function, including growth-based selections. High-throughput DNA sequencing measures the frequency of each variant during the selection experiment. The barcodes areused to create a database by performing deep sequencing and the frequency ofeach variant is identified by deep sequencing the barcode only. The associationbetween the barcodes and the enzyme sequences needs to be established individually through Sanger sequencing, limiting the capacity of the number of variants. Fowler DM et al. are silent about using high-throughput long-read sequencing such as through the Nanopore or PacBio platforms for establishing such association. WO 2022 / 125686 A1 provides a process for determining a coding DNA sequence for a specific dose-response curve for a genetic sensor.Zhuoxing Wu et al. (Expression level is a major modifier of the fitness landscapeof a protein coding gene. Nat Ecol Evol. 2022) describes an analysis of expressionlevel of the fitness landscape of protein coding gene. 82843PC01 3There is no disclosure for obtaining full-length (>100 amino acids) enzymesequence-activity data through a combination of 1) coupling of cell growth to the catalytic activity of an enzyme; 2) molecular barcoding; 3) long-read sequencing;4) deep sequencing, to obtain data for individual labeling of enzyme sequenceswith their fitness levels, nor is there any disclosure of using data obtained in sucha manner to predict enzyme sequences or sequence libraries with potentially improved activity.Hence, an improved method for predicting enzyme fitness and thereby optimizinga nucleotide sequence encoding an enzyme variant would be advantageous, andin particular a more efficient and / or reliable method for estimating enzyme fitnessand / or optimizing a nucleotide sequence encoding an enzyme variant comprisinghigh quality data in large amounts, which is advantageous for predicting enzymesequences with improved fitness or activity through machine learning.Summary of the inventionThe invention presented relates to a method for estimating fitness of enzymevariants and / or for optimizing a nucleotide sequence encoding an enzyme variant.The method relates to extracting data relating to DNA sequence variants encodingenzyme variant and their fitness. The method comprises generating a DNA variantlibrary of a given enzyme and linking a unique DNA barcode on each DNA variant. The DNA variant library is transformed into a selected strain for non-selective amplification and expression of the library. The amplified DNA variant library isextracted and subjected to long-read sequencing for generating information thatassociates the unique barcodes to their corresponding enzyme variants. Further, the strain comprising the DNA variant library is subjected to a selectionpressure favoring enzyme variants with increased fitness as manifested byincreased catalytic activity. The growth rate of the strain under selection pressureis a reflection of enzyme fitness and activity. Subsequently, the DNA variantlibrary is extracted and subjected to short-read deep sequencing, wherein onlythe barcode region is sequenced. The frequency of each barcode is used as areflection of the fitness of the corresponding enzyme variant and thereby allowingoptimization of a nucleotide sequence encoding an improved enzyme variant. Thebarcode is used for identifying the corresponding enzyme variant. 82843PC01 4 The enzyme sequence-activity data generated from the invention will be of high quality comprising accurate full-length of the enzyme sequences, and in largeamounts being a capacity of more than 106 data points per run with currenttechnologies, which far exceeds the state-of-the-art. This will greatly improve theapplication of machine learning for enzyme activity-related tasks.Example 1 shows the construction of the selection strain, where the activity andfitness of SAM-dependent methyltransferases is coupled to cell growth byproviding precursors for cysteine synthesis. The SAM-dependentmethyltransferase used in this example is Arabidopsis thaliana N-acetylserotoninO-methyltransferase (atASMT), and the target catalytic activity is the conversion of protocatechuic acid (PCA) into vanillic acid.Example 2 shows the creation of a library of DNA variants linked with unique DNAbarcodes. The DNA library can be obtained through error-prone PCR or othermethods. The variants in the DNA library are linked with unique DNA barcodesthrough a one-cycle PCR process. The barcoded DNA library is subsequently assembled into a plasmid vector for expression in E. coli.Example 3 shows the construction of an in vivo library by transforming theplasmid library, containing the DNA library for enzyme variants, into the selectionstrain. The in vivo library is further amplified by non-selective growth for differentsubsequent uses.Example 4 shows Nanopore sequencing to generate data that associates the DNAvariant sequences with their associated DNA barcodes. This example shows extraction of the plasmid library after growing an aliquot of the strain from Example 3 with antibiotic selection for the plasmid, but without selection for the enzyme variants. The region including the target enzyme expression cassette and barcode is isolated by restriction digestion and sequenced through ampliconsequencing with Nanopore long-read sequencing.Example 5 shows growth of the selection strain comprising the DNA library withselection biased towards higher target enzyme activity. The cells proliferate with agrowth rate positively correlated with the activity of the target enzyme variant. 82843PC01 5Sampling during this process and the subsequent short-read sequencing tocharacterize the change of variant abundance / frequency, allow the calculation of the growth rate of the cell population harboring a specific enzyme variant.Example 6 shows short-read sequencing of the barcode region for obtaining thefrequency change of the DNA barcodes for cell samples over the course ofselective growth. Short-read sequencing of the barcode region is performed toobtain the frequency information of the barcodes for each cell sample.Example 7 shows processing of long-read sequencing data and short-readsequencing data obtained from Example 4 and 6, respectively, in order to generate the full DNA sequence-activity relationship. The DNA barcode is used tolink the full-length DNA sequence to the corresponding growth rate.Example 8 shows training of a machine learning model to predict amino acidsequence fitness using obtained sequence-fitness data. The top-5 most highlyranked variants of the original atASMT sequence were identified.Example 9 shows improvement of sample preparation to improve the size,diversity, and quality of the collected data.Example 10 shows improvement of the quality and output size of the consensussequence generated from Nanopore data, and to improve the statistical rigor of the fitness estimation for sequence variants.Example 11 shows generation of a series of variant sequences using the MoCHIsuite, which identifies beneficial and non-beneficial mutations by considering bothindividual and epistatic effects. The resulting library combines highly beneficialmutations and epistatic interactions while avoiding less beneficial ones.Example 12 shows experimental validation of the data obtained from method ofthe invention to evaluate the performance of sequences generated based onmachine learned information from the data. The estimated enzyme fitness levelsobtained in the high-throughput manner are positively correlated with results in the bioconversion assay. In addition, many generated sequences based on 82843PC01 6learned mutational effects shows significant activity. These results indicate thegreat value of using the high-throughput dataset obtained by the method of the present invention for enzyme engineering.Thus, an object of the present invention relates to the provision of an improvedmethod for estimating enzyme fitness and / or optimizing a nucleotide sequenceencoding an enzyme variant by combining barcodes, selective growth, long-read,and short-read sequencing.In particular, it is an object of the present invention to provide a method thatsolves the above-mentioned problems of the prior art with an improved methodfor determining enzyme fitness allowing optimization of a nucleotide sequenceencoding an enzyme variant, which provides data of high quality, such as comprising full-length enzyme sequence, and in large amounts, which is furthersuitable for application in machine learning for predicting enzyme fitness and / orobtaining optimized / improved enzyme variant.Thus, one aspect of the invention relates to a method for estimating fitness ofenzyme variants, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a 82843PC01 7 first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant (such as inthe entire population of enzyme variants); and g) optionally estimating the enzyme fitness of the specific enzyme variantfrom the determined frequency information.A further aspect of the invention relates to a method for optimizing a nucleotidesequence encoding an enzyme variant, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, 82843PC01 8 e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant;f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant; and g) obtaining the optimized nucleotide sequence encoding an enzyme variantfrom the determined frequency information.Another aspect of the present invention relates to use of at least one barcodedDNA sequence encoding an enzyme variant for predicting enzyme fitness of said enzyme variant, wherein the at least one barcoded DNA sequence is subjected to long-read sequencing and short-read sequencing. Yet another aspect of the present invention relates to use of at least one barcoded DNA sequence encoding an enzyme variant for optimizing said DNA sequence encoding said enzyme variant, wherein the at least one barcoded DNA sequence is subjected to long-read sequencing and short-read sequencing. In an aspect, the invention relates to a polypeptide according to SEQ ID NO. 3 comprising one or more of the following mutations P72S, K132R, D147E, I103N, P286L, R77C, G105S, K187R, M215K, P231S, E46D, S84I, and / or P302L. Another aspect of the present invention relates to processor system programmed to operate according to a machine learning (ML) algorithm for estimating fitness of enzyme variants, the machine learning (ML) algorithm being trained, and / or being trainable, on data obtained by a method according to the first aspect of the invention. A further aspect of the present invention relates to use of a machine learning (ML) algorithm trained on data comprising a correlation of a specific nucleotidesequence encoding an enzyme variant to frequency information for optimizing a 82843PC01 9 nucleotide sequence encoding an enzyme variant, or to guide the generation of animproved nucleotide sequence encoding an enzyme variant with given fitnessproperties. Another aspect of the invention relates to use of a machine learning (ML) algorithm trained on data obtained by a method according to the first aspect ofthe invention for optimizing a nucleotide sequence encoding an enzyme variant, orto guide the generation of an improved nucleotide sequence encoding an enzymevariant with given fitness properties.In yet another aspect, the invention relates to use of a machine learning (ML) algorithm trained on data obtained by a method according to the first aspect ofthe invention to predict fitness for any enzyme sequence, or to guide thegeneration of enzyme sequences with given fitness properties.An aspect of the invention relates to a system suitable for executing an algorithm,such as machine learning (ML) algorithm, for optimizing a nucleotide sequenceencoding an enzyme variant, the system being trained, and / or being trainable, on data obtained by a method according to the first aspect of the invention, the system being arranged for:- receiving a first set of information (1SI), such as a first database, said first ofinformation being obtained, or being obtainable, from d) subjecting the isolated DNA variant library grown under condition I.) to long-read sequencing, thereby obtaining said first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant,- receiving a second set of information (2SI), such as a second database, saidsecond of information being obtained, or being obtainable, from e) subjecting the isolated DNA variant library grown under condition II.) to short-read deep sequencing, thereby obtaining said second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; 82843PC01 10 wherein the system comprises: -a first comparison module (1CM) for comparing said first set of information (1SI) to said second set of information (2SI), thereby correlating specific enzyme variants to the frequency information, such as frequency of the variant, using the DNA barcode as an identifier; and- an optional second comparison module (2CM) for correlating determinedfrequency information to the enzyme fitness of the specific enzyme variant.Yet another aspect of the invention relates to a method for training a machinelearning (ML) system for optimizing a nucleotide sequence encoding an enzymevariant according to the invention, the method comprises the steps of:-receiving training data comprising a first set of information (1SI), such as a first database, and a second set of information (2SI), such as a second database, -training the ML system for optimizing a nucleotide sequence encoding an enzymevariant using said training data, and- validating the ML system using correlated specific enzyme variants to thefrequency information, such as frequency of the variant. A further aspect relates to a computer-readable storage medium comprising dataobtained or obtainable by a method according to the first aspect.An aspect of the invention relates to neural network model obtained or obtainableby a method for training a machine learning (ML) system according to theinvention. Brief description of the figuresFigure 1 shows a workflow of the library generation, barcoding, and in vivo librarypreparation, Figure 2 shows a sequence-fitness data collection workflow, Figure 3 shows distribution of relative enzyme fitness for the aggregated dataset, 82843PC01 11Figure 4 shows the distribution of growth rates for A) variants with prematurestop codons using the aggregated dataset, based on variants with at least 7 readsand B) variants with premature stop codons using the aggregated dataset, basedon variants with at least 15 reads. C) illustrates comparable levels of dispersion for the WT2,Figure 5 shows growth assay results. A strong positive correlation was observedbetween the estimated fitness in dataset and the measured fitness in the growth assay, with R2value of 0.87, andFigure 6 shows the bioconversion results. A) vanillic acid produced inbioconversion experiment by all cloned ASMT variants. The dataset includes ASMT variants identified from the selected population, and novel ASMT variantsgenerated using first- and second-order models. ASMT-WT refers to the originalwild-type ASMT sequence. ASMT1-ASMT37 refer to ASMT variants from the dataset. The rest of the ASMT sequences were generated with the second-order model. B) shows vanillic acid production plotted against predicted growth rate for all cloned ASMT variants identified during selection. Error bars show one standard deviation. 95% confidence bands for the best-fit line are depicted with dashedlines. The variance of the data is considered when calculating the R2 value.Significance of deviation of regression slope from zero: p<0.001 (α=0.05)(Graphpad Prism). C) shows vanillic acid production plotted against predictedgrowth rate for functional, cloned ASMT variants identified during selection with detectable levels of vanillic acid in bioconversion experiment. Error bars show one standard deviation. 95% confidence bands for the best-fit line are depicted with dashed lines. The variance of the data is considered when calculating the R2 value. Significance of deviation of regression slope from zero: p<0.001 (α=0.05)(Graphpad Prism). D) shows vanillic acid production plotted against predictedgrowth rate for functional, cloned ASMT variants identified during selection with detectable levels of vanillic acid in bioconversion experiment. Only data points where SD<0.5 for vanillic acid production and SD<0.0152 for predicted growth rate are included. Error bars show one standard deviation. 95% confidence bands for the best-fit line are depicted with dashed lines. The variance of the data is considered when calculating the R2value. Significance of deviation of regression slope from zero: p=0.003 (α=0.05) (Graphpad Prism). 82843PC01 12 The present invention will now be described in more detail in the following. Detailed description of the invention Definitions Prior to discussing the present invention in further details, the following terms and conventions will first be defined: Long-read sequencingLong-read sequencing is a group of DNA / RNA sequencing techniques that enablesthe sequencing of much longer DNA fragments, such as having more than 601bases to more than 20,000 bases, than traditional short-read deep sequencingmethods. Long-read sequencing tends to have lower accuracy and lower readdepth than short-read deep sequencing.Long-read sequencing technologies may be selected from the group consisting ofPacific Biosciences’ (PacBio) single-molecule real-time (SMRT) sequencing and Oxford Nanopore Technologies’ (ONT) nanopore sequencing, Illumina Complete Long Read sequencing, TELL-Seq. Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) are alsoknown as third generation sequencing technologies.Short-read sequencingShort-read sequencing is sequencing of shorter fragments of DNA or RNA. Inthese types of sequencing, DNA or RNA is broken into smaller fragments, beforebeing sequenced. With short read sequencing, DNA or RNA can be sequenced withhigh read number, and with length limited to 50-300 per direction. Short-read sequencing is a form of next-generation sequencing (NGS), wherein the platform for short-read sequencing may be selected from the group consistingof Illumina, such as HiSeq2000, HiSeq2500, HiSeq4000, HiSeqX10, andNovaSeq6000, and MGI, such as BGISEQ-500, and MGISEQ-T7.Short-read sequencing is also known as deep sequencing; hence the terms maybe used interchangeably. Enzyme Enzymes are biological molecules (such as proteins and RNAs) that catalyze chemical reactions. The molecules that enzymes act upon are called substrates, and the molecules that enzymes convert the substrates into are called products. 82843PC01 13 DNA variant libraryA DNA variant library is a pool of DNA variants encoding enzyme variants that areexpected to have different sequences with different fitness levels. The DNAvariants are obtained through various methods, such as error-prone PCR based ongiven templates, or assembly of DNA fragments that differ from each other in specific region or regions. Enzyme variants Enzyme variants are the protein product from translation of the DNA variants. Theenzyme variants are expected to have different fitness levels or enzyme activitydue to their different sequences. Enzyme fitness Enzyme fitness is a measure of how well a given enzyme performs a specificfunction. It can be used interchangeably with enzyme activity in the context ofthis invention. It encompasses properties such as expression level or ease of expression in a given host cell, turnover number (unimolecular rate constant, Kcat), Michaelis constant (KM), solubility, and stability, substrate specificity, product specificity, thermotolerance. Thus, in an embodiment, the enzyme fitness is selected from the group consisting of enzyme activity, expression level, turnover number such as unimolecular rate constant (Kcat), Michaelis constant (KM), enzyme solubility, enzyme stability, substrate specificity, product specificity, thermotolerance, and combinations thereof. Rate of changeThe rate of change is to be understood as the change of abundance or frequencyof the DNA / enzyme variants, or the change of abundance or frequency of the cellsthat express specific enzyme variants. DNA barcode The DNA barcode is (a) DNA fragment(s) that is attached to a DNA variant, often randomly. The DNA barcode is linked to the DNA variants as an extension of the DNA sequence of the DNA variant via phosphodiester bonds. As an example, a one-cycle PCR step can be performed on a DNA variant to attach a unique barcode to the DNA variant. 82843PC01 14DNA barcodes serve three purposes: 1) for generating accurate full-length DNAsequence reading from long-read sequencing based on consensus generation forthe DNA reads with the same barcodes; 2) for profiling the frequency change ofDNA variants during growth-based selection using short-read sequencing; and 3)for associating the frequency information obtained using short-read sequencing to the full-length DNA variant sequence information obtained using long-read sequencing.Selection pressureA selection strain comprising a DNA variant library encoding enzyme variants of aspecific template enzyme may be subjected to a selection pressure biased towardhigher target enzyme activity. The choice of specific selection pressure dependson the specific type of target enzyme. As an example, a selection strain can beconstructed by deleting endogenous gene(s) responsible for synthesizing cysteine in the host cells, while introducing an alternative pathway for synthesizing cysteine, part of which is the target enzyme to be selected. In this case, in a certain range, higher target enzyme activity will lead to improved growth of the selection strain in a given media where the introduced alternative pathway could be used.The selection strain may also be subjected to selection pressure with antibiotics,such as chloramphenicol, ampicillin, and the like, which does not bias towardhigher target enzyme activity. Genetic modification In this invention, genetic modification specifically refers to changes introduced to the genome of the host cells besides the introduction of the DNA variants. Geneticmodifications serve the purpose of rendering the host cells susceptible to a certainselection pressure. As one example, genetic modifications may be used to delete the endogenous cysteine synthesis pathway and introduce a heterologous cysteine synthesis pathway such that there is selection pressure for higher activity for the enzymes in the heterologous pathway. As another example, genetic modifications may be used to introduce a sensor that could sense the concentration of the product of a target enzyme, and control the expression of an essential gene, such that higher target enzyme activity leads to higher expression level of the essential gene. 82843PC01 15 Growth rateThe growth rate of the selection strain is determined by the change in the numberof cells over a specific period of time. Under selection pressure, the growth rate ofa cell variant of the selection strain signifies the activity or fitness of a specificenzyme variant that the cell variant contains. Hence, a high growth rate correlateswith a high enzyme activity or enzyme fitness. Optical density (OD) Optical density (OD) is a measure of how much light is absorbed by a material as it passes through it. It is defined as the logarithmic ratio of the intensity of incident light (Io) to the intensity of transmitted light (It) through the material.In the context of cell cultures, optical density OD is commonly used to estimatethe concentration or density of cells in a culture. This is typically done using a spectrophotometer, which measures how much light is scattered by the cells inthe sample. When light passes through a cell culture, cells scatter the light. Thespectrophotometer measures this scattering, which is reported as optical density. The more cells present, the higher the OD value. MoCHIMoCHI (Faure, Andre & Lehner, Ben. (2024). MoCHI: neural networks to fitinterpretable models and quantify energies, energetic couplings, epistasis andallostery from deep mutational scanning data) is a software tool that allows theparameterisation of arbitrarily complex models using deep mutational scanning(DMS) data. MoCHI simplifies the task of building custom models frommeasurements of mutant effects on any number of phenotypes. It allows the inference of free energy changes, as well as pairwise and higher-order interaction terms (energetic couplings) for specified biophysical models. When a suitable user-specified mechanistic model is not available, global nonlinearities (epistasis) can be estimated directly from the data. MoCHI also builds upon and leverages theory on ensemble (or background-averaged) epistasis to learn sparse predictive models that can incorporate higher-order epistatic terms and are informative of the genetic architecture of the underlying biological system. The combination of DMS and MoCHI allows biophysical measurements to be performed at scale, including the construction of complete allosteric maps of proteins. 82843PC01 16 Storage mediumA storage medium is to be understood as a physical material or device used tostore digital data. It allows data to be saved, retrieved, and managed. A storagemedium may be, but not limited to: -a magnetic storage, such as Hard Disk Drive (HDD), magnetic tape, floppydisk, zip disk, super disk, -flash memory, such as USB flash drives, SD cards, Solid State Drive (SSD),- optical disc, or- network storage, such as Cloud storage, Network-Attached Storage (NAS).The invention presented is in the framework of directed evolution, meaning thescreening and selection of enzyme variant libraries based on the metrics of theenzymes. The aim is not to simply find the best-performing variant, but also thefitness information of each enzyme variant in the libraries, in order to obtain an information-rich dataset useful for machine learning. It combines four essential parts: 1) Coupling cell growth to the catalytic activity of an enzyme. Thisenables the use of cell growth rate as a reflection of enzyme activity / fitness. 2) Attaching unique DNA barcodes to the DNA sequence of eachenzyme variant. A crucial step in the library creation procedure is the attachment of one or more unique DNA barcodes to the DNA encoding eachenzyme variant. 3) Using long-read sequencing technologies (e.g., Oxford Nanopore,PacBio) to sequence the variant library. This allows the creation of a DNA variant sequence-barcode database that shows, which barcodecorresponds to which specific DNA variant encoding a specific enzymevariant. The DNA barcode additionally serves the purpose of correcting the relatively high error rate of long-read sequencing: by reading multiple copies of a DNA variant-barcode fragment, different sequencing readscontaining the same DNA barcode sequence will be assumed to have the same sequence, and the alignment of these different reads containing 82843PC01 17 sequencing errors will generate a consensus sequence with high sequence accuracy. 4) Using deep sequencing technologies (e.g., Illumina sequencing) toprofile the frequency change of barcodes during the library selection process. The frequency, or the change rate of the frequency orabundance of each barcode, will be used as a proxy for the fitness andactivity of the corresponding enzyme. Through the DNA sequence-barcode database generated through step 3), the identity of the correspondingenzyme sequence will be obtained for each DNA barcode. This way, theactivity for each enzyme variant is obtained. With this workflow, the enzyme sequence-activity dataset will have important desirable features: A. A full-length sequence, which allows characterization of the effects of theinteractions of different mutations at different positions of the enzyme sequence, termed epistasis; B. High sequence accuracy through barcode-based consensus generation;C. Quantitative activity labelling of each individual enzyme variant;D. Given the available technologies, a capacity of 105 – 107 data points duringeach run. The combination of these features is highly valuable for machine learning-assisted directed evolution. A method for estimating fitness of enzyme variantsOne aspect of the present invention relates to a method for estimating fitness ofenzyme variants, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; 82843PC01 18 I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant (such as in the entire population of enzyme variants); andg) optionally estimating the enzyme fitness of the specific enzyme variantfrom the determined frequency information.Another aspect of the invention relates to a method for optimizing a nucleotidesequence encoding an enzyme variant, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; 82843PC01 19 I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant; and g) obtaining the optimized nucleotide sequence encoding an enzyme variantfrom the determined frequency information.In an embodiment, the method further comprises a step between step d) and e)of subjecting the isolated DNA variant library grown under condition II.) to long-read sequencing.The long-read sequencing may be performed of post-selection library afterselection, in addition to the long-read sequencing of pre-selection library. Whensequencing coverage is not sufficient compared to the library size, the long-readsequencing of post-selection library enriches the reads for those with highest activities, thereby enriching the data with high activity, which contains more information than those with less activities. 82843PC01 20 As shown in the Examples of the present invention, studies have been conducted, documenting that the method of the invention indeed works as claimed. Step A In an embodiment of the invention, the identification of the enzyme variant fitness is for determining the fitness of a portion of enzyme variants in the DNA variant library, preferably a portion of the variants with the highest activity.In an embodiment, the identification of the optimized nucleotide sequenceencoding an enzyme variant comprises determining the fitness of a portion ofenzyme variants in the DNA variant library, preferably a portion of the variants with the highest activity. As seen throughout the Examples, the optimized or improved enzyme variants are selected based on the fitness of the enzyme variants, preferably variants with the highest fitness / activity. Example 12 shows validation of selected enzyme variants with predicted improved fitness and activity. In an embodiment of the invention, the enzyme fitness is selected from the group consisting of enzyme activity, expression level, turnover number such as unimolecular rate constant (Kcat), Michaelis constant (KM), enzyme solubility, enzyme stability, substrate specificity, product specificity, thermotolerance, and combinations thereof. In an embodiment of the invention, the DNA variant library is prepared by the steps: a) providing at least one DNA sequence encoding an enzyme variant;b) linking one or more unique DNA barcodes on the at least one DNAsequence encoding an enzyme variant and obtaining a barcode-DNA library; and c) assembling the barcode-DNA library and obtaining a DNA variant library.As seen in Example 2, the library generation and barcoding are performed successfully as outlined above. One DNA sequence may be linked to more than one DNA barcode. 82843PC01 21 In an embodiment of the invention, the at least one DNA sequence of step a) is generated by PCR, preferably error-prone PCR based on given templates, or assembly of DNA fragments that differ from each other in specific region or regions. As seen in Example 2, the DNA library is generated via PCR successfully. However, for the present invention, the method for generating the DNA library, byintroducing nucleotide mutational changes in a template DNA sequence encodinga template enzyme, is not limited to PCR. In an embodiment of the invention, the assembly of the barcode-DNA library is by Golden Gate assembly. As seen in Example 2, the DNA library is assembled into a vector by Golden Gate assembly, however for the present invention, the method for assembling the DNA library into a vector is not limited to Golden Gate assembly, hence the DNA library may be assembled by various DNA assembly of PCR methods. In an embodiment of the invention, the unique DNA barcode is a random nucleotide sequence. DNA barcodes serve three purposes: 1) for generating accurate full-length DNA sequence reading from long-read sequencing based on consensus generation forthe DNA reads with the same barcodes; 2) for profiling the frequency change ofDNA variants during growth-based selection using short-read sequencing; and 3) for associating the frequency information obtained using short-read sequencing tothe full-length DNA variant sequence information obtained using long-readsequencing. In an embodiment of the invention, the unique DNA barcode has a length in therange 10-300 DNA nucleotides, preferably 15-100 DNA nucleotides, mostpreferably 20-30 DNA nucleotides. The length of the unique DNA barcode should have a sufficient length for the purpose of identifying the associated DNA sequence of an enzyme variant as well as determining the frequency change of the barcode of the enzyme variants. 82843PC01 22 In an embodiment of the invention, the DNA variant is linked to the unique DNA barcode via a covalent bond, such as an extension of the DNA variant coding for the enzyme variant, such as a phosphodiester bond. The unique DNA barcode may be linked to the DNA variant as an extension of the DNA variant, hence the DNA barcode may be attached to the DNA variant via a phosphodiester bond. The DNA variant may have more than one DNA barcode linked. In an embodiment of the invention, the DNA library is a vector library, such as a plasmid library. The DNA library is assembled into a vector for obtaining a vector library, which isexpressed in a host cell. As seen in Example 2, the barcoded DNA library isassembled onto a cloning vector for expression in E. coli. Alternatively, the DNAvariants can be assembled with homology arms for integration onto the genomeof the host cell. The type of vector is not limited to a specific vector and thegeneral way of assembling the DNA variants is not limited to a specific way for thepresent invention.In an embodiment, the DNA variant library comprises at least 100.000 DNAvariants, such as at least 300.000 DNA variants, preferably at least 800.000 DNAvariants, more preferably at least 1 million DNA variants.In an embodiment, the DNA variant library comprises 100.000-5 million DNAvariants, such as 300.000-5 million DNA variants, such as 500.000-4 million DNAvariants, preferably 800.000-4 million DNA variants, more preferably 1-3 millionDNA variants, most preferably 3 million DNA variants. As seen in Example 9, the DNA variant library comprises the above mentioned size.In an embodiment, the enzyme is an enzyme for which the enzyme activityand / or enzyme fitness correlates with growth of the host cell, such as growthduring selection pressure.The enzyme for the determination of the fitness of the enzyme or for optimizing anucleotide sequence encoding an enzyme variant according to the presentinvention may be any enzyme for which the enzyme activity and enzyme fitness 82843PC01 23correlates with growth of the host cell during growth-based selection pressurefavoring enzyme variants with increased activity. In an embodiment of the invention, the enzyme is selected from the groupconsisting of oxidoreductases, transferases, hydrolase, lyases, isomerases,ligases, translocases, or a plurality thereof.As shown in the Examples, the enzyme may be SAM-dependentmethyltransferase.In an embodiment, the enzyme is N-acetylserotonin O-methyltransferase (ASMT).As seen in the Examples, the enzyme is ASMT. Step B In an embodiment of the invention, the host cell is a strain selected from the group consisting of a bacterial cell, a yeast cell, and a eukaryotic cell. The host cell may be any cell, which can express the DNA library and be subjected to selection pressure. The appropriate host cell may be chosen based on the specific enzyme to be expressed. For the present invention, the host cell is not limited to a specific type. In an embodiment of the invention, the host cell is a bacterial cell, preferably E. coli.As shown in Example 2, the host cell may be E. coli, especially for the expressionof SAM-dependent methyltransferase as presented in the Examples.In an embodiment of the invention, the selection pressure dependency is achievedby genetic modification of the host genome, represented preferably by the following table: Selection pressure Rationale ExampleSensor-based approach The activity of the target Andon, James S., enzyme leads to ByungUk Lee, and Tina increased concentration Wang. "Enzyme directed of a molecule, which is evolution using sensed by another genetically encodable 82843PC01 24 molecule (for example, a biosensors." Organic & transcriptional regulator) Biomolecular Chemistry that controls the 20.30 (2022): 5891- concentration of a third 5906. molecule (for example, product of an essential gene) Metabolic coupling The activity of the targetLuo, Hao, et al. enzyme leads to "Coupling S- increased concentration adenosylmethionine– of a molecule which is a dependent methylation building block for cell to growth: Design and growth. uses." Plos biology 17.3 (2019): e2007050. Substrate stress relief The activity of the targetZhan, Chunjun, et al. enzyme leads to "Improved polyketide decreased concentration production in C. of a molecule which glutamicum by inhibits cell growth. preventing propionate- induced growth inhibition." Nature metabolism 5.7 (2023): 1127-1140. The selection pressure is dependent on the enzyme’s catalytic activity for converting substrates to products. Depending on the enzyme to be expressed in the host cell, the selection pressure is chosen accordingly to favor the enzyme variants with increased activity. In an embodiment, the host cell is grown under selection pressure at an optical density (OD) of 0.01-0.5, such as 0.3-0.5, preferably 0.05-0.5.As seen in Example 9, the host cells are grown at the mentioned optical density. 82843PC01 25In a further embodiment, the host cell is grown in a colorless media. In the non-selective and selective media, coloration of the media may be avoided, since colour of the media can result in interference with OD reading. Step D In an embodiment of the invention, the long-read sequencing is sequencing of at least 601 nucleotides, such as at least 1000 nucleotides, such as at least 5000 nucleotides, such as at least 10,000 nucleotides, such as at least 20,000 nucleotides. In an embodiment of the invention, the long-read sequencing is selected from the group consisting of third generation sequencing, such as Oxford Nanopore sequencing, PacBio sequencing, Illumina Complete Long Read sequencing, TELL- Seq. Long-read sequencing enables sequencing of the DNA barcode and the DNA variant encoding an enzyme variant, hence the whole barcode-DNA variant issequenced for obtaining the sequence of the DNA barcode for identifying theassociated enzyme variant. Step E In an embodiment of the invention, the short-read sequencing is sequencing of nucleotide sequences in the range 10-600 nucleotides, preferably in the range 30- 500, most preferably in the range 50-300 nucleotides. In an embodiment of the invention, the short-read sequencing is selected from the group consisting of deep sequencing, Next-Generation Sequencing (NGS), such as Illumina sequencing. Short-read sequencing is a high-accuracy sequencing method, and for the presentinvention, short-read sequencing is performed on the barcode region for obtaininginformation relating to the abundance or frequency of the barcode present during selection, and optionally at different time points. Hence, the enzyme activity can be determined based on the number of barcodes for each enzyme variant. Further, the rate of change of abundance of DNA variants can be determined from sampling host cells expressing the DNA library under selection pressure at 82843PC01 26 different time points and subsequently isolating the DNA for short-read sequencing (See Example 5). In an embodiment of the invention, the frequency information of the specific DNA variant coding for a specific enzyme variant is a reflection of the enzyme fitness of said specific enzyme variant. The frequency information from short-read sequencing of the DNA variants, or the rate of change of the variants that can be calculated based on the frequencyinformation, is used as a proxy for the fitness of the corresponding enzyme. Theenzyme fitness includes expression level or ease of expression in a given host cell,turnover number (unimolecular rate constant, Kcat), Michaelis constant (KM),solubility, and stability, substrate specificity, product specificity, thermotolerance. In an embodiment of the invention, the rate of change of the specific DNA variant coding for a specific enzyme variant is determined from the change of the amountof the specific DNA variant over a specific period of time. The rate of change canbe obtained based on the barcode frequency information and total cell abundance over time. Sampling of the host cells under selection pressure, and subsequent short-readsequencing of the barcode region to characterize the change of variantabundance / frequency, will allow the calculation of the growth rate of the cell population harboring a specific enzyme variant. This growth rate will be used as a proxy for the activity of the cognate enzyme variant in the cell population. Step F In an embodiment of the invention, in step f) the comparison of the first set of information from step d) to the second set of information from step e), is by correlating DNA barcodes from the second set of information to the enzyme variants of the first set of information via the enzyme variant specific DNA barcode, thereby obtaining a prediction of enzyme fitness of the enzyme variantsand / or an optimized nucleotide sequence encoding an enzyme variant via thefrequency of the DNA barcodes from the second set of information or the rate of change of the specific DNA variant coding for a specific enzyme variant. Via the first set of information, the identity of the cognate enzyme variant is known for each barcode. Hence, the second set of information, which comprises 82843PC01 27 the sequence of the barcodes and frequency of barcodes only, is compared to the first set of information, wherein each barcode of the second set of information is matched with each barcode of the first set of information, and thereby determining the corresponding enzyme variants for the barcodes of the second set of information. As the frequency of each barcode is known from the second set of information, the frequency of corresponding enzymes variants is obtained. Step G In an embodiment of the invention, the enzyme fitness of the specific enzymevariant and / or an optimized nucleotide sequence encoding an enzyme variant isreflected by a growth rate of the host cell. Growth rate is obtained by characterizing the change of variant abundance / frequency, which will allow the calculation of the growth rate of the cell population harboring a specific enzyme variant. This growth rate will be used as a proxy for the activity of the cognate enzyme variant in the cell population. In an embodiment of the invention, the growth rate is determined from growing the host cell from step b(I.) under selection pressure favoring enzyme variants with increased fitness or activity. As the selection pressure favors enzyme variants with increased fitness or activity, it is to be expected that host cells expressing the enzyme variants with highest fitness or activity will exhibit a higher growth rate. As mentioned above, the growth rate can be determined based on the rate of change of variant abundance.In an embodiment, the DNA nucleotide sequence coding for enzyme variantscomprises a promoter such as J23100 (SEQ ID NO. 15) or J23110 (SEQ ID NO.16). As seen in Examples 3 and 9, the promoters used are J23100 or J23110. Use of barcoded DNA sequence for predicting enzyme fitness A second aspect of the present invention relates to use of at least one barcoded DNA sequence encoding an enzyme variant for predicting enzyme fitness of saidenzyme variant and / or for optimizing said DNA sequence encoding said enzymevariant, wherein the at least one barcoded DNA sequence is subjected to long- read sequencing and short-read sequencing. 82843PC01 28As can be seen in the Examples, the DNA library encoding enzyme variants isbarcoded for later long-read and short-read sequencing. The long-read sequencing of the barcoded DNA results in a first set of information or firstdatabase, which is the association of the DNA barcode sequence and itscorresponding enzyme variant. The short-read sequencing of the DNA barcodeallows for quantification of the DNA barcode at specific time points and therebyobtaining a second set of information comprising the frequency of each barcode orchange rate of variant abundance. The comparison of the first set of informationand second set of information allows for assessing the activity or fitness of eachenzyme variant. Thus, the assessment of activity or fitness of each enzymevariant allows for optimizing DNA sequences encoding enzyme variants.In an embodiment of the invention, the at least one barcoded DNA sequence for short-read sequencing has been obtained from a host cell subjected to selection pressure favoring enzyme variants with increased fitness. The host cells being subjected to a selection pressure favoring enzyme variantwith increased fitness or activity will exhibit a higher growth rate when expressingenzyme variants with increased activity. When sampling the cells for isolation ofthe DNA for short-read sequencing, the abundance of each of the enzyme variantcan be determined. Polypeptide As shown in Example 8, 11 and 12, the inventing team has identified a number ofoptimized variants of atASMT having a high fitness score for the conversion of PCAto vanillic acid in E. coli.Thus, in an aspect, the invention relates to a polypeptide according to SEQ ID NO.3 comprising one or more of the following mutations P72S, K132R, D147E, I103N, P286L, R77C, G105S, K187R, M215K, P231S, E46D, S84I, and / or P302L. In an embodiment, the polypeptide comprises one or more of the following combinations of mutations: -P72S, K132R, and D147E,- P72S, I103N, and P286L,- R77C, G105S, and K187R, 82843PC01 29 -R77C, M215K, and P231S, and / or- E46D, S84I, and P302L.In yet an embodiment, the polypeptide is a polypeptide according to any of SEQ ID NOs 10-14. In another embodiment, the polypeptide is an isolated polypeptide. Nucleic acids, vectors and host cells An aspect of the invention relates to nucleic acid molecules encoding the polypeptides according to the invention. Yet an aspect relates to a vector, such as a plasmid comprising the nucleic acid molecule according to the invention. Yet another aspect relates to a host cell comprising the vector according to the invention. Software / AIA further aspect of the present invention relates to a processor systemprogrammed to operate according to a machine learning (ML) algorithm forestimating fitness of enzyme variants and / or for optimizing a nucleotide sequenceencoding an enzyme variant, the machine learning (ML) algorithm being trained,and / or being trainable on data comprising a correlation of a specific nucleotidesequence encoding an enzyme variant to frequency information.Another aspect of the present invention relates to a processor system programmed to operate according to a machine learning (ML) algorithm forestimating fitness of enzyme variants and / or for optimizing a nucleotide sequenceencoding an enzyme variant, the machine learning (ML) algorithm being trained, and / or being trainable, on data obtained by a method according to the first orsecond aspect of the invention.Yet a further aspect of the invention relates to use of a machine learning (ML) algorithm trained on data comprising a correlation of a specific nucleotidesequence encoding an enzyme variant to frequency information to predict fitness 82843PC01 30 for any enzyme sequence, or to guide the generation of enzyme sequences with given fitness properties. In yet another aspect, the invention relates to use of a machine learning (ML) algorithm trained on data obtained by a method according to the first aspect ofthe invention to predict fitness for any enzyme sequence, or to guide thegeneration of enzyme sequences with given fitness properties. Another aspect of the invention relates to use of a machine learning (ML) algorithm trained on data comprising a correlation of a specific nucleotidesequence encoding an enzyme variant to frequency information for optimizing anucleotide sequence encoding an enzyme variant, or to guide the generation of animproved nucleotide sequence encoding an enzyme variant with given fitnessproperties. A further aspect of the present invention relates to use of a machine learning (ML) algorithm trained on data obtained by a method according to the second aspect for optimizing a nucleotide sequence encoding an enzyme variant, or to guide thegeneration of an improved nucleotide sequence encoding an enzyme variant withgiven fitness properties.An aspect of the present invention relates to a system suitable for executing analgorithm (such as machine learning (ML) algorithm) for estimating fitness ofenzyme variants and / or for optimizing a nucleotide sequence encoding an enzymevariant, the (machine learning) system being trained, and / or being trainable, ondata provided according to the first or second aspect, the system being arrangedfor:- receiving a first set of information (1SI), such as a first database, said first ofinformation being obtained, or being obtainable, from d) subjecting the isolated DNA variant library grown under condition I.) to long-read sequencing, thereby obtaining said first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, 82843PC01 31- receiving a second set of information (2SI), such as a second database, saidsecond of information being obtained, or being obtainable, from e) subjecting the isolated DNA variant library grown under condition II.) to short-read deep sequencing, thereby obtaining said second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; wherein the system comprises: -a first comparison module (1CM) for comparing said first set of information (1SI) to said second set of information (2SI), thereby correlating specific enzyme variants to the frequency information, such as frequency of the variant; and- an optional second comparison module (2CM) for correlating determinedfrequency information to the enzyme fitness of the specific enzyme variant.In an embodiment, the algorithm is MoCHI. In Example 11, a series of variantenzyme sequences were generated using the MoCHI. However, in principle anymachine learning-based models of protein fitness can be used.Advantageously, the invention may also relate to a method for training a machinelearning (ML) system for estimating fitness of enzyme variants and / or foroptimizing a nucleotide sequence encoding an enzyme variant, such as accordingto the previous aspects. Thus, yet an aspect of the invention relates to a methodcomprises the steps of:-receiving training data comprising a first set of information (1SI), such as a firstdatabase, and a second set of information (2SI), such as a second database,-training the system for estimating fitness of enzyme variants using said training data, and -validating the system using correlated specific enzyme variants to the frequencyinformation, such as frequency of the variant.In an embodiment, the ML system comprises:^ a ML model configured to assign a score to a nucleotide sequence encodingan enzyme variant, 82843PC01 32 ^a comparison component configured to compare the score with a thresholdfor predicting a growth rate of a host cell expressing said nucleotidesequence encoding the enzyme variant, and^ a featuring component configured to receive an output from thecomparison component, wherein the output indicates a predicted fitness, such as activity, of said enzyme variant encoded by said nucleotide sequence. In an embodiment, if the score is above the threshold, it is indicative of an improved growth rate of a host cell expressing said enzyme variant compared to a growth rate of a host cell expressing a corresponding wild-type enzyme. In yet an embodiment, if the score is below the threshold, it is indicative of a decreased growth rate of a host cell expressing said enzyme variant compared to a growth rate of a host cell expressing a corresponding wild-type enzyme. In a preferred embodiment, the threshold is 0.Preferably the system and / or algorithm and / or method is implemented on acomputer, thus being computer-implemented. A further aspect relates to a computer-readable storage medium comprising datacomprising a correlation of a specific nucleotide sequence encoding an enzymevariant to enzyme fitness, such as enzyme activity. Another aspect relates to a computer-readable storage medium comprising data obtained by a method according to the invention. An aspect of the invention relates to a neural network model obtained orobtainable by the method for training a machine learning (ML) system accordingto a previous aspect. Below is given a list of some no-limiting types of algorithms that are particularlysuited for machine learning (ML) system and / or training of a (ML) system usingenzyme data: 82843PC01 33 1. Deep Learning Algorithms: Deep learning is a subset of machine learning where artificial neural networks, algorithms inspired by the human brain, learn from large amounts of data. Deep learning algorithms are capable of learning to represent the world as a nested hierarchy of concepts, with each concept defined in relation to simpler concepts, and more abstract representations computed interms of less abstract ones. The skilled reader is referred to for exampleUniversity of Illinois at Urbana-Champaign; "AI predicts enzyme function better than leading tools." ScienceDaily. ScienceDaily, 30 March 2023. <www.sciencedaily.com / releases / 2023 / 03 / 230330172121.htm>. 2. Contrastive Learning: This is a type of unsupervised learning approach that trains models to learn similar features from similar data points and differentfeatures from different data points. An AI tool named ‘CLEAN’ was recentlyreported to use this algorithm to predict enzyme function, cf. Gupta, R.,Srivastava, D., Sahu, M. et al. Artificial intelligence to deep learning: machineintelligence approach for drug discovery. Mol Divers 25, 1315–1360 (2021) formore details. 3. Artificial Neural Networks (ANNs): ANNs are computing systems vaguely inspired by the biological neural networks that constitute animal brains. An ANN is based on a collection of connected units or nodes called artificial neurons, which loosely model the neurons in a biological brain. 4. Support Vector Machines (SVMs): SVMs are supervised learning models with associated learning algorithms that analyze data for classification and regression analysis. 5. Generative Adversarial Networks (GANs): GANs are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks contesting with each other in a zero-sum game framework. These algorithms can be used individually or in combination, depending on thespecific requirements of the type of enzyme data. The invention according to this 82843PC01 34aspect can be implemented by means of hardware, software, firmware or anycombination of these. The invention or some of the features thereof can also be implemented as software running on one or more data processors and / or digital signal processors. The individual elements of an embodiment of the invention may be physically, functionally and logically implemented in any suitable way such as in a single unit, in a plurality of units or as part of separate functional units. The invention may be implemented in a single unit, or be both physically and functionally distributed between different units and processors. It should be noted that embodiments and features described in the context of one of the aspects of the present invention also apply to the other aspects of theinvention. Thus, for example the embodiments / claims relating to the method ofthe invention also apply to the embodiments relating to the use-embodiments / claims. Hence, individual features mentioned in different claims,may possibly be advantageously combined, and the mentioning of these features in different claims does not exclude that a combination of features is not possible and advantageous. Although the present invention has been described in connection with the specified embodiments, it should not be construed as being in any way limited to the presented examples. The scope of the present invention is to be interpreted in the light of the accompanying claim set. In the context of the claims, the terms “comprising” or “comprises” do not exclude other possible elements or steps. Also, the mentioning of references such as “a” or “an” etc. should not be construed as excluding a plurality. The use of reference signs in the claims with respect to elements indicated in the figures shall also not be construed as limiting the scope of the invention. All patent and non-patent references cited in the present application, are hereby incorporated by reference in their entirety. The invention will now be described in further details in the following non-limiting examples. 82843PC01 35 ExamplesExample 1 – Coupling enzyme activity to cell growthAim of study The goal of this step is to link the growth rate of a cell to the activity of the target enzyme, such that the growth rate of a cell is positively or negatively correlated to the activity of the enzyme variant that the cell contains. This enables the use of growth rate as a metric for the enzyme activity. Materials and methods In this example, the selection strain is constructed based on WO2018037098A1, where the activity of SAM-dependent methyltransferases is coupled to cell growth by providing precursors for cysteine synthesis. The specific strain used in thisexample is E. coli BW25113 with FolE (T198I) YnbB (V197A) ΔtnaA ΔcysE ΔcfaompW-Δ(26 bp intergenic region)>yciE::cys3-cys4. Results and conclusionThe SAM-dependent methyltransferase used here is Arabidopsis thaliana N-acetylserotonin O-methyltransferase (atASMT) (SEQ ID NO: 3), and the targetcatalytic activity is the conversion of protocatechuic acid (PCA) into vanillic acid. (EC 2.1.1.).Example 2 – Library generation and barcodingAim of study This step creates a library of DNA variants with unique DNA barcodes attached. The DNA variants encode enzyme variants that are expected to have different fitness levels. The barcodes will serve two purposes in later steps: 1) for accurate full-length sequence reading from long-read sequencing; 2) for using the short barcodes to profile the frequency change of variants during growth-based selection. (Figure 1) Materials and methods Error prone PCR Error prone PCR was performed on the template DNA using GeneMorph II Random Mutagenesis Kit according to manufacturer’s instructions, with 1 ng of template 82843PC01 36DNA (SEQ ID NO: 2) and 30 cycles. The primers used may be found in SEQ IDNO: 4 and 5. The forward primer additionally contains sequences used for theassembly of the variants onto the cloning vector in a later step. Several reactions were run in parallel to increase the amount of products. The PCR products were gel-purified to remove residual primers. Barcoding with one-cycle PCR A one-cycle PCR step was performed on the error prone PCR products to attachunique barcodes to the DNA variants. Only one primer is used (SEQ ID NO: 6),which contains: 1) Sequence corresponding to the reverse primer for the DNA variants. 2) A region with 25 nt of fully random sequences, which serves as barcodes. The possibility space of this region is 425, or ~1015, which ensures a high chance that the barcode attached to each variant will be different from all other attached barcodes in a library of less than 1010variants. 3) Sequences used for assembly of the variants with the cloning vector. The reaction is performed in one cycle so that each sequence will only have onecorresponding barcode. All of purified “error prone PCR” products are used in thereaction (separated into several reactions as needed). 2X Q5 hot-start master mix is used. The reaction is done with 150s of 98 °C for denaturation, 150 seconds at the annealing temperature (68 °C in this case), and 8 min of 72 °C for extension. The prolonged time for each step is to maximize the extent of the reaction for all molecules. The one-cycle PCR product is purified through gel purification, which also serves the purpose of removing the leftover single strand in the process.Assembly of the barcoded library with cloning vector (SEQ ID NO: 1) to obtainthe plasmid library. A control library comprising non-mutated template DNA mayalso be constructed (SEQ ID NO: 7). The cloning vector, pSD221 (Rep101 ori, cat marker for chloramphenicol resistance), was constructed with a linker for Golden Gate assembly. The linker has two BsaI digestion sites on the two ends of the linker region. The barcoded library is ligated to pSD221 via Golden Gate assembly in several parallel reactions.Afterwards, the reactions were pooled together and purified with AMPure XP beads 82843PC01 37 (Beckman Coulter) according to manufacturer’s PCR purification protocol with 0.9bead-to-reaction volume. The entire library is eluted to ~25 ^L of water.Results and conclusion The DNA library is obtained through error prone PCR based on an existing enzyme template. The DNA library can alternatively be generated through other means, e.g., DNA synthesis with heterogenous components. All the PCR products are tagged with unique barcodes through a one-cycle PCR process. The barcoded library is subsequently assembled onto the vector for expression in E. coli. In other cases, the library could be assembled with homology arms for integration onto the chromosome.Example 3 – Construction and amplification of in vivo libraryAim of studyIn this step, an in vivo library is constructed by transforming the plasmid library(containing the DNA library for enzyme variants) into the selection strain. The invivo library is further amplified by non-selective growth for subsequent uses.(Figure 1) Materials and methods Transformation of the plasmid library Electrocompetent cells for the selection strain were prepared according to commonly used protocols. Several electroporations were done in parallel to transform the full amount of the plasmid library. Immediately after electroporation, the cells are added to 2x YT media (~1:10 volume ratio) without antibiotics and incubated for 1 hour at 37 °C for recovery. After recovery, the transformant library is stocked in glycerol stock to -80°C. A small volume of the library is serially diluted and plated onto Chloramphenicol plates to estimate the library size per volume of the stocked library.In vivo library amplificationGiven the technologies in 2023, the library size that can be characterized by this invention is one to several millions per run, limited by long-read sequencing capacity. Supposing a limit of 1 million, if the full library size exceeds this limit, only a portion of the full library, corresponding to 1 million variants, should be 82843PC01 38 used for subsequent steps; if the full library size is below 1 million, then the full library can be used for subsequent steps. An amount of the library based on the consideration above is grown in 2x YT media with chloramphenicol selection overnight for amplification so that aliquots of the same library can be used in subsequent steps. The amount of media should allow sufficient amplification of the library (>100x to ensure that aliquots will be representative of the full library) and sufficient number of aliquots as seeding cultures for subsequent steps. The amplified library is aliquoted and stocked as glycerol stocks at -80°C. Results and conclusionWith these steps, the barcoded enzyme library now exists as in vivo library in theselection strain, and aliquots of the same library can be used for the next steps.Example 4 – Long read sequencingAim of study This step uses long-read sequencing technologies (exemplified with Oxford Nanopore, PacBio, etc) to generate data that associates the variant sequences with their associated barcodes. Within a certain library size limit and sufficient read depth, most variant-barcode fragments will be read multiple times, which allows the generation of high-accuracy consensus sequences for the variants based on the clustering of the reads with the same barcode sequences. (Figure 2) Materials and methods Plasmid extraction and digestionOne aliquot of the in vivo library was grown in rich media (2x YT) withchloramphenicol in large volume at 37 °C with shaking. The large volume ensures sufficient plasmid output for Nanopore sample preparation. The plasmid wasextracted and purified through maxiprep (ZymoPURE™ II Plasmid Maxiprep Kit)according to manufacturer’s instructions. The extracted plasmid was digested with SalI and SpeI (both from NEB), two unique restriction sites on the plasmid sequence. SalI is 71 bp upstream of the expression promoter, while SpeI is 46 bp downstream of the barcode region. Thedigestion reaction was run on an agarose gel, and the fragment with lengthcorresponding to the target enzyme gene-barcode cassette was purified. 82843PC01 39 Nanopore sequencing This step was done with the R10.4.1 flow cell on a GridION instrument, with sample preparation done using Ligation Sequencing Kit V14 according to the protocol “Ligation sequencing amplicons V14 (SQK-LSK114)” from Oxford Nanopore. Additional materials and instruments were used according to the protocol: NEBNext® Companion Module for Oxford Nanopore Technologies® Ligation Sequencing, Quibit fluorometer and the associated consumables, AMPure XP beads and magnetic rack for 1.5 ml Eppendorf tubes, 1.5 ml Eppendorf DNA LoBind tubes. After sample preparation, the sample was loaded onto a GridION and run for 39 hours with 9.48 M reads generated (longer run time can be done for higher read output). 100% of the reads were called, and 15.76 Gb bases were called (based on min Q score of 7, though the overall Q score is generally above 12 throughout the run). Results and conclusion The plasmid library was extracted after growing an aliquot of the library with antibiotic selection for the plasmid, but without selection for the enzyme variants. The lack of selection in the growth condition avoids bias for different enzyme variants, which maximized the coverage of the library for sequence-barcode association. The region including the target enzyme expression cassette and barcode was isolated by restriction digestion and sequenced through ampliconsequencing with Nanopore sequencing. The sequenced expression cassetteincludes the promoter and RBS region, which allows the effect of potential mutation in the promoter and RBS on expression level during the enzyme librarypreparation process to be accounted for.The long-read Nanopore sequencing generates data that associates the DNA variant sequences with their associated DNA barcodes.Example 5 – Growth-based selection and samplingAim of studyIn this step, the selection strain comprising the DNA library is grown withselection bias toward higher target enzyme activity. Cells will proliferate with a growth rate positively correlated with the activity of the target enzyme variant. 82843PC01 40 Sampling during this process, and the subsequent deep sequencing to characterize the change of variant abundance / frequency, will allow the calculation of the growth rate of the cell population harboring a specific enzyme variant. This growth rate will be used as a proxy for the activity of the cognate enzyme variant in this cell population. (Figure 2) Materials and methods Prepare media Prepare non-selective and selective minimal media. The non-selective media consists of 1X M9 salts, 0.1 mM CaCl2, 2 mM MgSO4, 0.2% glucose, 1X trace minerals solution, 1X Wolfe’s vitamin solution, 25 µg / mL chloramphenicol, 50 mg / l cysteine. The selective media differs from the non-selective media by containing no cysteine, but 1 mM protocetachuic acid and 1 mM methionine. Growth-based selection and sampling An aliquot of the library is grown overnight in non-selective media at 37 °C with shaking. The next day, the culture is washed twice with the selection media, and inoculated into the selection media at an initial OD below 0.1, at 30 °C. Another portion of the overnight culture is spun down, washed twice with water, and the cell pellet is stored at -20 °C, which serves as sample 0 and reflects the initial frequency distribution of the variants. During growth, three other samples were taken at OD of 0.12, 0.22, 0.39. The cells were washed with water twice, and stored at -20 °C, serving as samples 1, 2, 3. Results and conclusion An aliquot of the library is grown overnight non-selectively and seeded into selection media after washing. Samples were taken at time 0, and during the exponential growth phase. The growth rate of cells with a specific variant isassumed to remain constant in the exponential phase and correlate positively withthe activity of the enzyme variant. The distribution of variants in each sample can be captured in the next step through deep sequencing. The OD information will allow the calculation of the expected total cell number given constant volume. Combined, they will allow thecalculation of growth rate enabled by each variant. 82843PC01 41Example 6 – Deep sequencingAim of study This step prepares the cell pellet samples for deep sequencing by amplifying the barcode region of the variants cassettes, and deep sequencing of the barcode region is performed to obtain the frequency information of the barcodes for each sample. (Figure 2) Materials and methods Sample preparation for deep sequencing The cell pellets are thawed, and water is added to the cell pellets to a final concentration around 0.3 OD per 10 ^L. Thermolysis is performed to lyse the cells by heating the cell solution at 95 °C for 10 min. The resulting mixture is spun down, and the supernatant is kept as PCR template for barcode amplification.The forward primer (SEQ ID NO: 8) is directly upstream of the barcode region.The reverse primer (SEQ ID NO: 9) is downstream of the barcode region for atotal length of 352 bp PCR product. The PCR reaction (50 ^L) is set up as follows:25 ^L Kapa HiFi Hotstart ReadyMix, 1.5 ^L of 10 ^M forward primer, 1.5 ^L of 10^M reverse primers, 5 ^L lysis supernatant, 1.5 ^L DMSO, 15.5 ^L water. Thereactions were run 22 cycles according to manufacturer’s instructions. The PCR products were purified with AMPure XP beads with 0.9 bead-to-reaction volume ration. The purified products were quality checked with Qubit. Deep sequencing In this example Deep sequencing is exemplified with Illumina sequencing.>1 ^g was sent to GENEWIZ Germany GmbH (Azenta Life Sciences) for Illuminasequencing. The sequencing platform was Illumina NovaSeq 2x150 bp sequencing, with estimated data output of ~29M paired-end reads (incl. 30% PhiX spike-in) per sample. Results and conclusion Through deep sequencing of the barcode region, the frequency change of the barcodes for samples over the course of growth is captured. 82843PC01 42Example 7 – Bioinformatics processing to determine the sequence-activity relationship Aim of study The long-read sequencing data and deep sequencing data are processed in this step to generate the full sequence-activity relationship. Specifically, it includes: 1) the processing of deep sequencing data to obtain growth rate (proxy for enzyme activity) for each barcode; 2) the processing of long-read sequencing data to generate consensus sequence for each variant by clustering the reads with the same barcode sequence, assuming the same barcode corresponds to the same variant sequence. The consensus sequence alleviates the high error rate issue ofcurrent long-read sequencing methods and allows high accuracy of the variantsequence data. The barcode is used to link the full-length consensus sequence to the corresponding growth rate. (Figure 2) Materials and methods Barcode-based clustering of long-read sequencing dataAll Nanopore files are consolidated into a single FASTQ file. The barcode of eachsequence is identified by 8 bp upstream and downstream of the barcode region. All barcodes are collected into a single file. Given the bidirectional nature of Nanopore reads, barcodes are extracted for both the forward and reverse complements. Length validation of 25 bases is again performed for each barcode.All barcodes extracted from the reverse complement are changed to conform tothe top strand and the 5’ to 3’ end direction, meaning that the reversecomplement of these barcodes was taken. The output is two files, one containingall sequences and the other containing all barcodes from the Nanopore data set is generated. Sequences with the same barcodes (accounting for reverse complements) are grouped together as a cluster. Each corresponding barcode is also saved in a separate file for technical purposes. Barcode count from deep sequencing data For each raw FASTQ file, barcodes are counted by parsing all sequences in the file. The barcode is identified by starting 31 nt upstream of the barcode and 4 nt downstream of the barcode. A length validation is performed to ensure that the 82843PC01 43 extracted sequence is precisely 25 nt. The output is the count of occurrences of each distinct barcode for each Illumina sample. Each experiment has counts extracted at two exponential growth time points. These are combined into one csv file with the counts at time point 1 and 2 for each barcode. Consensus sequence generation for barcodes The consensus sequence for each barcode is generated with medaka using r1041_e82_400bps_sup_v4.2.0 (https: / / github.com / nanoporetech / medaka). If compute is limited, consensus sequence generation is done only for barcodes of interest based on the Illumina sequence processing results. Amino acid sequence generation and mutation analysis The full variant sequence is segregated into promoter-RBS sequences, and the coding sequences. The coding sequence is translated into amino acid sequence using BioPython. The mutations at the nucleotide and amino acid levels for each consensus sequence is subsequently generated. Growth rate calculation for each variant sequence OD measurements at each time point are combined with barcode counts and the time differential (Δt) to calculate the growth rate (μ) for each variant: x is the amount of OD for the variant at a given time point. Results and conclusion 82843PC01 44 The frequency of each barcode can be obtained from the Illumina data. Combined with the time and OD information (OD provides information about the total cell amount), the growth rate is calculated for each barcode. The Nanopore data is processed by barcode-based clustering and consensus sequence generation. The consensus sequence will have higher accuracy than the raw reads after error correction based on the multiple raw reads. Combining all the information above, the growth rate and full-length sequence information is obtained for individual variants.Example 8 – Training machine learning model with sequence-fitness dataand predicting fitness of sequences not seen in the data Aim of study A machine learning model is trained to predict amino acid sequence fitness using the sequence-fitness data obtained from the procedure of the previous Examples. This model incorporates a pretrained embedding model and a specialized ensemble of regressors that is trained on the embeddings of the sequence-fitness data. Materials and methodsModel Training Process:1) Embedding: Each sequence from the sequence-fitness dataset was processed using the pretrained embedding model 'esm1_t34_670M_UR50S', sourced from a public Github respository [https: / / github.com / facebookresearch / esm#available- models]. 2) Ensemble Training: A diverse ensemble of regression models, including neural networks, Bayesian regression, and decision trees, was trained on a selected subset of data points. Each data point comprises an embedded amino acid sequence paired with its fitness. The code to run the ensemble is from the public github repository [https: / / github.com / fhalab / MLDE]. 3) Selection: The top 5 models in the ensemble, selected based on their generalization error, are retained for inference. 82843PC01 45 Inference Process: 1) Embedding: A set of variant sequences of atASMT are embedded using the same pretrained model. Each variant contains three substitution mutations of the original sequence. These substitution mutations are chosen independently andrandomly from the mutations seen in the training data. A total of 1000 variantsare generated. 2) Fitness Prediction: The prediction of fitness is conducted by the trained ensemble of regression models. These predictions are then averaged to yield a consolidated fitness estimate for each variant. 3) Selecting a useful subset as a library: The variants are ranked by the predicted fitness score. A subset of the highest ranked sequences is then selected, yieldingadvantageous variants of the original atASMT sequence (SEQ ID NO. 3).Results and conclusionThe top-5 most highly ranked variants of the 1000 examined sequences are SEQID NO. 10-14. These sequences are predicted to have the highest fitness levelgiven the experimental conditions, based on the machine learning model trained with the data obtained through the procedure in the preceding examples.The top-5 most highly ranked variants (SEQ ID NO. 10-14) comprise thefollowing mutations on the original atASMT sequence (SEQ ID NO. 3): SEQ ID NO. MutationsSEQ ID NO. 10 P72S, K132R, D147ESEQ ID NO. 11 P72S, I103N, P286LSEQ ID NO. 12 R77C, G105S, K187RSEQ ID NO. 13 R77C, M215K, P231SSEQ ID NO. 14 E46D, S84I, P302LExample 9 – Modified sample collection procedureAim of study This study shows several steps during sample preparation to improve the size, diversity, and quality of the collected data. 82843PC01 46 Materials and methods Library generation and barcoding Compared with Example 2, the template used in this case is the plasmid DNA extracted from the selected culture from Example 5, rather than the WT sequence. This is to increase the diversity of the template, as well as to have more sequences with higher activities, since the template sequences has a higher proportion of templates with improved initial activities compared with WT. Othersteps are done in the same manner as in Example 2.Construction and amplification of in vivo libraryCompared with Example 3, the promoter controlling the expression of the variantswere changed from J23100 (SEQ ID NO. 15) to a weaker J23110 (SEQ ID NO.16) in order to raise the selection ceiling of the selection range. The size of thelibrary used for subsequent steps is chosen as around 3 million, rather than 1 million. This exceeds the capacity of one Nanopore flow cell to obtain close to 10 reads per sequence but will allow the collection of more high-activity sequence data based on a subsequent step. Other steps were performed in the same way as Example 3. Growth-based selection and sampling Compared with Example 5, several changes were made to improve the selection process. In the non-selective and selective media, trace minerals solution and Wolfe’s vitamin solution were not added, to avoid the coloration of the media and interference with OD reading when protocatechuic acid is present. The initial OD of the selection culture was set at 0.1 rather than lower, so as to avoid extended selection time a high enough OD could be reached for sample collection. One dilution was made when the OD reached ~0.35, to a second initial OD of 0.05. Samples were taken at time point 0 hr (OD 0.1, denoted time point 0), 43 hr (OD ~0.24, denoted time point 1), 50 hr (OD ~0.35, denoted time point 2; dilution into 0.05 initial OD was done at this point), 67 hr (OD ~0.27, denoted time point 3), 71 hr (OD ~0.45, denoted time point 4). All were done in triplicates. Other steps were performed in the same manner as in Example 5. 82843PC01 47 Long read sequencing Due to limited data output from single Nanopore runs, three runs were performed on the unselected library to obtain sufficient reads. In addition, plasmids from one post-selection culture were also processed and sequenced in one Nanopore run. This enriches the reads for the top variants, which eventually leads to more data points for variants with higher activities. The resulting data from these runs were combined for processing. Other steps were performed in the same manner as Example 4. Deep sequencing Compared with Example 6, the amplicons were barcoded with custom barcodes according to the TruSeq UDI scheme. Other conditions were done in the same way as seen in Example 6. Results and conclusion Nanopore sequencing data and Illumina sequencing data were obtained for the samples. 20.8 million reads in total were obtained from Nanopore sequencing, including 19.3 million in non-selected library, and 1.5 million in selected library. 20-33 million Illumina reads were obtained for each sample.Example 10 – Optimized bioinformatics processing to determine thesequence-fitness relationship Aim of study Compared with Example 7, in this example, additional procedures were taken to improve the quality and output size of the consensus sequence generated from Nanopore data, and to improve the statistical rigor of the fitness estimation for sequence variants. Materials and methods Mapping Sequence to Growth Rate Using custom scripts, Illumina reads with a mean -q score below 30 were removed. The barcodes were extracted from the reads producing a CSV file with 82843PC01 48 the abundances of each unique barcode across all time points. Subsets of the dataset were generated each containing barcode abundances at time point pairs 0–1, 0–2, 0–3, and 0–4. The DiMSum suite was used to compute the growth rates for each of the dataset subsets with the following parameters –maxSubstitutions 100 – mixedSubstitutions T –indels all –startStage 4 –stopStage 5. The result is a mapping of each barcode to a set of growth rates. ONT reads were basecalled using the Dorado v5.0.0 sup-accuracy basecaller. Reads containing multiple concatenated coding regions were split into multiple reads using a custom script. Filtering based on read quality and read length was performed using Chopper with parameters -q 9 -1800 –maxlength 2000. UMIC-seq was then used to extract the barcodes from the filtered ONT reads. Separately, barcodes extracted from Illumina reads were clustered using mmseqs2 easy-linclust with parameters –min-seq-id 0.88 -c 0.88, and the most abundant barcode in each cluster was collected in a table. Using vsearch – usearch_global with parameters –query_cov 0.84 and –id 0.88, barcodes derived from ONT reads were clustered against the barcodes in the table. For clusters containing seven or more IDs, the reads were collected in a FASTA file. Using a custom pipeline, a consensus sequence was generated for each cluster. minimap2 was used to align the reads in each cluster to the reference sequence, and racon was used to generate a draft assembly. In three additional rounds, minimap2 and racon were used with the generated draft assembly replacing the reference sequence as the template. The final draft assembly was then polishedwith medaka (ONT) using one round of medaka_consensus, producing aconsensus sequence for each cluster. The consensus sequences, along with their corresponding barcodes, were compiled into a CSV file. The growth rate scores were then linked to the consensus sequences using the barcode as the identifier. Nanopore read count requirement evaluation The error rate of the coding region depends in part on the number of reads in the cluster used to produce it. A lower bound for the coding region error rate can be provided based on the read count of the cluster. Self-consistency was estimatedby computing the coding region of a set of 700 clusters with 25 reads or more.Having 25 reads or more in a cluster implies the consensus sequence will be 82843PC01 49 highly accurate. Sub-sampling the clusters, clusters were generated ranging from7 to 15 in size. For each cluster size between 7 and 15, 5 subsets were created.The coding regions computed from the subsets were compared to those computed from the full cluster. The frequency of mismatched coding regions was then estimated for cluster sizes 7 to 15, respectively. Post Processing the Dataset The coding region and promoter region were extracted using a custom script. The consensus sequence was aligned to the reference sequence, and all bases between the first and last base of the promoter and coding regions were extracted, respectively.MoCHI requires the information of the template sequence (wild type) from whichthe library is derived from. Since the template for the construction of the librarywas a mixture of variants, wildtype (WT2) coding sequence in this example was defined as a sequence 6 nucleotide mutations distant from the coding region of the original WT. These mutations were chosen as they are shared by the majority of the variants. An aggregated version of the dataset was created in which, for each variant v, an aggregated growth rate λv ‾ is produced from the set of growth rates. Using the equation below, the aggregated growth rate for a variant v was computed using only the time- step pairs Tv for which the variant has both values. Results and conclusions The Nanopore read count requirement evaluation results show around 3.5% error rate for clusters of size 7, decreasing significantly to 1.9% for clusters of size 10 and 1.3% for clusters of size 15. The resulting dataset contains 571,043 records for variants with at least 7 reads. The distribution of relative enzyme fitness for the aggregated dataset is illustrated 82843PC01 50in Figure 3. Relative enzyme fitness is defined as the growth rate normalized tothe median wild-type (WT2) growth rate, quantifying each variant’s growth as a multiple of this baseline. A non-aggregated version of the dataset was also used, where growth rate derives only from time step pair 0-4. Using a specific time step pair has the advantage that all growth rates have an associated variance score. Time step pair 0-4 was used as it showed the highest correlation for predicted growth rate of variants between the replicates, with an average correlation of 0.917. The dataset contains 464,014 records. The number of variants were found with a statistically significant higher growth rate than the WT2. To have a measure of growth rate variance, the non- aggregated dataset corresponding to time step pair 0-4 was used. Variants with the WT2 coding and promoter region were used to compute a single score for the WT2 growth rate and variance. Non-WT2 variants were then compared with the WT2 using one-tailed p-values, calculated based on the z-scores, with the False Discovery Rate (FDR) correction applied to account for multiple testing. Using the critical value of 0.01, a total of 12,115 variants were found to have a higher growth rate than the WT2. The quality of the sequence-growth rate relationship was assessed in part by plotting the growth rate distribution of variants that are inactive. Any variant is assumed to be inactive if its coding region has a stop codon before or at the active site. Only variants with a non-mutated coding and promoter region wereconsidered to be WT2. In Figure 4A below, the distribution of growth rates isshown for variants with premature stop codons using the aggregated dataset, based on variants with at least 7 reads.In Figure 4B below, the distribution of growth rates is shown for variants withpremature stop codons using the aggregated dataset, based on variants with at least 15 reads. The results show reasonably symmetrical distributions around 0 with few outliers. Comparable levels of dispersion are observed for the WT2, as illustrated in Figure 4C.Example 11 - Learning from the data to generate sequence variantsAim of study 82843PC01 51 The data from Example 10 was used to generate a series of variant sequencesusing the MoCHI suite, which identifies beneficial and non-beneficial mutations byconsidering both individual and epistatic effects. The resulting library combines highly beneficial mutations and epistatic interactions while avoiding less beneficial ones. Materials and methods Training a Second-Order Model Using MoCHI MoCHI (Faure, Andre & Lehner, Ben. (2024). MoCHI: neural networks to fit interpretable models and quantify energies, energetic couplings, epistasis and allostery from deep mutational scanning data) was used to fit a second-order model predicting variant growth rates based on additive effects of individual mutations and pairwise epistatic interactions. The model training parameters were set to –min_observed 4, –max_interaction_order 2, –learn_rate 0.02, and – transformation ReLU, with 10-fold cross-validation. A filtered version of the non- aggregated dataset, corresponding to time step pair 0-4, was used as the training set to ensure an accurate measure of variance. Variants containing indels, promoter region changes, or stop codons before the last active site were excluded, resulting in a dataset of 103,731 records. MoCHI provides a score for each single mutation and pairwise epistatic interaction inferred from the training set, as well as a score for the WT2. Using the defined parameters, a positive score indicates a beneficial effect on the growth rate, while a negative score indicates a detrimental effect. Generating Variants Using the Second-Order MoCHI Model Scores obtained for mutations and pairwise epistatic interactions were used togenerate a library of high-growth-rate variants. The variants are generated atamino acid level. A greedy iterative sampling algorithm was employed to identifya set of mutations that, when applied to the WT2, maximize the predicted growth rate according to the MoCHI model. To generate variants, the sum of additivetraits were optimized. A function f is defined as the sum of the mean singlemutation terms S and the mean pairwise epistasis terms E, as shown in equationsbelow: 82843PC01 52 Since the second-order model was trained using ten-fold cross-validation, up to ten different scores are available for each single mutation or pairwise epistasis term. The sampling algorithm operates in two phases. In the first phase, an empty set of mutations (the nominated set) is initialized. During each iteration, a position not yet occupied by a mutation in the nominated set is randomly sampled. All mutations with an available score (candidate mutations) are then evaluated using equation below: ^^^ℎ^^^1(^^^^ , ^) = ^(^ ∪ {^^^^}) − ^(^)From the set of candidates with a positive ^^^ℎ^^^1 (beneficial mutations), one mutation was sampled and added to the nominated set. The probability distribution for sampling was determined by applying the softmax function to the ^^^ℎ^^^1 values of all beneficial mutations, using a temperature Tmax. This process was repeated until the desired number of mutations was reached. In the second phase, mutations in the nominated set are iteratively removed and replaced with better ones, if available. During each iteration, a position not yet occupied by a mutation in the nominated set is randomly sampled. All possiblecombinations of adding a candidate mutation and removing an existing one fromthe nominated set are evaluated using equation below: For each candidate ^^^^, the mutation swap ^^^^ → ^^^^with the highest^^^ℎ^^^2 was added to the set of beneficial swaps, provided ^^^ℎ^^^2was greater than 0. From this set, one swap was sampled, adding^^^^and removing ^^^^from the nominated set. The probability distribution was determined by applying the softmax function to the ^^^ℎ^^^2 values of all beneficial swaps. Sampling was performed using a temperature Ti, which followed a linear schedule from Tmax to Tmin. The iteration process stopped when for all positions no more beneficial swaps were available or the maximum number of iterations, Imax, was reached. A subset of the generated variants was selected for synthesis. Variants with a negative pairwise epistatic interaction term below a threshold—set at half the 82843PC01 53 negative score of the WT2—were excluded. Additionally, variants were removed if less than one-third of their potential epistatic effects were accounted for by the MoCHI model. From the filtered pool, those offering the best balance of diversity and growth rate were selected for synthesis. 200 sequences were generated each with 7 and 15 mutations using Imax =1000, and 50 sequences each with 25 and 35 mutations using Imax =1500. The temperature parameters, Tmax and Tmin, were set to 5 and 0.5, respectively. Duplicate sequences were removed, and filters were applied, reducing the total number of sequences to 115. 3 sequences were selected each with 7 and 15 mutations, and 2 sequences each with 25 and 35 mutations. To convert amino acid sequences to DNA sequences for synthesis, the coding region of the WT2 was used as the template, and introduced the amino acid mutations using the most frequent codon for each desired amino acid in E. coli. Deterministic Naive Variant Generation Using the Second-Order Model As a comparison, a library with minimal consideration of epistasis was generated. Sequences were generated by sequentially adding mutations from the set of available mutations, in decreasing order of score, to the nominated set until the desired number of mutations was reached. The mutation set included all single mutations and pairs provided by the MoCHI model. Single mutations were assigned the score provided by MoCHI, while pair scores were calculated by combining the individual mutation scores and the epistatic term, based on the equation below: Pairwise epistasis with mutations already in the nominated set was ignored. When a pair was selected, both mutations were added to the set. In the final iteration, only single mutations were considered. Sequences were designed according to the algorithms described above. Results and conclusions In total, the second-order model produced through MoCHI processing provided 3,386 single mutation terms and 90,831 pairwise epistasis terms, achieving an R2 82843PC01 54 of 0.74, implying the model explains 74% of the variance in growth rate on the held-out test data for each fold.Example 12 - Experimental validation of the data and the generatedsequences Aim of study Experimental work was done to evaluate how well the data obtained from the invented method aligns with assays with lower throughput, as well as to evaluate the performance of sequences generated based on learned information from the data. Materials and methods Strain constructions Multiple variants with different estimated fitness levels from the dataset in Example 10 were selected for validation experiments. They were obtained by PCR amplification from the miniprepped plasmids of non-selected or selected libraries. They were subsequently assembled into pSD293 under the control of J23110 (SEQ ID NO. 16). Sequences were verified by Sanger sequencing after the strains containing the resulting expression plasmids were obtained. For growth assays, strains were constructed by introducing the expression plasmids into the selection strain. For bioconversion assays, strains were constructed with E. coli BW25113 instead. Sanger sequencing for sequence validation A series of variants with generated consensus sequences in the high throughput dataset chosen to be Sanger sequencing-verified. 9 out of 10 with indels, and 32 out of 37 without indels were successfully cloned, which were sent for sequencing. Growth assay The selection strain expressing variants of ASMT under the control of J23110(SEQID NO. 16) were grown overnight in 2xYT media supplemented with cysteinesupplemented with 25 µg / mL chloramphenicol. These precultures were washed with equal volume of 0.85% NaCl. 5 µl of the washed precultures were inoculatedinto 245 µl of the same selection media as in Example 9. The cultures were grown 82843PC01 55 in a Growth Profiler and growth was monitored over time. Growth rate was estimated with the built-in software of Growth Profiler. Bioconversion assayE. coli BW25113 strains expressing variants of ASMT under the control of J23110(SEQ ID NO. 16) were grown overnight in 2xYT media supplemented with 25µg / mL chloramphenicol in 96-deepwell plates placed in a MaxQ 8000 incubator set to 30 °C and 300 rpm. On the next day, each overnight culture was used to inoculate two new 500 µL cultures in new 96-deelwell plates with 2xYT supplemented with chloramphenicol to an initial OD of 0.05. The new plates were placed back in the MaxQ 8000 incubator at 30 °C, 300 rpm for 5 hours when the OD600 of the cultures was measured to be between 3.0 and 5.3. The plates were put on ice, and cell mass was normalized by transferring between 262 and 410 µL culture to a new plate, ensuring a final UOD600 of 1.23 for all wells. The plateswere then centrifuged at 1500 ×g for 15 minutes to pellet the cells. Thesupernatant was removed, and the cell pellets were stored at -70 °C overnight. Next day, the cell pellets were equilibrated to room temperature, and 70 µL of room-temperature B-PER Complete Bacterial Protein Extraction Reagent was then added to each pellet. The resuspended pellets were incubated at room temperature with gentle rocking (40 rpm) for 15 minutes. After lysis, 50 µL of B-PER lysate was mixed with 50 µL of bioconversion buffer consisting of an aqueous solution of 1 mM protocatechuic acid (PCA), 2 mM L- methionine (L-met) and 2 mM ATP disodium salt. The 100 µL bioconversion reactions were incubated in sealed microtiter plates for 120 minutes at 30 °C and 246 rpm in a Labnet 311DS Environmental Shaking Incubator. Reactions were stopped by heating the microtiter plates in a 90 °C water bath for 5 minutes.The microtiter plates were centrifuged at 1500 ×g for 15 minutes to pellet celldebris. The supernatant was filtered into HPLC sample plates through AcroPrep™ Advance Plates, Short Tip (Cytiva, 97052-096) and diluted 1:1 with blank buffer to ensure sufficient volume for HPLC. Blank buffer components were 0.5 mM PCA, 1 mM L-met, 1 mM ATP disodium salt. Samples were analyzed on an Ultimate 3000 HPLC with sPFP column coupled to a DAD-3000 module. The sample injection volume was 10 µL, and the eluents were 10 mM ammonium formate and 100% acetonitrile. The ratio of acetonitrile was initially 5%, then 60% at 5 minutes, 90% at 5.5 minutes, and 5% at 7.6 minutes. 82843PC01 56 The flow rate was constant a 0.7 ml / min. The length of the method was 10 minutes. Vanillic acid was observed as a peak in absorbance at 277 nm at a retention time of 3.7 minutes. Graphpad Prism 10.1.12 was used to generate graphs and perform linear regression and related analyses. Results Sanger sequencing validations The consensus sequences obtained based on Nanopore reads were perfectly aligned with the Sanger sequencing results in all 41 cases, including 9 with indels and 32 without indels. Growth assay results Strong positive correlation was observed between the estimated fitness in dataset and the measured fitness in the growth assay, with R2value of 0.87, as shown in Figure 5. Bioconversion assay resultsThe bioconversion results are shown in Figures 6A-6D and in Table 1.ASMT-WT (SEQ ID NO. 2) refers to the original wild-type ASMT sequence.ASMT1-ASMT37 refer to ASMT variants from the dataset as seen in Example 10.The rest of the ASMT sequences were generated with the second-order model.It can be seen in Figure 6A that the original ASMT WT led to a vanillicconcentration of 0.66 mg / L (95% CI [0.50, 0.81], SD=0.11, n=2). The highest vanillic acid concentration observed for the tested variants in the dataset was with ASMT17 at 3.18 mg / L (95% CI [2.33, 4.03], SD=0.61, n=2), 380% times higher than WT (p=0.13, with two-tailed t-test). The highest titers of vanillic acid were observed for the ASMT variant “seq39_7muts” generated with the second-order model. The mean concentration of vanillic acid in the bioconversion supernatant was 4.23 mg / L (95% CI [3.08, 5.38], SD=0.83, n=2). This is 541% higher than the WT (p=0.09, with one-tailed t-test), and 33% higher (p=0.49, with one-tailed t-test) than the best tested sequence from the dataset. 82843PC01 57 Seq39_7muts is 7 AA mutations from the WT2, 11 AA from the original WT, and at least 5 AA mutations from any sequence in the training dataset used for MoCHI.Figure 6B shows vanillic acid production plotted against predicted growth rate forall cloned ASMT variants identified during selection.Figure 6C shows vanillic acid production plotted against predicted growth rate forfunctional, cloned ASMT variants identified during selection with detectable levels of vanillic acid in bioconversion experiment.Figure 6D shows vanillic acid production plotted against predicted growth rate forfunctional, cloned ASMT variants identified during selection with detectable levels of vanillic acid in bioconversion experiment. Table 1: Estimated growth rate and vanillic acid production in bioconversion experiment by all cloned ASMT variants. The dataset includes ASMT variants identified from the selected population ASMT1-ASMT37, and novel ASMT variants generated using the second-order model. Standard deviation (SD) is indicated. N / A is non-available data. Variant name EstimatedSD for Vanillic acid SD for vanillic growth estimated production acid rate [1 / h] growth [mg / L] production rate [1 / h] [mg / L] ASMT-WT N / A N / A 0.66 0.11ASMT1 0.1667 0.00152 2.02 0.38ASMT2 0.1619 0.00088 2.54 0.32ASMT3 0.1623 0.00076 2.59 0.22ASMT4 0.1596 0.00093 2.23 1.12ASMT5 0.1321 0.00075 1.41 0.09ASMT6 0.1086 0.00116 1.56 0.72ASMT7 0.0953 0.00112 1.17 0.14 82843PC01 58ASMT8 0.0689 0.00096 2.24 0.11ASMT9 0.0265 0.00013 1.68 0.59ASMT10 0.0184 0.00094 1.84 0.46ASMT11 0.0008 0.00171 1.37 0.94ASMT12 -0.0082 0.00232 0.00 0.00ASMT15 0.0034 0.00387 0.00 0.00ASMT17 0.1591 0.00096 3.18 0.61ASMT18 0.1432 0.00144 2.06 0.07ASMT20 0.1017 0.00106 2.29 0.83ASMT21 0.0672 0.00083 1.91 1.11ASMT22 0.0631 0.00131 2.47 1.54ASMT24 0.0376 0.00147 2.24 0.56ASMT25 0.0323 0.00129 1.28 0.88ASMT26 0.0265 0.00013 1.10 0.02ASMT27 0.0299 0.00179 1.36 0.05ASMT28 0.0265 0.00013 1.29 0.31ASMT29 0.0128 0.00139 0.76 0.16ASMT30 0.0157 0.00213 0.90 0.07ASMT31 0.0095 0.00141 0.92 0.07ASMT32 -0.0027 0.00681 0.00 0.00ASMT33 -0.0085 0.00323 0.00 0.00ASMT34 -0.0057 0.00437 0.00 0.00 82843PC01 59 ASMT35 -0.0088 0.00419 0.00 0.00ASMT36 -0.0118 0.00452 0.00 0.00ASMT37 -0.0121 0.00371 0.00 0.00seq9_7muts N / A N / A 2.12 0.17seq29_25muts N / A N / A 0.00 0.00seq31_15muts N / A N / A 0.30 0.04seq31_35muts N / A N / A 0.00 0.00seq32_7muts N / A N / A 1.47 0.25seq39_7muts N / A N / A 4.23 0.83seq48_25muts N / A N / A 0.49 0.04seq103_15muts N / A N / A 0.63 0.04seq139_15muts N / A N / A 0.75 0.26seq149_15muts N / A N / A 0.68 0.07seq49_35muts N / A N / A 0.54 0.09seq69_7muts N / A N / A 0.00 0.00Conclusion The validation experiment results for the samples data points in the dataset show that, 1) The consensus sequences generated from Nanopore sequencingreads with the same barcodes have very low error rates (0% based on the 41 that were checked). 2) The growth rate data obtained in the high-throughput manner ishighly correlated with the actual growth rate data (R2= 0.87). 3) The estimated fitness levels obtained in the high-throughputmanner are positively correlated with results in the bioconversion 82843PC01 60 assay. The overall correlation was strong. Those with lowpredicted fitness (below 0.01) led to no detectable product formation, while those above that threshold led to significant product formation in almost all cases. In addition, many generated sequences based on learned mutational effects showed significant activity, despite not having been seen in the dataset. In particular, one generated sequence showed potentially higher product formation than any tested variants. These results indicate the great value of using the high- throughput dataset obtained with the method for enzyme engineering. Sequence listing SEQ ID NO. 1 Plasmid backbone SEQ ID NO. 9 Reverse primer forbarcode amplification SEQ ID NO. 2 atASMT DNA SEQ ID NO. 10 atASMT P72S, K132R,(ASMT-WT) D147E SEQ ID NO. 3 atASMT amino acid SEQ ID NO. 11 atASMT P72S, I103N,(ASMT-WT) P286L SEQ ID NO. 4 Forward primer for SEQ ID NO. 12 atASMT R77C,atASMT G105S, K187R SEQ ID NO. 5 Reverse primer for SEQ ID NO. 13 atASMT R77C,ASMT M215K, P231S SEQ ID NO. 6 One-cycle PCR SEQ ID NO. 14 atASMT E46D, S84I,primer P302L SEQ ID NO. 7 Library plasmid SEQ ID NO. 15 J23100 promoterwith non-mutated atASMT SEQ ID NO. 8 Forward primer for SEQ ID NO. 16 J23110 promoterbarcode amplification 82843PC01 61Items of the invention1. A method for estimating fitness of enzyme variants, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant; and g) optionally estimating the enzyme fitness of the specific enzyme variantfrom the determined frequency information. 2. A method for optimizing a nucleotide sequence encoding an enzyme variant, the method comprising 82843PC01 62 a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant;f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant; and g) obtaining the optimized nucleotide sequence encoding an enzyme variantfrom the determined frequency information.3. The method according to any of items 1 or 2, wherein the method furthercomprises a step between step d) and e) of subjecting the isolated DNA variantlibrary grown under condition II.) to long-read sequencing. 82843PC01 634. The method according to any of items 1 or 3, wherein identification of theenzyme variant fitness comprises determining the fitness of a portion of enzymevariants in the DNA variant library, preferably a portion of the variants with the highest activity.5. The method according to any of items 2 or 3, wherein identification of theoptimized nucleotide sequence encoding an enzyme variant comprisesdetermining the fitness of a portion of enzyme variants in the DNA variant library, preferably a portion of the variants with the highest activity. 6. The method according to any of items 1-5, wherein the enzyme fitness is selected from the group consisting of enzyme activity, expression level, turnover number such as unimolecular rate constant (Kcat), Michaelis constant (KM), enzyme solubility, enzyme stability, substrate specificity, product specificity, thermotolerance, and combinations thereof. 7. The method according to any of the preceding items, wherein the DNA variant library is prepared by the steps: a) providing at least one DNA sequence encoding an enzyme variant;b) linking one or more unique DNA barcodes on the at least one DNAsequence encoding an enzyme variant and obtaining a barcode-DNA library; and c) assembling the barcode-DNA library and obtaining a DNA variant library.8. The method according to item 7, wherein the at least one DNA sequence ofstep a) is generated by PCR, preferably error prone PCR.9. The method according to items 7 or 8, wherein the assembly of the barcode-DNA library is by Golden Gate assembly. 10. The method according to any of the preceding items, wherein the unique DNA barcode is a random nucleotide sequence. 82843PC01 64 11. The method according to any of the preceding items, wherein the unique DNA barcode has a length in the range 10-300 DNA nucleotides, preferably 15-100 DNA nucleotides, most preferably 20-30 DNA nucleotides. 12. The method according to any of the preceding items, wherein the DNA variant is linked to the unique DNA barcode via a covalent bond, such as an extension of the DNA variant coding for the enzyme variant, such as a phosphodiester bond. 13. The method according to any of the preceding items, wherein the DNA library is a vector library, such as a plasmid library. 14. The method according to any of the preceding items, wherein the enzyme isan enzyme for which the enzyme activity and / or enzyme fitness correlates withgrowth of the host cell, such as growth during selection pressure.15. The method according to any of the preceding items, wherein the enzyme is selected from the group consisting of oxidoreductases, transferases, hydrolase, lyases, isomerases, ligases, translocases, or a plurality thereof. 16. The method according to any of the preceding items, wherein the host cell is a strain selected from the group consisting of a bacterial cell, a yeast cell, and a eukaryotic cell. 17. The method according to any of the preceding items, wherein the host cell is a bacterial cell, preferably E. coli. 18. The method according to any of the preceding items, wherein the selection pressure is selected from the group consisting of sensor-based approach, metabolic coupling, and substrate stress relief. 19. The method according to any of the preceding items, wherein the long-read sequencing is sequencing of at least 601 nucleotides, such as at least 1000 nucleotides, such as at least 5000 nucleotides, such as at least 10,000 nucleotides, such as at least 20,000 nucleotides. 82843PC01 65 20. The method according to any of the preceding items, wherein the long-read sequencing is selected from the group consisting of third generation sequencing, such as Oxford Nanopore sequencing, PacBio sequencing, Illumina Complete Long Read sequencing, TELL-Seq. 21. The method according to any of the preceding items, wherein the short-read sequencing is sequencing of nucleotide sequences in the range 10-600 nucleotides, preferably in the range 30-500, most preferably in the range 50-300 nucleotides. 22. The method according to any of the preceding items, wherein the short-read sequencing is selected from the group consisting of deep sequencing, Next- Generation Sequencing (NGS), such as Illumina sequencing. 23. The method according to any of the preceding items, wherein the frequency information of the specific DNA variant coding for a specific enzyme variant is a reflection of the enzyme fitness of said specific enzyme variant. 24. The method according to any of the preceding items, wherein the rate of change of the specific DNA variant coding for a specific enzyme variant is determined from the change of an amount of the specific DNA variant over a specific period of time. 25. The method according to any of the preceding items, wherein in step f) the comparison of the first set of information from step d) to the second set of information from step e), is by correlating DNA barcodes from the second set of information to the enzyme variants of the first set of information via the enzyme variant specific DNA barcode, thereby obtaining a prediction of enzyme fitness ofthe enzyme variants and / or an optimized nucleotide sequence encoding anenzyme variant via the frequency of the DNA barcodes from the second set ofinformation or the rate of change of the specific DNA variant coding for a specific enzyme variant. 82843PC01 66 26. The method according to any of the preceding items, wherein the enzymefitness of the specific enzyme variant and / or an optimized nucleotide sequenceencoding an enzyme variant is reflected by a growth rate of the host cell.27. The method according to item 22, wherein the growth rate is determined fromgrowing the host cell from step b(I.) under selection pressure favoring enzyme variants with increased fitness or activity.28. The method according to any of the preceding items, wherein the DNAnucleotide sequence coding for enzyme variants comprises a promoter such asSEQ ID NO. 15 or SEQ ID NO.16.29. The method according to any of the preceding items, wherein the DNA variant library comprises 100.000-5 million DNA variants, such as 300.000-5 million DNA variants, such as 500.000-4 million DNA variants, preferably 800.000-4 millionDNA variants, more preferably 1-3 million DNA variants, most preferably 3 millionDNA variants. 30. The method according to any of the preceding items, wherein the host cell is grown under selection pressure at an optical density (OD) of 0.01-0.5, such as 0.3-0.5, preferably 0.05-0.5. 31. The method according to any of the preceding items, wherein the host cell isgrown in a colorless media.32. Use of at least one barcoded DNA sequence encoding an enzyme variant for predicting enzyme fitness of said enzyme variant, wherein the at least one barcoded DNA sequence is subjected to long-read sequencing and short-read sequencing. 33. Use of at least one barcoded DNA sequence encoding an enzyme variant for optimizing said DNA sequence encoding said enzyme variant, wherein the at least one barcoded DNA sequence is subjected to long-read sequencing and short-read sequencing. 82843PC01 6734. The use according to any of items 32 or 33, wherein the at least one barcodedDNA sequence for short-read sequencing has been obtained from a host cell subjected to selection pressure favoring enzyme variants with increased fitness. 35. A polypeptide according to SEQ ID NO. 3 comprising one or more of the following mutations P72S, K132R, D147E, I103N, P286L, R77C, G105S, K187R, M215K, P231S, E46D, S84I, and / or P302L.36. The polypeptide according to item 35 comprising one or more of the followingcombinations of mutations: -P72S, K132R, and D147E,- P72S, I103N, and P286L,- R77C, G105S, and K187R,- R77C, M215K, and P231S, and / or- E46D, S84I, and P302L.37. The polypeptide according to any of items 35 or 36, the polypeptide is apolypeptide according to any of SEQ ID NOs 10-14. 38. A processor system programmed to operate according to a machine learning (ML) algorithm for estimating fitness of enzyme variants, the machine learning(ML) algorithm being trained, and / or being trainable on data comprising acorrelation of a specific nucleotide sequence encoding an enzyme variant tofrequency information. 39. A processor system programmed to operate according to a machine learning (ML) algorithm for estimating fitness of enzyme variants, the machine learning (ML) algorithm being trained, and / or being trainable, on data obtained by amethod according to any of items 1-31.40. A processor system programmed to operate according to a machine learning (ML) algorithm for optimizing a nucleotide sequence encoding an enzyme variant,the machine learning (ML) algorithm being trained, and / or being trainable on datacomprising a correlation of a specific nucleotide sequence encoding an enzymevariant to frequency information. 82843PC01 68 41. A processor system programmed to operate according to a machine learning (ML) algorithm for optimizing a nucleotide sequence encoding an enzyme variant, the machine learning (ML) algorithm being trained, and / or being trainable, ondata obtained by a method according to any of items 1-31.42. Use of a machine learning (ML) algorithm trained on data comprising acorrelation of a specific nucleotide sequence encoding an enzyme variant tofrequency information to predict fitness for any enzyme sequence, or to guide thegeneration of enzyme sequences with given fitness properties. 43. Use of a machine learning (ML) algorithm trained on data obtained by amethod according to any of items 1-31 to predict fitness for any enzymesequence, or to guide the generation of enzyme sequences with given fitness properties. 44. Use of a machine learning (ML) algorithm trained on data comprising acorrelation of a specific nucleotide sequence encoding an enzyme variant tofrequency information for optimizing a nucleotide sequence encoding an enzymevariant, or to guide the generation of an improved nucleotide sequence encodingan enzyme variant with given fitness properties.45. Use of a machine learning (ML) algorithm trained on data obtained by amethod according to any of items 1-31 for optimizing a nucleotide sequenceencoding an enzyme variant, or to guide the generation of an improved nucleotidesequence encoding an enzyme variant with given fitness properties.46. A system suitable for executing an algorithm, such as machine learning (ML)algorithm, for estimating fitness of enzyme variants, the system being trained,and / or being trainable, on data obtained by a method according to any of items 1- 31, the system being arranged for:- receiving a first set of information (1SI), such as a first database, said first ofinformation being obtained, or being obtainable, from d) subjecting the isolated DNA variant library grown under condition I.) to long-read sequencing, thereby 82843PC01 69 obtaining said first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant,- receiving a second set of information (2SI), such as a second database, saidsecond of information being obtained, or being obtainable, from e) subjecting the isolated DNA variant library grown under condition II.) to short-read deep sequencing, thereby obtaining said second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; wherein the system comprises: -a first comparison module (1CM) for comparing said first set of information (1SI) to said second set of information (2SI), thereby correlating specific enzyme variants to the frequency information, such as frequency of the variant; and- an optional second comparison module (2CM) for correlating determinedfrequency information to the enzyme fitness of the specific enzyme variant. 47. A system suitable for executing an algorithm, such as machine learning (ML)algorithm, for optimizing a nucleotide sequence encoding an enzyme variant, thesystem being trained, and / or being trainable, on data obtained by a method according to any of items 1-31, the system being arranged for:- receiving a first set of information (1SI), such as a first database, said first ofinformation being obtained, or being obtainable, from d) subjecting the isolated DNA variant library grown under condition I.) to long-read sequencing, thereby obtaining said first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant,- receiving a second set of information (2SI), such as a second database, saidsecond of information being obtained, or being obtainable, from e) subjecting the isolated DNA variant library grown under condition II.) to short-read deep sequencing, thereby obtaining said second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA 82843PC01 70 variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; wherein the system comprises: -a first comparison module (1CM) for comparing said first set of information (1SI) to said second set of information (2SI), thereby correlating specific enzyme variants to the frequency information, such as frequency of the variant, using the DNA barcode as an identifier; and- an optional second comparison module (2CM) for correlating determinedfrequency information to the enzyme fitness of the specific enzyme variant.48. The system suitable for executing an algorithm according to any of items 46or 47, wherein the algorithm is MoCHI. 49. A method for training a machine learning (ML) system for estimating fitness ofenzyme variants according to item 46, the method comprises the steps of:-receiving training data comprising a first set of information (1SI), such as a first database, and a second set of information (2SI), such as a second database, -training the ML system for estimating fitness of enzyme variants using said training data, and -validating the ML system using correlated specific enzyme variants to thefrequency information, such as frequency of the variant. 50. A method for training a machine learning (ML) system for optimizing a nucleotide sequence encoding an enzyme variant according to item 47, the method comprises the steps of: -receiving training data comprising a first set of information (1SI), such as a first database, and a second set of information (2SI), such as a second database, -training the ML system for optimizing a nucleotide sequence encoding an enzyme variant using said training data, and -validating the ML system using correlated specific enzyme variants to thefrequency information, such as frequency of the variant, using a barcode as an identifier. 82843PC01 7151. The method for training a machine learning (ML) system according to any ofitems 49 or 50, wherein the ML system comprises:^ a ML model configured to assign a score to a nucleotide sequence encodingan enzyme variant, ^a comparison component configured to compare the score with a thresholdfor predicting a growth rate of a host cell expressing said nucleotidesequence encoding the enzyme variant, and^ a featuring component configured to receive an output from thecomparison component, wherein the output indicates a predicted fitness,such as activity, of said enzyme variant encoded by said nucleotidesequence.52. The method for training a machine learning (ML) system according to item 51,wherein if the score is above the threshold, it is indicative of an improved growth rate of a host cell expressing said enzyme variant compared to a growth rate of a host cell expressing a corresponding wild-type enzyme.53. The method for training a machine learning (ML) system according to item 51,wherein if the score is below the threshold, it is indicative of a decreased growth rate of a host cell expressing said enzyme variant compared to a growth rate of a host cell expressing a corresponding wild-type enzyme.54. The method for training a machine learning (ML) system according to any ofitems 50-53, wherein the threshold is 0.55. The processor system according to any of items 38-41, the use according toany of items 42-45, the system according to any of items 46-48, and / or themethod according to any of items 49-54, being implemented on a computer, suchas being computer-implemented.56. A computer-readable storage medium comprising data comprising acorrelation of a specific nucleotide sequence encoding an enzyme variant toenzyme fitness, such as enzyme activity. 82843PC01 72 57. A computer-readable storage medium comprising data obtained by a methodaccording to any of items 1-31.58. A neural network model obtained or obtainable by a method according to any of items 49-54. 59. A method for converting a substrate into a product comprising: a) selecting a parent enzyme capable of enzymatically converting thesubstrate into the product; b) evaluating the fitness of variants of the parent enzyme with respect toenzymatically converting the substrate into the product using the method, algorithm, system, model and / or storage medium of any of the preceding items; c) identifying and selecting a nucleotide sequence encoding an enzymevariant having improved fitness compared to the parent enzyme; d) transforming the nucleotide sequence into a host cell;e) expressing the nucleotide sequence in the host cell to produce the variantenzyme; f) optionally isolating the variant enzyme; andg) contacting the enzymes variant with the substrate, optionally in the hostcell, at conditions to allow the variant enzyme to convert the substrate into the product. 60. The method according to item 2, wherein -the nucleotide sequence from step g) is cloned into an expression hostcell; -the enzyme variant encoded by the nucleotide sequence is expressedfrom the host cell and optionally purified; and- the expressed enzyme variant is incubated with its enzymatic substrateto convert the substrate into the product of the enzymatic reaction.

Claims

82843PC01 73 Claims 1. A method for optimizing a nucleotide sequence encoding an enzyme variant, the method comprising a) providing a DNA variant library, preferably a plasmid library, the librarycomprises DNA variants coding for enzyme variants, wherein each DNA variant is linked to one or more unique DNA barcodes, preferably through a one-cycle PCR process; b) transforming the DNA variant library into a host cell for expressing theenzyme variants from the library; I. growing the host cell comprising the DNA variant library, preferablyunder conditions in the absence of a selection pressure favoring enzyme variants with increased activity; and II. growing the host cell comprising the DNA variant library, underselection pressure favoring enzyme variants with increased catalytic activity; c) isolating DNA variant library from the host cells grown under condition I.)and isolating at least the barcode region of the DNA variant library from the host cells grown under condition II.); d) subjecting the isolated DNA variant library grown under condition I.) tolong-read sequencing, thereby obtaining first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant, e) subjecting the isolated DNA variant library grown under condition II.) toshort-read deep sequencing, thereby obtaining a second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; f) comparing the first set of information from step d) to the second set ofinformation from step e), thereby correlating specific enzyme variants tothe frequency information, such as the frequency of the variant; and g) obtaining the optimized nucleotide sequence encoding an enzyme variantfrom the determined frequency information.82843PC01 742. The method according to claim 1, wherein the method further comprises a stepbetween step d) and e) of subjecting the isolated DNA variant library grown undercondition II.) to long-read sequencing.

3. The method according to any of claims 1 or 2, wherein identification of theoptimized nucleotide sequence encoding an enzyme variant comprisesdetermining the fitness of a portion of enzyme variants in the DNA variant library, preferably a portion of the variants with the highest activity.

4. The method according to claim 3, wherein the enzyme fitness is selected fromthe group consisting of enzyme activity, expression level, turnover number such as unimolecular rate constant (Kcat), Michaelis constant (KM), enzyme solubility, enzyme stability, substrate specificity, product specificity, thermotolerance, and combinations thereof.

5. The method according to any of the preceding claims, wherein the DNA variant library is prepared by the steps: a) providing at least one DNA sequence encoding an enzyme variant;b) linking one or more unique DNA barcodes on the at least one DNAsequence encoding an enzyme variant and obtaining a barcode-DNA library; and c) assembling the barcode-DNA library and obtaining a DNA variant library.

6. The method according to claim 5, wherein the at least one DNA sequence ofstep a) is generated by PCR, preferably error prone PCR.

7. The method according to claims 5 or 6, wherein the assembly of the barcode-DNA library is by Golden Gate assembly.

8. The method according to any of the preceding claims, wherein the unique DNA barcode is a random nucleotide sequence.

9. The method according to any of the preceding claims, wherein the unique DNA barcode has a length in the range 10-300 DNA nucleotides, preferably 15-100 DNA nucleotides, most preferably 20-30 DNA nucleotides.82843PC01 75 10. The method according to any of the preceding claims, wherein the DNA variant is linked to the unique DNA barcode via a covalent bond, such as an extension of the DNA variant coding for the enzyme variant, such as a phosphodiester bond.

11. The method according to any of the preceding claims, wherein the DNA library is a vector library, such as a plasmid library.

12. The method according to any of the preceding claims, wherein the enzyme is selected from the group consisting of oxidoreductases, transferases, hydrolase, lyases, isomerases, ligases, translocases, or a plurality thereof.

13. The method according to any of the preceding claims, wherein the host cell is a strain selected from the group consisting of a bacterial cell, a yeast cell, and a eukaryotic cell.

14. The method according to any of the preceding claims, wherein the host cell is a bacterial cell, preferably E. coli.

15. The method according to any of the preceding claims, wherein the selection pressure is selected from the group consisting of sensor-based approach, metabolic coupling, and substrate stress relief.

16. The method according to any of the preceding claims, wherein the long-read sequencing is sequencing of at least 601 nucleotides, such as at least 1000 nucleotides, such as at least 5000 nucleotides, such as at least 10,000 nucleotides, such as at least 20,000 nucleotides.

17. The method according to any of the preceding claims, wherein the long-read sequencing is selected from the group consisting of third generation sequencing, such as Oxford Nanopore sequencing, PacBio sequencing, Illumina Complete Long Read sequencing, TELL-Seq.82843PC01 76 18. The method according to any of the preceding claims, wherein the short-read sequencing is sequencing of nucleotide sequences in the range 10-600 nucleotides, preferably in the range 30-500, most preferably in the range 50-300 nucleotides.

19. The method according to any of the preceding claims, wherein the short-read sequencing is selected from the group consisting of deep sequencing, Next- Generation Sequencing (NGS), such as Illumina sequencing.

20. The method according to any of the preceding claims, wherein the frequency information of the specific DNA variant coding for a specific enzyme variant is a reflection of the enzyme fitness of said specific enzyme variant.

21. The method according to any of the preceding claims, wherein the rate of change of the specific DNA variant coding for a specific enzyme variant is determined from the change of an amount of the specific DNA variant over a specific period of time.

22. The method according to any of the preceding claims, wherein in step f) the comparison of the first set of information from step d) to the second set of information from step e), is by correlating DNA barcodes from the second set of information to the enzyme variants of the first set of information via the enzyme variant specific DNA barcode, thereby obtaining an optimized nucleotide sequenceencoding an enzyme variant via the frequency of the DNA barcodes from thesecond set of information or the rate of change of the specific DNA variant coding for a specific enzyme variant.

23. The method according to any of the preceding claims, wherein the enzyme fitness of the specific enzyme variant is reflected by a growth rate of the host cell.

24. The method according to claim 23, wherein the growth rate is determinedfrom growing the host cell from step b(I.) under selection pressure favoring enzyme variants with increased fitness or activity.82843PC01 77 25. The method according to any of the preceding claims, wherein the DNAnucleotide sequence coding for enzyme variants comprise a promoter such as SEQID NO. 15 or SEQ ID NO.16.

26. The method according to any of the preceding claims, wherein the DNAvariant library comprises 100.000-5 million DNA variants, such as 300.000-5million DNA variants, such as 500.000-4 million DNA variants, preferably800.000-4 million DNA variants, more preferably 1-3 million DNA variants, mostpreferably 3 million DNA variants.

27. The method according to any of the preceding claims, wherein the host cell isgrown under selection pressure at an optical density (OD) of 0.01-0.5, such as0.3-0.5, preferably 0.05-0.

5.

28. The method according to any of the preceding claims, wherein the host cell isgrown in a colorless media.

29. Use of at least one barcoded DNA sequence encoding an enzyme variant for optimizing said DNA sequence encoding said enzyme variant, wherein the at least one barcoded DNA sequence is subjected to long-read sequencing and short-read sequencing.

30. The use according to claim 29, wherein the at least one barcoded DNAsequence for short-read sequencing has been obtained from a host cell subjected to selection pressure favoring enzyme variants with increased fitness.

31. A polypeptide according to SEQ ID NO. 3 comprising one or more of the following mutations P72S, K132R, D147E, I103N, P286L, R77C, G105S, K187R, M215K, P231S, E46D, S84I, and / or P302L.

32. The polypeptide according to claim 31 comprising one or more of the followingcombinations of mutations: -P72S, K132R, and D147E,- P72S, I103N, and P286L,- R77C, G105S, and K187R,82843PC01 78 -R77C, M215K, and P231S, and / or- E46D, S84I, and P302L.

33. The polypeptide according to any of claims 31 or 32, the polypeptide is apolypeptide according to any of SEQ ID NOs 10-14.

34. A processor system programmed to operate according to a machine learning (ML) algorithm for optimizing a nucleotide sequence encoding an enzyme variant,the machine learning (ML) algorithm being trained, and / or being trainable on datacomprising a correlation of a specific nucleotide sequence encoding an enzymevariant to frequency information.

35. A processor system programmed to operate according to a machine learning (ML) algorithm for optimizing a nucleotide sequence encoding an enzyme variant, the machine learning (ML) algorithm being trained, and / or being trainable, ondata obtained by a method according to any of claims 1-28.

36. Use of a machine learning (ML) algorithm trained on data comprising acorrelation of a specific nucleotide sequence encoding an enzyme variant tofrequency information for optimizing a nucleotide sequence encoding an enzymevariant, or to guide the generation of an improved nucleotide sequence encodingan enzyme variant with given fitness properties.

37. Use of a machine learning (ML) algorithm trained on data obtained by amethod according to any of claims 1-28 for optimizing a nucleotide sequenceencoding an enzyme variant, or to guide the generation of an improved nucleotidesequence encoding an enzyme variant with given fitness properties.

38. A system suitable for executing an algorithm, such as machine learning (ML)algorithm, for optimizing a nucleotide sequence encoding an enzyme variant, thesystem being trained, and / or being trainable, on data obtained by a method according to any of claims 1-28, the system being arranged for:- receiving a first set of information (1SI), such as a first database, said first ofinformation being obtained, or being obtainable, from d) subjecting the isolated82843PC01 79 DNA variant library grown under condition I.) to long-read sequencing, thereby obtaining said first set of information, such as a first database, linking a specific DNA barcode to a specific DNA variant coding for a specific enzyme variant,- receiving a second set of information (2SI), such as a second database, saidsecond of information being obtained, or being obtainable, from e) subjecting the isolated DNA variant library grown under condition II.) to short-read deep sequencing, thereby obtaining said second set of information, such as a second database, linking a specific DNA barcode to a frequency of the specific DNA variant coding for a specific enzyme variant or a rate of change of the specific DNA variant coding for a specific enzyme variant; wherein the system comprises: -a first comparison module (1CM) for comparing said first set of information (1SI) to said second set of information (2SI), thereby correlating specific enzyme variants to the frequency information, such as frequency of the variant, using the DNA barcode as an identifier; and- an optional second comparison module (2CM) for correlating determinedfrequency information to the enzyme fitness of the specific enzyme variant.

39. The system suitable for executing an algorithm according to claim 38, whereinthe algorithm is MoCHI.

40. A method for training a machine learning (ML) system for optimizing anucleotide sequence encoding an enzyme variant according to any of claims 38 or39, the method comprises the steps of: -receiving training data comprising a first set of information (1SI), such as a first database, and a second set of information (2SI), such as a second database, -training the ML system for optimizing a nucleotide sequence encoding an enzymevariant using said training data, and- validating the ML system using correlated specific enzyme variants to thefrequency information, such as frequency of the variant.

41. The processor system according to any of claims 34 or 35, the use accordingto any of claim 36 or 37, the system according to any of claims 38 or 39, and / or82843PC01 80the method according to claim 40, being implemented on a computer, such asbeing computer-implemented.

42. A computer-readable storage medium comprising data obtained or obtainableby a method according to any of claims 1-28.

43. A neural network model obtained or obtainable by a method according toclaim 40.

Images