Spectroscopy process modelling and monitoring

WO2024209273A3PCT designated stage expired Publication Date: 2024-11-14CLARA FOODS
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
PCT/IB2024/000213
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-04-07
Filing Date
2024-04-08
Publication Date
2024-11-14

AI Technical Summary

Technical Problem

Conventional methods for monitoring fermentation processes are labor-intensive and time-consuming, limiting the frequency and accuracy of data collection for proteins of interest (POI) and exopolysaccharides (EPS), which hampers the real-time optimization of fermentation dynamics.

Method used

A Fourier Transform Infrared (FTIR) spectroscopy-based monitoring platform that uses machine learning techniques to analyze FTIR data from fermentation processes, enabling the prediction of POI presence and titer, as well as fermentation metrics like EPS concentration, allowing for real-time process optimization.

Benefits of technology

The platform reduces processing time and increases data collection frequency, providing accurate and timely insights into fermentation progress, enabling more efficient and optimized fermentation processes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure IB2024000213_14112024_PF_FP_ABST
    Figure IB2024000213_14112024_PF_FP_ABST
Patent Text Reader

Abstract

A fermentation monitoring platform obtains data regarding a biological sample collected during a fermentation process. The data comprises Fourier transform infrared (FTIR) data describing the content of the biological sample. The fermentation monitoring platform determines one or more proteins of interest (POI) present in the biological sample by applying a first vector representation of the FTIR data to a first machine-learning model. For each POI present in the biological sample, the fermentation monitoring platform determines a titer of the POI by applying a second vector representation of the FTIR data to a second machine-learning model. The fermentation monitoring platform identifies one or more process excursions during the fermentation process based on the POIs present in the biological sample and the titer of each present POI and modifies instructions regarding the fermentation process based on the one or more identified process excursions.
Need to check novelty before this filing date? Find Prior Art

Description

SPECTROSCOPY PROCESS MODELLING AND MONITORINGInventors: A Samuel Pottinger Zachary Friar Varsha Gopalakrishnan Dane Jacobson Ranjan PatnaikCROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 494,822, filed on April 07, 2023, which is incorporated by reference in its entirety.BACKGROUND

[0002] Understanding fermentation requires tracking of many variables and the labor involved in collecting all the data may limit an engineer’ insights into fermentation systems. Proteins of interest (POI) and exopolysaccharide (EPS) titers may indicate fermentation progress and success. Conventional methods for acquiring this data may require sampling and submitting samples for HPLC, sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) gels, or Bradford or Bicinchonic acid (BCA) total protein assays to get a sense of fermentation dynamics. However, processing times for these approaches may be significant. For example, collecting samples for protein of interest (POI) measures and exopolysaccharides (EPS) measures may limit the number of titer measures collected per fermentation to accommodate the capacities of the and their instruments. Similarly, contract manufacturers may lack the time and instrumentation to run these assays at short intervals with quick turnaround times.SUMMARY

[0003] Disclosed is a method for monitoring and optimizing a fermentation process. The fermentation process may be used to generate molecules of interest via expression in a recombinant cell. The molecules of interest may be amino acids, peptides, polypeptides, or proteins. A fermentation monitoring platform obtains data from a fermentation process at one or more sites. The data may comprise Fourier transform infrared (FTIR) data. For example, the data may comprise one or more spectra generated by FTIR analysis of biological samples. The FTIR spectra may be raw or unprocessed FTIR spectra. The FTIR data may be collected using a spectrophotometer that is operably coupled to one or more fermentation reactors. The1 41522-58976 / WOFTIR data may be collected on one or more pre-downstream (pre-DSP) (e.g. prior to deriving a protein product from a biological sample) fermentation samples.

[0004] The fermentation monitoring platform applies one or more machine learning techniques to analyze the FTIR data. The fermentation monitoring platform inputs a feature vector representation of the FTIR data to a predictive machine-learning model. In some embodiments, the feature vector representation comprises absorbance values collected by the FTIR probe at specific wavelengths. The machine learning models may comprise classification and / or regression models. The machine learning models may be linear or nonlinear models. The models may comprise one or more neural networks or deep neural networks. The fermentation monitoring platform inputs features extracted from raw or pre- processed FTIR data, for example amplitudes at corresponding wavenumbers from FTIR spectra and time points associated with each FTIR spectrum. Based on the input features of FTIR data, the fermentation monitoring platform generates one or more predicted or estimated outputs. The fermentation monitoring platform stores a first machine-learning model trained to predict whether a protein of interest (or another suitable molecule of interest) is present in a biological sample based on features of FTIR data. The first machinelearning model may be a trained classification model. The fermentation monitoring platform may store a second machine-learning model trained to predict a titer or concentration estimate of the protein of interest (or another suitable molecule of interest). The second machine-learning model may be a trained regression predictive model. In some embodiments, the machine-learning models are trained to predict a presence of a plurality of molecules of interest (e.g. of one or more groups of proteins) or a titer or concentration estimate of a plurality of molecules of interest (e.g. of one or more groups of proteins).

[0005] In some embodiments, the fermentation monitoring platform stores a third machine-learning model trained to predict one or more fermentation metrics, for example a predicted concentration of one or more analytes of interest. For example, for a fermentation process aimed at producing a specific protein, accurately predicting the target protein’s presence may serve as an indicator of the process’s success. Additionally, the concentration of all proteins produced during the fermentation process may provide insight into the purity of fermentation product relative to byproduct host proteins created by the cell. The one or more analytes of interest may include methanol, ammonia, potassium, exopolysaccharide (EPS) and / or hydrogen peroxide. The one or more analytes of interest may be organic acids, vitamins, carbohydrates, minerals, or fats. The one or more analytes of interest may comprise EPS, and the predictive model for EPS concentration is based on using phosphate as a latent2 41522-58976 / WOvariable. In some embodiments, the third machine-learning model is trained on a first training set for EPS and methanol and a second training set for ammonia and potassium. The fermentation monitoring platform may further use the one or more predicted or estimated outputs to enable inline monitoring and / or modification of the fermentation process at the one or more sites in real-time.

[0006] In some embodiments, the fermentation monitoring platform performs one or more pre-processing techniques on the FTIR data prior to machine learning (ML) analysis. The FTIR data may be un-augmented. The fermentation monitoring platform may augment the FTIR data with synthetically generated data comprising of simulated FTIR spectra and / or simulated noise. When applying data augmentation, a predictive model may be trained using a training dataset with a fewer number of samples compared to an un-augmented FTIR dataset. For example, the number of initial samples required for training may be reduced by at least 25% to 50% when using augmented FTIR data compared to using un-augmented FTIR data. The fermentation monitoring platform may generate simulated FTIR spectra by averaging or adding a plurality of randomly chosen spectra with a shared POI. A validation median absolute error (MdAE) of the predictive model may be reduced using augmented FTIR data compared to un-augmented FTIR data. In other embodiments, the fermentation monitoring platform performs data compression and / or dimensionality reduction.

[0007] In one embodiment, the fermentation monitoring platform obtains data regarding a biological sample collected during a fermentation process. The data comprises Fourier transform infrared (FTIR) data describing the content of the biological sample. The fermentation monitoring platform determines one or more proteins of interest (POI) present in the biological sample by applying a first vector representation of the FTIR data to a first machine-learning model. The first machine-learning model is trained to predict a confidence score representing a likelihood that each of the one or more proteins are present in the biological sample. For each POI present in the biological sample, the fermentation monitoring platform determines a titer of the POI by applying a second vector representation of the FTIR data to a second machine-learning model, the second machine-learning model is trained to predict a titer measurement for a POI. The fermentation monitoring platform identifies one or more process excursions during the fermentation process based on one or more of the POIs present in the biological sample and the titer of each POI present in the biological sample. The fermentation monitoring platform modifies instructions regarding the fermentation process based on the one or more identified process excursions.3 41522-58976 / WOBRIEF DESCRIPTION OF DRAWINGS

[0008] FIG. 1 illustrates a system for monitoring or modifying a fermentation process based on analysis of Fourier transform infrared spectroscopy (FTIR) data, in accordance with one or more embodiments.

[0009] FIG. 2A illustrates a process 200 for monitoring a fermentation process used to generate proteins, in accordance with one or more embodiments.

[0010] FIG. 2B illustrates an example system architecture for a fermentation monitoring platform 140, in accordance with one or more embodiments.

[0011] FIG. 2C is a flowchart illustrating a process for predicting one or more proteins of interest in a biological sample and titers of each protein of interest in the biological sample, in accordance with one or more embodiments.

[0012] FIG. 2D is a flowchart illustrating a process for training a machine-learning model to predict one or more proteins of interest in a biological sample, in accordance with one or more embodiments.

[0013] FIG. 2E is a flowchart illustrating a process for training a machine-learning model to predict titers of one or more proteins of interest in a biological sample, in accordance with one or more embodiments.

[0014] FIG. 2F is a flowchart illustrating a process for identifying one or more process excursions, in accordance with one or more embodiments.

[0015] FIG. 3 A illustrates an overlay of multiple spectra from biological samples, in accordance with one or more embodiments.

[0016] FIG. 3B illustrates amide bonds and atoms corresponding to the spectra of FIG.3 A, in accordance with one or more embodiments.

[0017] FIG. 4 provides a plot showing test accuracy by fermentation timepoints of a classifier trained on training and validation set, in accordance with one or more embodiments.

[0018] FIG. 5 provides a plot showing POI composition by fermentation timepoints of a test set, in accordance with one or more embodiments.

[0019] FIG. 6 illustrates a sample method for simulating data, in accordance with one or more embodiments.

[0020] FIG. 7 illustrates a plot of a method for improving a data augmentation process, in accordance with one or more embodiments.

[0021] FIG. 8 illustrates a set of histograms of model diagnostics of an elastic net regression model, in accordance with one or more embodiments.4 41522-58976 / WO

[0022] FIG. 9A illustrates a set of histograms of model diagnostics of an AdaBoost regression model, in accordance with one or more embodiments.

[0023] FIG. 9B illustrates a set of histograms of model diagnostics of the elastic net regression model trained on the amide region, in accordance with one or more embodiments.

[0024] FIG. 9C illustrates a set of histograms of model diagnostics of the PLS mode trained on the amide region, in accordance with one or more embodiments.

[0025] FIG. 9D illustrates a set of histograms 3800 of model diagnostics of the feedforward neural network model trained on the amide region, in accordance with one or more embodiments.

[0026] FIG. 10 illustrates a plot showing absolute error of each unaugmented model by fermentation time, in accordance with one or more embodiments.

[0027] FIG. 11 illustrates a set of histograms of model diagnostics of a support vector regression model, in accordance with one or more embodiments.

[0028] FIG. 12 illustrates a set of histograms of model diagnostics of a random forest regression model, in accordance with one or more embodiments.

[0029] FIG. 13 illustrates a histogram showing absolute error of each composite augmented model by fermentation time, in accordance with one or more embodiments.

[0030] FIG. 14 illustrates a histogram of absolute validation error for the baseline augmented elastic net model, in accordance with one or more embodiments.

[0031] FIG. 15 illustrates a histogram of absolute validation error for the baseline augmented random forest model, in accordance with one or more embodiments.

[0032] FIG. 16 illustrates a histogram of absolute error of each composite augmented model by fermentation time, in accordance with one or more embodiments.

[0033] FIG. 17 illustrates a histogram of absolute validation error for the fully augmented elastic net model, in accordance with one or more embodiments.

[0034] FIG. 18 illustrates a histogram of absolute validation error for the fully augmented random forest model, in accordance with one or more embodiments.

[0035] FIG. 19 illustrates a histogram showing absolute error of each fully augmented model by fermentation time, in accordance with one or more embodiments.

[0036] FIG. 20 illustrates a bar graph of data augmentation power analysis using a random forest model, in accordance with one or more embodiments.

[0037] FIG. 21 illustrates a bar graph of data power analysis without any data augmentation using random forest model, in accordance with one or more embodiments.5 41522-58976 / WO

[0038] FIG. 22 shows a computer system that is programmed or otherwise configured to implement methods provided herein.

[0039] FIG. 23 illustrates a bar plot showing median absolute training, and validation error, and the training and validation R squared for the augmented and unaugmented datasets.

[0040] FIG. 24 illustrates a scatter plot of predicted test YSI probe ammonia values and true values, in accordance with one or more embodiments.

[0041] FIG. 25 illustrates a histogram showing absolute test error between YSI probe ammonia values and true values, in accordance with one or more embodiments.

[0042] FIG. 26 illustrates a histogram showing test error between YSI probe ammonia values and true values broken down by fermentation time point, in accordance with one or more embodiments.

[0043] FIG. 27 illustrates a scatter plot of predicted test YSI probe potassium values and true values, in accordance with one or more embodiments.

[0044] FIG. 28 illustrates a scatter plot showing absolute test error between YSI probe potassium values and true values, in accordance with one or more embodiments.

[0045] FIG. 29 illustrates a scatter plot showing median test error between YSI probe potassium values and true values broken down by fermentation time point, in accordance with one or more embodiments.

[0046] FIG. 30 illustrates a bar plot showing absolute validation error between YSI probe methanol values and true values, in accordance with one or more embodiments.

[0047] FIG. 31 illustrates a bar plot showing median test error between YSI probe methanol values and predicted values broken down by fermentation time point, in accordance with one or more embodiments.

[0048] FIG. 32 illustrates a bar plot showing absolute error for the test set between lead ligand HPLC EPS predicted values and true values, in accordance with one or more embodiments.

[0049] FIG. 33 illustrates a bar plot showing median absolute test error between EPS HPLC titer values and predicted values broken down by fermentation time point, in accordance with one or more embodiments.

[0050] FIG. 34 illustrates a bar chart showing all metrics for five analytes.

[0051] FIG. 35 illustrates a truncated bar plot of frequency of absolute deviance for YSI probe methanol values, and predicted values, in accordance with one or more embodiments.6 41522-58976 / WO

[0052] FIG. 36 illustrates the segmented distance analysis showing the Euclidean distance between various spectra at different steps of the fermentation and downstream process, in accordance with one or more embodiments.DETAILED DESCRIPTION

[0053] While various embodiments of the disclosure have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions may occur to those skilled in the art without departing from the disclosure. It should be understood that various alternatives to the embodiments described herein may be employed.

[0054] Whenever the term “at least,” “greater than,” or “greater than or equal to” precedes the first numerical value in a series of two or more numerical values, the term “at least,” “greater than” or “greater than or equal to” applies to each of the numerical values in that series of numerical values. For example, greater than or equal to 1, 2, or 3 is equivalent to greater than or equal to 1, greater than or equal to 2, or greater than or equal to 3.OVERVIEW

[0055] The systems and methods disclosed herein may be applied to generate large numbers of molecules of interest through expression in recombinant cells. The systems and methods disclosed herein may use machine learning techniques to enhance monitoring and improvement of fermentation processes. The systems and methods disclosed herein may use machine learning analysis to process Fourier transform infrared (FTIR) spectrograms, either to predict the presence of one or more molecules of interest (e.g. proteins of interest (POIs)) or to predict a concentration (e.g. titer) of one or more molecules of interest (e.g. POIs).

[0056] Fourier transform infrared spectroscopy (FTIR) may determine the content of a biological sample (e.g. a sample comprising one or more proteins) by measuring the spectrum of light that the sample absorbs. In some embodiments, performing FTIR involves using an instrument (e.g. a spectrometer) to shine a beam comprising many frequencies of light at once on the sample and measure how much of that beam is absorbed. The beam may be iteratively modified to shine different combinations of wavelengths of light. The light absorption of the sample per wavelength of light may be measured by taking the Fourier transform of the raw measurement data to create an interferogram.7 41522-58976 / WOSAMPLE

[0057] Embodiments of the systems and methods described herein may involve performing FTIR on a biological sample. In some embodiments, the sample may comprise one or more proteins. The biological samples may be collected before, during, or after a fermentation process (e.g. during upstream or downstream process steps). In some embodiments, biological samples may be collected at multiple points throughout the fermentation process (e.g. periodically). The biological samples may be collected from a fermentation reactor or another vessel used for fermentation.

[0058] In some embodiments, the proteins are expressed in a cell. In some embodiments, the cell is a recombinant cell. In some embodiments, the recombinant cell may be a methanotroph, for example Komagataella pastoris and Komagataella phaffii are preferable (also known as Pichia pastoris). Examples of strains in the Komagataella genus include Pichia pastoris strains. Examples can include NRRL Y-l 1430, BG08, BG10, NRRL Y-l 1430 GS115 (NRRL Y-15851), GS190 (NRRL Y-18014), PPF1 (NRRL Y 18017), PPY120OH, YGC4, and strains derived therefrom. Other examples of P. pastoris strains that may be used as host cells include but are not limited to CBS7435 (NRRL Y-l 1430), CBS704 (DSMZ 70382) or derivatives thereof. Other examples of methanol -utilizing yeast include yeasts belonging to Ogataea (Ogataea polymorpha), Candida (Candida boidinii), Torulopsis (Torulopsis) or Komagataella.

[0059] In some embodiments, the recombinant cell is a microbial cell. In some embodiments, the microbial cells may be collected from organisms from a Komagataella species, a Saccharomyces species, a Trichoderma species, a Pseudomonas species, an Aspergillus species, or an E. coli species.

[0060] Further examples of suitable host cell organisms include but are not limited to eukaryotic cells such as: Arxula spp. Arxula adeninivorans, Kluyveromyces spp. Kluyveromyces lactis, Komagataella spp. Pichia angusta, Pichia pastoris, Saccharomyces spp. Saccharomyces cerevisiae, Schizosaccharomyces spp. Schizosaccharomyces pombe, Yarrowia spp. Yarrowia lipolytica, Agaricus spp. Agaricus bisporus, Aspergillus spp. Aspergillus awamori, Aspergillus fumigatus, Aspergillus nidulans, Aspergillus niger, Aspergillus oryzae, Colletotrichum spp. Colletotrichum gloeosporiodes, Endothia spp. Endothia parasitica, Fusarium spp. Fusarium graminearum, Fusarium solani, Mucor spp. Mucor miehei, Mucor pusillus, Myceliophthora spp. Myceliophthora thermophila, Neurospora spp. Neurospora crassa, Penicillium spp. Penicillium camemberti, Penicillium canescens, Penicillium chrysogenum, Penicillium (Talaromyces) emersonii, Penicillium funiculosum, Penicillium8 41522-58976 / WOpurpurogenum, Penicillium roqueforti, Pleurotus spp. Pleurotus ostreatus, Rhizomucor spp. Rhizomucor miehei, Rhizomucor pusillus, Rhizopus spp. Rhizopus arrhizus, Rhizopus oligosporus, Rhizopus oryzae, Trichoderma spp. Trichoderma altroviride, Trichoderma reesei, Trichoderma vireus, Aspergillus oryzae, Bacillus subtilis, Escherichia coli, Myceliophthora thermophila, Neurospora crassa, Pichia pastor is, Pichia Pastoris “MutS” strain (Graz University of Technology (CBS7435MutS) or Biogrammatics (BG11)), Komagatella phaffi, and Komagatella pastoris.

[0061] In some cases, a bacterial host cell such as Lactococcus lactis, Bacillus subtilis or Escherichia coli may be used as the host cells. Other host cells include bacterial hosts such as, but not limited to, Lactococci sp. Lactococcus lactis, Bacillus subtilis, Bacillus amyloliquefaciens, Bacillus licheniformis and Bacillus megaterium, Brevibacillus choshinensis, Mycobacterium smegmatis, Rhodococcus erythropolis and Corynebacterium glutamicum, Lactobacilli sp. Lactobacillus fermentum, Lactobacillus casei, Lactobacillus acidophilus, Lactobacillus plantarum, Pseudomonas sp. Pseudomonas luorescens.

[0062] In some embodiments, the proteins in the sample are animal or animal egg proteins. Egg proteins may naturally be expressed by birds and / or reptiles including poultry, fowl, waterfowl, game bird, chicken, quail, turkey, turkey vulture, hummingbird, duck, ostrich, goose, gull, guineafowl, pheasant, emu, crocodile, owl, finch, pigeon, or penguin eggs. Examples of egg proteins include, but are not limited to, ovalbumin, ovomucoid, ovotransferrin, lysozyme proteins, ovomucin, ovoglobulin G2, ovoglobulin G3, ovoinhibitor, ovoglycoprotein, flavoprotein, ovomacroglobulin, ovostatin, cystatin, avidin, ovalbumin related protein X, or ovalbumin related protein Y, or any combination thereof.SYSTEM

[0063] FIG. 1 illustrates a system for monitoring or modifying a fermentation process based on analysis of Fourier transform infrared spectroscopy (FTIR) data, in accordance with one or more embodiments.

[0064] The system includes a physical FTIR probe 110 (or instrument) that performs FTIR on sample data. The FTIR probe 110 may be collected on one or more pre-downstream process (pre-DSP) fermentation samples, for example prior to deriving a protein product from a biological sample. The FTIR probe 110 may produce raw spectra files as output. In one embodiment, the FTIR probe 110 is a spectrometer, for example a ReactIR® probe or a MATRIX-MF FT-IR® probe. In another embodiment, the FTIR probe 110 may be spectrophotometer operably coupled to one or more fermentation reactors.9 41522-58976 / WO

[0065] The FTIR probe 110 may comprise a light source, an interferometer, a sample compartment, a detector, amplifier, analog-to-digital (A / D) convertor, and / or a computer. The light source may generate radiation, which the interferometer processes prior to interacting with the sample and the detector. The interferometer may produce a signal, which is amplified by an amplifier and converted into a digital signal by the A / D converter. The converted digital signal may be transferred to a computer for processing by an interferogram. In one embodiment, the interferometer is a Michelson interferometer.

[0066] The raw spectra files may be, for example, a text or comma-separated value file comprising wavenumbers with corresponding transmittance percentages, or other types of files derived from the signals produced by the FTIR probe. For example, the data may comprise one or more spectrograms generated by FTIR analysis of biological samples. The FTIR spectra data may be raw or unprocessed FTIR spectra files. The spectra data platform 120 may produce FTIR spectra metadata from raw spectra files. In some embodiments, a person or entity (e.g. a collector) may interact with the spectra data platform 120 to generate FTIR spectra metadata from raw spectra files. FTIR spectra metadata may include proteins of interest (POI), dates, processes, and genetics. For example, metadata may comprise a location of a fermentation process. The location of the fermentation process may affect water-derived background artifacts within the FTIR spectra metadata. The spectra data platform 120 may use POI information to characterize protein content of the sample, which influences the shape of the FTIR peaks. The spectra data platform 120 may use strain information to characterize a strain, which influences non-POI proteins in the sample. FTIR spectra metadata may be processed by one or more machine learning models used by the system. For example, the one or more machine-learning models may process metadata relating to a protein of interest (POI) in the sample and the stage of processing (e.g. upstream or downstream) at which the sample was collected.

[0067] In some embodiments, the spectra metadata platform 120 may be integrated with a web application (not shown) to enable a user of the system to, via a network, submit or upload data (e.g. FTIR spectra) and / or metadata for computer processing (e.g. processing by one or more machine learning models). The web application may be a browser-based application or a software application that may have access to a computer network (e.g. the Internet). The web application may be accessible via one or more client devices (e.g. the client device 130). As described herein, the client device 130 is a computing device, such as a desktop computer, laptop computer, smartphone, or tablet computer.10 41522-58976 / WO

[0068] The fermentation monitoring platform 140 applies trained machine-learning models to FTIR spectra metadata (generated by the spectra data platform 120) to monitor a fermentation process. In some embodiments, the fermentation monitoring platform 140 may predict how effectively a fermentation process produces a protein of interest using predictions generated by machine-learning models regarding the presence or absence of a protein of interest in a sample or a titer of a protein of interest in the sample.

[0069] The fermentation monitoring platform 140 may include a model training pipeline (not shown). As described herein, the model training pipeline is a computing unit that facilitates training of the one or more machine learning algorithms. The model training pipeline stores a training dataset of entries containing feature vectors and ground truth labels. During training, the model training pipeline inputs data to a machine-learning model, analyzes outputs of the machine-learning model, and / or configures the param eters / hyperparameters of the machine-learning model. During training, the model training pipeline inputs the feature vector of each entry in the training dataset to a machinelearning model. The model training pipeline compares the output of the model to the ground truth label associated with the entry and tunes the parameters / hyperparameters based on the comparison. The model training pipeline may include or more pre-processing systems for modifying data in the training dataset prior to training. The model training pipeline may control various parameters, for example a frequency at which the model is to be trained, a learning rate for the machine-learning model, and / or a type of convergence condition to use (e.g. a loss function) to complete training of an ML model. Accordingly, the model training pipeline provides insights into how the machine-learning models applied by the fermentation monitoring platform improve over time and compare to a baseline PLS model. The model training pipeline may also provide insight into how any feature engineering or other local experiments compare to a current model.

[0070] FTIR spectroscopy data captured by the FTIR probe 110 and analyzed by the spectra data platform 120 and predictions generated by machine-learning models of the fermentation monitoring platform 140 may be communicated via a network architecture 130. For example, raw spectra data captured by the FTIR probe 110 may be transmitted to the spectra data platform 120 via the network architecture for storage and analysis. In some embodiments, the network architecture may comprise a local network. The local network may be a mesh network where devices communicate with each other without a centralized device, such as a hub, switch, or router. In some embodiments, the network may be a local11 41522-58976 / WOarea network (LAN), wide area network (WAN), mesh network, Wi-Fi network, Bluetooth network, wired network (e.g. Ethernet), cellular network, or another type of network.

[0071] Although not shown, the system and components of the system may comprise one or more databases for storing data transmitted through the network 130. A database may be one or more memory devices configured to store data. Additionally, the databases may also, in some embodiments, be implemented as a computer system with a storage device. In one aspect, the databases (e.g. local databases and cloud databases) may be used by components of the web application 130 to perform one or more operations consistent with the disclosed embodiments.

[0072] Databases may comprise cloud databases and / or local databases. One or more cloud databases and local databases of the platform may utilize any suitable database techniques. For instance, a structured query language (SQL) or “NoSQL” database may be utilized for storing data, including raw spectral data, metadata, and trained or untrained ML models or algorithms. Some of the databases may be implemented using various standard data-structures, such as an array, hash, (linked) list, struct, structured text file (e.g. XML), table, JavaScript Object Notation (JSON), NoSQL and / or the like. Such data-structures may be stored in memory and / or in (structured) files.

[0073] In another alternative, an object-oriented database may be used. Object databases can include several object collections that are grouped and / or linked together by common attributes; they may be related to other object collections by some common attributes. Object- oriented databases perform similarly to relational databases with the exception that objects are not just pieces of data but may have other types of functionalities encapsulated within a given object.

[0074] In some embodiments, the database may include a graph database that uses graph structures for semantic queries with nodes, edges and properties to represent and store data. If the database is implemented as a data- structure, the use of the database may be integrated into another component such as the component of the present invention. Also, the database may be implemented as a mix of data structures, objects, and relational structures. Databases may be consolidated and / or distributed in variations through standard data processing techniques. Portions of databases, e.g. tables, may be exported and / or imported and thus decentralized and / or integrated.

[0075] FIG. 2A illustrates a process 200 for monitoring a fermentation process used to generate proteins, in accordance with one or more embodiments. In some embodiments, the method disclosed may be used to monitor a process for generating proteins through12 41522-58976 / WOexpression in recombinant cells. For example, the fermentation monitoring platform 140 may use the machine-learning models to analyze FTIR spectroscopy data to identify the types of proteins in a sample and to predict the titer (e.g. concentration) of specific proteins in the sample.

[0076] In the illustrated embodiment, the fermentation monitoring platform 140 obtains 210 fermentation data from a lab or an industrial process. In one embodiment, fermentation data is collected through a process of inserting genetic material for a protein into a recombinant cell and / or feeding microorganisms comprising genetic code for the protein of interest with feedstock (e.g. by growing the microorganisms in a bioreactor). The fermentation data may be derived from an interferogram generated by FTIR spectroscopy of a sample comprising one or more proteins. In one embodiment, the fermentation monitoring platform 140 performs normalization techniques such as standard scaler or normalization to a specific spectrum peak to compare spectra from different biological samples The fermentation monitoring platform 140 analyzes FTIR spectra to determine concentrations of other molecules of interest. In one embodiment, a spectrophotometer is operably coupled to one or more fermentation reactors to enable more efficient processing of biological samples. The FTIR data may be collected on one or more pre-down stream processing (pre-DSP) fermentation samples (e.g. prior to purification and / or recovery from fermentation broth). Additionally, the fermentation monitoring platform 140 may perform pre-processing techniques using principal component analysis (PCA) or multidimensional scaling (MDS) or classical spectroscopy techniques like Multiplicative Scatter Correction (MSC).

[0077] The FTIR data may be raw or unprocessed FTIR data. In some embodiments, the fermentation monitoring platform 140 may pre-process the FTIR data. In one embodiment, the fermentation monitoring platform 140 trims the dataset by eliminating data associated with one or more wavelengths determined to not provide information uninformative to the ML analysis. In some embodiments, the fermentation monitoring platform 140 augments the FTIR data with synthetic data. The synthetic data may comprise simulated FTIR spectra and / or simulated noise. The fermentation monitoring platform 140 may generate simulated FTIR spectra by averaging or adding a plurality of randomly chosen spectra having the same POI. Using augmented data may serve to improve the predicted output of the models disclosed herein by reducing a validation median absolute error (MdAE). Using augmented data may enable the model to be trained using fewer samples. For example, the training set may be reduced in size by at least 25% or at least 50% when using augmented data compared to using un-augmented data.13 41522-58976 / WO

[0078] The fermentation monitoring platform 140 may process 220 the obtained FTIR data using one or more machine learning models to generate one or more predicted outputs (also referred to herein as “predictions”). In one embodiment, the fermentation monitoring platform 140 using a machine-learning model to predict a protein of interest (POI) in a biological sample. In another embodiment, the fermentation monitoring platform 140 uses a machine-learning model to predict the titer of a POI in a biological sample.

[0079] Depending on the task, the fermentation monitoring platform 140 may use one or more classification models or prediction models (e.g. regression models). For example, the platform 140 may use one or more classification methods to identify a POI in a sample, while the platform 140 may use one or more predictive models to determine titer or concentration of one or more proteins present in the sample. In the context of FTIR spectroscopy, the platform 140 may perform classification using a visual inspection model to analyze FTIR plots (e.g. a CNN). The platform 140 may input a plurality of wavenumbers within the FTIR data and time points to the classification model. Classification models may include tree methods, such as AdaBoost. Classification models may be selected based at least in part on test or validation accuracy. For example, the system may select a classification model that yields a test validation accuracy of at least 70%. The platform 140 may use a predictive model to receive a plurality of wavenumbers within the FTIR data and an identity of the one or more POIs as inputs. Prediction models may be trained at least in part using high performance liquid chromatography (HPLC) quantification of POI for ground truth data. A predictive model may comprise partial least squares (PLS) regression, principal components regression (PCR), Elastic net regression, Support Vector Regression, AdaBoost, Random Forest Regression, a generalized additive model (GAM), a CNN, or multivariate adaptive regression splines (MARS™).

[0080] Machine learning methods herein may use neural networks, such as feed-forward neural networks (FNNs), convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Machine learning methods may use deep learning (e.g. deep neural networks). Regression methods may include elastic net regression, multivariate adaptive regression splines (MARS™), support vector regression (SVR), random forest regression, and / or generative additive models (GAMs), AdaBoost with regression trees, partial least squares (PLS), and principal components regression (PCR).

[0081] The fermentation monitoring platform 140 may use 230 the one or more predicted or estimated outputs to enable inline monitoring and / or modification of the fermentation process or another process for producing POIs in recombinant organisms and / or cells. For14 41522-58976 / WOexample, the predictions generated by the machine-learning models may inform traditional, biomass, and / or precision fermentation processes. For example, the POI predictions or titer predictions produced using the methods disclosed herein may inform processes for growing microorganisms and designing bioreactors. Additionally, POI predictions or titer predictions may be used to monitor for contamination and / or process excursions during fermentation. For example, a process excursion could be found if a fermentation process produces a lower POI titer than the titer predicted by the system. Additionally, predictions generated by the machine-learning models may be used to identify fermentation conditions for maximizing a fermentation titer. In one embodiment, the fermentation monitoring platform 140 ranks candidate fermentation conditions to optimize for the highest end-of-fermentation titers by predicted titers.

[0082] In addition to POI presence or POI titer, the machine-learning model(s) may predict one or more fermentation metrics. As described herein, fermentation metrics may include predicted concentrations for one or more analytes of interest. Analytes of interest may include methanol, ammonia, potassium, exopolysaccharide (EPS), organic acids, vitamins, carbohydrates, minerals, fats, and / or hydrogen peroxide. The fermentation monitoring platform 140 may use predictive models to generate one or more predicted concentrations for one or more of the analytes of interest. In one embodiment, the predicted concentrations of these analytes have a test error below 1.5 mg / mL. The one or more machine learning models may be trained on a first training set for EPS and methanol, and a second training set for ammonia and potassium. A predictive model for EPS may be based at least in part on using phosphate as a latent variable.FERMENTATION MONITORING PLATFORM

[0083] FIG. 2B illustrates an example system architecture for a fermentation monitoring platform 140, in accordance with one or more embodiments. The fermentation monitoring platform 140 includes a POI prediction module 310, a POI titer prediction module 320, a fermentation metric prediction module 330, a training dataset 340, a model training pipeline 350, an FTIR data processing module 360, and a fermentation modification module 370. Computer components such as web servers, network interfaces, security functions, load balancers, failover servers, management and network operations consoles, and the like are not shown so as to not obscure the details of the system architecture. Alternative embodiments may include more, fewer, or different components from those illustrated in FIG. 2B, and the15 41522-58976 / WOfunctionality of each component may be divided between the components differently from the description below.

[0084] Components of the fermentation monitoring platform 140 described herein apply machine-learning techniques to monitor and / or improve a fermentation process and / or other processes to generate proteins through expression in a recombinant cell or organism. A person of ordinary skill in the art would appreciate that the fermentation monitoring platform 140 may implement a number of different machine-learning models (or algorithms) including, but not limited to, Bayes, Markov, Gaussian processes, clustering algorithms, generative models, kernel and neural network algorithms) without exceeding the scope described herein. A person of ordinary skill in the art would recognize that typical neural networks employ one or more layers of nonlinear activation functions to predict an output for a received input and may include one or more hidden layers in addition to the input and output layers. The output of each hidden layer in some of these networks is used as input to the next layer in the network. Neural networks may include, by way of example and not limitation, convolutional neural networks and recurrent neural networks.

[0085] Examples of various suitable machine-learning models are described herein. AdaBoost is a machine-learning model which combines the outputs of other learning algorithms (“weak learners”) into a weighted sum, representing the final output of the algorithm. AdaBoost may be used for binary classification or multiclass classification tasks. AdaBoost is an adaptive algorithm where, as new weak learners are added, they are modified in favor of instances misclassified by previous classifiers. Each weak learner may produce an output hypothesis which may fix a prediction for each sample in the training set. At each iteration, the algorithm may select a weak learner and assign the learner with a coefficient, such that the total training error of the classifier is minimized. Additionally at each iteration of the training process, the fermentation monitoring platform 140 may assign weights to each sample in the training set equal to the current error on that sample. These weights may be used during the training process. The learners may be decision tree algorithms. The fermentation monitoring platform 140 may tune the depth of the trees in the ensemble and the maximum number of estimators to use for the learning process (after this number has been reached, the process is terminated). AdaBoost may use shallow trees (e.g. with depths of 1, less than 5, less than 10, less than 15, or less than 20). The fermentation monitoring platform 140 may select a range (e.g. between 10 and 110) trees to use in the ensemble. The AdaBoost-based model may be trained on training and validation data to determine how well it may perform on never-encountered (“test”) data.16 41522-58976 / WO

[0086] An elastic net model may use a weighted combination (e.g. a linear combination) of LI and L2 regularization. The elastic net regression may add a quadratic part to the penalty function of the least absolute shrinkage and selection operator (LASSO) method. This quadratic penalty, when used alone, may be indicative of ridge regression. Thus, elastic net regression may be thought of as a combination of ridge regression and LASSO regression. Use of this quadratic penalty term may make the loss function strongly convex, thus having a unique minimum. Elastic net regression may be configured by tuning the li ratio and the alpha parameters. The alpha parameter may define how strong the Lagrangian penalties are and may be tuned with equally spaced values from 1 to 20. The li ratio may represent the ratio of LASSO regression to ridge regression and may be tuned over the range from 0 to 1.

[0087] A multivariate adaptive regression splines (MARS) classifier is a non-parametric regression classifier that can automatically model nonlinearities and interactions between variables. It can add “kinks” into the regression function. A MARS classifier may build a model that is a weighted sum of basis functions. A basis function may be a constant, a hinge function, or a product of two or more hinge functions. A hinge function may partition data into disjoint independent regions, with a constant (known as a knot) at the hinge of the regions. The product of two or more hinge functions may model interaction between two or more variables. The MARS classifier may build a model in two phases: a forward pass, followed by a backward pass. In the forward pass, the model may start with an intercept function, adding an intercept term (which is the mean of response values), and then may find basis functions (optimal knot functions) to add to the model. The classifier may continue to add basis functions until a residual error becomes too small to overcome. A backward pass is then performed in which the classifier attempts to remove basis functions one-by-one to delete the least effective terms. MARS may offer flexible regression modeling useful for high-dimensional data. By making use of the hinge function as spline basis functions, MARS may automatically search for interactions and non-linear relationships in a high-dimensional basis space.

[0088] Support vector machines (SVM) are a machine-learning technique typically used to classify data. An SVM-based classifier may represent a data point as a p-dimensional vector and classify the data points by separating these vectors with a hyperplane. For a binary SVM classifier, a reasonable choice for hyperplane may represent the largest separation between two classes of data. This hyperplane may be found by maximizing the distance from the hyperplane to each of the nearest data points of the two classes. SVMs may also perform nonlinear classification using the kernel trick, implicitly mapping inputs into high-dimension17 41522-58976 / WOfeature spaces. The SVM technique may also be used in regression analysis. SVM’s adapted to the regression problem may inherit some of the beneficial properties from an SVM classifier. SVR still may make use of support vectors but, the inverse of the classification case, they define the regressed hyperplane by being far from the hyperplane. Thus, extreme data points in the dataset may help in defining the hyperplane. Furthermore, SVR may support non-linearity via kernelization. As a hyperparameter in tuning, this study considers the polynomial kernel at various degrees and the radial basis function (RBF). Additionally, two more hyperparameters tuned are considered: epsilon which specifies the region within which no penalty is associated with points predicted within a distance epsilon from the actual value and inverse regularization (“cost").

[0089] A random forest regression is an extension of the decision tree model that tends to yield more robust predictions by stretching the use of the training data partition. Whereas a decision tree may make a single pass through the data, a random forest regression may bootstrap 50% of the data (e.g. with replacement) and build many trees. Rather than using all explanatory variables as candidates for splitting, a random subset of candidate variables may be used for splitting, which may enable trees that have completely different data and different variables (hence the term random). The predictions from the trees, collectively referred to as the “forest,” may be then averaged together to produce the final prediction. Many trees (e.g. one hundred trees) may be included in a random forest model, with a number (e.g. 3, 6, 10, etc.) of terms sampled per split, a minimum of number (e.g. 1, 2, 4, 10, etc.) of splits per tree, and a minimum split size (e.g. 16, 32, 64, 128, 256, etc.). Random forests may be trained in a similar way as decision trees. Specifically, training a random forest may include the following operations: (1) select randomly k features from the total number of features; (2) create a decision tree from these k features using the same operations as for generating a decision tree; and (3) repeat the previous two operations until a target number of trees is created. Random forest regression may use an ensemble of regression trees, each with a randomly limited number of features. Systems and methods disclosed herein may tune the depths of the trees as well as the number of trees in the ensemble, as well as how many randomly selected features may be available to split at each step, and the minimum number for a split.

[0090] A generalized additive model (GAM) is a generalized linear model in which the linear predictor may depend linearly on unknown smooth functions of some predictor variables, and interest may focus on inference about these smooth functions. The model may relate a univariate response variable to a plurality of predictor variables. An exponential18 41522-58976 / WOfamily distribution is specified for (for example, normal, binomial or Poisson distributions) along with a link function g (for example the identity or log functions) relating the expected value of univariate response variable to the predictor variables. The functions may have a specified parametric form (for example a polynomial, or an un-penalized regression spline of a variable) or may be specified non-parametrically, or semi-parametrically, simply as “smooth functions”, to be estimated by non-parametric means. A typical GAM might use a scatterplot smoothing function, such as a locally weighted mean. This flexibility to allow non-parametric fits with relaxed assumptions on the actual relationship between response and predictor, may provide the potential for better fits to data than purely parametric models, but arguably with some loss of interpretability. Any multivariate function can be represented as sums and compositions of univariate functions. The GAM may drop the outer sum and may demand instead that the function belong to a simpler class.

[0091] The original GAM fitting method estimated the smooth components of the model using non-parametric smoothers (for example smoothing splines or local linear regression smoothers) via the backfitting algorithm. Backfitting works by iterative smoothing of partial residuals and provides a very general modular estimation method capable of using a wide variety of smoothing methods to estimate the terms. Many modem implementations of GAMs and their extensions are built around the reduced rank smoothing approach, because it allows well founded estimation of the smoothness of the component smooths at comparatively modest computational cost, and also may facilitate implementation of a number of model extensions in a way that is more difficult with other methods. This approach may replace the unknown smooth functions in the model with basis expansions. Smoothing bias may complicate interval estimation for these models, and the simplest approach may involve a Bayesian approach. Understanding this Bayesian view of smoothing may also help to understand the restricted maximum likelihood (REML) and full Bayes approaches to smoothing parameter estimation. At some level smoothing penalties may be imposed.

[0092] Multiple linear regression (MLR) may be used when controlled variables are few and not significantly redundant. However, as MLR assumes features are independent, identically distributed, and clearly relevant to predicting the supervisor, Partial Least Squares (PLS) and Principal Components Regression (PCR) attempt dimensional reduction techniques before regressing. Though both extract latent features that account for as large of a proportion of variation among the features as possible, PLS may account for both the features and supervisors when making its dimensional reduction. As a result, PLS may out-perform PCR in datasets where the supervisor is strongly associated with axes that have low variance.19 41522-58976 / WOIn both algorithms, the primary hyperparameter is the number of components to describe the feature space for which this study considers 2 to 20.

[0093] Partial Least Squares (PLS) may find a linear regression model by projecting predicted variables and the observable variables to a new space. PLS may be used to find the fundamental relations between two matrices (X and Y), i.e. a latent variable approach to modeling the covariance structures in these two spaces. A PLS model may try to find the multidimensional direction in the X space that explains the maximum multidimensional variance direction in the Y space. PLS regression may be particularly suited when the matrix of predictors has more variables than observations, and when there is multicollinearity among X values. By contrast, standard regression may fail in these cases (unless it is regularized).

[0094] In statistics, principal component regression (PCR) is a regression analysis technique that is based on principal component analysis (PCA). More specifically, PCR may be used for estimating the unknown regression coefficients in a standard linear regression model. In PCR, instead of regressing the dependent variable on the explanatory variables directly, the principal components of the explanatory variables may be used as regressors. PCR may use a subset of all the principal components for regression, making PCR a kind of regularized procedure and a type of shrinkage estimator. Often the principal components with higher variances (the ones based on eigenvectors corresponding to the higher eigenvalues of the sample variance-covariance matrix of the explanatory variables) are selected as regressors. However, for the purpose of predicting the outcome, the principal components with low variances may also be important.

[0095] The PCR method may be broadly divided into three major steps. First, PCA is performed on the observed data matrix for the explanatory variables to obtain the principal components, and then (usually) select a subset, based on some appropriate criteria, of the principal components so obtained for further use. Second, the observed vector is regressed of outcomes on the selected principal components as covariates, using ordinary least squares regression (linear regression) to get a vector of estimated regression coefficients (with dimension equal to the number of selected principal components). Third, the vector of estimated regression coefficients is transformed back to the scale of the actual covariates, using the selected PCA loadings (the eigenvectors corresponding to the selected principal components) to get the final PCR estimator (with dimension equal to the total number of covariates) for estimating the regression coefficients characterizing the original model.

[0096] A trained convolutional neural network (CNN) (e.g. a feed forward network), takes input data into convolutional layers (also referred to as hidden layers) and applies a20 41522-58976 / WOseries of trained weights or filters to the input data in each of the convolutional layers. The output of the first convolutional layer is an activation map (not shown), which is the input to the second convolution layer, to which a trained weight or filter (not shown) is applied, where the output of the subsequent convolutional layers results in activation maps that represent more and more complex features of the input data to the first layer. After each convolutional layer a non-linear layer (not shown) is applied to introduce non-linearity into the problem, which nonlinear layers may include tanh, sigmoid or ReLU. In some cases, a pooling layer (not shown) may be applied after the nonlinear layers, also referred to as a downsampling layer, which basically takes a filter and stride of the same length and applies it to the input, and outputs the maximum number in every sub-region the filter convolves around. Other options for pooling are average pooling and L2-norm pooling. The pooling layer reduces the spatial dimension of the input volume reducing computational costs and to control overfitting. The final layer(s) of the network is a fully connected layer, which takes the output of the last convolutional layer and outputs an n-dimensional output vector representing the quantity to be predicted. The output could be a scalar value data point being predicted by the network, a stock price for example. Trained weights may be different for each of the convolutional layers, as will be described more fully below. To achieve this real-world prediction / detection (e.g. it's a boat), the neural network needs to be trained on known data inputs or training examples resulting in trained CNN. To train CNN many different training examples are input into the model. A skilled artisan in neural networks will fully understand the description above provides a somewhat simplistic view of CNNs to provide some context for the present discussion and will fully appreciate the application of any CNN alone or in combination with other neural networks will be equally applicable and within the scope of some embodiments described herein.

[0097] The number of convolutional layers can be increased or decreased, as well as the number of fully-connected layers. In general, the optimal number and proportions of convolutional vs. fully-connected layers can be set experimentally, by determining which configuration gives the most optimal validation set performance on a given dataset. The number of convolutional layers could be decreased to 0, leaving a fully-connected network. The number of convolutional filters and width of each filter can also be increased or decreased. The output of the neural network may be a single, scalar value, corresponding to an exact prediction for the primary time sequence. Alternatively, the output of the neural network could be a logistic regression, in which each category corresponds to a specific range21 41522-58976 / WOor class of primary time sequence values, are any number of alternative outputs readily appreciated by the skilled artisan.

[0098] As described above, the fermentation monitoring platform 140 uses machinelearning models to identify POI’s present in a biological sample, to determine a titer of the POI’s in the biological sample, and to determine other fermentation metrics for the biological sample. Accordingly, the modules 310, 320, 330 are software modules that implement one or more machine-learning models. A “machine-learning model,” as used herein, comprises one or more machine-learning models that perform the described functionality. Machine-learning models may be stored on one or more computer-readable media with a set of weights. These weights are parameters used by the machine-learning model to transform input data received by the model into output data. The weight for each parameter (also referred to herein as “parameter values”) is determined through a training process, whereby the machine-learning model is trained based on a set of training examples and labels associated with the training examples. The weights may be stored on one or more computer-readable media, and are used by a system when applying the machine-learning model to new data.

[0099] Machine-learning models are applied in two phases: a training phase and a prediction phase. During the training phase, a machine-learning model is applied to training data. During the training phase, the model training pipeline 350 constructs a transformation function for generating a prediction of a target outcome, for example predicting a POI or predicting a POI titer from FTIR data and / or any features or metrics derived from metadata. The machine-learning model dynamically learns what characteristics of input data are most predictive of the target outcome. For example, a machine-learning model applied by the module 310 dynamically learns what characteristics of input data are predictive in determining a POI using the FTIR data and / or any features or metrics computed from or derived from FTIR metadata. As another example, a machine-learning model applied by the module 320 dynamically learns what characteristics of input data are predictive in determining POI titer using the FTIR data and / or any features or metrics computed from or derived from FTIR metadata. During the prediction phase (described above), the transform function optimized during the training phase is applied to generate a prediction. For example, the machine-learning model applied by the module 310 is applied to FTIR data and / or any features or metrics computed from or derived from metadata to predict the POI.

[0100] The POI prediction module 310 applies a trained machine-learning model to a sample of FTIR spectroscopy data to determine whether a POI is present in a biological sample. As described herein, such a model is referred to as a “POI presence model 315.”. In22 41522-58976 / WOone embodiment, the POI presence model 315 is trained to predict whether one or more proteins of interest are present in a biological sample. The POI prediction module 310 encodes.

[0101] FTIR spectroscopy data into a feature vector. As described herein, a feature vector is a representation of the FTIR data, which may be processed by a machine-learning model (e.g. the POI presence model) to generate a prediction regarding a POI. In particular, the POI presence model 315 processes a feature vector encoded from FTIR data to generate a prediction regarding whether a POI is present in a biological sample corresponding to the FTIR data. Each feature vector may include values describing or characterizing the attributes (features) of the FTIR data state. In some implementations, the feature vector may correspond to a data element in a multi-dimensional vector space, where each dimension represents a feature. Machine-learning algorithms operate within this multi-dimensional vector space to learn patterns and relationships among feature vectors and components of feature vectors. The exact composition of the feature vector may vary depending on the context and data available from the FTIR data. In some embodiments, the feature vector may include encoded features corresponding to the FTIR data. The encoded feature vector is input to the POI presence model 315. Examples of input features extracted from FTIR samples that are encoded into a feature vector include, the full FTIR spectrum and / or regions of the FTIR spectrum.

[0102] The POI presence model 315 predicts whether one or more POI (or any other suitable molecule of interest) are present in a biological sample corresponding to the vectorized FTIR sample based on inputs of an encoded feature vector. Examples of other molecules of interest include methanol, EPS, phosphate, glucose, etc. The POI presence model 315 may be a mathematical function of another more complex logical structure, trained using a combination of features stored in the training in the data store 340 to determine a set of parameter values stored in advance and used as part of the prediction generation. In some embodiments, the POI presence model 315 is a classification model, for example any of the classification models described above. In some embodiments, the POI presence model 315 is trained to predict the presence of one or more groups of proteins (or any other suitable group of molecules). Examples of groups of molecules include total carbohydrates, organic acids, vitamins, etc.

[0103] The prediction generated by the POI presence model 315 is an indication whether a given POI is present in a sample of FTIR data. In some embodiments, the POI presence model 315 generates a prediction for multiple POI’s in a sample of FTIR data. In some23 41522-58976 / WOembodiments, the POI presence model 315 generates a binary classification indicating whether a given POI is present in a biological sample. For example, a classification of 0 indicates that a POI is not present in a sample and a classification of 1 indicates that a POI is present in the sample. In other embodiments, the POI presence model 315 generates a classification indicating a confidence level that a POI is present in a biological sample. For example, the POI presence model 315 may generate confidence values ranging from 0 to 1, where 0 indicates that a POI is not present in the biological sample, 1 indicates that a POI is present in the sample, and the values between 0 and 1 represent a probability that the POI is present in the sample (organizing in ascending order of likelihood). In some embodiments, the POI presence model 315 may apply a confidence threshold and compare the confidence value determined for each POI to the confidence threshold. In such embodiments, the POI presence model may identify POI’s satisfying the confidence threshold as “present” in the FTIR data.

[0104] For each POI identified as present in the FTIR data, the POI titer prediction module 320 applies a trained machine-learning model to a set of FTIR spectroscopy data to determine a titer of the POI in the biological sample. As described herein, a titer prediction is an estimate of the concentration of the POI in the biological sample (or another suitable molecule of interest). As described herein, such a model is referred to as a “POI titer model 325.” In one embodiment, the POI titer model 325 is trained to predict a titer of a POI present in a biological sample. The POI titer prediction module 320 encodes FTIR spectroscopy data into a feature vector using the techniques described above. In particular, the POI titer model 325 processes a feature vector encoded from FTIR to generate a prediction regarding a titer of a POI present in a biological sample corresponding to the FTIR data. In some embodiments, the feature vector may include encoded features corresponding to the FTIR data. The encoded feature vector is input to the POI titer model 325.

[0105] Examples of input features extracted from the FTIR samples that are encoded into a feature vector include, but are not limited to the FTIR wavenumbers, timepoint or hour at which a biological sample was taken, scale of the sample measurement, the type of strain used for the fermentation run, and an identity of one or more present POI’s.

[0106] The POI titer model 325 predicts a titer of a POI identified as present in a sample corresponding to the vectorized FTIR sample based on inputs of an encoded feature vector. The POI titer model 325 may be a mathematical function of another more complex logical structure, trained using a combination of features stored in the training in the data store 340 to determine a set of parameter values stored in advance and used as part of the prediction24 41522-58976 / WOgeneration. In some embodiments, the POI presence model 325 is a predictive model, for example any of the predictive models described above, or a neural network, for example any of the neural networks described above. In some embodiments, the particular predictive model may be selected using HPLC quantification of POI as a response variable. In some embodiments, the POI titer model 325 is trained to predict a titer of one or more groups of proteins (or any other suitable group of molecules).

[0107] In some embodiments, the POI titer prediction module 320 implements a probability threshold to tune the sensitivity of the trained POI titer prediction module. The POI titer prediction module 320 determines whether a predicted POI titer matches the actual POI titer. For example, the probability threshold can be tuned to a sensitivity of 1%, 2%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 55%, 60%, 65%, 70%, 75%, 80%, 85%, 90%, 95%, 98% or 99%. In some embodiments, the POI titer prediction module 320 adjusts the probability threshold if the accuracy, sensitivity or specificity falls below a predefined adjustment threshold. In some embodiments, POI titer prediction module 320 applies the adjustment threshold to determine the parameters of the training period. For example, if the accuracy of the probability threshold falls below the adjustment threshold, the POI titer prediction module 320 can extend the training period and / or require additional FTIR data and / or POI titers. In some embodiments, the training data includes additional measurements and / or POI titers, which may be used to refine the training dataset.

[0108] The prediction generated by the POI titer model 325 is a measure of the titer of a POI in a given sample of FTIR data. In some embodiments, the POI titer model 325 generates a prediction for multiple POFs present in a biological sample corresponding to the FTIR data. The prediction generated by the POI titer model 325 is a predicted titer measurement for a POI present in the biological sample. The prediction is an estimate of POI titer that would be present in an actual biological sample under the same conditions encoded in the feature vector. Accordingly, the POI titer model 325 outputs a single value representing a predicted titer measurement for a given POI.

[0109] The fermentation metric module 330 applies a trained machine-learning model to a set of FTIR spectroscopy data to determine one or more fermentation metrics based on the POFs present in the FTIR data and the corresponding titer of each present POI. As described herein, such a model is referred to as a “fermentation metric model 335.”. The one or more fermentation metrics may include, but is not limited to, a predicted concentration for one or more analytes of interest. Examples of analytes of interest include, but are not limited to, methanol, ammonia, potassium, exopolysaccharide (EPS), hydrogen peroxide, organic acids,25 41522-58976 / WOvitamins, carbohydrates, minerals, or fats. In one embodiment, the fermentation metric model is trained to predict a titer (e.g. concentration) of one or more analytes of interest. The fermentation metric model 335 encodes FTIR spectroscopy data into a feature vector. The fermentation metric module 330 encodes FTIR spectroscopy data into a feature vector using the techniques described above. In particular, the fermentation metric model 335 processes a feature vector encoded from FTIR to generate a prediction regarding a titer of an analyte of interest. In some embodiments, the feature vector may include encoded features corresponding to the FTIR data. The encoded feature vector is input to the POI titer model 325.

[0110] Examples of input features extracted from the FTIR samples that are encoded into a feature vector include, but are not limited to the FTIR wavenumbers, timepoint or hour at which a biological sample was taken, scale of the sample measurement, the type of strain used for the fermentation run, and an identity of one or more present POFs. In one embodiment, the fermentation metric model 335 is trained using a first training set of samples corresponding to EPS and methanol and a second training set corresponding to ammonia and potassium. To predict a titer for EPS, the fermentation metric module 330 may encode a feature vector with a feature representing phosphate levels in the sample as a latent variable in the feature vector.[OHl] The fermentation metric model 315 predicts a titer of one or more analytes of interest in a biological sample corresponding to the vectorized FTIR sample based on inputs of an encoded feature vector. The fermentation metric model 335 may be a mathematical function of another more complex logical structure, trained using a combination of features stored in the training in the data store 340 to determine a set of parameter values stored in advance and used as part of the prediction generation. In some embodiments, the fermentation metric model 335 is a predictive model, for example any of the predictive models described above, or a neural network, for example any of the neural networks described above.

[0112] The prediction generated by the fermentation metric model 335 is a measure of the titer of an analyte of interest in a given sample of FTIR data. In some embodiments, the fermentation metric model 335 generates a prediction for multiple analytes of interest in a sample of FTIR data. In some embodiments, the fermentation mode

[0113] The prediction generated by the fermentation metric model 335 is a measure of the titer of an analyte of interest in a given sample of FTIR data. In some embodiments, the fermentation metric model 325 generates a prediction for multiple analytes present in a26 41522-58976 / WObiological sample corresponding to the FTIR data. The prediction generated by the fermentation metric model 325 is a predicted titer measurement for an analyte of interest present in the biological sample. The prediction is an estimate of analyte titer that would be present in an actual biological sample under the same conditions encoded in the feature vector. Accordingly, the fermentation metric model 325 outputs a single value representing a predicted titer measurement for a given analyte of interest.

[0114] The model training pipeline 350 trains the machine-learning software module. The techniques described herein for training a machine-learning model may be applied to one or more of the machine-learning models applied by the modules 310, 320, and 330. The model training pipeline 350 is configured to provide training data in the training dataset 340 to the machine-learning software. The training dataset 340 includes training entries of previously labeled FTIR spectroscopy data and one or more ground truth labels. The training dataset 340 includes FTIR data representing a wide range of conditions that provide a broad training dataset for the machine-learning models. For example, the training dataset 340 includes a prescribed number of measurements collected during a set period. These measurements can additionally be prescribed as having a set amount of time between measurements. In one embodiment, each entry of the training dataset 340 includes a ground truth label identifying POIs present in the FTIR data. In another embodiment, each entry of the training dataset 340 additionally or alternatively includes a ground truth label identifying a titer of the POI. Some entries in the training dataset 340 further include a ground truth label identifying an exopolysaccharide (EPS) titer. Examples of training data described herein include, but are not limited to, encoded data, encoded features, or metrics derived from FITR data.

[0115] In one embodiment, the training data consists of entries of Fourier transform infrared spectroscopy (FTIR) data labeled with ground truths corresponding to different target outputs. For example, a single entry of FTIR spectroscopy data may include one ground truth label corresponding to the identified POIs and a second ground truth label corresponding to the titer of each identified POI. In one embodiment, the training data is organized into a lookup table, where each entry includes features of the FTIR data and corresponding ground truth labels. In some embodiments, said training dataset 340 includes FTIR data and exopolysaccharide (EPS) titer. During training, each model training pipeline 350 utilizes automatic statistical analysis of data to determine which features to extract and / or analyze from FTIR data. In this manner, the model training pipeline 350 tunes parameters and hyperparameters of the machine-learning model to output an accurate prediction.27 41522-58976 / WO

[0116] Training data input to the machine-learning model is used, in some embodiments, to construct a hypothesis function to determine an output, for example to determine a predicted POI or titer of POI. The output of the model either reinforces or contradicts the hypothesis function with respect to the data upon which the hypothesis function was constructed. Depending on how close the output tends to be to an outcome determined by the hypothesis function, the machine learning algorithm will either adopt, adjust, or abandon the hypothesis function with respect to the data upon which the hypothesis function was constructed. As such, the machine learning algorithm described herein dynamically learns through the training phase what characteristics of an input (e.g. data) are most predictive in determining whether the features of FTIR data are associated with a particular POI or POI titer.

[0117] Training begins when the model training pipeline 350 applies a machine learning model (e.g. the models of the modules 310, 320, and 330) to FTIR data of a training data and asks the model to generate a prediction (e.g. predict POI or POI titer). During training, the model training pipeline 350 determines parameter values for each feature vector of the FTIR sample by analyzing and recognizing correlations between the features associated with the sample and the prediction generated by the model. Each entry of the training data includes a ground truth label characterizing the target prediction (e.g. the actual protein identity or the actual POI titer). The model training pipeline 350 compares the prediction generated by the model to the ground truth label corresponding to the entry of the FTIR data. During training, one or more optimization techniques (e.g. gradient descent and backpropagation) are used to update the weights in each layer of the model to produce closer agreement between the prediction generated by the machine learning label and the actual label. In some embodiments, the model training pipeline 350 trains each of the models 315, 325, and 335 using a loss value (e.g. using a loss function, cost function, etc.). The loss value indicates how well the prediction matches the true label of a training example. To generate a loss value, the model training pipeline 350 compares the prediction generated for an FTIR entry of the training dataset 340 with the labeled ground truth prediction for the entry. Each model 315, 325, 335 adjusts the model’s parameters (e.g. weights in neural networks) to minimize the loss function such that the prediction generated by the model more closely approximates the desired target value. Each model adjusts the parameters and hyperparameters of the model based on the loss value and the parameter values are stored for later use in generating model predictions. For example, the model training pipeline 350 defines a target for the POI presence model 315 as POI identified as present within each entry in the dataset. The model28 41522-58976 / WOtraining pipeline 350 calculates the difference between the ground truth POI’s present in the sample and the POI’s present in the prediction generated by the model 315 and modifies the parameters of the model 315 such that the model’s prediction more closely approximates the ground truth label. As another example, the model training pipeline 350 defines a target for the POI presence model 315 as the POI titer determined for each entry in the test set. The model training pipeline 350 calculates the difference between the ground truth POI titer for the sample and the POI titer predicted by the model 320 and modifies the parameters of the model 320 such that the model’s prediction more closely approximates the ground truth label. This training process is performed iteratively with new FTIR data until the model generates predictions with a desired level of accuracy.

[0118] In some embodiments, the model training pipeline 350 trains the 315, 325, and 335 through supervised learning. During supervised learning, the model training pipeline 350 divides the training dataset 340 into a training set, a test set, and a validation set. In other embodiments, the model training pipeline 350 divides the training dataset 340 into a training set and a validation set. During training, the model training pipeline 350 specifies a target that contains the correct classification of each entry in the test set. For example, the model training pipeline 350 defines a target for the POI presence prediction model of the module 310 as the POI presence determined for each entry in the test set. The model training pipeline 350 presents a training set of FTIR data (stored in the training dataset 340) to the POI presence prediction model. For each sample presented during training, the output generated by the model is compared with the desired target (e.g. the ground truth label of the entry). The model training pipeline 350 calculates the difference between the target and the set of input samples and modifies the POI presence prediction model such that the output more closely approximates the desired target value. When new input data, not used during training, is presented to a machine trained machine-learning model, the model generates an output classification value indicating which of the categories the new sample is most likely to fall into. Accordingly, the machine learning software module is said to be able to “generalize” from its training to new, previously unseen input samples. This feature of a machine learning software module allows it to be used to classify almost any input data which has a mathematically calculable relationship to the category to which it should be assigned.

[0119] In some embodiments of the machine training software module described herein, the modules 310, 320, and 330 utilize a simulated training model. As described herein, a simulated training model is a machine-learning model trained at least in part on simulated FTIR. In some embodiments, the model training pipeline 350 adjusts the use of training29 41522-58976 / WOmodels as the availability of FTIR data changes. For instance, a simulated training model may be used if there are insufficient quantities of FTIR data available for training the machine training software module to a desired accuracy. In some embodiments, additional data becomes available, the model training pipeline 350 changes one training model to a global or individual model. In other embodiments, the model training pipeline 350 uses a mixture of training models to train the machine-learning models, 315, 325, and 335. For example, the model training pipeline 350 may use the simulated and global training model, utilizing a mixture of laboratory-obtained FTIR data and simulated data to meet training data requirements.

[0120] In some embodiments, the model training pipeline 350 trains the models 315, 325, and 335 through unsupervised learning. Unsupervised learning, in some embodiments, includes applying input data to a machine-learning model and allowing the model to extract features from the input data. The extracted features may be used for visualization, for classification, for subsequent supervised training, and more generally for representing the input for subsequent storage or analysis. In the embodiments described herein, each training case may consist of a plurality of FTIR data. Examples of machine-learning models commonly used for unsupervised training include, but are not limited to, k-means clustering, mixtures of multinomial distributions, affinity propagation, discrete factor analysis, hidden Markov models, Boltzmann machines, restricted Boltzmann machines, autoencoders, convolutional autoencoders, recurrent neural network autoencoders, and long short-term memory autoencoders. While there are many unsupervised learning models, all unsupervised learning models require a training set consisting of FTIR spectra data of biological samples without associated labels.

[0121] As predictions generated by the models 315, 325, and 335 are verified by manual operators, the training dataset 340 may be continuously updated with entries pertaining to newly recorded FTIR data. Additionally, the training dataset 340 may be periodically updated with entries of novel FTIR samples or novel features extracted from existing FTIR samples in the training dataset 340. Accordingly, the models 315, 325, and 335 continue to learn and refine its parameter values based on the new and updated dataset 340. Iteratively re-training models 315, 325, and 335 in the manner described above allows the models to more accurately generate predictions regarding an FTIR sample.

[0122] In some embodiments, the model training pipeline 350 trains the models 315, 325, and 335 on the entire mid-infrared region of the FTIR spectra, which ranges from approximately 3000 - 650 cm’1. In other embodiments, the model training pipeline 350 trains30 41522-58976 / WOthe models 315, 325, and 335 on specific regions of the FTIR spectra, for example the amide region which ranges from approximately 1800 - 1200 cm’1, the carbohydrate region which ranges from approximately 1100 - 900 cm’1, or the fingerprint region of the FTIR spectrum. The model may perform with better median absolute test error when trained on a particular region of the FTIR spectrum compared to a model trained on the entire mid-infrared spectrum. Accordingly, in some embodiments, the models 315, 325, and 335 are trained using the amide region of the FTIR spectra in the context of zero-short transfer learning, for example to predict titers at a new manufacturing location or using new genetics for our host organism.

[0123] The FTIR data processing module 360 may process FTIR spectroscopy data prior to encoding a feature vector to be input to the models 315, 325, or 335. In some embodiments, the FTIR data processing module 360 performs compression or dimensionality reduction. As described herein, data compression is performed using autoencoder techniques whereas dimensionality reduction refers to matrix-based techniques involving singular value decomposition. The FTIR data processing module 360 performs compression and / or dimensionality reduction by tuning hyperparameters of each machine-learning model to identify the most relevant components of each model.

[0124] In some embodiments, the FTIR data processing module 360 may augment the FTIR data with synthetically generated data comprising simulated FTIR spectra and / or simulated noise. In some embodiments, the FTIR data processing module determines whether to augment the FTIR data with synthetically generated data depending on the volume of data in a training set. The FTIR data processing module 360 may compare the volume of training data (e.g., the number of entries in the training dataset) to a threshold and synthetically generate additional FTIR data if the training data does not satisfy the threshold. Accordingly, the FTIR data processing module 360 may leverage the data augmentation techniques described herein to improve the efficiency with which a machine-learning model is trained. The FTIR data processing module 360 may generate synthetic FTIR spectra data by averaging or adding a plurality of randomly chosen spectra for a common POI. Augmenting FTIR data with synthetically generated data may reduce a validation mean absolute error (MdAE) of the models 315, 325, and 335. The models described herein may be capable of being trained using fewer samples using augmented FTIR data compared to unaugmented FTIR data. For example, the number of samples required for training is reduced by at least 25% to 50% when using augmented FTIR data compared to using un-augmented FTIR data.31 41522-58976 / WO

[0125] In other embodiments, the FTIR data processing module 360 generates synthetic data to enhance the uniformity of the value distribution within the dataset. For example, the FTIR data processing module 360 may identify areas in the FTIR data with gaps in the observed values relevant to the target prediction generated by the model. The FTIR data processing module 360 may fill these gaps by generating synthetic data to ensure a more balanced representation of the target prediction.

[0126] In some embodiments, the FTIR data processing module 360 may compare FTIR data from one or more downstream processing (DSP) steps across one or more fermentation processes. For example, the FTIR data processing module 360 may compare FTIR spectra data at a given DSP step across two fermentation processes. In such embodiments, the FTIR data processing module 360 compares FTIR data collected from different fermentation processes by performing segmented distance analysis (SDA) on the FTIR data. As described herein, SDA is a technique that can be applied to determine a level of similarity between FTIR spectrum data from two different fermentation processes. The FTIR data processing module 360 compares FTIR data collected from a fermentation process to FTIR data collected from a baseline fermentation process to identify high variability regions of FTIR data. In some embodiments, the baseline fermentation process may be defined manually based on parameters set by engineers or operators. In another embodiment, the baseline fermentation process may be defined based on a previous fermentation process that satisfied one or more target outcomes. In another embodiment, the FTIR data processing module 360 defines the baseline fermentation process based on the one or more protein products produced by a fermentation run.

[0127] The FTIR data processing module 360 interprets such high variability regions of FTIR data as being indicative of a possible process excursion (e.g., an extra host protein or an additional protein). As described herein, a process excursion is a deviation from the target outcome of a fermentation process, for example high quantities of host proteins, a presence of EPS, or a compositional difference in the protein of interest. To identify high variability regions, the FTIR data processing module 360 performs a distance comparison of FTIR data from a fermentation process to FTIR data from baseline fermentation process (e.g., computing a Euclidean distance, a Manhattan distance, or a Mahlanobis distance), which characterizes the similarity between two spectra. Described differently, the computed distance represents a measure of deviation between the FTIR data collected from the fermentation process relative to the baseline fermentation process. In some embodiments, the FTIR data processing module 360 normalizes FTIR data using the normalization techniques32 41522-58976 / WOdescribed above and compares the normalized FTIR data to the baseline fermentation process. In some embodiments, the FTIR data processing module 360 compares the computed distance to a threshold distance. If the computed distance exceeds the threshold distance, the FTIR data processing module 360 identifies the region as indicative of a potential process excursion.

[0128] In some embodiments, the FTIR data processing module 360 uses the SDA technique to identify regions of FTIR data collected for a fermentation process to be encoded into feature vectors. For example, the FTIR data processing module 360 may identify a first region of FTIR data that exceeds the distance threshold and a second region of FTIR data within the distance threshold. The FTIR data processing module 360 instructs the modules 310, 320, and 330 to extract feature vectors from the second region that is closer in similarity to the baseline fermentation process.

[0129] The fermentation modification module 370 uses the predictions generated by the models 315, 325, 335 to detect contaminations and / or process excursions in the fermentation process. For example, a process excursion may arise where a fermentation process exhibits high levels of methanol. Examples of process excursions include deviations in methanol amounts, pH shifts, protein degradations, spikes in DO, and generation of unexpected byproducts. In other embodiments, the fermentation modification module 370 uses the predictions generated by the model to determine whether the biological sample should be subject to further DSP steps during the downstream process. For example, if the POI presence model 315 generates a prediction identifying an extra POI that is not expected to be present at the DSP stage, the fermentation modification module 370 may generate and execute a set of instructions to apply further processing steps to the biological sample to remove the extra POI. As another example, if the POI titer model 325 generates a prediction identifying a POI at a higher or lower titer than expected at the DSP stage, the fermentation modification module 370 may generate and execute a set of instructions to apply further processing steps to the biological sample to adjust the titer of the POI. For example, if a biological sample is determined to have additional host proteins or ingredients at the end of a microfiltration step, the fermentation modification module 370 may modify the downstream process such that the sample undergoes additional filtration prior to being spray dried. Alternatively, if POI presence model 315 does not identify any extra POIs and / or the POI titer model 325 predicts an acceptable titer for each identified POI, the fermentation modification module 370 determines that the fermentation process may proceed to completion.33 41522-58976 / WO

[0130] In some embodiments, the fermentation modification module 370 applies the outputs of the models 315, 325, and 335 to monitor the progress of a fermentation process and take corrective actions, for example adjusting set points or ending the fermentation process. In other embodiments, the fermentation modification module 370 may input the titer predictions generated by the module 325 to improve fermentation conditions. The fermentation modification module 370 may additionally apply the titer predictions in a feedback loop to iteratively automate control of the fermentation process. In some embodiments, the fermentation modification module 370 may apply the POI predictions generated by the model 315 and the titer predictions generated by the model 325 to onboard new manufacturing sites.

[0131] FIG. 2C is a flowchart illustrating a process for predicting one or more proteins of interest in a biological sample and titers of each protein of interest in the biological sample, in accordance with one or more embodiments. The fermentation monitoring platform 140 accesses 405 FTIR spectroscopy data collected for a biological sample, for example using an FTIR spectroscope. The fermentation monitoring platform 140 encodes 410 a feature vector by extracting attributes from the FTIR data. The fermentation monitoring platform 140 applies 415 a first machine-learning model to identify one or more proteins of interest in the biological sample. The first machine-learning model outputs a confidence score characterizing a likelihood that each protein of interest is present in the biological sample. The fermentation monitoring platform 140 applies 420 a second machine-learning model to an extracted feature to predict a titer of each protein of interest present in the biological sample. The second machine-learning model outputs a predicted titer measurement of each identified protein of interest based on the features encoded in the feature vector.

[0132] FIG. 2D is a flowchart illustrating a process for training a machine-learning model to predict one or more proteins of interest in a biological sample, in accordance with one or more embodiments. The fermentation monitoring platform 140 trains 425 a machine-learning model to identify proteins of interest in a biological sample based on FTIR data collected for the biological sample. The fermentation monitoring platform 140 trains the model using a training dataset comprising entries of previously collected FTIR data. Each entry is labeled with proteins of interest present in each entry. When insufficient training data is available, the fermentation monitoring platform 140 generates synthetic FTIR data to augment the available training data. The fermentation monitoring platform 140 determines 430 a set of parameter values based on labels assigned to entries of FTIR data in the training set. Each parameter value describes a weight associated with features of the FTIR data. The fermentation34 41522-58976 / WOmonitoring platform 140 applies 435 the trained machine-learning model to a test set of FTIR data to identify one or more proteins of interest in the test set.

[0133] FIG. 2E is a flowchart illustrating a process for training a machine-learning model to predict titers of one or more proteins of interest in a biological sample, in accordance with one or more embodiments. The fermentation monitoring platform 140 trains 440 a machinelearning model to predict titers for one or more proteins of interest in a biological sample based on FTIR data collected for the biological sample. The fermentation monitoring platform 140 trains the model using a training dataset comprising entries of previously collected FTIR data. Each entry is labeled with titers of each protein of interest present in each entry. When insufficient training data is available, the fermentation monitoring platform 140 generates synthetic FTIR data to augment the available training data. The fermentation monitoring platform 140 determines 445 a set of parameter values based on labels assigned to entries of FTIR data in the training set. Each parameter value describes a weight associated with features of the FTIR data. The fermentation monitoring platform 140 applies 450 the trained machine-learning model to a test set of FTIR data to predict titers of one or more proteins of interest in the test set.

[0134] FIG. 2F is a flowchart illustrating a process for identifying one or more process excursions, in accordance with one or more embodiments. The fermentation monitoring platform 140 obtains 455 FTIR data collected during a baseline fermentation process and FTIR data collected during a test fermentation process. The baseline fermentation process is a fermentation process that satisfies one or more target objectives for efficacy. The fermentation monitoring platform 140 determines 460 a distance measurement between a region in the FTIR data collected from the test fermentation process and a corresponding region in the FTIR data collected during a baseline fermentation process. The fermentation monitoring platform 140 determines a Euclidean distance, a Manhattan distance, or a Mahlanobis distance. Accordingly, the distance measurement characterizes a deviation of the test fermentation process from the baseline fermentation process. The fermentation monitoring platform 140 identifies 465 one or more process excursions in the test fermentation process based on a comparison of the distance measurement to a threshold. If the distance measurement exceeds the threshold, the fermentation monitoring platform 140 identifies one or more process excursions. Otherwise, the fermentation process proceeds as planned.35 41522-58976 / WOEXAMPLE EXPERIMENT AND RESULTS - FTIR ON PRE-D SP SAMPLES

[0135] Protein of interest (POI) and exopolysaccharide (EPS) titers may indicate fermentation progress and success. Present methods for acquiring this data may require pulling and submitting samples for HPLC, sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) gels, Bradford total protein assays or Bichiconic Acid Assay (BCA) total protein assay to get a sense of fermentation dynamics. Processing times for these approaches may be significant.

[0136] The following experiment demonstrates that applying machine-learning techniques to Fourier transform infrared red data is an effective approach for performing protein-specific tasks. FTIR not only supports discernment of protein type from spectral data but can predict the titer of specific proteins to within around 2.5 mg / mL of reference high performance liquid chromatography (HPLC) results. The results described herein justify real time monitoring using the ReactIR™ system and suggest that the systems and techniques described herein may provide a fast, inexpensive, and "scalable" alternative to current and conventional methods. The experiment described herein analyzed two proteins in particular, labeled herein as Protein A and Protein B. To understand the context of this experiment, the experiment further explores the feasibility of training the machine learning model only on the amide region (1800 - 1200 cm’1) as compared to the model trained on the entire mid-infrared spectrum region which is from approximately 3000 to 650 wavenumbers (3000 - 650 cm’1).

[0137] The ReactIR spectrophotometer may provide an "in-line" perspective into the composition of samples throughout a fermentation process, which may serve as a proxy to titer and other similar measures. Furthermore, encompassing more than POI and EPS quantification, Fourier transform infrared spectrometry (FTIR) could also potentially attribute sensory changes to physical components and possibly inform understanding of cell efficiency, driving fermentation parameter optimization. FTIR could provide concentrations of all molecules of interest that absorb infrared light, which could provide even more metrics for deciding if a fermentation process is progressing as anticipated.

[0138] This experiment considers the following hypotheses to understand the feasibility of using FTIR in protein-specific tasks. This experiment explores the feasibility of the ReactIR spectrophotometer with the application of many different machine learning algorithms trained on data in the mid-IR region (e.g., 3000-650 cm’1wavenumbers).

[0139] Hypothesis 1: FTIR on pre-down stream process (DSP) samples from multiple POIs can identify the POI in the sample. If possible, this would suggest that FTIR can provide POI-specific signals resilient to noise.36 41522-58976 / WO

[0140] Hypothesis 2: FTIR on pre-DSP samples from multiple POIs can predict POI titer. If possible, this would suggest signal extraction across multiple proteins resilient to noise for protein-specific measures.

[0141] This analysis uses FTIR spectra paired with HPLC chromatograms from various time points and reactors for two-liter fermentations. Additionally, this experiment includes one fermentation at the 250-mL scale across all time points. Despite a wide variety of strains and conditions, the dataset skews towards (Protein A) at 78.6% of samples. In total, this experiment uses 589 spectra with associated HPLC data. Twenty-five randomly chosen Protein B spectra and 25 randomly chosen Protein A spectra are set aside for a validation set with a similar procedure for a hidden test set, leaving 489 samples for training.

[0142] Traditional chemometric approaches to making models based on FTIR "trim down" datasets, selecting the most featured wavelengths and eliminating those deemed by an expert as uninformative. However, this experiment models without pre-processing to "learn" this expertise. This in mind, consider that the samples collected are relatively simple: they are all fermentation samples with more or less the same major components, and they are collected paying extra mind to collecting quality data with no baseline shifts. Of course, scale may prohibit this cleanliness. That in mind, this experiment centers on an exploration of feasibility on "clean" samples to start but future work may focus on these kinds of "dirty" samples.1 CLASSIFICATION MODELS

[0143] With 589 wavenumbers and the time point serving as inputs, both AdaBoost and an intuitive visual inspection address Hypothesis 1.1.A ADABOOST

[0144] With classification trees as the "weak learners," the AdaBoost ensemble method lends higher weights to incorrectly classified samples to encourage focus on better performance in these hard to classify observations. From each weak learner trained, the resulting weighted error rate is computed and, from this, individual weights of each of the observations are updated for the next weak learner. The two main hyperparameters to tune are the depth of each tree in the ensemble — that is how many features a binary decision is made upon — and the maximum number of estimators at which boosting is terminated. When choosing the depths of trees to be used in AdaBoost it is recommended to use shallow trees. Specifically, this experiment tunes over trees with depths of 1, 5, 10, 15, and 20. For the number of trees in the ensemble, this experiment selects a range of 10 to 110 trees to tune37 41522-58976 / WOover. With a "best" classification model chosen, the model is trained on both the training and validation data. Using this complete model, the hidden test estimates how this model will perform on data it has never seen before.1.B SYNTHETIC DATA GENERATION

[0145] As described above, this experiment finds POI imbalance in its dataset, possibly causing issues in classification algorithms. To understand this problem, consider that a model could achieve a "decent" accuracy of 78.6% by always predicting the majority class, possibly preventing actual learning by getting "stuck" on this unhelpful model. This in mind, consider two classic solutions to class imbalance: collect more data of the underrepresented class (infeasible due to cost and time of new experiments) or employ an oversampling undersampling technique. Due to limited dataset size, this experiment focuses on oversampling and specifically Synthetic Minority Over-sampling Technique (SMOTE) which creates synthetic minority class examples by examining the k nearest neighbors in the "feature space" before choosing a random point between those neighbors. Applying SMOTE increases the training set of 489 data points (of mostly Protein A) to 752 data points of equal proportions of Protein A and Protein B. This experiment does not apply SMOTE to the validation and test sets which are balanced by selection.2 REGRE S SIGN MODEL S

[0146] To address the second hypothesis, this experiment employs several types of regression models of varying complexity with the 589 wavenumbers and the identity of the POI served as inputs.2. A ELASTIC NET REGRESSION

[0147] A combination of both Ridge Regression and Least Absolute Shrinkage and Selection Operator (LASSO) Regression, Elastic Net can both regularize coefficients and perform feature selection. The hyper-parameters that are relevant to tune for the Elastic net are the LI ratio and the alpha parameter. The alpha parameter defines how strong the Lagrangian penalties are. The Li ratio is tuned over the range of 0 to 1 and the alpha parameter are tuned with equally spaced values from 0 to 20.2.B MULTIVARIATE ADAPTIVE REGRESSION SPLINES (MARS)

[0148] This method offers flexible regression modeling useful for high dimensional data. By making use of the hinge function as spline basis functions, MARS™ automatically searches for interactions and non-linear relationships in a high dimensional basis space.38 41522-58976 / WO2.C SUPPORT VECTOR REGRESSION (SVR)

[0149] Support Vector Machines adapted to the regression problem inherit some of the beneficial properties from Support Vector Machine classifier. SVR still makes use of support vectors but, the inverse of the classification case, they define the regressed hyperplane by being far from the hyperplane. Thus, extreme data points in the dataset help in defining the hyperplane. Furthermore, SVR supports non-linearity via kernelization. As a hyperparameter in tuning, this experiment considers the polynomial kernel at various degrees and the radial basis function (RBF). Additionally, two more hyperparameters tuned are considered: epsilon which specifies the region within which no penalty is associated with points predicted within a distance epsilon from the actual value and inverse regularization ("cost").2.D RANDOM FOREST REGRESSION

[0150] Random Forest Regression is an ensemble of regression trees each with a randomly limited number of features, decorrelating the trees and leading to further variance reduction. This experiment tunes hyperparameters: the depth of the trees and the number of trees in the ensemble as well as how many randomly selected features are available to split on at each step and the minimum number for a split.2.E GENERALIZED ADDITIVE MODELS (GAM)

[0151] GAMs add flexibility to linear regression models via a linear combination of "nonparametric" functions. Expressing each of these nonlinear functions as an expansion of basis functions, GAM can be fit via least squares regression. Though they offer a couple of hyperparameters beyond the scope of this experiment, this experiment only tunes the number of splines (4 to 12).2.F ADABOOST WITH REGRESSION TREES

[0152] Like the AdaBoost for classification, AdaBoost for regression employs regression trees as weak learners. In addition to the depth and the number of trees in the ensemble, the learning rate is also tuned for this set of models.2.G PARTIAL LEAST SQUARE (PLS) AND PRINCIPAL COMPONENTS REGRESSION (PCR)

[0153] Multiple linear regression (MLR) may help when controlled variables are few and not significantly redundant. However, as MLR assumes features are independent, identically distributed, and clearly relevant to predicting the supervisor, Partial Least Squares (PLS) and Principal Components Regression (PCR) attempts dimensionality reduction techniques before regressing. This experiment considers both PCR and PLS. Though both extract latent features39 41522-58976 / WOthat account for as large of a proportion of variation among the features as possible, PLS accounts for both the features and supervisors when making its dimensional reduction. As a result, PLS may outperform PCR in datasets where the supervisor is strongly associated with axes that have low variance. In both algorithms, the primary hyperparameter is the number of components to describe the feature space for which this experiment considers 2 to 20.2.H CONVOLUTIONAL NEURAL NETWORK (CNN)

[0154] State-of-the-art neural networks, CNNs show success in many computer vision applications. When performing image processing, the problems are usually in a very high dimensional space — even more so than the FTIR data. Additionally, CNNs (like other neural networks) "learn " lower dimensional representations of the features based on the data. Neural networks in general have a lot of hyperparameters that can be tuned from the architecture of the network to things like learning rate. This experiment focuses on: kernel size used by the CNN, learning rate of the optimizer, types and severity of regularization (dropout and L2), the number of convolutional and fully connected layers.

[0155] Accordingly, this experiment keeps the batch size constant at 64 and uses the Adam optimization algorithm (an extension of stochastic gradient descent). Similarly, the activations for the inner layers are LeakyReLU and the activation of the final layer is linear.3 METRICS FOR MODEL SELECTION

[0156] One of the fundamental choices that needs to be made with either a classification or a regression model is the metric that will be used to evaluate a given model.3.A CLASSIFICATION METRIC

[0157] The aim of the classification model is to give a sense of the feasibility of being able to identify protein type based solely on FTIR spectra and time point. Therefore, this experiment uses validation accuracy for model selection.3.B REGRESSION METRIC

[0158] This experiment uses HPLC quantification of POI as the response variable. As further explored in the discussion, this requires some interpretation as the HPLC reference values may themselves carry some error.4 RESULTS

[0159] This experiment evaluates each model individually.40 41522-58976 / WO4. A CLASSIFICATION MODELS

[0160] The classification model provides a basic task to evaluate if FTIR spectra provide protein-specific information.VISUAL INSPECTION RESULTS

[0161] To support Hypothesis 1, this experiment starts with a visual inspection of Protein B and Protein A FTIR spectra at the final timepoint of an in-house two-liter (2L) fermentation.

[0162] FIG. 3 A illustrates an overlay of FTIR spectra, including traces 510, 515, 520 and 525, in accordance with one or more embodiments.

[0163] This display subtracts water out of all the FTIR spectra for ease of visualization, but the actual modeling does not need to perform this preprocessing step. The amide bonds are labeled based on FIG. 3B.

[0164] FIG. 3B illustrates theoretical amide bonds 360, 370, 380 and their atoms 365, 375 corresponding to the stretches, in accordance with one or more embodiments. The axes in FIGs. 3 A and 3B and are reversed.

[0165] Protein B shows all three amide stretches but Protein A arguably only shows the amide two stretches. With this clear visual difference, protein-specific modeling seems plausible.ADABOOST RESULTS

[0166] Table 1 shows that the top five models give over 85% validation accuracy with most models exceeding 95% training accuracy.Table 1: Summary of Adab post Classifier hyperparameter tunins41 41522-58976 / WO

[0167] A hidden test set determines expected accuracy on a new dataset. When running the AdaBoost classifier with 100 estimators and a maximum depth of 20 trained on both the validation and training set, this experiment achieves a test set accuracy of 78%. It is expected that there would be a drop in accuracy going from the validation set accuracy to the test set accuracy as information "leaks" from the validation set to the training set when training many models. Either way, this model lends evidence for Hypothesis 1.4.B REGRESSION MODELS

[0168] This experiment finds a viable regressor, lending evidence for Hypothesis 2.

[0169] Though squared error loss is a canonical choice, this experiment prefers median absolute error ("MdAE") as a more intuitive metric for model selection not biased by outliers. Specifically, post-hoc analysis finds that most models see relatively uniform error with a consistently small number of outliers at large HPLC reported POI titer. Given that it remains unclear if FTIR or HPLC may be reporting incorrect values at these outliers, median provides a more realistic estimate of performance.ELASTIC NET REGRESSION

[0170] Table 2 illustrates results using Elastic net regression. Using Elastic Net, the top five models give a training median absolute error of around 4.34 mg / mL error. The validation median absolute error is close to 3.67 mg / mL off for all the top penalized regression models.

[0171] Interestingly, / i ratio is small for most of the best models, potentially because of many multicollinear wavenumbers and Ridge Regression being apt in dealing with multicollinearity.42 41522-58976 / WOSUPPORT VECTOR REGRESSION

[0172] Table 3 illustrates results using SVR. Using SVR and tuning over the hyperparameters, the top five models give a training error of around 4.5 mg / mL. The validation accuracy was closer to 4 mg / mL off for all the top penalized regression models.

[0173] Like with Elastic Net, the top five models see similar performance and moderate costs (slight regularization). When tuning with median absolute validation error all the "best" models picked the RBF kernel to add non-linearity. Again, the training median absolute error is slightly higher than the validation median absolute error.RANDOM FOREST REGRESSION

[0174] Table 4 illustrates results using Random Forest regression. Using Random Forest Regression and tuning over the hyperparameters, the top five models give a training error ranging from around 4 to 7 mg / mL. The validation median absolute error is closer to 4.5 mg / mL for the best model.Table 4: Summary of Random Forest Regression hyperparameter tuning43 41522-58976 / WO

[0175] Although a higher training MAE compared to Elastic Net, the validation MAE is only slightly higher using Random Forest Regression. It is also notable that there is less agreement between the validation and training MAE but the training MAE values are not all above the validation MAE values as is the case with the previous models. Further, in addition to using just single estimators, all top five Random Forest Regression models use depths ranging from 21 to 91 splits. Interestingly, some one-split trees (stumps) are used in the top five models as well. This is different from initial intuitions expected from this model (the first set of tuning parameters using depth 50 to 250).GENERALIZED ADDITIVE MODELS

[0176] Table 5 illustrates results using GAM models.ADABOOST

[0177] Table 6 illustrates results from using AdaBoost. Tuning over the hyperparameters for the AdaBoost algorithm using regression trees as the weak learners shows similar performance to the Random Forest Regression. In addition to training the depth and the number of estimators, this experiment also tunes learning rate.Table 6: Summary of AdaBoost hyperparameter tuning44 41522-58976 / WO

[0178] AdaBoost gives better validation of MdAE values than the random forest regression did; what is interesting is that for three of the five top models only one estimator was used. This would imply that boosting offers limited value to the tree method.PARTIAL LEAST SQUARES REGRESSION

[0179] Table 7 illustrates results from PLS regression. For PLS' only tuning parameter, this experiment considers between 1 and 20 components. Though a small number of components, this starting point gives an idea for FTIR compressibility. This in mind, PLS achieves lower error for the training and the validation sets than any other model, but it is not drastically lower. That PLS performs well makes sense as it is a standard modeling tool for a variety of spectroscopic methods.

[0180] However, taking the traditional approach to PLS may require one or two components on pre-processed signals, as the experiment attempts to predict one component. This statement uses the idea of latent variables, but interestingly this experiment suggests that keeping more components would be advantageous.PRINCIPAL COMPONENTS REGRESSION

[0181] Table 8 illustrates results from PCR. PCR performs similarly to PC A and the only hyperparameter is again the number of principal components.45 41522-58976 / WOTable 8: Summary of PCR hyperparameter tuninsCONVOLUTIONAL NEURAL NETWORKS

[0182] Table 9 illustrates results from convolutional neural networks (CNNs). CNNs offer much flexibility in terms of hyperparameters and this experiment provides a large sweep over hundreds of models. The top five use 200 epochs, a batch size of 64, LeakyReLU, dropout of 0.1, and learning rate of 10'5.Table 9: Summary of CNN hyperparameter tunins

[0183] The next experiment explores the feasibility of the ReactlR spectrophotometer with the application of many different machine learning algorithms trained on data in the amide region (1800 - 1200 cm’1) for protein A alone.

[0184] The experiment further considers using the model trained on protein A to predict the titer on protein of interest for protein B using a transfer learning approach where in the pre-train models is used to predict titer on a new protein.

[0185] Further, this experiment does not include examples of protein B in the training and validation sets, and spectra from protein B are only considered in the test set.46 41522-58976 / WO

[0186] In one embodiment, this analysis uses FTIR spectra from the amide region (1800 - 1200 cm’1) for protein A paired with HPLC chromatograms from various time points and reactors for two-liter fermentations, 250-mL fermentations, and larger scale fermentations from contract manufacturing locations. In total, this experiment uses 180 spectra with associated HPLC data for protein A.

[0187] 36 randomly chosen Protein A spectra are set aside for a validation set, leaving144 samples for training.

[0188] In one embodiment, the transfer learning method includes FTIR spectra from the amide region for protein B paired with HPLC chromatograms for 23 samples in the hidden test set.

[0189] Furthermore, the experiment considers training machine learning models that are agnostic to the type of protein, using one or more machine learning models.

[0190] In one embodiment, the protein agnostic machine learning model trained on the amide region consistent of total of 768 FTIR spectra for both protein A and protein B paired with HPLC chromatograms from various time points and reactors for two-liter fermentation, 250-mL fermentations, and larger scale fermentations from contract manufacturing locations. In total, this experiment uses 393 spectra with associated HPLC data for protein A, 67 spectra with associated HPLC data for protein B and 308 spectra associated with other protein.

[0191] 77 randomly chosen spectra are set aside for a validation set with a similar procedure for a hidden test set (77 randomly chose spectra), leaving 614 samples for training.6 RESULTS

[0192] This experiment evaluates each model individually.TRANSFER LEARNING RESULTS

[0193] This experiment finds that transfer learning where in a pre-trained model trained on protein A and used to predict on protein B results in a model with median absolute error of 2.29 mg / mL and with an R2of 0.746 for the HPLC POI measures.PARTIAL LEAST SQUARES REGRESSION

[0194] Table 10 illustrates results from PLS regression for the model trained to predict HPLC POI trained based on just the amide region of the FTIR spectra. For PLS' only tuning parameter, this experiment considers between 1 and 20 components. Though a small number of components, this starting point gives an idea for FTIR compressibility. This in mind, PLS achieves lower error for the training and the validation sets than any other model, but it is not47 41522-58976 / WOdrastically lower. That PLS performs well makes sense as it is a standard modeling tool for a variety of spectroscopic methods.Table 10: Summary of PLS hyperparameter tuning

[0195] However, taking the traditional approach to PLS may require one or two components on pre-processed signals, as the experiment attempts to predict one component. This statement uses the idea of latent variables, but interestingly this experiment suggests that keeping more components would be advantageous.ELASTIC NET REGRESSION

[0196] Table 11 illustrates results using Elastic net regression. Using Elastic Net, the top five models give a training median absolute error of around 1.46 mg / mL error. The validation median absolute error is close to 1.29 mg / mL off for all the top penalized regression models.Table 11: Summary of Elastic Net hyperparameter tuning48 41522-58976 / WOFEED FORWARD NEURAL NETWORK

[0197] Table 12 illustrates results from the feed forward neural network. Using neural networks offers much flexibility in terms of hyperparameters and this experiment provides a large sweep over hundreds of models. The top five use 1500 epochs, a learning rate of 0.0001, dropout of 0.05 with an MSE (mean squared error) loss criterion..Table 12: Summary of hyperparameter tuning for Feed Forward Neural Network5 DISCUSSION

[0198] In general, these results support both Hypotheses 1 and 2, suggesting that modeling on FTIR can address protein-specific tasks. Still, to understand these findings, this experiment further evaluates the results to provide direction for later work.5.A CLASSIFICATION PERFORMANCE

[0199] The final test set accuracy of 78% for the POI classifier lends evidence for Hypothesis 1 but post-hoc analysis may provide lessons for future modeling.POST-HOC ANALYSIS OF CLASS STABILITY

[0200] Breaking down the test accuracy by type of POI finds some disparity. However, note that Protein B is oversampled and these results may suggest that data augmentation may have limited viability compared to "organic" real data.49 41522-58976 / WOTable 13

[0201] Table 13 shows the predictive ability of AdaBoost model trained on both the training and validation sets. The data was generated from a hidden test set.POST-HOC ANALYSIS OF TIMEPOINT STABILITY

[0202] No test accuracy values drop below 60% at any timepoint. FIG. 4 provides a plot 400 showing test accuracy by fermentation timepoints of the AdaBoost classifier trained on training and validation set, in accordance with one or more embodiments. This disparity does not appear to arise from class imbalance.POI COMPOSITION BY FERMENTATION TIMEPOINT

[0203] FIG. 5 provides a plot 500 showing POI composition by fermentation timepoints of the test set, in accordance with one or more embodiments. For this experiment, differences between these timepoints do not seem to negatively impact modeling.RESPONDING TO HYPOTHESIS 1

[0204] Hypothesis 1 asks if FTIR can distinguish between proteins within FTIR data. For that purpose, these results show that these data can yield protein specific information, lending evidence for Hypothesis 1. This increases hope that FTIR can get more detailed composition information beyond total protein.5.B REGRESSION PERFORMANCE

[0205] Table 14 shows a summary of the models from each of the classes of regression models ordered by validation performance. For the sake of brevity, the top model of each class will only be referred to as the class of model.Table 14: Summary of best of each class of regression model50 41522-58976 / WOADDING TIME AS A FEATURE

[0206] One may also consider adding ancillary features alongside the raw FTIR absorbance values. In the above experiment, the only extra feature is the class of the POI to enable the model to learn faster. However, this experiment also considers adding time an ancillary feature instead of POI. The hyperparameter tuning results are in the following table.Table 15: Summary of PLS hyperparameter titling with timepoint as additional feature

[0207] Table 15 illustrates a summary of PLS hyperparameter tuning, with time points as the only added feature other than raw FTIR absorbances. Ordered in increasing validation MSE.

[0208] Interestingly, these two very similar models chose a different number of components to include when tuning the hyperparameters. There is not a huge difference in terms of the mean absolute error, but what this experiment suggests is that the time points do matter in some way for this model.51 41522-58976 / WODIMENSIONALITY REDUCTION

[0209] The more successful regression models perform dimensionality reduction which may suggest that many wavenumbers associate with compounds that are not POI or total protein. This much dimensional reduction in the univariate regression case may arise because it only requires attention to a small portion of the information in the FTIR signal. There are stretches related to alcohols, acids, saccharides, and other fermentation compounds that absorb infrared radiation, a great benefit to the FTIR signal for multiple uses but, in the above experiment, these extra features may serve merely as noise variables that the model must learn to ignore in predicting HPLC concentrations.REGRESSOR TEST SET PERFORMANCE

[0210] For the purposes of comparison with the more standard MSE metric, PLS on the test set yields an MSE of 22.83, only slightly higher than the validation set MSE of 15.22. This represents an absolute difference of around 2 mg / mL that suggests good generalization.RESPONSE VARIABLES

[0211] This experiment demonstrates that the FTIR spectra contains enough information to predict protein titer with the "true values" coming from HPLC. However, based on previous work, HPLC may include more proteins beyond the POI or have limits in its precision. Indeed, the error observed in the final regressor could possibly reflect some imprecision or inaccuracy in the HPLC and not a modeling limitation.FUTURE RESPONSE VARIABLE

[0212] Other methods may be tracked in real time with this technology: EPS, methanol, ethanol, and glucose may be fermentation "levers" that should be monitored but others such as vitamins or organic acids may prove helpful.6 CONCLUSION

[0213] This experiment demonstrates that FTIR can isolate individual proteins or groups of proteins to automatically generate protein-specific calibration to accurately estimate POI- specific measurements (in this case HPLC-based POI titer). This suggests that FTIR could accurately predict a substantial number of metrics that are labor intensive, doing so as automated "online" data. In fact, these results suggest that it already can do that for the HPLC today with one model serving multiple proteins. Though preliminary, this experiment fundamentally changes what the company believes is possible to monitor in real time with52 41522-58976 / WOlow labor and cost. This experiment also justifies the use of the FTIR to do real time monitoring.EXAMPLE EXPERIMENT AND RESULTS - FTIR DATA AUGMENTATION

[0214] Data augmentation enables more robust predictive FTIR models with smaller datasets. Therefore, in addition to exploring methods for making more robust models, this experiment examines if data augmentation techniques can reduce this dataset requirement. Specifically, this experiment finds that data simulation cuts the number of training samples required by nearly half and that noise generation builds models robust to baseline drift, making models much more resilient with less manual adjustment and input data. Finally, training a Random Forest model with both additive and averaged composite simulated spectra along with noise yields a validation median absolute error of 1.4 mg / mL.

[0215] The literature commonly reports the benefit of FTIR spectral pre-processing before analysis with many techniques proposed for baseline correction. Often researchers use multiplicative scatter correction (MSC) or a standardized normal variate (SNV) approach to correcting baselines. MSC requires "clean" spectra of the components to be a reference for this correction while SNV does not. These techniques aim to address three critical baseline problems common to FTIR spectra: baseline offset, baseline slope, and baseline curvature. This experiment considers MSC and SNV given their previous success.

[0216] This experiment explores the feasibility of data augmentation techniques through application of many different machine learning algorithms. To understand sample set requirements and model resilience to noise, the experiment considers the following hypothesis. These hypotheses outline the feasibility of using data augmentation and noise generation to both improve model performance and to require fewer costly samples for training.

[0217] Hypothesis 1: Augmenting very small training sets achieves good performance in a non-augmented hidden set.

[0218] Hypothesis 2: The simulation of noise by perturbation of spectra baselines yields models robust to noise.

[0219] Hypothesis 3: Simulation of FTIR data provides a robust means of generating data to increase the statistical power of models without additional data collection.1 DATASET

[0220] This analysis uses FTIR spectra paired with HPLC chromatograms from various time points and reactors for internal two-liter fermentations. Additionally, this experiment53 41522-58976 / WOincludes one fermentation at the 250-mL scale across all time points. Despite a wide variety of strains and conditions, the dataset skews towards Protein A at 78.6% of samples. In total, this experiment uses 589 spectra with associated HPLC data. Unless otherwise specified, models predict HPLC POI titer from FTIR spectra.

[0221] This experiment sets aside 25 randomly chosen Protein B spectra and 25 randomly chosen Protein A spectra for a validation set with a similar procedure for a hidden test set, leaving 489 samples for training.

[0222] Many of the Protein A samples at timepoint 3 see titers below the HPLC's limit of quantification (LOQ). As mentioned in the previous FTIR experiment, this is called left censored data as the value is no greater than the LOQ but rather somewhere between zero and the LOQ. As in the previous work, this dataset will have a random concentration that is generated from a uniform distribution between 0 mg / mL and the LOQ of the method.2 DATA AUGMENTATION PIPELINE

[0223] Sweeps examine several data augmentation pipeline permutations. Note that this sweep only selects zero or one algorithm from each part of the pipeline except the preprocessing where search allows for up to two methods.2. A PRE-PROCESSING OF SPECTRAL DATA

[0224] Nearly all chemometric modeling pipelines include pre-processing of spectral data and this experiment examines two approaches. The first, multiplicative scattering correction (MSC) aims to separate the chemical infrared absorption from the physical light scattering during sample collection. This technique usually requires a "reference" spectrum but, as in this experiment, also allows for use of a "mean" spectrum of all the spectra in the dataset. This technique combats baseline shifting and noise generated during a sample collection. Second, both as a standalone pre-processing step and in conjunction with MSC, this experiment considers standard normal variate (SNV) transformation which the literature suggests as equivalent to the MSC transform differing only in scaling factors.2.B NORMALIZATION OF SPECTRAL DATA

[0225] Often in machine learning, the normalization and scaling of data ensures comparability and can help in training, especially for models inherently comparing distances. This experiment considers two methods of normalization. First, the standard scaler works like creating a z-score in statistics: the algorithm subtracts sample mean from each of the observations before dividing by sample variance. Second, the minimum-maximum scaler54 41522-58976 / WOtranslates each feature individually such that all values "scale" between zero and one. Often models will use the minimum-maximum scaler when standard scaling fails.2.B DATA SIMULATION

[0226] This experiment considers three techniques for generating data to reduce median error:Synthetic Minority Over-Sampling Technique for Regression with Gaussian Noise (SMOGN)

[0227] This experiment attempts SMOGN to increase the coverage of the support of the supervisor to be more uniform.

[0228] Addition: Taking advantage of the additive nature of FTIR, this algorithm adds two randomly chosen FTIR spectra (and their response variable) with the same POI. Technically, this method contains a slight error as it does not subtract water, causing an inflated water background. However, as this experiment aims to find a model that is resilient to noise, modeling may still learn from the augmentation.

[0229] Averaging: Like the additive approach, except it will generate samples closer to the middle of the domain.

[0230] Notably, addition and averaging trivially make seemingly unlimited samples from small datasets. Furthermore, these two related data simulation techniques target different parts of the domain: the additive simulation extends the range of the domain while averaging bolsters the central region.2.C ADDITION OF SYNTHETIC NOISE

[0231] As the ReactlR transitions from hand collected single point collections to an "inline" measure, the device may see an increase in noise or baseline shifting.

[0232] When installing the ReactlR in a fermentation tank, there is a lot of mechanical agitation — stirring — that could cause the baseline to shift.

[0233] Fermentation runs last for nearly seven days and, as with most analytical instrumentation, longer running times mean more probability of encountering baseline shift.

[0234] Therefore, in addition to resilience to human error, the ability to manage noise enables broader use of FTIR. With that in mind, this experiment tests adding synthetic noise by creating slope and translations for all the points in the spectrum to improve model resilience.55 41522-58976 / WO2.D DIMENSIONALITY REDUCTION

[0235] Prior internal work with FTIR data saw partial least squares and principal components regression out-perform many other models. Notably, both use dimensionality reduction before regression. Therefore, this experiment explores principal components analysis (PCA), singular value decomposition (SVD), and t-distributed stochastic neighbor embedding (t-SNE).3 MACHINE -LEARNING MODELS

[0236] The following techniques were explored for FTIR:Partial Least Squares: 1 to 20 components.Elastic Net: alpha from 0.001 to 10 and L2 ratio from 0.001 to 1.Random Forest Regression: 1 to 250 estimators with max depths from 1 to 100. AdaBoost: 1 to 250 estimators with max depths from 1 to 100.Convolutional Neural Network: All the various parameters explored in prior work are revisited.

[0237] The metric for model selection can strongly influence results. FTIR data show extreme skew in validation and training error for all the models. For this reason, this experiment uses median absolute error (MdAE) for model and pipeline configuration selection.

[0238] Notably, this experiment's FTIR and HPLC POI titers dataset span a six-month timeframe. In general, generating hundreds of FTIR spectra paired with the appropriate "offline" measure requires large company investment. To determine the "required" number of samples in hope of reducing cost of future models, this experiment randomly chooses samples from the training set to drop. These samples are then augmented and the resulting dataset put through the noise generation function if selected. From these augmented data, this experiment trains a random forest model from which it compares different simulated input dataset sizes through validation median absolute error, looking for the "elbow" to determine the minimum number of samples needed for the training set.4 RESULTS

[0239] This experiment finds that success of data augmentation depends on model type being explored. However, with full data augmentation, this experiment recommends a Random Forest model given its lowest MdAE of 1.41 mg / mL.56 41522-58976 / WO4. A TOP PERFORMING PIPELINES

[0240] First, this experiment considers performance of the pipelines on several machine learning models.PARTIAL LEAST SQUARESTable 16: Summary of best Partial Least Squares pipelines. Ordered in decreasing validation MdAE

[0241] Interestingly, the best improvement comes from a pipeline containing only an additive data generation step. However, the second best model suggests using both SNV and MSC with the addition of noise and a standard scaler. Note that SMOGN does appear in the top ten model pipelines but shows only a 3.33% improvement in validation MdAE.RANDOM FOREST REGRESSION

[0242] Random forest shows notable success, especially in robustness to overfitting.Table 17: Summary of best Random Forest pipelines. Ordered in decreasing validation MdAE.57 41522-58976 / WO

[0243] This method benefits in general from the pre-processing and data generation though with less dependence on scaling. This intuitively makes sense based on how a tree model works: if one scales the dataset by a factor of seven, the same splits will still occur internally just with inflated split values. Notably, not every modeling approach seems to provide this robustness.ADABOOSTTable 18: Summary of best AdaBoost pipelines. Ordered in decreasing validation MdAE.

[0244] However, AdaBoost does not benefit from any of the pipelines with only a modest 1.37% decrease in error compared to the model with no data augmentation for the top 15 models. Only the top five are shown in the above table for brevity. Given so many models tied for first, this model may have hit a "cap" of MdAE without collecting more real data.4.B PERFORMANCE OF DATA AUGMENTATION

[0245] This experiment uses both elastic net and random forest to explore different data augmentation techniques.58 41522-58976 / WOUNAUGMENTED DATA

[0246] Starting with unaugmented data, elastic net with a validation MdAE of 3.67 mg / mL shows reasonable performance throughout but especially good performance on lower titers. However, the model may report negative titers.

[0247] FIG. 8 illustrates a set of histograms 800 of model diagnostics of the Elastic net regression model, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 1.32 mg / mL for the training and validation sets, and 1.56 mg / mL for the test set. The histograms additionally show high R2values (0.94 for training, 0.96 for validation, and 0.91 for test), indicating strong predictive ability for the Elastic net regression model.

[0248] Random forest performs less well in unaugmented data with about 1 mg / mL higher MdAE. In particular, Random Forest predictions seem more variable compared to the Elastic Net though it does not predict negative titers.

[0249] FIG. 9A illustrates a set of histograms 900 of model diagnostics of the AdaBoost regression model, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 0.01 mg / mL for the training set, 1.36 mg / mL for the validation set, and 1.72 mg / mL for the test set. The histograms additionally show high R2values (0.88 for training, 0.89 for validation, and 0.88 for test), indicating strong predictive ability for the AdaBoost regression model.

[0250] FIG. 9B illustrates a set of histograms of model diagnostics of the elastic net regression model trained on the amide region, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 2.066 mg / mL for the training set, 1.805 mg / mL for the validation set, and 2.084 mg / mL for the test set. The histograms additionally show R2values (0.934 for training, 0.954 for validation, and 0.933 for test), indicating strong predictive ability for the Elastic net regression model.

[0251] FIG. 9C illustrates a set of histograms of model diagnostics of the PLS mode trained on the amide region, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 2.05 mg / mL for the training set, 1.849 mg / mL for the validation set, and 2.087 mg / mL for the test set. The histograms additionally show R2values (0.941 for training, 0.956 for validation, and 0.94 for test), indicating strong predictive ability for the PLS model.59 41522-58976 / WO

[0252] FIG. 9D illustrates a set of histograms 3800 of model diagnostics of the feedforward neural network model trained on the amide region, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 1.0.35 mg / mL for the training set, 1.024 mg / mL for the validation set, and 0.973 mg / mL for the test set. The histograms additionally show R2values (0.974 for training, 0.974 for validation, and 0.957 for test), indicating strong predictive ability for the feed forward neural network.

[0253] FIG. 10 illustrates a plot 1000 showing absolute error of each unaugmented model by fermentation time, in accordance with one or more embodiments.

[0254] This divergence may suggest model performance differences in different kinds of compositions. All this in mind, consider that Random Forest offers lower validation error at most timepoints. Additionally, as further examined below, note Random Forest's excellent performance at timepoint six.COMPOSITE AUGMENTATIONS

[0255] For composite augmentations, 400 average and 400 additive spectra are added to the dataset for a total training size of 1245. Starting again with Elastic Net, note that the highest POI titer is predicted to be about 10 % lower after using the augmentations.

[0256] FIG. 11 illustrates a set of histograms 1100 of model diagnostics of a support vector regression model, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 1.55 mg / mL for the training set, 2.14 mg / mL for the validation set, and 1.65 mg / mL for the test set. The histograms additionally show high R2values (0.92 for training, 0.95 for validation, and 0.92 for test), indicating strong predictive ability for the AdaBoost regression model.

[0257] FIG. 12 illustrates a set of histograms 1200 of model diagnostics of a random forest regression model, in accordance with one or more embodiments. The histograms show the error in mg / ml of the training set, validation set, and test set. The histograms show errors of 1.07 mg / mL for the training set, 1.46 mg / mL for the validation set, and 2.20 mg / mL for the test set. The histograms additionally show high R2values (0.95 for training, 0.95 for validation, and 0.92 for test), indicating strong predictive ability for the AdaBoost regression model.

[0258] FIG. 13 illustrates a histogram 1300 showing absolute error of each composite augmented model by fermentation time, in accordance with one or more embodiments.60 41522-58976 / WO

[0259] Random forest with composite augmentations see slightly lower error across all times with more consistency compared to unaugmented data. However, Elastic Net model performs similarly with unaugmented data.NOISE GENERATION AUGMENTATIONS

[0260] This experiment also explores adding noise to models. Starting with Elastic Net, noise appears to cause more conservative titer estimates but performance looks slightly improved.

[0261] FIG. 14 illustrates a histogram 1400 of absolute validation error for the baseline augmented Elastic Net model, in accordance with one or more embodiments. Random forest also predicts lower titers in the presence of noise and, for "true" titers below a threshold, this model sees much better agreement compared to training with unaugmented data. That said, despite overall improved error and density primarily between 0 and 6 mg / mL, some individual instances see more extreme error.

[0262] FIG. 15 illustrates a histogram 1500 of absolute validation error for the baseline augmented Random Forest model, in accordance with one or more embodiments.

[0263] FIG, 16 illustrates a histogram 1600 of absolute error of each composite augmented model by fermentation time, in accordance with one or more embodiments.

[0264] By timepoint, both models perform slightly better with noise, likely due to regularization effect.COMBINED DATA AUGMENTATIONS

[0265] Next, this experiment combines augmentations, introducing 400 additive and 400 average spectra to the training set before baseline noise to yield 3,735 training examples.

[0266] FIG. 17 illustrates a histogram 1700 of absolute validation error for the fully augmented Elastic Net model, in accordance with one or more embodiments.

[0267] FIG. 18 illustrates a histogram 1800 of absolute validation error for the fully augmented Random Forest model, in accordance with one or more embodiments.

[0268] FIG. 19 illustrates a histogram 1900 showing absolute error of each fully augmented model by fermentation time, in accordance with one or more embodiments.

[0269] Although the mean performance looks essentially the same, MdAE for the Random Forest model on the fully augmented dataset sees strong improvements in error with the best fully augmented model having an MdAE of 1.406 mg / mL.61 41522-58976 / WOCOMBINED DATA AUGMENTATIONS AND CNN

[0270] Prior work observes CNNs with just an error of 3.67 mg / mL MdAE. In general, neural networks perform well with large amounts of training examples so this experiment revisits the approach on the fully augmented dataset. Tuning over the hyperparameters, the top five models use 100 epochs of training, a kernel size of 3, a batch size of 64, LeakyReLU, three dense layers, four convolution layers, and learning rate of 10'5. Three out of the top five models show any improvement using the fully augmented training set.Table 20: Summary of CNN hyperparameter tuning with fully augmented training dataPOWER ANALYSIS

[0271] Finally, this experiment turns to simulations for power analysis experiments to find minimum training set sizes for FITR models.

[0272] FIG. 20 illustrates 2000 a bar graph of data augmentation power analysis using Random Forest model, in accordance with one or more embodiments.

[0273] Having fewer than 65 samples appears to cause higher median absolute validation error though with lots of noise. Due to randomness of augmentation also consider relative performance with it removed.NUMBER OF COLLECTED SAMPLES IN SET

[0274] FIG. 21 illustrates 2100 a bar graph of data power analysis without any data augmentation using Random Forest model, in accordance with one or more embodiments. Without augmentation, the suggested training set size reaches closer to 125 samples and the reduction in validation MdAE appears slower on the dataset without augmentation, suggesting less efficiency. Of course, these experiments only consider the training set.Researchers still need a validation and a test set and this experiment recommends a minimum of 35 samples in each, bringing the total sample size requirements for new modeling to 125 samples all together.62 41522-58976 / WO7 DISCUSSION

[0275] Data augmentation offers a sharply lower MdAE compared to unaugmented data with more efficient error reduction in smaller sample sizes. That in mind, this experiment examines the relative performance of these different approaches.7. A DIMENSIONALIT Y REDUCTION OFFERS LITTLE BENEFIT

[0276] Though some models prefer different pipelines compared to others, the data suggest that learning in general does not benefit a dimensionality reduction. Beer's Law suggests the model should use only a handful of features per task; these findings challenge the traditional approach of "cleaned and trimmed" data for FTIR modeling.7.B SWEEP PREFERS MULTI- ALGORITHM BASELINE CORRECTION

[0277] Despite expectations that the sweep would only pick one or the other as these two techniques generally perform the same operation up to a scaling factor, many models benefit from performing both SVN and MSC as pre-processing steps. Considering the engineering ease of performing both algorithms, this experiment recommends their combined use in preprocessing.7.C DATA GENERATION OFFERS BROAD BENEFITS

[0278] Synthetic data generation like adding and averaging spectra broadly offers benefit with Random Forest seeing striking performance improvements. Therefore, this experiment recommends using composite data generations to add density to middle and upper parts of the model domain, especially with composite additive spectra encouraging models to predict higher titers with less imbalance between "high" / "low" values.7.D RANDOM FOREST FEATURE SELECTION

[0279] Notably, the best Random Forest uses only one estimator, effectively choosing a simple regression tree. Exploring the structure of this tree offers some model explainability: The first feature that is split upon is wavenumber 1580, a stretch in the amide two region. The next split two splits are far from the amide stretch region at 2780 wavenumbers and 824 wavenumbers, respectively. A researcher using the traditional trimming approach to just look at the amide regions would ignore these two splits. After the top of the tree, the model keeps selecting features within the amide bond region before focusing on other parts of the spectra. This type of "jumping around" appears to be happening throughout all the splits.

[0280] These splits suggest that traditional methods likely ignore information useful for regression problems but also require some caution as it remains less clear what specific63 41522-58976 / WOinformation the model extracts from those other areas. Regardless, this may explain why dimensionality reduction failed to offer benefit. This experiment calls for future investigation into the utility of these other regions.7.E SAMPLE SIZE REQUIREMENTS

[0281] In addition to helping with regularization and generalization, data augmentation techniques appear to reduce required sample sizes by about half to achieve similar performance (65 vs 125 samples). Therefore, this experiment recommends targeting around 65 training samples to start FTIR modeling for new metrics (125 in total due to validation and test sets).7.F CONCENTRATED SAMPLE TITER ERROR

[0282] The random forest model has a lower correspondence to HPLC data when samples have higher POI titers. FIG. 23 illustrates a bar plot 2300 showing, median absolute training, and validation error, and the training and validation R squared for the augmented and unaugmented datasets. Due to the way the FTIR data are trained, this could possibly (though certainly not yet demonstrated) under certain assumptions indicate that host protein could account for some of the difference between FTIR and HPLC as well. Even without ML, one certainly can make a strong case that the current HPLC approach includes host protein in its POI titer estimation. It was unexpected that this type of behavior is seen in the FTIR model and future work will focus on this discrepancy.8 CONCLUSION

[0283] Using FTIR HPLC POI titer prediction as an example, this experiment demonstrates that data augmentation can fortify experimentally collected data. Specifically, while still requiring resource allocation for data collection prior to modeling, this experiment recommends at least 125 total samples, a much smaller threshold than prior work's nearly 600 samples. This allows less resources to be spent in data collection, offering faster model development at lower cost. Therefore, this experiment recommends application of these techniques on exopolysaccharide and other infrared active analytes of interest. Beyond reduced dataset requirements, this experiment also suggests that these techniques improve model generalization: not only do noise and simulated data reduce the validation MdAE but they also serve as another means of regularization. This experiment selects unexpected configurations for those techniques, notably with combined MSC and SVN for baseline correction, suggesting the utility of these sweeps in future problems. All that said,64 41522-58976 / WOaugmentation does not impact all models equally. Some like the CNN or Partial Least Squares remain fairly unaffected by the augmentation whereas Random Forest sees sharp benefits, allowing it to achieve the best observed HPLC POI titer prediction validation MdAE of 1.40 mg / mL compared to Partial Least Squares' validation MdAE of 1.94 mg / mL. For context, this 1.4 mg / mL error may be as small as 5% of a final Protein B titer. Regardless, data augmentation provides a path forward not for modeling with fewer physical samples but also creating more robust estimators. Like the AAA model, there is deviation between FTIR and HPLC at very high titers. Although data generation through additive composite spectra helps allay this issue, it is persistent.EXAMPLE EXPERIMENT AND RESULTS - MULTIPLE ANALYTE PREDICTION

[0284] One of the more promising aspects of using vibrational spectroscopy in fermentation processing is that a myriad of potential metrics can be measured simultaneously from one source. This experiment presents at least five titers that can be measured using Fourier transform infrared (FTIR) spectra, but there has yet to be an exhaustive exploration of the analyte space. Compounds such as organic acids, vitamins, carbohydrates, minerals, and fats are theoretically FTIR active and of fermentation interest.

[0285] With the modeling and data augmentation intuition from the previous two experiments, this experiment explores predicting high performance liquid chromatography (HPLC) and monitoring probe quantification of fermentation analytes of interest. The analytes contained in this experiment will be methanol, ammonia, potassium, and exopolysaccharide (EPS) and they will be measured in milligrams per milliliter, unless otherwise mentioned. For this experiment, only "upstream" fermentation samples will be considered.1 DOWNSTREAM PROCESSING

[0286] Preliminary testing of the Random Forest model shows that this model is not agnostic to the type of process it is monitoring. This is altogether not that surprising as the matrix effects between fermentation samples and sample in the protein isolation processing are likely to be great. To wit, the goal of "downstream" processing is to purify the product, and thus remove any interfering compounds that add noise to the signal in this problem.

[0287] This experiment explores the feasibility of data augmentation techniques through application of many different machine learning algorithms. To understand the utility of these data and characterize sensitivity to noise, consider the following hypotheses. In short, these65 41522-58976 / WOhypotheses help understand the feasibility of using FTIR to predict a variety of fermentation analytes of interest.

[0288] Hypothesis: A similar hypothesis to the upstream process was also explored for downstream processing.2 DATASET

[0289] Additionally, this experiment includes two fermentations at the 250-milliliter scale across all time points. Using the guidance of a previous experiment this experiment creates two training sets, one for EPS and methanol, and the other for the remaining analytes. This experiment sets aside 25 randomly chosen Protein B spectra and 25 randomly chosen Protein A spectra for a validation set with a similar procedure for a hidden test set, leaving 60 samples for training in each of the datasets.

[0290] Many of the Protein A samples at time point 3 see titers below the HPLC's limit of quantification (LOQ). As mentioned in the previous FTIR experiment, this is called left censored data as the value is no greater than the LOQ but rather somewhere between zero and the LOQ. This dataset will have a random concentration that is generated from a uniform distribution between 0 mg / mL and the LOQ of the method.

[0291] The methanol probe data contained no missing data, while both the ammonia and the potassium probes contained around 15% missing data points. The sample sizes were already just below the recommended number of samples needed to model, so no more data could be spared. To deal with the missingness of these data, median imputation was deployed. Median was used over mean as the distributions of each probes data was right skewed.3 DATA AUGMENTATION PIPELINE

[0292] Summarized below, the best aspects of the data augmentation findings of the previous experiment were used in this experiment. Both multiplicative scattering correction (MSC) and the standard normal variate transformation (SNV) were used in series during the pre-processing part of the pipeline.

[0293] Thirty additive composite samples and 30 averaged samples were simulated before noise was added to the dataset. Thus, before noise was added, the experiment used equal proportions of simulated and experimental data in the training set. After data augmentation was conducted on these samples, the resulting number of training examples was 300. This is five times greater the initial number of samples.66 41522-58976 / WO4 MACHINE -LEARNING MODELS

[0294] The following techniques were used:Elastic Net: alpha from 0.001 to 10 and L2 ratio from 0.001 to 1.Random Forest Regression: 1 to 250 estimators with max depths from 1 to 100.AdaBoost: 1 to 250 estimators with max depths from 1 to 100.Partial Least Squares: 1 to 15 components; used only on EPS.

[0295] Of course, the metric for model selection can strongly influence results. As discussed in prior work, FTIR data show extreme skew in validation and training error for all the models. For this reason, this experiment uses median absolute error (MdAE) for model and pipeline configuration selection.4. A HIDDEN TEST SET ANALYSIS

[0296] The "best" model and hyperparameters are chosen using MdAE for each of the analytes of interest before being trained on the combined training and validation set. Using these models trained on all the data, hidden test set analysis is performed to get a sense of how the model performs on new and unseen data. First, the distribution of absolute error and a correspondence plot will be made between the real and predicted analyte titers.Additionally, the model’s performance will be evaluated across different time points and types of POI. This will help detect any systematic error in our model.5 RESULTS

[0297] This experiment finds that the success of data augmentation depends on model type being explored as well. This experiment finds that the best models trained on all the validation and training data all have test MdAE less than 1.5 mg / mL.5. A AMMONIA QUANTIFICATION

[0298] FTIR shows promise in ammonia quantification.Table 21: Best training and validation MdAE values for each model family for the quantification of ammonia.67 41522-58976 / WO

[0299] This dataset only contained in-house, two-liter bioreactor samples with FTIR spectra paired with YSI ammonia broth titers.5.B POTASSIUM QUANTIFICATION

[0300] This dataset only contained in-house, two-liter bioreactor samples with FTIR spectra paired with YSI potassium broth titers.Table 22: Best training and validation MdAE values for each model family for the quantification of potassium

[0301] The Elastic Net and the Random Forest models were set up as bivariate models, where AdaBoost was set up as two univariate models. This is the reason why the MdAE values are the same for Elastic Net and Random Forest in Table 21. It is amazing the impact of data simulation on validation MdAE for the quantification of potassium. Simulated data was only used in the training set, but AdaBoost performed above three times better on validation MdAE. Simulation did not have as large effect on the other models. All that said, the same issue of modeling potassium titer from supernatant samples is the same here as with ammonia.5. C METHANOL QUANTIFICATION

[0302] Pichia pastoris is a methylotrophic yeast, which can be induced with methanol to produce a desired protein using synthetic biology techniques. As one can imagine, this lever of induction is quite an important fermentation metric to monitor. Currently, some of the two- liter bioreactors are equipped with methanol broth concentration sensors, which this data was trained upon.Table 23: Best training and validation MdAE values for each model family for the quantification of methanol.68 41522-58976 / WO

[0303] It is important to keep in mind that methanol only took on a range of values from 0 to 1 mg / mL broth methanol titers. Thus, AdaBoost having 0.093 mg / mL validation MdAE does not imply that the percentage error on methanol is small. In the middle of the range at 0.5 mg / mL, the average MdAE would be 18.60 % of the value.5 D EPS QUANTIFICATION

[0304] An analytical method of measuring EPS is an HPLC method that takes about 30 minutes per samples, resulting in long turnaround times and heavy HPLC utilization.Table 24: Best trainins and validation MdAE values for each model family for the quantification of EPS.

[0305] Partial least squares was considered for EPS as it was successfully used in the past when purchasing the ReactlR. However, a big caveat to this modeling is that the EPS peaks are obfuscated by the presence of phosphate in our sample. This experiment attempted to use phosphate concentration as a latent variable, instead of estimating its concentration and subtracting spectra as was done in previous work.

[0306] The result of treating phosphate as a latent variable worked easier than planned, and no extra work accounting for this needed to be done. At the middle of this range the best validation MdAE from AdaBoost would give an error of less than 0.5 %. At worst, this is only a 10 % error, which is modest.5.E MULTIPLE ANALYTE HIDDEN TEST SET ANALYSIS

[0307] FIG. 24 illustrates a scatter plot 2400 of predicted test YSI probe ammonia values and true values, in accordance with one or more embodiments.

[0308] From the scatter plot shown in FIG. 24, there is a strong correspondence between the predicted and true ammonia concentrations. There is one point that was predicted to have69 41522-58976 / WOnearly no ammonia in solution, when the YSI probe gave a reading of over 2 mg / mL. This suggests that the model performs a bit better on large ammonia concentrations.

[0309] FIG. 25 illustrates a histogram 2500 showing absolute test error between YSI probe ammonia values and true values, in accordance with one or more embodiments.

[0310] Looking at the histogram it can be seen that 84 % of the test set had an absolute error less than 0.5 mg / mL. Additionally, the test error here was lower than the validation error observed earlier. This implies that gains in test MdAE likely could be made as more experimental data is collected.

[0311] FIG. 26 illustrates a histogram 2600 showing test error between YSI probe ammonia values and true values broken down by fermentation time point, in accordance with one or more embodiments.

[0312] The larger errors are a symptom of ammonia concentrations changing over the time course more than the error increasing.

[0313] The best model for potassium broth concentration prediction was an AdaBoost model with a learning rate of 0.0005, a maximum depth of 13, and an ensemble size of 1 estimator. The test MdAE for this model was 0.590 mg / mL.

[0314] FIG. 27 illustrates a scatter plot 2700 of predicted test YSI probe potassium values and true values, in accordance with one or more embodiments.

[0315] In the middle of the range of values that potassium takes on, this would be a 16.86 % error, likely related to the lack of theoretical underpinnings to measure this compound.

[0316] What is likely occurring is the AdaBoost algorithm finds a compound that is highly correlated with potassium concentration (e.g. its counterion) which is FTIR active.

[0317] FIG. 28 illustrates a scatter plot 2800 showing absolute test error between YSI probe potassium values and true values, in accordance with one or more embodiments.

[0318] FIG. 29 illustrates a scatter plot 2900 showing median test error between YSI probe potassium values and true values broken down by fermentation time point, in accordance with one or more embodiments.METHANOL QUANTIFICATION

[0319] The best model for methanol broth concentration prediction was an AdaBoost model with a learning rate of 0.0005, a maximum depth of 25, and an ensemble size of 21 estimators. The test MdAE for this model was 0.100 mg / mL. In the middle of the range of values that methanol takes on, this would be a 18.60 % error, which is a higher amount of error than an acceptable error.70 41522-58976 / WO

[0320] FIG. 30 illustrates a bar plot 3000 showing absolute validation error between YSI probe methanol values and true values, in accordance with one or more embodiments.

[0321] The skew in error comes mainly from low titer readings. However, there are a couple of high deviation points that appear in larger methanol titers. In general, there appears to be a moderate amount of correlation between these two measures, with lower points having higher variance.

[0322] FIG. 31 illustrates a bar plot 3100 showing median test error between YSI probe methanol values and predicted values broken down by fermentation time point, in accordance with one or more embodiments.EPS QUANTIFICATION

[0323] FIG. 32 illustrates a bar plot 3200 showing absolute error for the test set between HPLC EPS predicted and true values, in accordance with one or more embodiments.

[0324] The best model for EPS broth concentration prediction was an AdaBoost model with a learning rate of 0.005, a maximum depth of 25, and an ensemble size of 41 estimators. The test MdAE for this model was 1.169 mg / mL. In the middle of the range of values that EPS takes on, this would be a 3.59 % error.

[0325] There is a strong correlation that spans the prediction space, at least for the "upstream" process (USP). It will be interesting to see how long this linearity last in the more concentrated samples. A typical DSP sample will generally be at both very low and very high EPS titers by design. If this type of correspondence remains for larger valued titers, DSP sample collection should be able to be done without having to perform any dilutions on the samples further removing busy work for the engineer.

[0326] FIG. 33 illustrates a bar plot 3300 showing median absolute test error between EPS HPLC titer values and predicted values broken down by fermentation time point, in accordance with one or more embodiments.5 F POI CLASS TEST ERRORS

[0327] For this experiment two different datasets were used with very different proportions of Protein B to Protein A. The training set for both the EPS and methanol supervisors, have observations that are mainly Protein A, at 70%. Conversely, the test set that was used for the ammonia and potassium supervisors was mainly Protein B, at 75%. Experimentation with the training and validation set did not show a boost in performance by dealing with this class imbalance.71 41522-58976 / WO

[0328] Regardless, it appears as if this steep imbalance in the training set sizes affected the test set. The test sets were equally stratified by design, but the dominant POI in the training tended to outperform in the test set. As expected, the models that had a larger proportion of a given POI tended to perform better in the test set, even though the test sets were stratified by POI type.6 DISCUSSION

[0329] In conjunction with the previous FTIR modeling experiments, this experiment shows that at least five fermentation analytes could be measured during a fermentation in a live fashion. Additionally, this experiment suggests that EPS could be measured at scale. This experiment examines the relative performance of each of these metrics and then examines establishing a limit of quantification for the methanol FTIR model.6. A FIVE ANALYTES FROM ONE SPECTRUM

[0330] There are five metrics that can be obtained using one FTIR spectrum live during a fermentation. Two of these measures, EPS and POI, are costly HPLC methods that would not be feasible to obtain at our partner manufacturing sites. Below is a summary of the hidden test set error using the best model for the analyte.

[0331] FIG. 34 illustrates a bar chart 3400 showing all metrics for five analytes. Median error in previously unseen upstream samples. Some metrics were trained on small training sets. All these values are below 1.5 mg / mL, FTIR can effectively perform "online" fermentation monitoring today, in accordance with one or more embodiments.

[0332] As previously mentioned, these metrics have a range of deviance from 3 to 18.60 % from in the middle of the concentration range. Most of the metrics are below 10 % deviance, however. The two metrics that are showing large deviation are potassium at 16.86% error in the middle of the range and methanol having nearly 20 % error in the middle of the range. The results follow the YSI broth potassium readings.6.B METHANOL LIMIT OF QUANTIFICATION

[0333] Examining the truncated histogram of methanol prediction deviance from true methanol broth titers above 0.4 mg / mL shows a much better correspondence. The mean absolute error falls to 0.0878 mg / mL. In the middle of the new titer range this is 12.54 % error from the observed YSI broth methanol reading.

[0334] FIG. 35 illustrates a truncated bar plot 3500 of frequency of absolute deviance for YSI probe methanol values, and predicted values, in accordance with one or more72 41522-58976 / WOembodiments. This marked decrease in mean absolute error suggests that the model performance improves without lower titer true methanol values. Further investigation is needed into lower titer spectra, but for now this experiment suggests left-censoring any data that the FTIR predicts below 0.4 mg / mL. There will still be a couple of extraneous deviations, but for the most part this bolsters the model’s performance.

[0335] Scaling offers some positive effect for the Partial Least Squares model and does not hurt the tree-based models. That in mind, a majority of the top performing tree-based models do not use any scaling. Therefore, except for Partial Least Squares, this experiment does not recommend inclusion of scaling.6.C TIMEPOINT SIX

[0336] Many models see a large disparity between techniques at fermentation timepoint six. To understand this, consider that the T6 titers fall higher on the high end of what would be expected for Protein B and one T6 Protein A sample sees over double some of the other titers for the same POI.

[0337] In contrast to elastic net, the mechanics of random forest may offer more flexibility with these data points and can capture the "extreme" values. Although the included table only represents the fully augmented predictions, Random Forest consistently out preforms the Elastic Net at this timepoint regardless of augmentation.6.D RANDOM FOREST FEATURE SELECTION

[0338] Notably, the best Random Forest uses only one estimator, effectively choosing a simple regression tree. Exploring the structure of this tree offers some model explainability.EXAMPLE EXPERIMENT AND RESULTS - SEGMENTED DISTANCE ANALYSIS

[0339] Segmented Distance Analysis (SDA) is a novel but simple technique to perform classical modeling on FTIR spectra against a new metric. In some embodiments, only some segments of the FTIR spectra are truly of interest. In some embodiments, this is used to compare the spectra between different spectra by normalizing the entire spectra based on Amide I peak and then performing simple distance comparisons to use as a proxy for “how similar these two spectra are”.

[0340] FIG. 39 illustrates the segmented distance analysis showing the Euclidean distance between various spectra at different steps of the fermentation and downstream process, in accordance with one or more embodiments. Plot 3900 illustrates three fermentation processes: Process 1, Process 2, Process 3. At wavenumber 1500, FTIR73 41522-58976 / WOspectrum data for Timepoint 2 and Timepoint 3 deviates from the Timepoint 1, indicating a possible contamination or process excursion.COMPUTER SYSTEMS

[0341] The present disclosure provides computer systems that are programmed to implement methods of the disclosure. FIG. 22 shows a computer system 2201 that is programmed or otherwise configured to predict POI or POI titer from FTIR data. The computer system 2201 can regulate various aspects of predicting POI or POI titer of the present disclosure, such as, for example, implementing one or more machine learning algorithms. The computer system 2201 can be an electronic device of a user or a computer system that is remotely located with respect to the electronic device. The electronic device can be a mobile electronic device.

[0342] The computer system 2201 includes a central processing unit (CPU, also “processor” and “computer processor” herein) 2205, which can be a single core or multi core processor, or a plurality of processors for parallel processing. The computer system 2201 also includes memory or memory location 2210 (e.g. random-access memory, read-only memory, flash memory), electronic storage unit 2215 (e.g. hard disk), communication interface 2220 (e.g. network adapter) for communicating with one or more other systems, and peripheral devices 2225, such as cache, other memory, data storage and / or electronic display adapters. The memory 2210, storage unit 2215, interface 2220 and peripheral devices 2225 are in communication with the CPU 2205 through a communication bus (solid lines), such as a motherboard. The storage unit 2215 can be a data storage unit (or data repository) for storing data. The computer system 2201 can be operatively coupled to a computer network (“network”) 2230 with the aid of the communication interface 2220. The network 2230 can be the Internet, an internet and / or extranet, or an intranet and / or extranet that is in communication with the Internet. The network 2230 in some cases is a telecommunication and / or data network. The network 2230 can include one or more computer servers, which can enable distributed computing, such as cloud computing. The network 2230, in some cases with the aid of the computer system 2201, can implement a peer-to-peer network, which may enable devices coupled to the computer system 2201 to behave as a client or a server.

[0343] The CPU 2205 can execute a sequence of machine-readable instructions, which can be embodied in a program or software. The instructions may be stored in a memory location, such as the memory 2210. The instructions can be directed to the CPU 2205, which can subsequently program or otherwise configure the CPU 2205 to implement methods of the74 41522-58976 / WOpresent disclosure. Examples of operations performed by the CPU 2205 can include fetch, decode, execute, and writeback.

[0344] The CPU 2205 can be part of a circuit, such as an integrated circuit. One or more other components of the system 2201 can be included in the circuit. In some cases, the circuit is an application specific integrated circuit (ASIC).

[0345] The storage unit 2215 can store files, such as drivers, libraries, and saved programs. The storage unit 2215 can store user data, e.g. user preferences and user programs. The computer system 2201 in some cases can include one or more additional data storage units that are external to the computer system 2201, such as located on a remote server that is in communication with the computer system 2201 through an intranet or the Internet.

[0346] The computer system 2201 can communicate with one or more remote computer systems through the network 2230. For instance, the computer system 2201 can communicate with a remote computer system of a user (e.g. a mobile device). Examples of remote computer systems include personal computers (e.g. portable PC), slate or tablet PC’s (e.g. Apple® iPad, Samsung® Galaxy Tab), telephones, Smart phones (e.g. Apple® iPhone, Android-enabled device, Blackberry®), or personal digital assistants. The user can access the computer system 2201 via the network 2230.

[0347] Methods as described herein can be implemented by way of machine (e.g. computer processor) executable code stored on an electronic storage location of the computer system 2201, such as, for example, on the memory 2210 or electronic storage unit 2215. The machine executable or machine-readable code can be provided in the form of software.During use, the code can be executed by the processor 2205. In some cases, the code can be retrieved from the storage unit 2215 and stored on the memory 2210 for ready access by the processor 2205. In some situations, the electronic storage unit 2215 can be precluded, and machine-executable instructions are stored on memory 2210.

[0348] The code can be pre-compiled and configured for use with a machine having a processor adapted to execute the code or can be compiled during runtime. The code can be supplied in a programming language that can be selected to enable the code to execute in a pre-compiled or as-compiled fashion.

[0349] Aspects of the systems and methods provided herein, such as the computer system 2201, can be embodied in programming. Various aspects of the technology may be thought of as “products” or “articles of manufacture” typically in the form of machine (or processor) executable code and / or associated data that is carried on or embodied in a type of machine readable medium. Machine-executable code can be stored on an electronic storage unit, such75 41522-58976 / WOas memory (e.g. read-only memory, random-access memory, flash memory) or a hard disk. “Storage” type media can include any or all the tangible memory of the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide non-transitory storage at any time for the software programming. All or portions of the software may at times be communicated through the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer into the computer platform of an application server. Thus, another type of media that may bear the software elements includes optical, electrical, and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry such waves, such as wired or wireless links, optical links, or the like, also may be considered as media bearing the software. As used herein, unless restricted to non-transitory, tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates in providing instructions to a processor for execution.

[0350] Hence, a machine readable medium, such as computer-executable code, may take many forms, including but not limited to, a tangible storage medium, a carrier wave medium or physical transmission medium. Non-volatile storage media include, for example, optical or magnetic disks, such as any of the storage devices in any computer(s) or the like, such as may be used to implement the databases, etc. shown in the drawings. Volatile storage media include dynamic memory, such as main memory of such a computer platform. Tangible transmission media include coaxial cables; copper wire and fiber optics, including the wires that comprise a bus within a computer system. Carrier-wave transmission media may take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media therefore include for example: a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a CD-ROM, DVD or DVD-ROM, any other optical medium, punch cards paper tape, any other physical storage medium with patterns of holes, a RAM, a ROM, a PROM and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave transporting data or instructions, cables or links transporting such a carrier wave, or any other medium from which a computer may read programming code and / or data. Many of these forms of computer readable media may be76 41522-58976 / WOinvolved in carrying one or more sequences of one or more instructions to a processor for execution.

[0351] The computer system 2201 can include or be in communication with an electronic display 2235 that comprises a user interface (UI) 2240 for providing, for example, a POI titer. Examples of UI’s include, without limitation, a graphical user interface (GUI) and web-based user interface.

[0352] Methods and systems of the present disclosure can be implemented by way of one or more algorithms. An algorithm can be implemented by way of software upon execution by the central processing unit 2205. The algorithm can, for example, predict a POI or POI titer.OTHER CONSIDERATIONS

[0353] The foregoing description of the embodiments has been presented for the purpose of illustration; it is not intended to be exhaustive or to limit the scope of the disclosure. Many modifications and variations are possible in light of the above disclosure.

[0354] Some portions of this description describe the embodiments in terms of algorithms and symbolic representations of operations on information. These algorithmic descriptions and representations are commonly used by those skilled in the data processing arts to convey the substance of their work effectively to others skilled in the art. These operations, while described functionally, computationally, or logically, are understood to be implemented by computer programs or equivalent electrical circuits, microcode, or the like. Furthermore, it has also proven convenient at times, to refer to these arrangements of operations as modules, without loss of generality. The described operations and their associated modules may be embodied in software, firmware, hardware, or any combinations thereof.

[0355] Any of the steps, operations, or processes described herein may be performed or implemented with one or more hardware or software modules, alone or in combination with other devices. In one or more embodiments, a software module is implemented with a computer program product comprising one or more computer-readable media containing computer program code or instructions, which can be executed by a computer processor for performing any or all of the steps, operations, or processes described. In one or more embodiments, a computer-readable medium comprises one or more computer-readable media that, individually or together, comprise instructions that, when executed by one or more processors, cause the one or more processors to perform, individually or together, the steps of the instructions stored on the one or more computer-readable media. Similarly, a processor77 41522-58976 / WOcomprises one or more processors or processing units that, individually or together, perform the steps of instructions stored on a computer-readable medium.

[0356] Embodiments may also relate to an apparatus for performing the operations herein. This apparatus may be specially constructed for the required purposes, or it may comprise a computing device selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a non-transitory, tangible computer readable storage medium, or any type of media suitable for storing electronic instructions, which may be coupled to a computer system bus. Furthermore, any computing systems referred to in the specification may include a single processor or may be architectures employing multiple processor designs for increased computing capability.

[0357] Embodiments may also relate to a product that is produced by a computing process described herein. Such a product may comprise information resulting from a computing process, where the information is stored on a non-transitory, tangible computer readable storage medium and may include any embodiment of a computer program product or other data combination described herein.

[0358] The description herein may describe processes and systems that use machinelearning models in the performance of their described functionalities. A “machine-learning model,” as used herein, comprises one or more machine-learning models that perform the described functionality. Machine-learning models may be stored on one or more computer- readable media with a set of weights. These weights are parameters used by the machinelearning model to transform input data received by the model into output data. The weights may be generated through a training process, whereby the machine-learning model is trained based on a set of training examples and labels associated with the training examples. The weights may be stored on one or more computer-readable media, and are used by a system when applying the machine-learning model to new data.

[0359] The language used in the specification has been principally selected for readability and instructional purposes, and it may not have been selected to delineate or circumscribe the inventive subject matter. It is therefore intended that the scope of the patent rights be limited not by this detailed description, but rather by any claims that issue on an application based hereon. Accordingly, the disclosure of the embodiments is intended to be illustrative, but not limiting, of the scope of the patent rights, which is set forth in the following claims.

[0360] As used herein, the terms “comprises,” “comprising,” “includes,” “including,” “has,” “having,” or any other variation thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, article, or apparatus that comprises a list of78 41522-58976 / WOelements is not necessarily limited to only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Further, unless expressly stated to the contrary, “or” refers to an inclusive “or” and not to an exclusive “or.” For example, a condition “A or B” is satisfied by any one of the following: A is true (or present) and B is false (or not present), A is false (or not present) and B is true (or present), and both A and B are true (or present). Similarly, a condition “A, B, or C” is satisfied by any combination of A, B, and C having at least one element in the combination that is true (or present). As a not-limiting example, the condition “A, B, or C” is satisfied by A and B are true (or present) and C is false (or not present). Similarly, as another not-limiting example, the condition “A, B, or C” is satisfied by A is true (or present) and B and C are false (or not present).79 41522-58976 / WO

Claims

WHAT IS CLAIMED:

1. A method comprising: obtaining data regarding a biological sample collected during a fermentation process, wherein the data comprises Fourier transform infrared (FTIR) data describing the content of the biological sample; determining one or more proteins of interest (POI) present in the biological sample by applying a first vector representation of the FTIR data to a first machine-learning model, wherein the first machine-learning model is trained to predict a confidence score representing a probability that each of the one or more proteins are present in the biological sample; for each POI present in the biological sample, determining a titer of the POI by applying a second vector representation of the FTIR data to a second machine-learning model, wherein the second machine-learning model is trained to predict a titer measurement for a POI; identifying one or more process excursions during the fermentation process based on one or more of the POIs present in the biological sample and the titer of each POI present in the biological sample; and modifying instructions regarding the fermentation process based on the one or more identified process excursions.

2. The method of claim 1, wherein the FTIR data is collected from one or more predownstream process fermentation samples using a FTIR spectrometer.

3. The method of claim 1, wherein the first machine-learning model and the second machine-learning model are trained using a training dataset comprising entries of previously collected FTIR data labeled with POIs present in each entry and titer measurements of each POI present in the entry.

4. The method of claim 3, wherein training the first machine-learning model comprises: for each entry in the training dataset, generating a prediction of POIs present in the entry by applying the first machine-learning model to the entry in the training dataset;80 41522-58976 / WOcomparing the prediction of POIs present in the entry to the label describing the POIs present in the entry to determine a difference between the prediction and the label; and adjusting one or more parameters of the first machine-learning model to reduce the difference between the prediction and the label.

5. The method of claim 3, wherein training the second machine-learning model comprises: for each entry in the training dataset, predicting a titer of each POI present in the entry by applying the second machine-learning model to the entry in the training dataset; comparing the predicted titer of each POI present in the entry to the label describing titer measurements of each POI present in the entry to determine a difference between the prediction and the label; and adjusting one or more parameters of the second machine-learning model to reduce the difference between the prediction and the label.

6. The method of claim 1, further comprising: determining a titer of one or more analytes of interest in the biological sample by applying a third vector representation of the FTIR data to a third machinelearning model, wherein the third machine-learning model is trained to predict a titer measurement for an analyte of interest.

7. The method of claim 1, wherein the third machine-learning model is trained using the training dataset comprising entries of previously collected FTIR data and each entry of previously collected FTIR data is further labeled with analytes of interest present in each entry and titer measurements of each analyte present in the entry.

8. The method of claim 1, further comprising: comparing a number of entries in the training dataset to a threshold; responsive to determining the number of entries does not satisfy the threshold, generating synthetic training data to augment the training dataset.81 41522-58976 / WO9. The method of claim 8, wherein generating synthetic training comprises: averaging a plurality of randomly chosen FTIR spectra points for a common POI; or averaging a plurality of randomly chosen FTIR spectra points for a common POI.

10. The method of claim 1, further comprising: comparing a region of the FTIR data collected for the biological sample to a corresponding region of FTIR data collected from a baseline fermentation process by computing a distance measurement between FTIR data at the region and and the corresponding region, wherein the region of the FTIR data corresponding a step in the fermentation process; responsive to the distance measurement exceeding a threhsold distance, identifying a high variability region of FTIR data, wherein the high variabiltiy region represents a potential process excursion during the fermentation process.

11. The method of claim 1, further comprising: comparing a region of the FTIR data collected for the biological sample to a corresponding region of FTIR data collected from a baseline fermentation process by computing a distance measurement between FTIR data at the region and and the corresponding region, wherein the region of the FTIR data corresponding a step in the fermentation process; responsive to the distance measurement being less than a threshold distance, extracting one or more features of the FTIR data to be input to the first or second machine-learning models.

12. A non-transitory computer readable storage medium comprising stored instructions that when executed by one or more processors of one or more computing devices, cause the one or more computing devices to: obtain data regarding a biological sample collected during a fermentation process, wherein the data comprises Fourier transform infrared (FTIR) data describing the content of the biological sample; determine one or more proteins of interest (POI) present in the biological sample by applying a first vector representation of the FTIR data to a first machine-learning model, wherein the first machine-learning model is82 41522-58976 / WOtrained to predict a confidence score representing a probability that each of the one or more proteins are present in the biological sample; for each POI present in the biological sample, determine a titer of the POI by applying a second vector representation of the FTIR data to a second machine-learning model, wherein the second machine-learning model is trained to predict a titer measurement for a POI; identify one or more process excursions during the fermentation process based on one or more of the POIs present in the biological sample and the titer of each POI present in the biological sample; and modify instructions regarding the fermentation process based on the one or more identified process excursions.

13. A system comprising: one or more computer processors; and one or more computer-readable mediums storing instructions that, when executed by the one or more computer processors, cause the system to: obtain data regarding a biological sample collected during a fermentation process, wherein the data comprises Fourier transform infrared (FTIR) data describing the content of the biological sample; determine one or more proteins of interest (POI) present in the biological sample by applying a first vector representation of the FTIR data to a first machine-learning model, wherein the first machine-learning model is trained to predict a confidence score representing a probability that each of the one or more proteins are present in the biological sample; for each POI present in the biological sample, determine a titer of the POI by applying a second vector representation of the FTIR data to a second machine-learning model, wherein the second machinelearning model is trained to predict a titer measurement for a POI; identify one or more process excursions during the fermentation process based on one or more of the POIs present in the biological sample and the titer of each POI present in the biological sample; and modify instructions regarding the fermentation process based on the one or more identified process excursions.83 41522-58976 / WO

Citation Information

Patent Citations

  • Method of Characterization of Visible and / or Sub-Visible Particles in Biologics

    US20200241002A1

  • Real time monitoring of product purification

    US20210149361A1

  • Forecasting industrial aging processes with machine learning methods

    US20230028276A1

  • Systems for end-to-end optimization of precision fermentation-produced animal proteins in food applications

    WO2022246284A2