Just-in-time learning with variational autoencoders for monitoring and / or control of cell culture processes

Abstract

A method for monitoring and / or controlling a biopharmaceutical process includes querying an observation database containing observation data sets associated with past scans based on a first spectral scan vector of the biopharmaceutical process. Each of the observation data sets includes spectral data and corresponding actual analytical measurements. Querying the observation database includes determining a first parameter defining a set of distributions for the first spectral scan vector and selecting specific observation data sets from among the observation data sets as training data based on (i) the first parameter and (ii) other parameters defining the set of distributions for each of the observation data sets. The method also includes calibrating a local model specific to the biopharmaceutical process using the selected training data. The method also includes predicting analytical measurements of the biopharmaceutical process by using the local model to analyze spectral data generated during scanning the biopharmaceutical process.

Description

[Technical Field]

[0001] This application relates generally to monitoring and / or control of biopharmaceutical processes using spectroscopic techniques such as Raman spectroscopy, and more specifically to the use of just-in-time learning (JITL) models to predict or infer product quality attributes based on spectroscopic scans. [Background technology]

[0002] Stable production of biopharmaceutical processes (e.g., biotherapeutic proteins) generally requires bioreactors to maintain balanced and consistent parameters (e.g., cellular metabolite concentrations), which therefore necessitates rigorous process monitoring and control. To meet these demands, process analytical engineering (PAT) tools are increasingly being adopted. Online monitoring of cell culture pH, dissolved oxygen, and temperature are some examples of traditional PAT tools used in feedback control systems. In recent years, other in-process probes have been investigated and deployed for continuous monitoring of more complex species such as viable cell density (VCD), glucose, lactate, and other important cellular metabolites, including amino acids, titer, and critical quality attributes.

[0003] Raman spectroscopy is a common PAT tool widely used for online monitoring in biopharmaceutical manufacturing. It is an optical method capable of nondestructive analysis of chemical composition and molecular structure. In Raman spectroscopy, incident laser light is inelastically scattered by molecular vibrational modes. The frequency difference between the incident and scattered photons is referred to as the "Raman shift," and the vector of the Raman shift relative to the intensity level (referred to herein as the "Raman spectrum," "Raman scan," or "Raman scan vector") can be analyzed to identify the chemical composition and molecular structure of a sample. Over the past 30 years, advances in laser sampling and detector technology have led to a rapid increase in the application of Raman spectroscopy in polymer, pharmaceutical, and biopharmaceutical manufacturing and biomedical analysis. These technological advances have made Raman spectroscopy a practical analytical technique used both inside and outside the laboratory. Since the first reported application of in situ Raman measurements in biopharmaceutical manufacturing, they have been employed to provide online, real-time predictions of several critical process states, such as glucose concentration, lactate concentration, glutamate concentration, glutamine concentration, ammonium concentration, potassium concentration, sodium concentration, viability, VCD, osmolality, and titer. These predictions are typically based on calibration models or soft-sensor models built in an offline setting. The models are built using analytical measurements from analytical instruments. Partial least squares (PLS) and multiple linear regression modeling methods are commonly used to correlate Raman spectra with analytical measurements. These models typically require preprocessing (filtering) of the Raman scans before calibrating against analytical measurements. Once the calibration model is trained, it is implemented in a real-time setting to provide in situ measurements for process monitoring and / or control.

[0004] Because biopharmaceutical processes typically operate under strict constraints and regulations, Raman model calibration for biopharmaceutical applications is not straightforward. Raman model calibration in the biopharmaceutical industry traditionally involves running multiple test campaigns to generate relevant data used to correlate Raman spectra with analytical measurements. These tests are both costly and time-consuming, as each campaign can last 2–4 weeks in a laboratory environment. Furthermore, only limited samples may be available to analytical instruments (e.g., to ensure that a lab-scale bioreactor maintains a healthy population of viable cells). Indeed, it is not uncommon for only one or two measurements to be available daily from inline or offline analytical instruments. Further exacerbating the situation, models are tied to specific processes, specific formulations or profiles of bioreactor media, and specific operating conditions. Therefore, if any of the aforementioned variables change, the model must be recalibrated based on new data. Both Raman model calibration and model maintenance require significant resource allocation and are typically performed in an offline setting.

[0005] To address the issue of model performance degradation over time due to process variations, various soft-sensing techniques, such as moving window, time difference, and recursive modeling, have been implemented. See Qin et al., Comput. Chem. Eng., 22, 503-514, Recursive PLS Algorithms for Adaptive Data Modeling (1998); Kaneko et al., Ind. Eng. Chem. Res., 54, 700-704, Moving Window and Just-In-Time Soft Sensor Model Based on Time Differences Considering a Small Number of Measurements (2015). However, none of these techniques adequately account for or address sudden changes in industrial processes.

[0006] To better account for sudden changes, just-in-time learning (JITL) techniques have been proposed for automatic calibration and assessment of Raman models. See Tulsyan et al., AICHE Journal, e17210, Spectroscopic Models for Real-Time Monitoring of Cell Culture Processes Using Spatiotemporal Just-In-Time Gaussian Processes (2020); Tulsyan et al., Biotechnology and Bioengineering, 116(10), 2575-2586, A Machine Learning Approach to Calibrate Generic Raman Models for Real-Time Monitoring of Cell Culture Processes (2019); Tulsyan et al., Biotechnology and Bioengineering, 117(2), 406-416, Automatic Real-Time Calibration, Assessment, and Maintenance of Generic Raman Models for Online Monitoring of Cell Culture Processes (2020). JITL is an instant modeling platform based on local modeling and database sampling techniques. Unlike other machine learning methods, JITL generally assumes that all available observational data is stored in a central observational database, and local models are dynamically constructed in real time based on a query sample (e.g., a new Raman scan), ideally using the most "similar" or relevant data from the observational database. This allows for good approximation of complex process dynamics using relatively simple local models. Under the JITL framework, libraries can include not only spectral data for a single process operating under specific operating conditions, but also data for different processes, different media profiles, and / or different operating conditions.This can significantly reduce the time required to calibrate and maintain models, especially for pipeline drugs that may have little or no past manufacturing history.

[0007] Traditionally, in JITL, "similar" historical samples are identified in an observation database based on Euclidean distance, angle, or correlation. See Quan et al., Applied Soft Computing, 10, 562-566, "Weighted Least Squares Support Vector Machine Local Region Method for Nonlinear Time Series Prediction" (2010); Cheng et al., Chemical Engineering Science, 59(13), 2801-2810, "A New Data-Based Methodology for Nonlinear Process Modeling" (2004); Fujiwara et al., AIChE Journal, 55(7), 1754-1765, "Softsensor Development Using Correlation-Based Just-In-Time Modeling" (2009). All of these techniques find the most relevant historical samples to a query sample in a deterministic and point-to-point manner. See ZQGe et al., Chemometr. Intell. Lab. Syst., 104(13), 306-317, A Comparative Study of Just-In-Time-Learning Based Methods for Online Soft Sensor Modeling (2010). However, Raman scan results can reflect significant uncertainties that deterministic techniques do not take into account. Therefore, uncertainty in the sample / scan can lead to a relatively poor selection of "similar" historical samples, which in turn can lead to a relatively poor predictive performance of the JITL local model built based on the selected historical samples. [Prior art documents] [Non-patent literature]

[0008] [Non-Patent Document 1] Qin et al.,Comput.Chem.Eng.,22,503-514,Recursive PLS Algorithms for Adaptive Data Modeling(1998) [Non-patent document 2] Kaneko et al.,Ind.Eng.Chem.Res.,54,700-704,Moving Window and Just-In-Time Soft Sensor Model Based on Time Differences Considering a Small Number of Measurements(2015) [Non-patent document 3] Tulsyan et al.,AICHE Journal,e17210,Spectroscopic Models for Real-Time Monitoring of Cell Culture Processes Using Spatiotemporal Just-In-Time Gaussian Processes(2020) [Non-patent document 4] Tulsyan et al.,Biotechnology and Bioengineering,116(10),2575-2586,A Machine Learning Approach to Calibrate Generic Raman Models for Real-Time Monitoring of Cell Culture Processes(2019) [Non-patent document 5] Tulsyan et al.,Biotechnology and Bioengineering,117(2),406-416,Automatic Real-Time Calibration,Assessment,and Maintenance of Generic Raman Models for Online Monitoring of Cell Culture Processes(2020) [Non-patent document 6] Quan et al.,Applied Soft Computing,10,562-566,Weighted Least Squares Support Vector Machine Local Region Method for Nonlinear Time Series Prediction(2010) [Non-Patent Document 7] Cheng et al.,Chemical Engineering Science,59(13),2801-2810,A New Data-Based Methodology for Nonlinear Process Modeling(2004) [Non-patent document 8] Fujiwara et al., AIChE Journal, 55(7), 1754-1765, Softsensor Development Using Correlation-Based Just-In-Time Modeling(2009) [Non-Patent Document 9] ZQGe et al.,Chemometr.Intell.Lab.Syst.,104(13),306-317,A Comparative Study of Just-In-Time-Learning Based Methods for Online Soft Sensor Modeling(2010) Summary of the Invention [Means for solving the problem]

[0009] To address the aforementioned issues with just-in-time learning (JITL) techniques for biopharmaceutical applications, the systems and methods disclosed herein account for uncertainty in spectroscopic measurements by using distributions, rather than simply deterministic values, to identify "similar" historical samples (e.g., similar Raman scans) in an observation database. In particular, a computing system may use the historical samples to train / generate a variational autoencoder (VAE), including an encoder and a decoder. The encoder converts each input sample (scan) into a low-dimensional latent space representation that includes parameters (e.g., mean and variance) that define the input sample's distribution, and the decoder attempts to recreate the complete input sample by sampling from the distribution in the latent space. The computing system (or another computing system) can then use the encoder portion of the trained VAE to determine parameters that define a set of distributions for each historical sample (e.g., each historical Raman scan) and parameters that define a set of distributions for a sample of interest (e.g., a new real-time Raman scan), and use the determined parameters to identify / select those historical samples with distributions that are most similar to the sample of interest. The selection step may include, for example, using multivariate Kullback-Leibler (KL) divergence to identify the most similar historical samples. In some embodiments, the encoder includes exactly one hidden layer.

[0010] In addition to the benefits of JITL (e.g., preventing model degradation over time and tracking rapid changes in biopharmaceutical processes), integrating VAE into JITL helps identify more relevant JITL training samples based on sample distribution. By selecting better historical samples to train / build a local model, the local model can better predict analytical measurements such as metabolite concentrations, viable cell densities, etc.

[0011] Those skilled in the art will appreciate that the figures described herein are included for purposes of illustration and not limitation of the present disclosure. The figures are not necessarily to scale, emphasis instead being placed upon illustrating the principles of the present disclosure. It should be understood that in some instances, various aspects of the described implementations may be shown exaggerated or enlarged to facilitate understanding of the described implementations. In the figures, like reference numbers generally refer to functionally similar and / or structurally similar components throughout the various views. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a simplified block diagram of an example Raman spectroscopy system incorporating VAE-JITL to predict analytical measurements of a biopharmaceutical process for monitoring purposes. [Figure 2] FIG. 1 is a simplified block diagram of an example Raman spectroscopy system incorporating VAE-JITL to predict analytical measurements of a biopharmaceutical process for closed-loop control of glucose concentration. [Figure 3] 1 illustrates an example data flow that may occur when analyzing a biopharmaceutical process using a JITL technique such as VAE-JITL. [Figure 4] 1 or 2 to select a historical scan from the observation database of FIG. 1 or 2 based on a query sample. [Figure 5] 3 illustrates an example VAE-JITL process that may be implemented by the computing system of FIG. 1 or FIG. 2. [Figure 6A] 1A and 1B show an example downsampled Raman scan before and after baseline correction, respectively. [Figure 6B] 1A and 1B show an example downsampled Raman scan before and after baseline correction, respectively. [Figure 7A] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7B] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7C] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7D] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7E] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7F] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7G] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7H] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7I] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7J] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 7K] Examples of linear JITL prediction performance and VAE-JITL prediction performance against measurements are presented for various types of analytical measurements. [Figure 8] FIG. 1 is a flow diagram of an example method for analyzing a biopharmaceutical process using VAE-JITL. DETAILED DESCRIPTION OF THE INVENTION

[0013] The various concepts described introductory above and discussed in more detail below may be implemented in any of numerous ways, and the described concepts are not limited to any particular manner of implementation. Several example implementations are provided for illustrative purposes.

[0014] Figure 1 is a simplified block diagram of an example Raman spectroscopy system 100 that may be used to predict analytical measurements of a biopharmaceutical process. While Figure 1 illustrates system 100 implementing Raman spectroscopy techniques, it is understood that in other embodiments, system 100 may implement other spectroscopy techniques suitable for analyzing biopharmaceutical processes, such as, for example, near-infrared (NIR) and mass spectrometry. For ease of explanation, the following description (for all figures) refers specifically to the Raman spectroscopy embodiment.

[0015] The system 100 includes a bioreactor 102, one or more analytical instruments 104, a Raman analyzer 106 with a Raman probe 108, and a computing system 110. The bioreactor 102 may be any suitable vessel, device, or system that supports a biologically active environment that may include living organisms and / or materials derived therefrom (e.g., cell cultures) within the medium. The bioreactor 102 may contain recombinant proteins expressed by the cell cultures, for example, for research purposes, clinical use, commercial sale or other distribution, etc. Depending on the biopharmaceutical process being monitored, the medium may include a particular fluid (e.g., a "broth") and particular nutrients, and may have a target pH level or range, a target temperature or temperature range, etc. Collectively, the contents and parameters / characteristics of the medium are referred to herein as a "media profile."

[0016] The analytical instrument 104 may be any in-line, at-line, and / or offline instrument or instruments configured to measure one or more characteristics or parameters of the biologically active contents within the bioreactor 102 based on samples taken therefrom. For example, the analytical instrument 104 may measure one or more media component concentrations, such as metabolite levels (e.g., glucose, lactate, glutamate, glutamine, ammonium, pCO2, pO2, Na+, K+, etc.) and / or amino acid levels. Additionally or alternatively, the analytical instrument 104 may measure osmolality, viability, viable cell density (VCD), titer, critical quality attributes, cell state (e.g., cell cycle), and / or other characteristics or parameters associated with the contents of the bioreactor 102. As a more detailed example, a sample may be taken, centrifuged, purified through multiple columns, passed through a first analytical instrument 104 (e.g., a high performance liquid chromatography (HPLC) or ultra performance liquid chromatography (UPLC) instrument), and then passed through a second analytical instrument 104 (e.g., a mass spectrometer), which provide analytical measurements. One, some, or all of the analytical instruments 104 may use destructive analytical techniques.

[0017] The Raman analyzer 106 may include a spectrograph device coupled to the Raman probe 108 (or multiple Raman probes in some implementations). The Raman analyzer 106 may include a laser source that sends laser light to the Raman probe 108 via a fiber optic cable, and may also include a charge-coupled device (CCD) or other suitable camera / recording device that records signals received from the Raman probe 108, for example, via another channel of the fiber optic cable. Alternatively, the laser source may be integrated within the Raman probe 108 itself. The Raman probe 108 may be an immersion probe or any other suitable type of probe (e.g., a reflection probe and a transmission probe).

[0018] Collectively, the Raman analyzer 106 and Raman probe 108 are configured to non-destructively scan the biologically active content during a biopharmaceutical process in the bioreactor 102 by exciting, observing, and recording the molecular "fingerprint" of the biopharmaceutical process. The molecular fingerprint corresponds to vibrational, rotational, and / or other low-frequency modes of molecules within the biologically active content within the biopharmaceutical process when the bioreactor contents are excited by laser light delivered by the Raman probe 108. As a result of this scanning process, the Raman analyzer 106 generates Raman scan vectors, each representing intensity as a function of Raman shift (frequency).

[0019] The computing system 110 may be a single computing device or may include two or more co-located and / or distributed computing devices. The computing system 110 is coupled to the Raman analyzer 106 and the analytical instrument 104 and is generally configured to analyze Raman scan vectors generated by the Raman analyzer 106 to predict one or more analytical measurements of the biopharmaceutical process. For example, the computing system 110 may analyze the Raman scan vectors to predict the same types of analytical measurements performed by the analytical instrument 104. As a more detailed example, the computing system 110 may predict glucose concentrations, while the analytical instrument 104 actually measures the glucose concentrations. However, while the analytical instrument 104 may perform relatively infrequent “offline” analytical measurements of samples extracted from the bioreactor 102 (e.g., due to limited volumes of the biopharmaceutical process and / or the higher cost of performing such measurements), the computing system 110 may perform relatively frequent “online” predictions of analytical measurements in real time. In one embodiment, Raman scans are collected every 30 minutes and the analytical measurements are taken every 24 hours (i.e., so that exactly one Raman scan each day is performed simultaneously with the analytical measurements). As used herein, it is understood that terms such as "predict" or "forecast" do not necessarily refer to a determination or estimation of a future value, but may instead refer to an inference of a present value.

[0020] 1 , computing system 110 includes a processing unit 120, a network interface 122, a display 124, a user input device 126, and a memory 128. Processing unit 120 includes one or more processors, each of which may be a programmable microprocessor that executes software instructions stored in memory 128 to perform some or all of the functions of computing system 110 described herein. Alternatively, one, some, or all of the processors in processing unit 120 may be other types of processors (e.g., application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), etc.), and the functions of computing system 110 described herein may instead be implemented partially or entirely in hardware. Memory 128 may include one or more physical memory devices or units, including volatile and / or non-volatile memory. Any suitable type of memory may be used, such as read-only memory (ROM), random-access memory (RAM), solid-state drives (SSDs), hard disk drives (HDDs), etc.

[0021] Network interface 122 may include any suitable hardware (e.g., front-end transmitter and receiver hardware), firmware, and / or software configured to communicate with external devices and / or systems (e.g., analytical instrument 104, Raman analyzer 106, and / or observations database 136) over one or more networks using one or more communication protocols. For example, network interface 122 may be or include an Ethernet interface, and / or may include a wireless local area network (LAN) interface, etc.

[0022] Display 124 may use any suitable display technology (e.g., LED, OLED, LCD, etc.) to present information to a user, and user input device 126 may be a keyboard or other suitable input device. In some embodiments, display 124 and user input device 126 are integrated into a single device (e.g., a touchscreen display). Generally, display 124 and user input device 126 may be combined to allow a user to interact with a graphical user interface (GUI) provided by computing system 110, for purposes such as manually monitoring various processes running within system 100. However, in some embodiments, computing system 110 does not include display 124 and / or user input device 126, or one or both of display 124 and user input device 126 are included in another computer or system communicatively coupled to computing system 110 (e.g., in some embodiments where predictions are sent directly to a control system implementing closed-loop control).

[0023] The memory 128 stores instructions for one or more software applications, including a variational autoencoder (VAE) just-in-time learning (JITL) prediction application 130 (also referred to herein as the “VAE-JITL prediction application 130”). When executed by the processing unit 120, the VAE-JITL prediction application 130 is generally configured to select historical samples / scans from an observation database 136 using a VAE's encoder, generate / build / calibrate a local model 132 using the selected samples, and predict analytical measurements of a biopharmaceutical process in the bioreactor 102 by analyzing Raman scan vectors generated by the Raman analyzer 106 using the calibrated local model 132. Depending on how frequently the Raman analyzer 106 generates such scan vectors, the VAE-JITL prediction application 130 may predict analytical measurements periodically or on another suitable time basis. The Raman analyzer 106 may itself control when the scan vectors are generated, or the computing system 110 may trigger the generation of the scan vectors by sending commands to the Raman analyzer 106. VAE-JITL prediction application 130 may predict only a single type of analytical measurement (e.g., only glucose concentration) based on each scan vector, or may predict multiple types of analytical measurements (e.g., glucose concentration and viable cell density) based on each scan vector. In other embodiments, multiple different VAE-JITL prediction applications (e.g., each similar to VAE-JITL prediction application 130) each generate a different local model that predicts a different type of analytical measurement, all based on the same scan vector. VAE-JITL prediction application 130 and local model 132 are discussed in further detail below.

[0024] The observational database 136 stores historical observational datasets associated with past observations. The observations / datasets may be curated, for example, by removing outliers and imputing missing data. Each observational dataset in the observational database 136 may include spectral data (e.g., Raman scan vectors of the type generated by the Raman analyzer 106) and one or more corresponding analytical measurements (e.g., one or more measurements of the type generated by the analytical instrument 104). Depending on the embodiment and / or scenario, past observations may be collected for a number of different biopharmaceutical processes under a number of different operating conditions (e.g., different metabolite concentration set points) and / or with a number of different media profiles (e.g., different fluids, nutrients, pH levels, temperatures, etc.). In general, it may be desirable for the observational database 136 to represent a widely diverse array of processes, operating conditions, and media profiles. The observational database 136 may or may not store information indicative of these processes, cell lines, proteins, metabolites, operating conditions, and / or media profiles, depending on the embodiment (as discussed further below).

[0025] It will be understood that other architectural configurations and / or components may be used instead of that shown in Figure 1. For example, the observation database 136 may be maintained and / or accessed by a database server separate from the computing system 110, in which case some of the functions depicted in Figure 1 (e.g., some functions of the query unit 140) may be performed by the database server. Such an architecture may be desirable to collect and store a larger number of observational data sets in the observation database 136 (e.g., when the database server is connected to multiple computing systems, each similar to the computing system 110). As another example of a different architecture, a different computer (not shown in Figure 1) may send measurements provided by the analytical instrument 104 to the computing system 110.

[0026] During runtime operation of the system 100, the Raman analyzer 106 and the Raman probe 108 are used to scan (i.e., generate Raman scan vectors for) the biopharmaceutical process in the bioreactor 102, which are then transmitted from the Raman analyzer 106 to the computing system 110. The Raman analyzer 106 and the Raman probe 108 may provide scan vectors to support predictions (made by the VAE-JITL prediction application 130) according to a predetermined scheduled monitoring period, such as once per minute or once per hour. Alternatively or additionally, Raman scans may be collected and predictions made based on these scans at irregular intervals (e.g., in response to specific process-based triggers such as changes in measured pH level and / or temperature) such that each monitoring period has a variable or uncertain duration.

[0027] The query unit 140 of the VAE-JITL prediction application 130 uses the scan vectors received during the monitoring period to generate query points used to query the observation database 136. In some embodiments, the query points (i.e., the data defining the query points, also referred to below as "query samples") include only data representing the Raman scan vectors received from the Raman analyzer 106 (e.g., the intensity / frequency tuples that make up the scan vector). In other embodiments, the query points used to query the query observation database 136 also include one or more other types of information. For example, the query points may also include data representing operating conditions associated with the process (e.g., metabolite concentration setpoints in a control system or laser light wavelength and / or intensity associated with the Raman analyzer 106 or Raman probe 108, etc.), data representing a media profile for the biopharmaceutical process media (e.g., fluid type, nutrient type or concentration, pH level, etc.), and / or other data (e.g., indicators of cell lines, proteins, or metabolites associated with the biopharmaceutical process).

[0028] In general, query points may include data representing the same vectors, parameters, and / or classifications that local model 132 uses as inputs (i.e., as feature sets for local model 132). Using multiple different data types for the feature sets (e.g., the operating conditions, media profile data, etc. described above) may improve the accuracy of analytical measurement predictions made by local model 132. However, because each observational data set in observational database 136 generally must include the same vectors, parameters, and / or classifications as the feature set, it may be preferable to limit the feature sets / inputs of query points and local model 132 to include only Raman scan vectors. This may provide various advantages, such as enabling the collection of more information for storage in observational database 136 and / or simplifying the collection of that information. For example, if only Raman scan vectors are used, an observational data set may be included in observational database 136 even if little or nothing is known about the processes, cell lines, proteins, metabolites, operating conditions, and / or media profiles that existed when the data set was collected.

[0029] The query unit 140 then uses the generated query points to query the observation database 136. After receiving the query points / samples, the query unit 140 uses the query samples to select relevant observation data sets from the observation database 136 that are particularly useful as training data for the local model 132. In some embodiments, the VAE-JITL prediction application 130 pre-processes the Raman scan vectors to generate the query samples. For example, as discussed further below, the VAE-JITL prediction application 130 may generate the query samples by downsampling the Raman scan vectors, performing baseline correction on the downsampled vectors, and / or normalizing the downsampled, baseline-corrected vectors.

[0030] To determine which observational data sets are most “relevant” (e.g., most similar, most correlated, etc.) to the query sample, the query unit 140 uses an encoder of a variational autoencoder (VAE). The VAE may have been previously trained (e.g., by the VAE-JITL prediction application 130, or another application of the computing system 110, or another computing system) based on a large number of Raman scans (e.g., downsampled, baseline-corrected, and / or normalized Raman scans) from the observational database 136 and / or one or more other sources. Once the VAE is trained, the encoder layer of the VAE can capture features of the input Raman scan that represent the input Raman scan in a lower dimensional manner. Specifically, the encoder layer of the trained VAE (also referred to herein simply as the “encoder”) can generate parameters defining a large number of distributions that represent the Raman scan input to the encoder. For example, the encoder may generate a mean and variance for each of a large number (e.g., 2, 3, 5, 10, etc.) of normal (or nearly normal) distributions for a given input Raman scan. The parameters that define a distribution or set of distributions may be referred to herein as "distribution parameters."

[0031] Upon identifying the most relevant data set, the query unit 140 generates (1) distribution parameters for each of the plurality (e.g., all) Raman scan vectors stored in the observational database 136, and (2) distribution parameters for the query sample. It is understood that each Raman scan vector may be preprocessed (e.g., downsampled, baseline corrected, and / or normalized) before inputting the historical Raman scan vectors into the encoder and before inputting the Raman scan vector of interest into the encoder. The query unit 140 may generate the distribution parameters for the Raman scan vectors stored in the observational database 136 at any time (e.g., offline, before or during runtime operation of the bioreactor 102). However, the query unit 140 generates the distribution parameters for the query sample in real time (e.g., during runtime operation of the bioreactor 102 if the scans are provided by the Raman analyzer 106). Once the distribution parameters are available for both the historical scans and the new (query sample) scan, the query unit 140 can select a particular data set / scan from the observational database 136 based on those distribution parameters. For example, query unit 140 may select the most relevant / similar scans from observation database 136 by selecting the scans that have the lowest multivariate KL divergence with the query sample. VAE and multivariate KL divergence are discussed in more detail below.

[0032] In some embodiments, query unit 140 considers one or more other factors in addition to the distribution parameters when selecting relevant scans. For example, to better adapt to time-varying process changes, query unit 140 may further consider the timing or order of samples in observation database 136 (e.g., which sample is most recent). To this end, in addition to using VAE-JITL to select historical samples based on the distribution parameters, query unit 140 may incorporate the “adaptive” JITL (A-JITL) or “spatiotemporal” JITL (ST-JITL) approaches described in U.S. Patent Application Publication No. 2022 / 0128474 (Tulsyan, “Automatic Calibration and Automatic Maintenance of Raman Spectroscopic Models for Real-Time Predictions”), which is incorporated herein by reference. More generally, any of the techniques described in U.S. Patent Application Publication No. 2022 / 0128474 may be used, as long as they are compatible with, and used in addition to, the VAE-JITL techniques described herein.

[0033] In some embodiments, query unit 140 selects only a predetermined number of relevant observational datasets in response to a single query, or selects no more than some maximum allowable number of relevant observational datasets, ensuring that only a relatively small subset of all datasets in observational database 136 is searched. However, in other embodiments, query unit 140 can select any number of relevant observational datasets, as long as appropriate relevance criteria are met for each selected dataset (e.g., as long as the multivariate KL divergence is below a predetermined threshold).

[0034] After identifying related / similar observational data sets (each of which may or may not correspond to the same process conditions as the biopharmaceutical process in bioreactor 102 currently being monitored), query unit 140 provides or indicates those data sets (e.g., Raman scan vectors and corresponding analytical measurements) to local model generator 142. Local model generator 142 then uses the relevant data sets as training data to calibrate local model 132. That is, local model generator 142 uses the Raman scan vectors (and possibly other data) associated with each observational data set (possibly after some preprocessing) as a feature set and uses the analytical measurements associated with the same observational data set as labels for that feature set.

[0035] In some embodiments, the local model 132 constructed by the local model generator 142 is a Gaussian process model to efficiently capture complex nonlinear process dynamics and easily adapt to virtually any process change. Unlike partial least squares (PLS) and principal component regression (PCR) models, Gaussian process models use nonparametric methods and are much more likely to capture complex nonlinear correlations between Raman scan vectors and analytical measurements, even when using a very limited number of training samples. This can be particularly important in scenarios where a new product or process corresponds to only a limited number of data sets in the observation database 136. In such scenarios, a Gaussian process model can generally extract the most information from those limited data sets, along with other related data sets that the query unit 140 selects from the observation database 136. However, in other embodiments, the local model generator 142 may instead construct any other suitable type of machine learning model (e.g., a recurrent neural network, a convolutional neural network, etc.), so long as the training time does not exceed the minimum desired duration of the monitoring period. The local model generator 142 may construct the local model 132 such that the local model 132 is capable of outputting confidence limits or other suitable indicators of prediction confidence (e.g., confidence scores). Compared to at least PLS and PCR models, Gaussian process models are particularly well-suited for providing confidence limits around analytical measurement predictions. While various advantages of Gaussian process models over PLS and PCR models have been described, it will be understood that in some embodiments, the local model generator 142 may construct the local model 132 using PLS, PCR, or other modeling methods (e.g., to speed calibration or facilitate deployment in industrial environments, etc.).

[0036] The local model generator 142 may build the local model 132 in an online, real-time manner such that the prediction unit 144 can then use the trained local model 132 to predict one or more analytical measurements of the biopharmaceutical process by processing the same Raman scan vectors that the query unit 140 used to generate the query points. Indeed, in some embodiments, the query unit 140 may perform a new query, and the local model generator 142 may generate a new version of the local model 132, every time the Raman analyzer 106 provides a new Raman scan vector to the computing system 110. However, in other embodiments, the query unit 140 performs a new query (and the local model generator 142 generates a new version of the local model 132) on a less frequent basis, such as once every 10 prediction / monitoring periods or once every 100 prediction / monitoring periods.

[0037] The database maintenance unit 146 may also cause the analytical instrument 104 to periodically collect one or more actual analytical measurements significantly less frequently than the monitoring period of the Raman analyzer 106 (e.g., only once or twice a day, etc.). Measurements by the analytical instrument 104 may, in some embodiments, be destructive and may require permanent removal of the sample from the process in the bioreactor 102. At or about the time the database maintenance unit 146 causes the analytical instrument 104 to collect and provide an actual analytical measurement, the database maintenance unit 146 may also cause the Raman analyzer 106 to provide one or more Raman scan vectors. The database maintenance unit 146 may then store the Raman scan vectors as a new observation data set in the observation database 136. The observation database 136 may be updated according to any suitable timing, which may vary depending on the embodiment. If the analytical instrument 104 outputs an actual analytical measurement within seconds of measuring a sample, for example, the observation database 136 may be updated with a new measurement almost immediately after the sample is taken. However, in certain other embodiments, the actual analytical measurement may be the result of minutes, hours, or days of processing by one or more analytical instruments 104, in which case the observations database 136 is not updated until such processing is complete. In still other embodiments, different ones of the analytical instruments 104 may incrementally add new observational data sets to the observations database 136 as they complete their respective measurements.

[0038] Thus, the observation database 136 may provide a "dynamic library" of past observations that the local model generator 142 can draw from for model training. In some embodiments, the most recent analytical measurements are always added to the observation database 136, and the local model generator 142 may always use the most recent observation data set in the observation database 136 when calibrating the local model 132. This may enable the local model 132 to encode process information from the recent past and quickly adapt to new conditions, or to new process conditions for which it has no history.

[0039] In some embodiments, only a subset of the scans in the observation database 136 have corresponding actual analytical measurements, in which case the VAE may be trained using all or most of the datasets, and the local model 132 is trained using only selected datasets from among those datasets that have corresponding actual analytical measurements (the analytical measurements are used as labels in training the local model 132).

[0040] Some or all of the processes described above can be repeated many times over the life of a biopharmaceutical process in a bioreactor to continuously monitor the process using a local model that is fully automated and in real time, both in calibration and maintenance. Analytical measurements can be predicted for various purposes, depending on the embodiment and / or scenario. For example, certain parameters can be monitored (i.e., predicted) as part of a quality control process to ensure that the process remains compliant with relevant regulations. As another example, one or more parameters can be monitored / predicted to provide feedback in a closed-loop control system. For example, FIG. 2 illustrates a system 150 similar to system 100, but that attempts to control glucose concentrations in a biopharmaceutical process (i.e., attempts to match a predicted glucose concentration to a desired setpoint within some acceptable tolerance). It will be understood that in other embodiments, system 150 can instead (or similarly) be used to control process parameters other than glucose levels or to control glucose levels based on predictions of one or more other process parameters (e.g., lactate levels). In FIG. 2, the same reference numbers are used to refer to corresponding components of FIG. 1. For example, the VAE-JITL prediction application 130 of FIG. 2 may be the same as the VAE-JITL prediction application 130 of FIG. 1 (for clarity, the various units of the VAE-JITL prediction application 130 are not shown in FIG. 2).

[0041] 2 , in system 150, memory 128 also stores control unit 152. Control unit 152 is configured to control glucose pump 154, i.e., to cause glucose pump 154 ​​to selectively introduce additional glucose into the biopharmaceutical process in bioreactor 102. Control unit 152 may include, for example, software instructions executed by processing unit 120 and / or appropriate firmware and / or hardware. In some embodiments, control unit 152 implements model predictive control (MPC) techniques that use glucose concentration as an input in a closed-loop architecture. In embodiments in which local model 132 provides confidence limits or other confidence indicators with each prediction (e.g., certain embodiments in which local model 132 is a Gaussian process model), control unit 152 may also accept the confidence indicators as inputs. For example, the control unit 152 may only generate control instructions for the glucose pump 154 ​​based on glucose concentration predictions that have a sufficiently high reliability index (e.g., only based on predictions associated with reliability limits that do not exceed a certain percentage or absolute measurement range, or predictions associated with reliability scores above a certain minimum threshold score, etc.), or may increase and / or decrease the weight of certain predictions based on their reliability index, etc.

[0042] Referring now to FIG. 3 , an example data flow 250 that may occur when analyzing a biopharmaceutical process using a JITL technique, such as the VAE-JITL technique described herein, is shown. Data flow 250 may occur, for example, in system 100 of FIG. 1 or system 150 of FIG. 2 . In data flow 250, spectral data 252 is provided by a spectrometer / probe. For example, spectral data 252 may include Raman scan vectors or NIR scan vectors generated by Raman analyzer 106, etc. A query sample 254 is generated (e.g., by query unit 140) based on the spectral data 252 and used to query a global dataset 256, which may include, for example, all observation datasets in observation database 136. Query unit 140 may generate query sample 254 by pre-processing spectral data 252 or simply using spectral data 252 as the query sample 254. Based on the query, a local dataset 258 is identified within global dataset 256. In VAE-JITL, a local dataset 258 is selected using the VAE's encoder, as discussed above and in more detail below. The local dataset 258 is then used as training data (e.g., by the local model generator 142) to calibrate a local model 260 (e.g., the local model 132). The local model 260 is then used (e.g., by the prediction unit 144) to predict an output (analytical measurement) 262, such as a medium component (e.g., glucose or other metabolite or amino acid, etc.) concentration, viability, viable cell density, titer, etc., and optionally also output a confidence limit or another suitable confidence indicator. After the query is made, the process of building / calibrating the local model 260 (e.g., the local model 132) may be performed, for example, as mathematically described in U.S. Patent Application Publication No. 2022 / 0128474.

[0043] An autoencoder is a neural network trained to reconstruct an input through an encoder and a decoder layer. The encoder creates a latent space that represents the main structure of the input information. The decoder uses this information to reconstruct the input layer by minimizing the reconstruction error. However, training an autoencoder without information loss between the input and output layers results in severe overfitting to the dataset, which prevents the autoencoder from generating new content. To solve this problem, the query unit 140 uses a VAE with a regularized latent space, where the distribution is sampled instead of a fixed point.

[0044] FIG. 4 illustrates an example VAE 400. The VAE 400 includes an encoder 402 that operates on inputs (scan values) applied at an input layer 404 and includes a hidden layer 406 that transforms the inputs into a latent space 410. A decoder 420 of the VAE 400 attempts to reconstruct the inputs at an output layer 422 by processing samples 424 in the latent space 410 using a hidden layer 428. A VAE-JITL prediction application 130 or another application and / or computing system may construct the VAE 400, after which a query unit 140 may select similar historical samples / scans using the encoder 402 (without the decoder 420) as described herein. As seen in FIG. 4 , the encoder 402 generates / outputs parameters 426 that define the latent space distribution of the scans applied as input at the input layer 404 (e.g., the mean and variance of the latent space distribution in this example). The latent space distribution is Pr[Z|X], where Z is a sample representing the approximate information of the latent space 410. The desired objective in a VAE is to make this distribution as similar as possible to an intrinsic distribution, such as the standard normal distribution, i.e., N(0,I). The similarity between the latent space distribution Pr[Z|X] and the normal distribution N(0,I) is then obtained by calculating the multivariate Kullback-Leibler (KL) divergence, as described, for example, in Guo et al., Chemometr. Intell. Lab. Syst., 197, A Deep Learning Just-In-Time-Learning Modeling Approach for Soft Sensor Based on Variational Autoencoder (2020). Therefore, the VAE loss function consists of a reconstruction loss term along with a regularization term: Loss=||XD(Z)|| 2 +KL[N(μ,Σ),N,(0,I)] (Eq. 1)

[0045] By minimizing Equation 1 through VAE training based on historical Raman scans, the weights of the VAE400 network can be optimized. Subsequently, the query unit 140 can use the trained encoder 402 as a preprocessing step for feature extraction in VAE-JITL (i.e., before determining which historical Raman scan is most similar to the target Raman scan). As described above, Raman spectroscopy provides a signal with many features (each corresponding to a different Raman shift), but the Euclidean distance or distance between points of these features may not be the best criterion for finding the most similar samples. However, by integrating JITL with the VAE encoder 402, a better combination of features from the dataset samples can be selected for query points for model development

Number

Number

Number

[0046] ​In some embodiments, the VAE 400 may have an architecture different from that shown in Figure 4. For example, the encoder 402 and the decoder 420 may each include exactly one hidden layer. In other embodiments, the encoder 402 and the decoder 420 include no hidden layers, or each include two or more hidden layers. As another example, any of the layers shown may include more nodes than shown in Figure 2, and the encoder 402 may generate parameters 426 such as three or more distributions (e.g., 3, 4, 5, 10, etc.) shown in Figure 4.

[0047] The Raman scans near the query sample are extracted and stored as a local set (i.e.,

number

[0023] A local model 132 is constructed using the local model 132 (e.g., based on Gaussian process regression). See Tulsyan et al., AICHE Journal, e17210, Spectroscopic Models for Real-Time Monitoring of Cell Culture Processes Using Spatiotemporal Just-In-Time Gaussian Processes (2020); Williams, Learning in Graphical Models, 599-621 (1998). Finally, a prediction unit 144 uses the local model 132 to predict analytical measurements of the query sample.

[0048] Figure 5 illustrates an example VAE-JITL process 500 that may be implemented by computing system 110 of Figure 1 or Figure 2. For ease of explanation, VAE-JITL process 500 is described with reference to elements of Figures 1, 2, and 4. Process 500 may be implemented, for example, by VAE-JITL prediction application 130 of computing system 110.

[0049] In the exemplary process 500, when a new Raman scan vector (query scan 502) is acquired (e.g., during runtime operation of the bioreactor 102), the query unit 140 (or another unit or application) processes each of the multiple Raman scan vectors in the observation database 136 (historical scans 504) and the query scan using the same three pre-processing steps: 1) downsampling the scan (step 506), 2) baseline correcting the downsampled scan (step 508), and 3) normalizing the downsampled and baseline corrected scan (step 510). In other embodiments, one or more of steps 506, 508, 510 are omitted, additional steps are included, and / or steps 506, 508, and / or 510 occur in a different order than that shown in FIG.

[0050] An example of baseline correction, such as that which may occur in step 508, is shown in Figures 6A and 6B. Raman spectroscopy provides detailed spectroscopic fingerprint information about molecules in a bioreactor, such as bioreactor 102. However, uncertainties in the scanning process typically result in a blurred baseline, which impacts the accuracy of quantitative analysis. Baseline correction can be an important step for extracting more precise information from raw Raman spectra. However, due to overlapping peaks and strong nonlinearities in Raman data, capturing the baseline with polynomial fitting is a nontrivial task. Recently, a penalized spline smoothing algorithm combined with a vector transformation strategy has been proposed to preserve background signals. This algorithm fits a function as a combination of B-splines by minimizing a regularized least-squares objective function for the Raman data. See Yao-yi et al., Analytical Methods, 10, 3525-3533, Baseline Correction for Raman Spectra Using Penalized Spline Smoothing Based on Vector Transformation (2018).

[0051] FIG. 6A shows a downsampled Raman scan 600 (e.g., downsampled by a factor of 4 or 5), where the x-axis represents feature number (i.e., Raman shift after downsampling) and the y-axis represents intensity. The downsampling in step 506 may occur after query unit 140 truncates the Raman scan to a range that removes noise and / or other undesired phenomena (e.g., for 3344 features / Raman shifts in the original scan, selecting only a range of 500-3000, 600-1800, etc.). Query unit 140 applies a smoothing algorithm (e.g., penalized spline smoothing) to the downsampled Raman scan 600 to generate a baseline Raman scan 602. By subtracting the baseline Raman scan 602 from the Raman scan 600, query unit 140 generates the baseline-corrected Raman scan 610 shown in FIG. 6B.

[0052] Referring again to FIG. 5 , query unit 140 may apply normalization at stage 510 using, for example, min-max normalization. Min-max normalization may be based on minimum and maximum values ​​for a particular experiment or across all experiments. In other embodiments, other suitable types of normalization are used. Query unit 140 (or another unit or application) may also or instead apply other pre-processing steps not shown in FIG. 5 . For example, as described above, the Raman scan may be truncated before (or after) downsampling. As another example, a particular historical data set from observation database 136 is not used by query unit 140. This may be necessary, for example, if a particular experiment used a particular filter that significantly altered the Raman scan signal.

[0053] In step 512, query unit 140 applies the preprocessed data (respectively, historical scan 504 and query scan 502) as input to encoder 402 of VAE 400 to extract dominant features represented by distribution parameters 426. In step 514, query unit 140 uses the output of encoder 402 (i.e., distribution parameters 426) to find the scan of historical scan 504 (also after processing in steps 506, 508, 510, 512) that is most similar to query scan 502 (also after processing in steps 506, 508, 510, 512). In particular, query unit 140 determines similarity based on the respective distribution parameters 426 of historical scan 504 and the distribution parameters 426 of query scan 502. In some embodiments, query unit 140 achieves this using multivariate KL divergence (e.g., as in equation 2 above).

[0054] In step 516, the local model generator 142 generates / calibrates a local model 132 (e.g., a Gaussian process model) based on the K most similar samples, where K is any suitable positive integer. In step 518, the prediction unit 144 predicts an actual measurement (e.g., a particular metabolite concentration, or VCD, or titer, etc.) using the local model 132 and the query scan 502. That is, the prediction unit 144 applies the query scan 502 (after preprocessing in steps 506, 508, 510) as input to the trained / calibrated local model 132.

[0055] 5 occur during runtime operation of the bioreactor 102, e.g., both the query scan 502 and the historical scan 504 are processed (steps 506, 508, 510, 512) after a new query scan 502 is acquired by the Raman analyzer 106 and transmitted to the computing system 110. In other embodiments, the historical scan 504 is processed in steps 506, 508, 510, 512 in an offline manner (e.g., prior to runtime operation of the bioreactor 102), while all other operations (i.e., processing of the query scan 502 in steps 506, 508, 510, 512 and steps 514, 516, 518) are performed after a new query scan 502 is acquired by the Raman analyzer 106 and transmitted to the computing system 110.

[0056] 7A-J show examples of linear JITL ("Linear-JITL") prediction performance and VAE-JITL prediction performance (obtained using embodiments of the invention disclosed herein) against measurements for various types of analytical measurements. To compare the performance of Linear-JITL and (one embodiment of) VAE-JITL, validation tests were performed using the same historical data and the same local model. Both VAE-JITL and Linear-JITL used the same JITL framework, except for the technique used to identify similar scans / samples. Linear-JITL prediction uses a K-nearest neighbor approach with a Euclidean distance metric to find the historical scans most similar to the query scan, while VAE-JITL uses the distribution parameters generated by the VAE encoder and a multivariate KL divergence metric to find the most similar historical scans. In each of FIGS. 7A-J, the y-axis of the diagram is normalized to the maximum value of the prediction, and the x-axis represents the number of Raman scans in the order in which the scans were acquired.

[0057] The performance of both VAE-JITL and Linear-JITL was determined offline by calculating the root mean square error (RMSE) and the mean absolute percentage error (MAPE). These two metrics are standard methods for measuring the difference between actual analytical measurements and model predictions. RMSE is calculated as follows:

number

number

[0058] For VAE training and Gaussian process local model development, Raman features between specific (and identical) ranges were used, and the data was downsampled by 5 times. Furthermore, the VAE was selected as a two-layer neural network with 260 units / node in the input layer and 128 units / node in the (single) hidden layer. The activation function in the hidden layer was Relu, and the VAE was trained for 20 epochs. In other embodiments or applications, the VAE may be trained for a different number of epochs and / or a different activation function may be used. The number of nearest samples (i.e., K) used for Gaussian process local model development was 100.

[0059] Figure 7A shows the predictive performance of VAE-JITL relative to Linear-JITL for the key variable of viable cell density (VCD). Figure 7A shows that both algorithms have similar performance at the beginning of the batch. However, while the VAE-JITL predictions more closely track the actual analytical measurements, the Linear-JITL predictions deviate from the offline measurements midway through the batch (between scans 350 and 450). Around scan 500, Linear-JITL had better predictions than VAE-JITL, as at scan 470 both models did not match the offline measurements.

[0060] Figures 7C and 7K compare the predictive performance of other important variables, glucose concentration (GLC) and titer, respectively. In both cases, similar performance improvements can be observed for VAE-JITL predictions over Linear-JITL predictions. Figure 7C, in particular, clearly shows that VAE-JITL predictions are much closer to the actual analytical measurements, while Linear-JITL predictions track the measurements with undesirable bias. In Figure 7K, although performance is similar, both VAE-JITL and Linear-JITL are unable to "catch up" with the actual measurements near the end of the batch. The reason for this is that historical batches rarely have titer values ​​greater than 1, and as a result, both techniques were unable to extrapolate predictions.

[0061] Figures 7D, 7I, and 7J show the predicted trajectories of lactate concentration (LAC), potassium concentration (K), and osmolality (OSMO), respectively. Figures 7D and 7I confirm that VAE-JITL is better than Linear-JITL in predicting LAC and K. However, Figure 7J shows that VAE-JITL does not necessarily provide better predictions than Linear-JITL for all metabolites. Figure 7J also shows that both models provide good predictions of OSMO trajectories, but introduce biases relative to offline measurements.

[0062] Figures 7B, 7E, 7F, 7G, and 7H compare the prediction performance of viability (VIAB), glutamine concentration (GLN), glutamate concentration (GLU), ammonium concentration (NH), and sodium concentration (Na), respectively. To quantitatively compare the JITL algorithms, the RMSE and MAPE for the corresponding measures of prediction of specific bioprocess variables are presented in Table 1 below. These values ​​confirm that VAE-JITL generally provides better performance than Linear-JITL. However, the RMSE and MAPE of OSMO indicate that VAE-JITL does not outperform Linear-JITL in all cases.

[0063] [Table 1]

[0064] 8 is a flow diagram of an example method 800 for analyzing a biopharmaceutical process (e.g., for monitoring and / or control purposes) using VAE-JITL. Method 800 may be implemented by a computer (e.g., by processing unit 120 executing instructions of VAE-JITL prediction application 130), such as computing system 110 of FIG. 1 or FIG. 2.

[0065] In block 802, an observational database (e.g., observational database 136) containing multiple observational data sets associated with past scans of the biopharmaceutical process is queried based on first spectral scan vectors of the biopharmaceutical process acquired by a spectroscopic system (e.g., Raman analyzer 106 and Raman probe 108). Each of the observational data sets includes spectral data (e.g., Raman scan vectors) and corresponding actual analytical measurements (e.g., any one of the variables shown in Table 1). The query in block 802 includes determining, using an encoder of the VAE (e.g., encoder 402 of VAE 400), a first parameter defining a distribution set (e.g., mean and variance of a set of normal or approximately normal distributions) of the first spectral scan vector. The query also includes selecting particular observational data sets as training data from among the multiple observational data sets based on the first parameter and further based on other parameters defining the distribution sets of each of the multiple observational data sets (e.g., mean and variance of a set of normal (or approximately normal) distributions, where each distribution set corresponds to a different observational data set).

[0066] Selecting particular observation datasets as training data in block 802 may include calculating a multivariate KL divergence metric based on the first parameter and other parameters (e.g., as in Equation 2). In some embodiments, block 802 also includes preprocessing the first spectral scan vector (e.g., as in steps 506, 508, and / or 510 of FIG. 5 ) before determining the first parameter, where the spectral data of the multiple observation datasets are similarly preprocessed (i.e., before generating the distribution sets for the multiple observation datasets).

[0067] In block 804, the selected training data is used to calibrate a local model specific to the biopharmaceutical process (eg, local model 132).

[0068] In block 806, an analytical measurement of the biopharmaceutical process is predicted using the local model. The analytical measurement may be any one of the variables in Table 1, such as metabolite concentration or viability, VCD, osmolality, or titer, or another suitable type of analytical measurement, for example. The analytical measurement is of the same type as the measurements used as labels when training the local model in block 804.

[0069] In some embodiments, method 800 includes one or more additional blocks not shown in Figure 8. For example, method 800 may include an additional block in which at least one parameter of a biopharmaceutical process is controlled based at least in part on the predicted analytical measurement in block 806. Depending on the embodiment, the parameter may be of the same type as the predicted analytical measurement (e.g., controlling a glucose concentration based on a predicted glucose concentration) or may be of a different type. Model predictive control (MPC) techniques may be used, for example, to control the parameter (or parameters).

[0070] As another example, method 800 may include a first additional block that causes the predicted analytical measurements to be displayed in a user interface (e.g., presented on display 124). As yet another example, method 800 may include one or more sets of additional blocks, each similar to blocks 802-806. In each of these sets of additional blocks, a local model may be calibrated by querying the observational database (or other observational databases) and used to predict different types of analytical measurements.

[0071] Additional considerations related to this disclosure are now addressed.

[0072] Some of the drawings described herein show example block diagrams having one or more functional components. It will be understood that such block diagrams are for illustrative purposes, and that the devices described and shown may have more, fewer, or alternative components than those shown. Additionally, in various embodiments, the components (and the functionality provided by each component) may be associated with or otherwise integrated as part of any suitable component.

[0073] Embodiments of the present disclosure relate to non-transitory computer-readable storage media having computer code thereon for performing various computer-implemented operations. The term "computer-readable storage medium" is used herein to include any medium capable of storing or encoding a set of instructions or computer code for performing the operations, methods, and techniques described herein. The medium and computer code may be of a type specially designed and constructed for the embodiments of the present disclosure, or may be of a type known and available to those skilled in the computer software arts. Examples of computer-readable storage media include, but are not limited to, magnetic media such as hard disks, floppy disks, and magnetic tape; optical media such as CD-ROMs and holographic devices; magneto-optical media such as optical disks; and hardware devices specially configured to store and execute program code, such as ASICs, programmable logic devices ("PLDs"), and ROM and RAM devices.

[0074] Examples of computer code include machine code, such as produced by a compiler, and files containing high-level code executed by a computer using an interpreter or compiler. For example, embodiments of the present disclosure may be implemented using Python, Java, C++, or other object-oriented programming languages ​​and development tools. Further examples of computer code include encrypted and compressed code. Furthermore, an embodiment of the present disclosure may be downloaded as a computer program product, which may be transferred over a transmission channel from a remote computer (e.g., a server computer) to a requesting computer (e.g., a client computer or a different server computer). Another embodiment of the present disclosure may be implemented in hardwired circuitry instead of, or in combination with, machine-executable software instructions.

[0075] As used herein, the singular terms "a," "an," and "the" can include plural references unless the context clearly dictates otherwise.

[0076] As used herein, the terms "connect," "connected," and "connection" mean an operative coupling or linking. Connected components may be coupled to each other directly or indirectly, for example, through a set of other components.

[0077] As used herein, the terms "nearly," "substantially," "substantial," and "about" are used to describe and explain small variations. When used in conjunction with an event or circumstance, these terms can refer not only to instances in which the event or circumstance occurs exactly, but also to instances in which the event or circumstance occurs approximately. For example, when used in conjunction with a numerical value, these terms can refer to a variation range of ±10% or less of the numerical value, e.g., ±5% or less, ±4% or less, ±3% or less, ±2% or less, ±1% or less, ±0.5% or less, ±0.1% or less, or ±0.05% or less. For example, two numerical values ​​can be considered "substantially" the same if the difference between the numerical values ​​is ±10% or less of the mean value, e.g., ±5% or less, ±4% or less, ±3% or less, ±2% or less, ±1% or less, ±0.5% or less, ±0.1% or less, or ±0.05% or less.

[0078] Additionally, amounts, ratios, and other numerical values ​​may be presented herein in a range format. It should be understood that such range format is used for convenience and brevity and should be understood flexibly to include not only the numerical values ​​explicitly stated as the limits of a range, but also to include all individual numerical values ​​or subranges subsumed within that range, as if each numerical value and subrange were expressly stated.

[0079] While the present disclosure has been described and illustrated with reference to specific embodiments, these descriptions and illustrations are not intended to limit the disclosure. Those skilled in the art will recognize that various modifications may be made and equivalents may be substituted without departing from the true spirit and scope of the present disclosure, as defined by the appended claims. Illustrations may not necessarily be to scale. Differences between the artistic depictions in this disclosure and the actual devices may occur due to manufacturing processes, tolerances, and / or other reasons. There may be other embodiments of the present disclosure not specifically illustrated. The specification (other than the claims) and drawings should be considered illustrative, not restrictive. Modifications may be made to adapt a particular situation, material, composition, technique, or process to the objective, spirit, and scope of the present disclosure. All such modifications are intended to be within the scope of the claims appended hereto. While the technology disclosed herein has been described with specific operations performed in a particular order, it will be understood that these operations may be combined, subdivided, or reordered to form equivalent technology without departing from the teachings of the present disclosure. Thus, unless specifically indicated herein, the order and grouping of the operations is not a limitation of the present disclosure.

Claims

1. 1. A computer-implemented method for monitoring and / or controlling a biopharmaceutical process, comprising: and querying, by one or more processors and based on a first spectral scan vector of the biopharmaceutical process acquired by a spectroscopic system, an observational database including a plurality of observational data sets associated with past scans of the biopharmaceutical process, each of the observational data sets including spectral data and corresponding actual analytical measurements, wherein querying the observational database includes: determining first parameters defining a distribution set of the first spectral scan vectors; selecting, as training data, particular observational data sets from among the plurality of observational data sets based on (i) the first parameter and (ii) other parameters defining a set of distributions for each of the plurality of observational data sets; and calibrating, by the one or more processors and using the selected training data, a local model specific to the biopharmaceutical process, wherein the local model is trained to predict analytical measurements based on spectral data input; and predicting, by the one or more processors, an analytical measurement of the biopharmaceutical process, wherein predicting the analytical measurement of the biopharmaceutical process includes using the local model to analyze spectral data generated by the spectroscopy system when scanning the biopharmaceutical process; 10. A computer-implemented method comprising:

2. 2. The computer-implemented method of claim 1, wherein determining the first parameters comprises processing a query sample using an encoder of a variational autoencoder, the encoder outputting the first parameters.

3. The computer-implemented method of claim 2 , wherein the encoder includes exactly one hidden layer.

4. The computer-implemented method of claim 2 or 3, further comprising determining, by the one or more processors, the other parameters using the encoder of the variational autoencoder, the encoder outputting the other parameters.

5. 5. The computer-implemented method of claim 1, wherein selecting the particular observation data set comprises calculating a multivariate KL-divergence metric based on the first parameter and the other parameters.

6. 6. The computer-implemented method of claim 1, wherein calibrating the local model specific to the biopharmaceutical process comprises calibrating a Gaussian process machine learning model specific to the biopharmaceutical process.

7. Querying the observation database includes downsampling the first spectral scan vector; and The computer-implemented method of any one of claims 1 to 6, wherein analyzing the spectral data using the local model comprises downsampling the spectral data.

8. Querying the observation database includes baseline correcting the downsampled first spectral scan vector; and The computer-implemented method of claim 7 , wherein analyzing the spectral data using the local model includes baseline correcting the downsampled second spectral data.

9. Querying the observation database includes normalizing the downsampled and baseline corrected first spectral scan vector; and 9. The computer-implemented method of claim 8, wherein analyzing the spectral data using the local model comprises normalizing the downsampled and baseline-corrected spectral data.

10. 10. The computer-implemented method of claim 1, wherein analyzing the spectral data using the local model comprises analyzing the first spectral scan vector using the local model.

11. The computer-implemented method of any one of claims 1 to 10, wherein the predicted analytical measure of the biopharmaceutical process is a metabolite concentration.

12. The computer-implemented method of any one of claims 1 to 10, wherein the predicted analytical measurement of the biopharmaceutical process is osmolality, viability, viable cell density or titer.

13. The computer-implemented method of any one of claims 1 to 12, wherein the spectroscopy system is a Raman spectroscopy system.

14. 14. The computer-implemented method of any one of claims 1 to 13, further comprising controlling, by the one or more processors, at least one parameter of the biopharmaceutical process based on the predicted analytical measurements of the biopharmaceutical process.

15. The computer-implemented method of any one of claims 1 to 14, further comprising causing the one or more processors to display the predicted analytical measurements in a user interface.

16. 1. A spectroscopy system comprising: one or more spectroscopic probes collectively configured to (i) deliver source electromagnetic radiation to a biopharmaceutical process, and (ii) collect electromagnetic radiation while the source electromagnetic radiation is delivered to the biopharmaceutical process; a computing system configured to perform the method according to any one of claims 1 to 15; A spectroscopic system including:

17. A non-transitory computer readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of any one of claims 1 to 15.

Images