Instant learning of cell culture process monitoring and / or control with variational autoencoder

Abstract

A method for monitoring and / or controlling a biopharmaceutical process includes querying an observation database including an observation dataset associated with past scans based on a first spectral scan vector of the biopharmaceutical process. Each of these observation data sets includes spectral data and a corresponding actual analytical measurement. Query the observation database includes determining first parameters defining a distribution set of the first spectral scan vector, and selecting a particular observation dataset from the observation datasets as training data based on (i) the first parameters and (ii) other parameters defining a respective distribution set of the observation datasets. The method also includes calibrating a local model specific to the biopharmaceutical process using the selected training data. The method further includes predicting analytical measurements of the biopharmaceutical process by analyzing spectroscopic data generated when scanning the biopharmaceutical process using the local model.

Description

Technical Field

[0001] The present application relates generally to monitoring and / or controlling biopharmaceutical processes using spectroscopic techniques such as Raman spectroscopy, and more particularly to predicting or inferring product quality attributes based on spectral scans using just-in-time learning (JITL) models. Background Art

[0002] The stable production of biopharmaceutical processes (e.g., biotherapeutic proteins) generally requires bioreactors to maintain balanced and consistent parameters (e.g., cell metabolism concentrations), which in turn requires strict process monitoring and control. In order to meet these needs, process analytical technology (PAT) tools are increasingly being used. Online monitoring of cell culture pH, dissolved oxygen, and temperature are some examples of traditional PAT tools that have been used in feedback control systems. In recent years, other process probes have been studied and deployed to continuously monitor more complex species, such as viable cell density (VCD), glucose, lactate, and other key cell metabolites (including amino acids), titers, and key quality attributes.

[0003] Raman spectroscopy is a popular PAT tool that is widely used for online monitoring in biomanufacturing. Raman spectroscopy is an optical method that enables non-destructive analysis of chemical composition and molecular structure. In Raman spectroscopy, incident laser light is scattered inelastically due to molecular vibrational modes. The frequency difference between the incident photon and the scattered photon is called the "Raman shift", and the vector of the Raman shift and intensity level (referred to herein as the "Raman spectrum", "Raman scan" or "Raman scan vector") can be analyzed to determine the chemical composition and molecular structure of the sample. With the improvement of laser sampling and detector technology, the application of Raman spectroscopy in polymer, pharmaceutical, biomanufacturing and biomedical analysis has surged in the past three decades. Due to these technological advances, Raman spectroscopy has now become a practical analytical technique used in and outside the laboratory. Since the first reported application of in situ Raman measurements in biomanufacturing, in situ Raman measurements have been used to provide online real-time predictions of several key process states such as glucose concentration, lactate concentration, glutamate concentration, glutamine concentration, ammonium concentration, potassium concentration, sodium concentration, viability, VCD, osmotic pressure and titer. These predictions are typically based on a calibration model or a soft sensor model built in an offline environment. The model is built using analytical measurements from analytical instruments. Partial least squares (PLS) and multiple linear regression modeling methods are typically used to correlate the Raman spectra to the analytical measurements. These models typically require pre-processing (filtering) of the Raman scans before calibration against the analytical measurements. Once the calibration model is trained, it can be implemented in a real-time environment to provide in-situ measurements for process monitoring and / or control.

[0004] Since biopharmaceutical processes are usually operated under strict constraints and regulations, Raman model calibration for biopharmaceutical applications is very necessary. Raman model calibration in the biopharmaceutical industry usually involves performing multiple activity tests to generate relevant data for associating Raman spectra with (multiple) analytical measurement results. These tests are both expensive and time-consuming, because each activity may last two to four weeks in a laboratory environment. In addition, only limited samples can be used for analytical instruments (for example, to ensure that laboratory-scale bioreactors maintain a healthy amount of viable cells). In fact, it is not uncommon to only be able to perform one or two measurements per day in online or offline analytical instruments. To further exacerbate the situation, these models are related to specific processes, specific formulations or configurations of bioreactor culture media, and specific operating conditions. Therefore, if any of the above variables are to be changed, the model may need to be recalibrated based on new data. Both Raman model calibration and model maintenance require a lot of resource allocation and are usually performed in an offline environment.

[0005] To address the problem of model performance degradation over time due to process variability, a variety of soft sensor techniques such as moving windows, time differences, and recursive modeling have been implemented. See Qin et al., Computers and Chemical Engineering, 22, 503-514, Recursive PLS Algorithms for Adaptive Data Modeling (1998); Kaneko et al., Industrial and Engineering Chemistry Research, 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 fully considers or addresses sudden changes in industrial processes.

[0006] To better account for sudden changes, a just-in-time learning (JITL) technique is proposed for automatic calibration and evaluation 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[Automatic real-time calibration, evaluation and maintenance of generic Raman models for online monitoring of cell culture processes](2020). JITL is a just-in-time modeling platform based on local modeling and database sampling techniques. Unlike other machine learning approaches, JITL typically assumes that all available observations are stored in a central observation database, and local models are dynamically built in real time based on a query sample (e.g., a new Raman scan), ideally using the most "similar" or relevant data from that observation database. This allows complex process dynamics to be well approximated using relatively simple local models. Under the JITL framework, a library can contain more than just spectral data for a single process operating under specific operating conditions, but can also contain data for different processes, different media configurations, and / or different operating conditions. This can significantly reduce the time required to calibrate and maintain models, especially for drug products in development that may have little or no past production history.

[0007] Traditionally, in JITL, "similar" historical samples are identified in a database of observations 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, 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 seek historical samples that are most relevant to a query sample in a deterministic and point-to-point manner. See ZQ Ge et al., Chemometrics and Intelligent Laboratory Systems, 104(13), 306-317, A Comparative Study of Just-In-Time-Learning Based Methods for Online Soft Sensor Modeling (2010). However, Raman scan results may reflect considerable uncertainty that is not taken into account by deterministic techniques. Therefore, the uncertainty in the sample / scan may lead to a relatively poor selection of "similar" historical samples and, therefore, to relatively poor predictive performance of the JITL local model built based on the selected historical samples. Summary of the invention

[0008] To address the above-mentioned problems associated with just-in-time learning (JITL) techniques for biopharmaceutical applications, the systems and methods disclosed herein take into account uncertainty in spectral measurements by using distributions in addition to deterministic values ​​to identify "similar" historical samples (e.g., similar Raman scans) in a database of observations. In particular, a computing system can use historical samples to train / generate a variational autoencoder (VAE) including an encoder and a decoder. The encoder transforms each input sample (scan) into a lower-dimensional latent space representation that includes parameters (e.g., mean and variance) that define the distribution of the input sample, 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 the parameters that define the distribution set for each historical sample (e.g., each historical Raman scan) and the parameters that define the distribution set for a sample of interest (e.g., a new real-time Raman scan), and use the determined parameters to identify / select those historical samples that have a distribution that is most similar to the sample of interest. For example, the selection stage may include using multivariate KL (Kullback-Leibler) divergence to identify the most similar historical samples. In some embodiments, the encoder includes only one hidden layer.

[0009] In addition to the benefits of JITL (e.g., preventing model degradation over time and tracking sudden changes in biopharmaceutical processes), integrating VAE into JITL also helps to identify more relevant JITL training samples based on sample distribution. By selecting better historical samples to train / build local models, local models can better predict analytical measurements such as metabolite concentrations, viable cell density, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Those skilled in the art will appreciate that the drawings described herein are included for illustrative purposes and are not intended to limit the present disclosure. The drawings are not necessarily drawn to scale, but rather focus on illustrating the principles of the present disclosure. It should be understood that in some cases, various aspects of the described embodiments may be enlarged or magnified to aid in understanding the described embodiments. In the drawings, similar reference numerals generally refer to functionally similar and / or structurally similar components throughout the various drawings.

[0011] Figure 1 is a simplified block diagram of an example Raman spectroscopy system combined with VAE-JITL for predicting analytical measurement results of a biopharmaceutical process for monitoring purposes.

[0012] Figure 2 is a simplified block diagram of an example Raman spectroscopy system combined with VAE-JITL for predicting analytical measurements of a biopharmaceutical process for closed-loop control of glucose concentration.

[0013] Figure 3 Depicts an example data flow that might occur when analyzing a biopharmaceutical process using a JITL technique such as VAE-JITL.

[0014] Figure 4 Depicts an example VAE whose encoder part can be composed of Figure 1 or Figure 2 The VAE-JITL predictor application is used to query samples from Figure 1 or Figure 2 Select Historical Scan from the Observation Results Database.

[0015] Figure 5 Describes the Figure 1 or Figure 2 An example VAE-JITL process implemented by a computing system.

[0016] Fig. 6A and Figure 6B Example downsampled Raman scans before and after baseline correction are depicted, respectively.

[0017] 7A to 7J Examples of VAE-JITL prediction performance relative to linear JITL prediction performance and measured values ​​are plotted for various types of analytical measurements.

[0018] Figure 8 is a flowchart of an example method for analyzing a biopharmaceutical process using VAE-JITL. DETAILED DESCRIPTION

[0019] The various concepts introduced above and discussed in more detail below may be implemented in any of a variety of ways, and the concepts described are not limited to any particular implementation.For purposes of illustration, examples of implementation are provided.

[0020] Figure 1 is a simplified block diagram of an example Raman spectroscopy system 100 that can be used to predict analytical measurement results for a biopharmaceutical process. Figure 1 A system 100 implementing Raman spectroscopy is depicted, but it should be understood that in other embodiments, the system 100 may implement other spectroscopy techniques suitable for analyzing biopharmaceutical processes, such as near infrared (NIR) spectroscopy and mass spectrometry. For ease of explanation, the following description (applicable to all figures) specifically refers to the Raman spectroscopy embodiment.

[0021] 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. Bioreactor 102 can be any suitable container, device, or system that supports a biologically active environment, which can include living organisms in a culture medium and / or substances derived from these living organisms (e.g., cell cultures). Bioreactor 102 can contain recombinant proteins expressed by cell culture, for example, such as for research purposes, clinical use, commercial sales, or other distribution. Depending on the biopharmaceutical process being monitored, the culture medium can include specific fluids (e.g., "liquid culture medium") and specific nutrients, and can have a target pH level or range, a target temperature or temperature range, etc. The contents and parameters / characteristics of the culture medium are collectively referred to as "culture medium configuration" herein.

[0022] The analytical instrument(s) 104 can be any in-line, on-line, and / or off-line instrument(s) configured to measure one or more characteristics or parameters of the bioactive contents within the bioreactor 102 based on samples obtained from the bioactive contents. For example, the analytical instrument(s) 104 can 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(s) 104 can measure osmotic pressure, 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 specific example, a sample may be collected, spun down, purified through multiple columns, and may pass through a first analytical instrument 104 (e.g., a high performance liquid chromatography (HPLC) or ultra high performance liquid chromatography (UPLC) instrument), and then through a second analytical instrument 104 (e.g., a mass spectrometer), where both the first and second analytical instruments 104 provide analytical measurements. One, some, or all of the analytical instrument(s) 104 may use destructive analytical techniques.

[0023] The Raman analyzer 106 may include a spectrometer device coupled to a Raman probe 108 (or, in some embodiments, to multiple Raman probes). The Raman analyzer 106 may include a laser source that delivers 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 to record, for example, a signal received from the Raman probe 108 via another channel of the fiber optic cable. Alternatively, the laser source may be integrated into 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).

[0024] The Raman analyzer 106 and the Raman probe 108 are collectively configured to non-destructively scan the bioactive content during the biopharmaceutical process within the bioreactor 102 by exciting, observing, and recording the molecular "fingerprint" of the biopharmaceutical process. The molecular fingerprint corresponds to the vibrational, rotational, and / or other low-frequency modes of the molecules within the bioactive content within the biopharmaceutical process when the contents of the bioreactor are excited by the laser light delivered by the Raman probe 108. As a result of this scanning process, the Raman analyzer 106 generates Raman scan vectors, each of which represents intensity as a function of the Raman shift (frequency).

[0025] The computing system 110 may be a single computing device, or include more than one co-located and / or distributed computing devices. The computing system 110 is coupled to the Raman analyzer 106 and the analytical instrument(s) 104, and is generally configured to analyze the Raman scan vectors generated by the Raman analyzer 106 in order to predict one or more analytical measurement results of the biopharmaceutical process. For example, the computing system 110 may analyze the Raman scan vectors in order to predict the analytical measurement results of the same type(s) performed by the analytical instrument(s) 104. As a more specific example, the computing system 110 may predict the glucose concentration, while the analytical instrument(s) 104 actually measures the glucose concentration. However, while the analytical instrument(s) 104 may perform relatively infrequent "offline" analytical measurements of samples extracted from the bioreactor 102 (e.g., due to the limited number of biopharmaceutical processes and / or due to the high cost of performing such measurements, etc.), the computing system 110 may perform relatively frequent "online" predictions of the analytical measurement results in real time. In one embodiment, a Raman scan is collected every 30 minutes and an analytical measurement is taken every 24 hours (i.e., such that a Raman scan is taken exactly at the same time as an analytical measurement each day). It should be understood that as used herein, terms such as "predict" or "prediction" do not necessarily refer to the determination or estimation of future values, but can refer to the inference of current values.

[0026] exist Figure 1In the illustrated example embodiment, the computing system 110 includes a processing unit 120, a network interface 122, a display 124, a user input device 126, and a memory 128. The processing unit 120 includes one or more processors, each of which may be a programmable microprocessor that executes software instructions stored in the memory 128 to perform some or all of the functions of the computing system 110 as described herein. Alternatively, one, some, or all of the processors in the processing unit 120 may be other types of processors (e.g., an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc.), and the functions of the computing system 110 as described herein may be implemented in part or in whole in hardware instead. The memory 128 may include one or more physical memory devices or units including volatile and / or non-volatile memory. Any suitable one or more memory types may be used, such as read-only memory (ROM), random access memory (RAM), solid-state drive (SSD), hard disk drive (HDD), etc.

[0027] The 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(s) 104, Raman analyzer 106, and / or observation database 136) via one or more networks using one or more communication protocols. For example, the network interface 122 may be or include an Ethernet interface, and / or include a wireless local area network (LAN) interface, etc.

[0028] The display 124 may use any suitable display technology (e.g., LED, OLED, LCD, etc.) to present information to the user, and the user input device 126 may be a keyboard or other suitable input device. In some embodiments, the display 124 and the user input device 126 are integrated into a single device (e.g., a touch screen display). Typically, the display 124 and the user input device 126 may be combined to enable a user to interact with a graphical user interface (GUI) provided by the computing system 110, such as for purposes such as manually monitoring various processes being executed within the system 100. However, in some embodiments, the computing system 110 does not include the display 124 and / or the user input device 126, or one or both of the display 124 and the user input device 126 are included in another computer or system communicatively coupled to the computing system 110 (e.g., in some embodiments where predictions are sent directly to a control system implementing closed-loop control).

[0029] The memory 128 stores instructions for one or more software applications including a variational autoencoder (VAE) just-in-time learning (JITL) predictor application 130 (also referred to herein as the "VAE-JITL predictor application 130"). The VAE-JITL predictor application 130, when executed by the processing unit 120, is generally configured to predict analytical measurements of a biopharmaceutical process in the bioreactor 102 by selecting historical samples / scans from an observation database 136 using an encoder of the VAE, generating / building / calibrating a local model 132 using the selected samples, and analyzing Raman scan vectors generated by the Raman analyzer 106 using the calibrated local model 132. Depending on how often the Raman analyzer 106 generates such scan vectors, the VAE-JITL predictor application 130 may predict analytical measurements based on a periodic time or other suitable time. The Raman analyzer 106 itself may control when the scan vectors are generated, or the computing system 110 may trigger the generation of the scan vectors by sending a command to the Raman analyzer 106. The VAE-JITL predictor 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 predictor applications (e.g., each similar to the VAE-JITL predictor application 130) each generate a different local model to predict different types of analytical measurements all based on the same scan vector. The VAE-JITL predictor application 130 and the local models 132 are discussed in further detail below.

[0030] The observation database 136 stores historical observation data sets associated with past observations. Observations / data sets can be selected, for example, by removing outliers and inputting missing data. Each observation data set in the observation database 136 can include spectral data (e.g., such Raman scan vectors generated by the Raman analyzer 106) and one or more corresponding analytical measurements (e.g., such one or more measurements generated by (multiple) analytical instruments 104). Depending on the embodiment and / or scenario, past observations may have been collected for a variety of different biopharmaceutical processes, under a variety of different operating conditions (e.g., different metabolite concentration set points), and / or in a variety of different culture medium configurations (e.g., different fluids, nutrients, pH levels, temperatures, etc.). Typically, it may be desirable to have the observation database 136 represent a variety of processes, operating conditions, and culture medium configurations. However, depending on the embodiment (as further discussed below), the observation database 136 may or may not store information indicating those processes, cell lines, proteins, metabolites, operating conditions, and / or culture medium configurations.

[0031] It should be understood that other architectures, configurations and / or components may be used instead. 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, Figure 1 Some of the functions represented in (e.g., some of the functions of query unit 140) may instead be performed by the database server. This architecture may be desirable for collecting a larger number of observation result data sets for storage in observation result database 136 (e.g., if the database server is connected to multiple computing systems that are each similar to computing system 110). As another example of a different architecture, different computers ( Figure 1 (not shown) the measurement results provided by the analytical instrument(s) 104 may be transmitted to the computing system 110.

[0032] During run-time operation of the system 100, the biopharmaceutical process in the bioreactor 102 is scanned using the Raman analyzer 106 and the Raman probe 108 (i.e., a Raman scan vector of the biopharmaceutical process is generated), and the Raman scan vector(s) 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 (performed by the VAE-JITL predictor application 130) according to a predetermined monitoring period schedule (such as once per minute or once per hour, etc.). Alternatively or additionally, Raman scans may be collected, and predictions may be made based on these scans at irregular intervals (e.g., in response to some process-based trigger, such as a change in measured pH level and / or temperature), such that each monitoring period has a variable or indeterminate duration.

[0033] The query unit 140 of the VAE-JITL predictor application 130 uses the scan vector(s) received during the monitoring period to generate query points that will be used to query the observation database 136. In some embodiments, the query points (i.e., the data defining the query points, also referred to herein as "query samples") include only data representing the Raman scan vectors received from the Raman analyzer 106 (e.g., including the intensity / frequency tuples of the scan vectors). 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 set points in the control system, or laser wavelengths and / or intensities associated with the Raman analyzer 106 or Raman probe 108, etc.), data representing the media configuration of 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).

[0034] Typically, these query points may include data representing the same vectors, parameters, and / or classifications used as inputs by the local model 132 (i.e., as feature sets for the local model 132). Using a variety of different data types (e.g., operating conditions, culture medium configuration data, etc., as described above) for feature sets can improve the accuracy of predictions of analytical measurements made by the local model 132. However, since each observation data set in the observation database 136 will typically need to include the same vectors, parameters, and / or classifications as the feature set, it may be preferred to limit the query points, and the feature set / inputs of the local model 132, to only include Raman scan vectors. This can provide various benefits, such as allowing more information to be collected for storage in the observation database 136, and / or simplifying the collection of this information. For example, if only Raman scan vectors are used, these observation data sets may be included in the observation database 136 even if little or nothing is known about the process, cell line, protein, metabolite, operating condition, and / or culture medium configuration that existed when the observation data set was collected.

[0035] The query unit 140 then uses the generated query points to query the observation database 136. After receiving the query point / query sample, the query unit 140 uses the query sample to select a relevant observation data set from the observation database 136 to be used specifically as training data for the local model 132. In some embodiments, the VAE-JITL predictor application 130 pre-processes the Raman scan vectors to generate the query samples. For example, as discussed further below, the VAE-JITL predictor application 130 may generate the query samples by down-sampling the Raman scan vectors, performing baseline correction on the down-sampled vectors, and / or normalizing the down-sampled and baseline-corrected vectors.

[0036] To determine which observation datasets are most "correlated" (most similar, most relevant, 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 a VAE-JITL predictor application 130 or another application of the computing system 110 or another computing system) based on multiple Raman scans (e.g., downsampled, baseline-corrected, and / or normalized Raman scans) from the observation 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 scans that represent the input Raman scans in a lower dimension. Specifically, the encoder layer of the trained VAE (also referred to herein as an "encoder") can generate parameters that define multiple distributions representing the Raman scans input to the encoder. For example, the encoder can generate a mean and variance for each of multiple (e.g., 2, 3, 5, 10, etc.) normal (or approximately normal) distributions for a given input Raman scan. The parameters that define a distribution or set of distributions are sometimes referred to herein as "distribution parameters."

[0037] Upon identifying the most relevant data set, the query unit 140 (1) generates distribution parameters for each of a plurality (e.g., all) of the Raman scan vectors stored in the observation database 136, and (2) generates distribution parameters for the query sample. It should be appreciated that each Raman scan vector may be pre-processed (e.g., down-sampled, baseline corrected, and / or normalized) prior to inputting the historical Raman scan vectors into the encoder and prior to 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 observation database 136 at any time (e.g., offline, prior to run-time operation of the bioreactor 102, or during run-time operation). However, the query unit 140 generates the distribution parameters for the query sample in real-time (e.g., during run-time operation of the bioreactor 102 as scans are provided by the Raman analyzer 106). Once the distribution parameters for both the historical scans and the new (query sample) scans are available, the query unit 140 may select a particular data set / scan from the observation database 136 based on these distribution parameters. For example, query unit 140 may select the most relevant / similar scans from observation database 136 by selecting those scans that have the lowest multivariate KL divergence with the query sample. VAE and multivariate KL divergence are discussed in further detail below.

[0038] In some embodiments, the query unit 140 considers one or more other factors in addition to the distribution parameters when selecting relevant scans. For example, in order to better adapt to time-varying process changes, the query unit 140 may further consider the timing or order of samples in the observation result database 136 (e.g., which samples are the most recent). To this end, in addition to using VAE-JITL to select historical samples based on distribution parameters, the query unit 140 may also combine the "adaptive" JITL (A-JITL) method or the "spatiotemporal" JITL (ST-JITL) method described in the following literature: U.S. Patent Publication No. 2022 / 0128474 (Tulsyan, "Automatic Calibration and Automatic Maintenance of Raman Spectroscopic Models for Real-Time Predictions"), the entire contents of which are incorporated herein by reference. More generally, any of the techniques described in U.S. Patent 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.

[0039] In some embodiments, query unit 140 selects only a predetermined number of relevant observation datasets in response to a single query, or selects no more than a certain maximum allowed number of relevant observation datasets to ensure that only a relatively small subset of all datasets in observation database 136 are retrieved. However, in other embodiments, query unit 140 may select any number of relevant observation datasets as long as each selected dataset satisfies a suitable relevance criterion (e.g., as long as the multivariate KL divergence is below a predetermined threshold).

[0040] After identifying relevant / similar observation data sets (where each observation data set may or may not correspond to the same process conditions as the currently monitored biopharmaceutical process in the bioreactor 102), the query unit 140 provides or indicates these data sets (e.g., Raman scan vectors and (multiple) corresponding analytical measurements) to the local model generator 142. The local model generator 142 then uses these relevant data sets as training data to calibrate the local model 132. That is, the local model generator 142 uses the Raman scan vectors (and possibly other data) associated with each observation data set (possibly after some pre-processing) as a feature set, and uses the (multiple) analytical measurements associated with the same observation data set as labels for this feature set.

[0041] 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 almost any process changes. Unlike partial least squares (PLS) models and principal component regression (PCR) models, the Gaussian process model uses a non-parametric approach and is able to capture complex nonlinear correlations between Raman scan vectors and analytical measurement results even when using a very limited number of training samples. This may be particularly important when new products or new processes correspond to only a limited number of data sets in the observation database 136. In this case, the Gaussian process model is usually able to extract the most information from those limited data sets in combination with other related data sets selected by the query unit 140 from the observation database 136. However, in other embodiments, as long as the training time does not exceed the minimum expected duration of the monitoring period, the local model generator 142 can alternatively construct any other suitable type of machine learning model (e.g., recursive neural network, convolutional neural network, etc.). The local model generator 142 can also construct the local model 132 so that the local model 132 can output a credibility boundary, or some other suitable prediction confidence indicator (e.g., confidence score). At least compared to PLS models and PCR models, Gaussian process models are particularly suitable for providing confidence boundaries around predictions of analytical measurement results. Although various advantages of Gaussian process models relative to PLS models and PCR models have been described, it should be understood that in some embodiments, local model generator 142 can use PLS, PCR, or other modeling methods to construct local model 132 (e.g., to speed up calibration, or to make it easier to deploy in an industrial environment, etc.).

[0042] The local model generator 142 can build the local model 132 in an online, real-time manner, so that the prediction unit 144 can then use the trained local model 132 to predict one or more analytical measurement results of the biopharmaceutical process by processing the same Raman scan vectors that have been used by the query unit 140 to generate the query points. In fact, in some embodiments, each time the Raman analyzer 106 provides a new Raman scan vector to the computing system 110, the query unit 140 can execute a new query and the local model generator 142 can generate a new version of the local model 132. However, in other embodiments, the query unit 140 executes new queries less frequently (and the local model generator 142 generates a new version of the local model 132 less frequently), such as once every 10 prediction / monitoring cycles, once every 100 prediction / monitoring cycles, etc.

[0043] The database maintenance unit 146 may also cause the (multiple) analytical instruments 104 to periodically collect one or more actual analytical measurement results at a significantly lower frequency than the monitoring cycle of the Raman analyzer 106 (e.g., only once or twice a day, etc.). In some embodiments, the (multiple) measurement results of the (multiple) analytical instruments 104 may be destructive and require permanent removal of the sample from the process in the bioreactor 102. At or near the time when the database maintenance unit 146 causes the (multiple) analytical instruments 104 to collect and provide (multiple) actual analytical measurement results, 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 cause the observation database 136 to store the (multiple) Raman scan vectors as (multiple) new observation data sets. The observation database 132 may be updated according to any suitable timing, which may vary according to the embodiment. For example, if the (multiple) analytical instruments 104 output the actual analytical measurement results within a few seconds of measuring the sample, the observation database 132 may be updated with the new measurement results almost immediately when sampling. However, in certain other embodiments, the actual analytical measurements may be the result of several minutes, hours, or even days of processing by one or more of the analytical instruments 104, in which case the observation database 132 is not updated until such processing has been completed. In still other embodiments, new observation data sets may be added to the observation database 132 in an incremental manner as different analytical instruments in the analytical instruments 104 complete their respective measurements.

[0044] Thus, the observation database 136 may provide a “dynamic library” of past observations that may be used by the local model generator 142 for model training. In some embodiments, the most recent analytical measurement(s) are always added to the observation database 136, and the local model generator 142 may always use the most recent observation data set(s) in the observation database 136 when calibrating the local model 132. This may allow the local model 132 to encode process information from the recent past and quickly adapt to new conditions, or quickly adapt to new process conditions for which there is no history.

[0045] In some embodiments, only a subset of the scans in the observation database 136 have corresponding actual analytical measurements, in which case all or most of the dataset can be used to train the VAE, while only a dataset selected from those datasets with corresponding actual analytical measurements is used to train the local model 132 (wherein these analytical measurements are used as labels when training the local model 132).

[0046] Throughout the life cycle of a biopharmaceutical process in a bioreactor, some or all of the above processes may be repeated multiple times to continuously monitor the process using a local model whose calibration and maintenance are fully automatic and in real time. Depending on the embodiment and / or scenario, the analytical measurement(s) may be predicted for various purposes. For example, as part of a quality control process, certain parameters may be monitored (i.e., predicted) to ensure that the process is still within the relevant specifications. As another example, one or more parameters may be monitored / predicted to provide feedback in a closed-loop control system. For example, Figure 2 A system 150 is depicted that is similar to system 100, but that attempts to control glucose concentration in a biopharmaceutical process (i.e., attempts to match a predicted glucose concentration to a desired set point within some acceptable tolerance). It should be understood that in other embodiments, system 150 may alternatively (or also) be used to control a process parameter other than glucose level, or to control glucose level based on a prediction of one or more other process parameters (e.g., lactate level). Figure 2 The same reference numerals are used to indicate Figure 1 For example, Figure 2 The VAE-JITL predictor application 130 can be used with Figure 1 The VAE-JITL predictor application 130 is the same as the VAE-JITL predictor application 130 (wherein, for clarity, the individual elements of the VAE-JITL predictor application 130 are not shown in FIG. Figure 2 ).

[0047] like Figure 2 As can be seen in the figure, within the system 150, the memory 128 also stores a control unit 152. The control unit 152 is configured to control the glucose pump 154, that is, to cause the glucose pump 154 ​​to selectively introduce additional glucose into the biopharmaceutical process within the bioreactor 102. The control unit 152 may include, for example, software instructions executed by the processing unit 120, and / or appropriate firmware and / or hardware. In some embodiments, the control unit 152 uses the glucose concentration as an input in a closed-loop architecture to implement a model predictive control (MPC) technique. In embodiments where the local model 132 provides a confidence boundary or other confidence indicator for each prediction (for example, in certain embodiments where the local model 132 is a Gaussian process model), the control unit 152 may also accept a confidence indicator as input. For example, the control unit 152 may only generate control instructions for the glucose pump 154 ​​based on the glucose concentration prediction having a sufficiently high confidence indicator (e.g., only based on predictions associated with confidence boundaries that do not exceed a certain percentage or absolute measurement result range, or only based on predictions associated with confidence scores that exceed a certain minimum threshold score, etc.), or may increase and / or decrease the weight of a given prediction based on the confidence indicator of the glucose concentration prediction, etc.

[0048] Now go to Figure 3 , shows an example data flow 250 that may occur when using JITL techniques (such as the VAE-JITL techniques described herein) to analyze a biopharmaceutical process. For example, data flow 250 may occur in Figure 1 system 100, or Figure 2 The process of performing the above-mentioned process occurs within the system 150 of FIG. 1 . In the data stream 250 , spectral data 252 is provided by a spectrometer / probe. For example, the spectral data 252 may include a Raman scan vector generated by the Raman analyzer 106 , or a NIR scan vector, etc. A query sample 254 is generated (e.g., by the query unit 140 ) based on the spectral data 252 and is used to query a global dataset 256 , which may include, for example, all observation datasets in the observation database 136 . The query unit 140 may generate the query sample 254 by preprocessing the spectral data 252 or by directly using the spectral data 252 as the query sample 254 . Based on the query, a local dataset 258 is identified within the global dataset 256 . As discussed above and discussed in further detail below, in VAE-JITL, the local dataset 258 is selected using an encoder of the VAE. The local dataset 258 is then used (e.g., by the local model generator 142 ) as training data to calibrate the 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 outputs (analytical measurements) 262, such as medium component (e.g., glucose or other metabolites, or amino acids, etc.) concentration, viability, viable cell density, titer, etc., and may also output a confidence boundary or another suitable confidence indicator. For example, after executing the query, a process for building / calibrating the local model 260 (e.g., the local model 132) can be performed, as mathematically described in U.S. Patent Publication No. 2022 / 0128474.

[0049] An autoencoder is a neural network trained to reconstruct the input through an encoder layer and a decoder layer. The encoder layer creates a latent space that represents the main structured part 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 layer and the output layer will lead to severe overfitting of the dataset, thereby preventing the autoencoder from generating new content. To solve this problem, the query unit 140 uses a VAE with a regularized latent space, in which a distribution is sampled instead of a fixed point.

[0050] Figure 4An example VAE 400 is depicted. The VAE 400 includes an encoder 402 that includes hidden layers 406 that operate on inputs (scan values) applied at an input layer 404 and transform those inputs to a latent space 410. A decoder 420 of the VAE 400 attempts to reconstruct inputs at an output layer 422 by processing samples 424 in the latent space 410 using hidden layers 428. The VAE-JITL predictor application 130 or another application and / or computing system can construct the VAE 400, after which the query unit 140 can use the encoder 402 (without the decoder 420) to select similar historical samples / scans as described herein. Figure 4 As can be seen in , the encoder 402 generates / outputs parameters 426 that define the scanned latent space distribution (e.g., in this example, the mean and variance values ​​of the latent space distribution) applied as input at the input layer 404. The latent space distribution is Pr[Z|X], where Z is a sample representing approximate information of the latent space 410. The desired goal of the VAE is to make this distribution as similar as possible to a unique distribution, such as a 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 KL (Kullback-Leibler) divergence, for example as described in the following literature, Guo et al., Chemometrics and Intelligent Laboratory Systems, 197, A Deep Learning Just-In-Time-Learning Modeling Approach for Soft Sensor Based on Variational Autoencoder [Soft Sensor Deep Learning Just-In-Time Learning Modeling Method Based on Variational Autoencoder] (2020). Therefore, the VAE loss function consists of a reconstruction loss term and a regularization term together, as shown below:

[0051] Loss = ‖XD(Z)‖ 2 +KL[N(μ,Σ),N,(0,I)] (Equation 1)

[0052] By minimizing Equation 1 through VAE training based on historical Raman scans, the weights of the VAE 400 network can be optimized. Thereafter, 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 scans are most similar to the Raman scan of interest). As pointed out above, although Raman spectra provide signals with many features (each feature corresponding to a different Raman shift), the Euclidean distance or point-to-point distance of these features may not be the best criterion for finding the most similar samples. However, integrating JITL with the VAE encoder 402 can allow for a better combination of features to be selected from the dataset samples for model development. To find the sample most similar to the query sample, the encoder 402 maps each input to the latent space where l < d. Thereafter, by generating a multivariate normal distribution Z ∼ N(μ, Σ) for each Raman scan in the latent space, the most similar scan will have the minimum multivariate KL divergence (MKL) from the query sample. Thus, the distance function to be minimized is:

[0053] Dist(x i , x q ) = MKL[N(μ i , Σ i ), N(μ q , Σ q k )] (Equation 2)

[0054] In some embodiments, the VAE 400 can have an architecture different from that shown in Figure 4 . For example, the encoder 402 and the decoder 420 can each include only one hidden layer. In other embodiments, neither the encoder 402 nor the decoder 420 includes a hidden layer, or each includes more than one hidden layer. As other examples, any of the shown layers can include more nodes than the nodes shown in Figure 2 , the encoder 402 can produce the parameters 426 of more than two distributions (e.g., 3, 4, 5, 10, etc. distributions) shown in Figure 4 , and so on.

[0055] Once the Raman scans close to the query sample are extracted and collected into a local set (i.e., where K is the number of samples required for model development), the local model generator 142 uses the scan x k and the corresponding target variable y kTo construct a 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, the prediction unit 144 uses the local model 132 to predict the analytical measurement results of the query sample.

[0056] Figure 5 Describes the Figure 1 or Figure 2 Example VAE-JITL process 500 implemented by the computing system 110 of FIG. For ease of explanation, reference will be made to Figure 1 , Figure 2 and Figure 4 The VAE-JITL process 500 is explained with reference to the elements of FIG. 501 . For example, the process 500 may be implemented by the VAE-JITL predictor application 130 of the computing system 110 .

[0057] In the example process 500, when a new Raman scan vector (query scan 502) is captured (e.g., during run-time operation of the bioreactor 102), the query unit 140 (or another unit or application) processes each of the plurality of Raman scan vectors (historical scans 504) in the observation database 136 and the query scan using the same three pre-processing steps: 1) down-sampling the scan (stage 506); 2) baseline correcting the down-sampled scan (stage 508); and 3) normalizing the down-sampled and baseline corrected scan (stage 510). In other embodiments, one or more of stages 506, 508, and 510 are omitted, additional stages are included, and / or stages 506, 508, and / or 510 are normalized to the same level as the baseline corrected scan. Figure 5 The order shown occurs in a different order.

[0058] Fig. 6A and Figure 6BAn example of baseline correction (such as that which may occur at stage 508) is shown in . Raman spectroscopy provides detailed spectral fingerprint information about molecules in a bioreactor (such as bioreactor 102). However, uncertainties in the scanning process often result in a blurred baseline, which affects the accuracy of quantitative analysis. In order to extract more refined information from the raw Raman spectrum, baseline correction can be an important step. However, due to overlapping peaks and strong nonlinearities in Raman data, it is not easy to capture the baseline using polynomial fitting. Recently, a combination of penalized spline smoothing algorithms and vector transformation strategies has been proposed to retain background signals. The algorithm fits a function as a combination of B-splines to Raman data by minimizing a regularized least squares objective function. See Yao-yi et al., "Analytical Methods", 10, 3525-3533, Baseline Correctionfor Raman Spectra Using Penalized Spline Smoothing Based on VectorTransformation [Baseline Correction for Raman Spectra Using Penalized Spline Smoothing Based on VectorTransformation] (2018).

[0059] Fig. 6A A downsampled Raman scan 600 (e.g., downsampled to 1 / 4 or 1 / 5) is shown, where the x-axis represents the number of features (i.e., the Raman shift after downsampling) and the y-axis represents the intensity. The downsampling at stage 506 may be performed after the query unit 140 truncates the Raman scan to a range that removes noise and / or other undesirable phenomena (e.g., for 3344 features / Raman shifts in the original scan, only the range of 500 to 3000 or 600 to 1800 is selected, etc.). The 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, the query unit 140 generates Figure 6B A baseline-corrected Raman scan 610 is shown in FIG.

[0060] Now return to reference Figure 5 For example, query unit 140 may apply the normalization at stage 510 using min-max normalization. The min-max normalization may be based on the minimum and maximum values ​​for a particular experiment or for 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 Figure 51. For example, as noted above, the Raman scan may be truncated before (or after) downsampling. As another example, the query unit 140 does not use certain historical data sets from the observation database 136. This may be desirable, for example, if certain experiments used specific filters that would significantly alter the Raman scan signal.

[0061] At stage 512, query unit 140 applies the pre-processed data (each of the historical scans 504 and the query scan 502) as input to encoder 402 of VAE 400 in order to extract the main features as represented by distribution parameters 426. At stage 514, query unit 140 uses encoder 402 output (i.e., distribution parameters 426) to find the scan in historical scans 504 (also after processing at stages 506, 508, 510, and 512) that is most similar to query scan 502 (after processing at stages 506, 508, 510, and 512). In particular, query unit 140 determines similarity based on distribution parameters 426 of each of the historical scans 504 and distribution parameters 426 of query scan 502. In some embodiments, query unit 140 uses multivariate KL divergence (e.g., as in Equation 2 above) to achieve this.

[0062] At stage 516, the local model generator 142 generates / calibrates the local model 132 (e.g., a Gaussian process model) based on the K most similar samples, where K is any suitable positive integer. At stage 518, the prediction unit 144 uses the local model 132 and the query scan 502 to predict the actual measurement results (e.g., specific metabolite concentration or VCD or titer, etc.). That is, the prediction unit 144 applies the query scan 502 (after preprocessing at stages 506, 508, and 510) as input to the trained / calibrated local model 132.

[0063] In some embodiments, Figure 5All of the stages shown in the process occur during run-time operation of the bioreactor 102, e.g., both the query scan 502 and the historical scan 504 are processed (stages 506, 508, 510, 512) after a new query scan 502 is acquired by the Raman analyzer 106 and sent to the computing system 110. In other embodiments, the historical scan 504 is processed at stages 506, 508, 510, and 512 in an offline manner (e.g., prior to run-time operation of the bioreactor 102), while all other operations (i.e., the processing of the query scan 502 at stages 506, 508, 510, 512 and stages 514, 516, 518) are performed after a new query scan 502 is acquired by the Raman analyzer 106 and sent to the computing system 110.

[0064] 7A to 7J Examples of VAE-JITL prediction performance (as obtained using an embodiment of the present invention disclosed herein) relative to Linear JITL (“Linear-JITL”) prediction performance and measurements are plotted for various types of analytical measurements. In order to compare the performance of Linear-JITL and (an embodiment of) VAE-JITL, a validity study was performed using the same historical data and the same local model. Both VAE-JITL and Linear-JITL use the same JITL framework, except for the techniques used to identify similar scans / samples. Linear-JITL predictions use a K-nearest neighbor approach with a Euclidean distance metric to find the historical scans that are 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 7A to 7J In each of the graphs, the y-axis of the graph is normalized relative to these predicted maximum values ​​and the x-axis represents the number of Raman scans in the sequence in which the scans were acquired.

[0065] The performance of both VAE-JITL and Linear-JITL was determined offline by calculating the root mean square error (RMSE) and 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

[0066]

[0067] And MAPE is calculated as

[0068]

[0069] For VAE training and Gaussian process local model development, Raman features between specific (and the same) ranges were used and the data was downsampled to 1 / 5. In addition, the VAE was selected as a two-layer neural network with 260 units / nodes in the input layer and 128 units / nodes in the (single) hidden layer. The activation function in the hidden layer was Relu, and the VAE was trained for 20 rounds. In other embodiments or applications, the VAE can be trained for a different number of rounds and / or a different activation function can be used. The number of recent samples (i.e., K) used for Gaussian process local model development is 100.

[0070] Fig. 7A The prediction performance of VAE-JITL relative to Linear-JITL for the key variable of viable cell density (VCD) is demonstrated. Fig. 7A In Figure 1, it can be seen that both algorithms have similar performance at the beginning of the batch. However, VAE-JITL predictions follow the actual analytical measurements more closely, while Linear-JITL predictions deviate from the offline measurements in the middle of the batch (between scans 350 and 450). At scan number 470, both models fail to match the offline measurements, and at scan number around 500, Linear-JITL predictions are better than VAE-JITL.

[0071] Figure 7C and Figure 7K The prediction performance for the other key variables of glucose concentration (GLC) and titer were compared separately. In each case, similar performance improvements of the VAE-JITL predictions were observed relative to the Linear-JITL predictions. In particular, Figure 7C It is clearly shown that the predictions of VAE-JITL are closer to the actual analytical measurements, while the predictions of Linear-JITL have an undesirable deviation from the measurements. Figure 7K In the , the performance is similar, but both VAE-JITL and Linear-JITL fail to “catch up” with the actual measurements near the end of the batch. The reason for this is that the titer values ​​in the historical batches are rarely above 1, so neither technique can extrapolate the predictions.

[0072] Fig.7D , Fig.7I and Figure 7J Represent the predicted trajectories of lactate concentration (LAC), potassium concentration (K), and osmotic pressure (OSMO), respectively. Fig.7D and Fig.7I It is confirmed that VAE-JITL is better than Linear-JITL in predicting LAC and K. However, Figure 7J This indicates that VAE-JITL does not necessarily give better predictions than Linear-JITL for all metabolites. Figure 7J It is also shown that both models provide good predictions of OSMO trajectories, but both suffer from deviations relative to offline measurements.

[0073] Figure 7B , Fig. 7E , Figure 7F , Figure 7G and Figure 7H The prediction performances for viability-to-life ratio (VIAB), glutamine concentration (GLN), glutamate concentration (GLU), ammonium concentration (NH4) and sodium concentration (Na) were compared, respectively.

[0074] To quantitatively compare the JITL algorithms, the RMSE and MAPE of the predictions of some biological process variables relative to their corresponding measurements are shown in Table 1 below. These values ​​confirm that VAE-JITL generally provides better performance than Linear-JITL. However, the RMSE and MAPE of OSMO suggest that VAE-JITL may not outperform Linear-JITL in all cases.

[0075]

[0076]

[0077] Table 1

[0078] Figure 8 is a flow chart of an example method 800 for analyzing a biopharmaceutical process (e.g., for monitoring and / or control purposes) using VAE-JITL. For example, the method 800 may be performed by, for example, Figure 1 or Figure 2 The method may be implemented by a computer such as a computing system 110 (e.g., by a processing unit 120 executing instructions of a VAE-JITL predictor application 130).

[0079] At block 802, an observation database (e.g., observation database 136) including a plurality of observation data sets associated with past scans of a biopharmaceutical process is queried based on a first spectral scan vector of the biopharmaceutical process obtained by a spectral system (e.g., Raman analyzer 106 and Raman probe 108). Each of the observation data sets includes spectral data (e.g., Raman scan vector) and corresponding actual analytical measurements (e.g., any one of the variables shown in Table 1). The query at block 802 includes determining, using an encoder of a VAE (e.g., encoder 402 of VAE 400), first parameters of a distribution set defining the first spectral scan vector (e.g., mean and variance of a normal or approximately normal distribution set). The query also includes selecting a particular observation data set from the plurality of observation data sets as training data based on these first parameters and further based on other parameters of corresponding distribution sets defining the plurality of observation data sets (e.g., mean and variance of a normal (or approximately normal) distribution set, where each distribution set corresponds to a different observation data set).

[0080] Selecting the particular observation data sets as training data at block 802 may include calculating a multivariate KL divergence measure (e.g., as in Equation 2) based on the first parameters and the other parameters. In some embodiments, block 802 also includes preprocessing the first spectral scan vector (e.g., as in Equation 2) before determining the first parameters. Figure 5 ), in which case the spectral data of the plurality of observation data sets are pre-processed in a similar manner (ie, before generating a distribution set for the plurality of observation data sets).

[0081] At block 804 , a local model specific to a biopharmaceutical process (eg, the local model 132 ) is calibrated using the selected training data.

[0082] At block 806, the local model is used to predict analytical measurements of the biopharmaceutical process. For example, the analytical measurement can be any variable in Table 1, such as metabolite concentration, or viability, VCD, osmotic pressure, or titer, or other suitable types of analytical measurements. The analytical measurement is the same type of measurement used as a label when training the local model at block 804.

[0083] In some embodiments, method 800 includes Figure 8800. For example, method 800 may include an additional block in which at least one parameter of the biopharmaceutical process is controlled based, at least in part, on the analytical measurement predicted at block 806. Depending on the embodiment, the parameter may be of the same type as the predicted analytical measurement (e.g., controlling glucose concentration based on the predicted glucose concentration), or may be of a different type. For example, a model predictive control (MPC) technique may be used to control the parameter(s).

[0084] As another example, the method 800 may include a first additional block in which a user interface (e.g., presented on the display 124) is caused to display the predicted analysis measurement results. As yet another example, the method 800 may include one or more additional groups of blocks, each similar to blocks 802 through 806. In each of these additional groups of blocks, the local model may be calibrated by querying the observation database (or another observation database) and used to predict different types of analysis measurement results.

[0085] Additional considerations related to the present disclosure will now be addressed.

[0086] Some figures described herein show example block diagrams with one or more functional components. It will be understood that such block diagrams are for illustrative purposes, and the described and illustrated devices may have additional, fewer, or alternative components than those shown. Additionally, in various embodiments, components (and the functions provided by the corresponding components) may be associated with any suitable components or otherwise integrated as a part thereof.

[0087] Embodiments of the present disclosure relate to non-transitory computer-readable storage media having computer codes for performing various computer-implemented operations on the non-transitory computer-readable storage media. The term "computer-readable storage medium" is used herein to include any medium capable of storing or encoding a series of instructions or computer codes for performing the operations, methods, and techniques described herein. The medium and computer code may be a medium and computer code specifically designed and constructed for the purposes of the embodiments of the present disclosure, or the medium and computer code may be of a type known and available to a person skilled in the art of computer software. Examples of computer-readable storage media include, but are not limited to: magnetic media, such as hard disks, floppy disks, and tapes; optical media, such as CD-ROMs and holographic devices; magneto-optical media, such as optical disks; and hardware devices specifically configured to store and execute program codes, such as ASICs, programmable logic devices ("PLDs"), and ROM and RAM devices.

[0088] The example of computer code includes machine code such as produced by a compiler, and files containing higher-level codes executed by a computer using an interpreter or a compiler. For example, Python, Java, C++ or other object-oriented programming languages ​​and development tools can be used to implement the embodiments of the present disclosure. Additional examples of computer code include encrypted code and compressed code. In addition, the embodiments of the present disclosure can be downloaded as a computer program product, which can be transmitted to a requesting computer (for example, a client computer or different server computers) from a remote computer (for example, a server computer) via a transmission channel. Another embodiment of the present disclosure can be implemented by replacing machine executable software instructions with hard-wired circuitry or in combination with machine executable software instructions.

[0089] As used herein, the singular terms "a," "an," and "the" may include plural referents unless the context clearly indicates otherwise.

[0090] As used herein, the terms "connect," "connected," and "connector" refer to an operational coupling or link. Connected components may be coupled to each other directly or indirectly, such as through another set of components.

[0091] As used herein, the terms "approximately", "substantially", "substantially" and "about" are used to describe and explain small changes. When used in conjunction with an event or situation, these terms can refer to situations where the event or situation just occurs and situations where the event or situation approximately occurs. For example, when used in conjunction with a numerical value, these terms can refer to a range of variation where the numerical value is less than or equal to ±10%, such as less than or equal to ±5%, less than or equal to ±4%, less than or equal to ±3%, less than or equal to ±2%, less than or equal to ±1%, less than or equal to ±0.5%, less than or equal to ±0.1%, or less than or equal to ±0.05%. For example, if the difference between the values ​​is less than or equal to ±10% of the average value of the values, such as less than or equal to ±5%, less than or equal to ±4%, less than or equal to ±3%, less than or equal to ±2%, less than or equal to ±1%, less than or equal to ±0.5%, less than or equal to ±0.1%, or less than or equal to ±0.05%, then two numerical values ​​can be considered to be "substantially" the same.

[0092] In addition, quantities, ratios, and other numerical values ​​are sometimes presented in a range format herein. It should be understood that this range format is used for convenience and brevity, and should be flexibly interpreted to include the values ​​explicitly specified as the limits of the range, but also include all individual values ​​and subranges contained within the range, just as if each value or subrange was explicitly specified.

[0093] Although the disclosure has been described and shown with reference to the specific embodiments of the disclosure, these descriptions and illustrations do not limit the disclosure. It should be understood by those skilled in the art that various changes can be made and equivalents can be substituted without departing from the true spirit and scope of the disclosure defined by the appended claims. These illustrations may not necessarily be drawn to scale. Due to manufacturing processes, tolerances and / or other reasons, there may be differences between the artistic reproduction in the disclosure and the actual device. There may be other embodiments of the disclosure that are not specifically shown. The specification (except for the claims) and the drawings should be regarded as illustrative rather than restrictive. Modifications may be made to adapt specific situations, materials, material compositions, technologies, or processes to the purpose, spirit and scope of the disclosure. All these modifications are intended to fall within the scope of the appended claims. Although the technology disclosed herein has been described with reference to specific operations performed in a specific order, it should be understood that these operations can be combined, subdivided, or reordered to form equivalent technologies without departing from the teachings of the disclosure. Therefore, unless there is a specific indication herein, the order and grouping of operations are not limitations on the disclosure.

Claims

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

2. The computer-implemented method of claim 1, wherein: Determining the first parameters includes processing the query sample using an encoder of a variational autoencoder, and wherein the encoder outputs the first parameters.

3. The computer-implemented method of claim 2, wherein: The encoder consists of only one hidden layer.

4. The computer-implemented method of claim 2 or 3, further comprising: The other parameters are determined by the one or more processors using an encoder of the variational autoencoder, wherein the encoder outputs the other parameters.

5. The computer-implemented method of any one of claims 1 to 4, wherein: Selecting the particular observation data sets includes computing a multivariate KL divergence metric based on the first parameters and the other parameters.

6. The computer-implemented method of any one of claims 1 to 5, wherein: Calibration of this local model specific to this biopharmaceutical process includes: Calibrate a Gaussian process machine learning model specific to this biopharmaceutical process.

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

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

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

10. The computer-implemented method of any one of claims 1 to 9, wherein: Analyzing the spectral data using the local model includes 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 predictive analytical measurement outcome for this biopharmaceutical process is metabolite concentration.

12. The computer-implemented method of any one of claims 1 to 10, wherein: The predictive analytical outcomes measured in this biopharmaceutical process are 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. The computer-implemented method of any one of claims 1 to 13, further comprising: At least one parameter of the biopharmaceutical process is controlled by the one or more processors based on the predictive analysis measurements of the biopharmaceutical process.

15. The computer-implemented method of any one of claims 1 to 14, further comprising: The one or more processors cause a user interface to display the predicted analytical measurement results.

16. 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 delivering the source electromagnetic radiation to the biopharmaceutical process; and A computing system configured to perform the method of any one of claims 1 to 15.

17. A non-transitory computer readable medium storing instructions which, 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