Deep Learning-Based Prediction for Drug Monitoring Using Spectroscopy

Abstract

A method (500) for monitoring and / or controlling a pharmaceutical process includes acquiring (502) one-dimensional spectral data generated by a spectroscopy system (e.g., a Raman spectroscopy system), converting (508) the one-dimensional spectral data to a two-dimensional spectral data matrix, and applying the two-dimensional spectral data matrix to an input layer of a deep learning model (e.g., a convolutional neural network). The deep learning model predicts (410) a parameter (e.g., a metabolite level) based on the two-dimensional spectral data matrix, for example, to monitor and / or control the pharmaceutical process.

Description

[Technical field]

[0001] The present invention relates generally to monitoring and / or control of pharmaceutical (e.g., biopharmaceutical) processes using spectroscopic techniques (e.g., Raman spectroscopy), and more particularly to the use of deep learning in conjunction with such spectroscopic techniques. [Background technology]

[0002] The stable production of biotherapeutic proteins through biopharmaceutical processes generally requires bioreactors to maintain balanced and consistent parameters (e.g., cellular metabolite concentrations), which in turn requires rigorous process monitoring and control. To meet these demands, process analytical engineering (PAT) tools are increasingly being adopted. Online monitoring of pH, dissolved oxygen, and cell culture temperature are some examples of traditional PAT tools that have been 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, amino acids, titers, and critical quality attributes.

[0003] In biopharmaceutical and other (e.g., small molecule) fields, advanced process control techniques typically rely on real-time, frequent measurements from the process being controlled. However, such measurements may not be available or may be cumbersome. For example, in the biopharmaceutical industry, real-time measurements are often not available, and instead, scientists rely on offline samples (e.g., taken once a day) to monitor biological processes. Increasing the number of offline samples to obtain a more holistic view of the process may not be feasible due to workload constraints or resource limitations, for example, on the size of the bioreactor.

[0004] To enable real-time trending of bioprocess cultures, tools such as Raman spectroscopy are often used. In this setup, an in-situ Raman probe is inserted into the bioreactor to collect Raman spectra. Raman spectroscopy is a common PAT tool widely used for online monitoring in biopharmaceutical manufacturing. It is an optical method that allows non-destructive 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 termed the "Raman shift," and the vector of the Raman shift (usually expressed in terms of wavenumbers) versus 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 surge 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 that is now used both inside and outside the laboratory. Since the application of in-situ Raman measurements in biopharmaceutical manufacturing was first reported, it has been employed to provide online, real-time predictions of several important process states, such as glucose, lactate, glutamate, glutamine, ammonia, and VCD. These predictions are usually based on calibration models or soft-sensor models that are constructed in an offline setting based on 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 usually require pre-processing filtering of the Raman scans before calibrating against the analytical measurements. Once the calibration models are trained, the models are implemented in a real-time setting to provide in-situ measurements for process monitoring and / or control.

[0005] Raman model calibration for biopharmaceutical applications is not straightforward, as biopharmaceutical processes typically operate under strict constraints and regulations. The current state-of-the-art approach for Raman model calibration in the biopharmaceutical industry is to first perform multiple campaign tests to generate relevant data that are used to correlate Raman spectra to analytical measurements. These tests are both costly and time-consuming, as each campaign may span, for example, 2-4 weeks in a laboratory environment. Furthermore, only limited samples may be available to the analytical instruments (e.g., to ensure that lab-scale bioreactors maintain a healthy amount of viable cells). In fact, it is not uncommon to only have one or two measurements available each day from in-line or offline analytical instruments. Further exacerbating the situation is that current best practices result in calibration models that are tied to a specific process, a specific formulation or profile of the bioreactor media, and specific operating conditions. Thus, if any of the aforementioned variables change, the model may need to be recalibrated based on new data. In fact, both Raman model calibration and model maintenance require significant resource allocation and are typically performed in an offline setting. Approaches have been proposed to adapt the model to new operating conditions (e.g., recursive, moving window, and time-lag methods), but these methods may not be able to adequately handle rapid process changes.

[0006] There are many publications that describe general Raman models based on traditional chemometric methods (e.g., PLS modeling) for multiple molecules. However, these general models assume that the processes use similar, if not identical, media formulations and / or running process conditions. The media and processes are usually platformed with little or no variation. The drawback of this kind of general model is that the general model loses precision and accuracy once the process deviates from the standard or the training data set includes too broad a process range to account for the variation between different molecules (e.g., media additives, process duration, and / or other process changes). Thus, these "general" models are only general within the strict boundaries described. See Mehdizaheh et al., Biotechnol.Prog.31(4):1004-1013,2015; Webster et al., Biotechnol.Prog.34(3):730-737,2018.

[0007] More recently, systems have been described that employ automatic calibration and maintenance of Raman spectroscopy models using just-in-time learning (JITL) for real-time predictions. See WO 2020 / 086635. However, when used in isolation, JITL typically requires ongoing (but infrequent) analytical measurements for recalibration, which may not be feasible (e.g., in small bioreactors), consumes time and other resources, and may provide different results when measurements are rerun. On the other hand, when recalibration is not performed (e.g., when "offline" JITL is used), results can vary widely depending on the modality and the amount and type of historical data available. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] International Publication No. 2020 / 086635 Brochure [Non-patent literature]

[0009] [Non-Patent Document 1] Mehdizaheh et al.,Biotechnol.Prog.31(4):1004-1013(2015) [Non-Patent Document 2] Powell et al.,Prog.34(3):730-737;2018) Summary of the Invention [Means for solving the problem]

[0010] The term "pharmaceutical process" refers to a process used in biopharmaceutical manufacturing, such as a cell culture process or a small molecule manufacturing process that produces a desired recombinant protein. For biopharmaceuticals, cell culture is carried out in a cell culture vessel, such as a bioreactor, under conditions that support the growth and maintenance of an organism engineered to express the protein. During recombinant protein production, process parameters such as media component concentrations, including nutrients and metabolites (e.g., glucose, lactate, glutamate, glutamine, ammonia, amino acids, Na+, K+ and other nutrients or metabolites), media conditions (pH, pCO2, pO2, temperature, osmolality, etc.), and cell and / or protein parameters (e.g., viable cell density (VCD), titer, cell condition, critical quality attributes, etc.) are monitored to control and / or maintain the cell culture process.

[0011] To address some of the above-mentioned limitations of current best industrial practice, embodiments described herein relate to systems and methods that improve upon conventional techniques for spectroscopic analysis of pharmaceutical processes, such as Raman spectroscopy. In particular, deep learning models, such as convolutional neural networks (CNNs), are used as an alternative modeling method to predict process-related parameters, such as metabolite concentrations. It is understood that the term "predict" (or "predicting", "prediction", etc.) is used broadly herein to refer to prediction and / or inference. CNNs are feed-forward neural networks specialized for image processing, for example, to perform object detection and classification. However, Raman and other (e.g., NIR, HPLC, etc.) spectroscopic measurements are not images and therefore are not natural candidates for CNN processing. Nevertheless, the systems and methods described herein generate "pseudo-images" from spectroscopic scans and process those pseudo-images using one or more CNNs (e.g., one CNN per metabolite or other process parameter of interest). Deep CNNs and Raman spectroscopic measurements can be used to create offline models, which can be non-product dependent, to predict one or more parameters or characteristics (e.g., product quality attributes) of the pharmaceutical process. This allows the model to be used in different processes without the need for recalibration or retraining. Another advantage of CNN is the weight sharing feature, which can significantly reduce the number of parameters compared to traditional deep neural networks. Furthermore, this allows CNN models to be trained using smaller training datasets.

[0012] Deep CNNs are general offline models that can be used to predict metabolite concentrations using spectroscopic measurements from any process and can be fine-tuned for a specific process to optimize performance. The models do not require prior knowledge of the process and are therefore a truly comprehensive spectroscopic modeling solution for all processes. Deep learning CNN approaches overcome many of the issues associated with chemometrics, such as the need for frequent analytical measurements, the inability to do so in small bioreactors, the time delay between sampling and measurement acquisition, and the potential for lack of repeatability in measurement runs.

[0013] In contrast to JITL platforms, which typically maintain dynamic libraries that are updated whenever new analytical measurements become available, the CNN approach does not require updating the model every time a new analytical measurement is made. Instead, an input scan is fed into a previously generated / trained CNN model. With the CNN approach, the CNN model can optionally be updated after prediction or process control is performed.

[0014] In contrast to Gaussian process models, which generally do not require pre-processing filtering of the spectral data (e.g., Raman scans), CNN models use pre-processing of Raman scans.

[0015] The deep learning (e.g., CNN) techniques described herein can be used in conjunction with JITL / PLS or other techniques for process monitoring and control, or can be used independently of such techniques.

[0016] Those skilled in the art will appreciate that the figures described herein are included for illustrative purposes and are not intended to limit the disclosure. The figures are not necessarily to scale, with emphasis instead being placed on illustrating the principles of the 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 drawings, like reference numbers generally refer to functionally similar and / or structurally similar components throughout the various views. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a simplified block diagram of an example system that can be used for process monitoring. [Diagram 2] FIG. 1 is a simplified block diagram of an exemplary system that can be used for closed-loop control of glucose concentration. [Diagram 3] A typical convolutional neural network (CNN) is shown. [Figure 4] 1 illustrates an example data flow that may be performed in the system of FIG. 1 to enable and perform analysis of pharmaceutical processes using deep learning models. [Diagram 5] 2 shows an example of pre-processing of spectral data that can be performed in the system of FIG. 1. [Figure 6] Another example of data flow that may occur in the system of FIG. 1 when using deep learning models to analyze pharmaceutical processes is shown. [Figure 7] 1 is a flow diagram of an example method for using techniques of this disclosure in combination with just-in-time learning (JITL). [Figure 8] We present experimental results of VCD prediction using the deep learning and preprocessing techniques described herein. [Figure 9] 1 shows experimental results of survival prediction using the deep learning and preprocessing techniques described herein. [Figure 10] We present experimental results of TCD prediction using the deep learning and preprocessing techniques described herein. [Figure 11] 1 shows experimental results of glucose prediction using the deep learning and preprocessing techniques described herein. [Figure 12] 1 shows experimental results of lactate prediction using the deep learning and preprocessing techniques described herein. [Figure 13] 1 shows experimental results of osmolarity prediction using the deep learning and preprocessing techniques described herein. [Figure 14]1 shows experimental results of glutamate prediction using the deep learning and preprocessing techniques described herein. [Figure 15] 1 shows experimental results of glutamine prediction using the deep learning and preprocessing techniques described herein. [Figure 16] 1 shows experimental results of potassium prediction using the deep learning and preprocessing techniques described herein. [Figure 17] 13 shows experimental results of sodium prediction using the deep learning and preprocessing techniques described herein. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] The various concepts described introductory above and discussed in more detail below can be implemented in any of numerous ways, and the concepts described are not limited to any particular implementation manner. Example implementations are provided for illustrative purposes.

[0019] Figure 1 is a simplified block diagram of an example system 100 that may be used to predict a parameter or characteristic of a biopharmaceutical process. Although Figure 1 illustrates system 100 performing Raman spectroscopy techniques of a biopharmaceutical process, it is understood that in other embodiments, system 100 may perform other suitable spectroscopic techniques (e.g., near-infrared (NIR) spectroscopy, high performance liquid chromatography (HPLC), ultra-performance liquid chromatography (UPLC) spectroscopy, mass spectrometry, etc.) and / or may perform such techniques with respect to non-biopharmaceutical processes (e.g., small molecule pharmaceutical processes).

[0020] The system 100 includes a bioreactor 102, one or more analytical instruments 104, a Raman analyzer 106 having a Raman probe 108, a computer 110, and a training server 112 coupled to the computer 110 via a network 114. The bioreactor 102 may be any suitable vessel, device, or system that supports a biologically active environment that may include living organisms in the medium and / or substances derived therefrom (e.g., cell cultures). The bioreactor 102 may include recombinant proteins expressed by the cell cultures, such as 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., "broth") and particular nutrients, and may have target medium state parameters, such as a target pH level or range, a target temperature or temperature range, etc. The medium may also include organisms and substances derived from the organisms, such as metabolites and recombinant proteins. Collectively, the contents and parameters / characteristics of the medium are referred to herein as a "media profile."

[0021] 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 constituent concentrations, such as nutrient and / or metabolite levels (e.g., glucose, lactate, glutamate, glutamine, ammonia, amino acids, Na+, K+, etc.) and media condition parameters (pH, pCO2, pO2, temperature, osmolality, etc.). Additionally or alternatively, the analytical instrument 104 may measure osmolality, 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 taken, centrifuged, purified through one or more columns, and passed through a first of the analytical instruments 104 (e.g., an HPLC or UPLC instrument) and then through a second of the analytical instruments 104 (e.g., a mass spectrometer), where the first and second analytical instruments 104 provide analytical measurements. One, some, or all of the analytical instruments 104 may use destructive analytical techniques.

[0022] The Raman analyzer 106 may include a spectrograph device coupled to the Raman probe 108 (or, in some implementations, multiple Raman probes). 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, for example, a charge-coupled device (CCD) or other suitable camera / recording device that records signals received from the Raman probe 108 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).

[0023] Collectively, the Raman analyzer 106 and the Raman probe 108 form a Raman spectroscopy system configured to non-destructively scan the biologically active content during a biopharmaceutical process in the bioreactor 102 by exciting, observing, and recording a 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 one or more Raman scan vectors, each of which represents intensity as a function of Raman shift (a frequency-related parameter). The Raman scan vectors can be, for example, intensity values ​​as a function of wavenumber (e.g., cm -1 unit).

[0024] More generally, system 100 may include any spectroscopy system that generates 1D spectral data (e.g., a Raman spectroscopy system, a NIR spectroscopy system, an HPLC spectroscopy system, etc.). As used herein, "1D spectral data" refers to values ​​of spectral data (e.g., intensity values) that are not arranged in a matrix format having two or more dimensions. For example, 1D spectral data may be a string / sequence of tuples, each having the format [wavenumber, intensity value]. As another example, 1D spectral data may simply be a string / sequence of intensity values, so long as the order of the intensity values ​​in the string follows a known / common knowledge format (e.g., each position in the string corresponds to each wavenumber). In some embodiments, 1D spectral data may be represented as a function of a spectral parameter other than wavenumber (e.g., wavelength or frequency).

[0025] The computer 110 is coupled to the Raman analyzer 106 and the analytical instrument 104 and is generally configured to analyze the Raman scan vectors generated by the Raman analyzer 106 to predict one or more properties or parameters of the biopharmaceutical process. For example, the computer 110 may analyze the Raman scan vectors to predict the same types of properties or parameters measured by the analytical instrument 104. As a more detailed example, the computer 110 may predict a glucose concentration while the analytical instrument 104 actually measures the glucose concentration. However, the analytical instrument 104 may perform relatively infrequent "offline" analytical measurements of samples extracted from the bioreactor 102 (e.g., due to limited amounts of media from the biopharmaceutical process and / or due to higher costs of making such measurements, etc.), whereas the computer 110 may perform relatively frequent "online" predictions of properties or parameters in real time. The computer 110 may also be configured to transmit the analytical measurements made by the analytical instrument 104 to the training server 112 via the network 114, as described in more detail below.

[0026] In the exemplary embodiment shown in FIG. 1, computer 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 computer 110 as described herein. Alternatively, one or more 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.). Memory 128 may include one or more physical memory devices or units, including volatile and / or non-volatile memory. Any suitable type or types of memory may be used, such as read only memory (ROM), solid state drives (SSDs), hard disk drives (HDDs), etc.

[0027] Network interface 122 may include any suitable hardware (e.g., front-end transmitter and receiver hardware), firmware, and / or software configured to communicate over network 114 using one or more communications protocols. For example, network interface 122 may be or may include an Ethernet interface. Network 114 may be a single communications network or may include multiple communications networks of one or more types (e.g., one or more wired and / or wireless local area networks (LANs) and / or one or more wired and / or wireless wide area networks (WANs), such as the Internet).

[0028] The display 124 may use any suitable display technology (e.g., LED, OLED, LCD, etc.) to present information to a 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). In general, the display 124 and the user input device 126 may be combined to allow a user to interact with a graphical user interface (GUI) provided by the computer 110, for purposes such as, for example, to manually monitor various processes running within the system 100. However, in some embodiments, the computer 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 that is communicatively coupled to the computer 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 and data and possibly other data or data structures used and / or output by such applications. In the example of FIG. 1, the memory 128 stores at least a deep learning (DL) model 130, a predictive application 132, a data cleaning software 134, and a database maintenance unit 136. The predictive application 132, when executed by the processing unit 120, is generally configured to use the DL model 130 to predict parameters (e.g., parameters of the kind that can be measured by the analytical instrument 104) of the biopharmaceutical process in the bioreactor 102 by processing Raman scan vectors generated by the Raman analyzer 106. Depending on how often the Raman analyzer 106 generates such scan vectors, the predictive application 132 may predict the properties or parameters on a periodic or other suitable time basis. The Raman analyzer 106 may itself control when the scan vectors are generated, or the computer 110 may trigger the generation of the scan vectors, for example, by sending a command to the Raman analyzer 106. The predictive application 132 may use a single DL model 130 to predict only a single property or parameter (e.g., only glucose concentration) based on each scan vector, or may use multiple DL models to predict multiple properties or parameters (e.g., glucose concentration and viable cell density) based on each scan vector. Predictive application 132 and DL models 130 are described in more detail below.

[0030] The data cleaning software 134 typically removes noise and / or outliers from the scan vectors or otherwise optimizes the scan vectors generated by the Raman analyzer 106 prior to processing by the predictive application 132. The database maintenance unit 136 typically updates the training data in the training database 138 by sending new Raman scan vectors and corresponding analytical measurements performed by the analytical instrument 104 to the training server 112. However, in some embodiments, the data cleaning software 134 and / or the database maintenance unit 136 are not included in the system 100.

[0031] The training server 112 may be remote from the computer 110 (e.g., such that a local facility may include only the bioreactor 102, the analytical instrument 104, the Raman analyzer 106 with the Raman probe 108, and the computer 110) and may include or be communicatively coupled to a training database 138 that stores observational data sets related to past observations, as seen in FIG. 1. Each observational data set in the training database 138 may include spectral data (e.g., one or more Raman scan vectors of the type generated by the Raman analyzer 106 or other 1D spectral data generated by a different type of spectroscopic system) and one or more corresponding analytical measurements (e.g., one or more measurements of the type generated by the analytical instrument 104). In some embodiments and / or scenarios, the past observations may have been 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 training database 138 to represent a broadly diverse collection of processes, operating conditions, and media profiles. However, depending on the embodiment, training database 138 may or may not store information indicative of these processes, cell lines, proteins, metabolites, operating conditions, and / or media profiles. In some embodiments, training server 112 is remotely coupled to multiple other computers similar to computer 110 via network 114 and / or other networks. This may be desirable to collect a large number of observational data sets for storage in training database 138.

[0032] The training server 112 trains the DL model 130. That is, the training server 112 uses the Raman scan vectors (and possibly other feature data) associated with each observational data set as a feature set, and the analytical measurements related to the same observational data set as labels for that feature set. The training server 112 then provides the DL model 130 to the computer 110 via the network 114. In other embodiments, the server 112 does not provide the DL model 130 to the computer 110, but instead operates the DL model 130 (and possibly the entire predictive application 132) as a cloud-based service. For example, the server 112 may store both the predictive application 132 and the DL model 130 locally, or may store only the DL model 130 locally (in which case the predictive application 132 at the computer 110 utilizes the DL model 130 via the network 114 and any suitable application programming interface). In yet other embodiments, the system 100 does not include the training server 112, and the computer 110 directly accesses the training database 138. For example, the training database 138 may be stored in the memory 128 .

[0033] It is understood that other configurations and / or components may be used instead of those shown in Figure 1. For example, a different computer (not shown in Figure 1) may transmit measurements provided by the analytical instruments 104 to the training server 112, one or more additional computing devices or systems may act as intermediaries between the computer 110 and the training server 112, some or all of the functions of the computer 110 as described herein may instead be performed remotely by the training server 112 and / or another remote server, etc. For ease of explanation, the remaining description will assume that the training database 138 is coupled to the training server 112 as shown in Figure 1. However, one of ordinary skill in the art will readily understand how the communication paths would be different if the training database 138 was instead local to the computer 110 or in another suitable location within the system architecture.

[0034] After the DL model 130 has been trained (e.g., by the training server 112) and during run-time operation of the system 100, the Raman analyzer 106 and the Raman probe 108 scan (i.e., generate Raman scan vectors for) the still-life pharmaceutical process in the bioreactor 102, and the Raman analyzer 106 transmits the Raman scan vectors to the computer 110. The Raman analyzer 106 and the Raman probe 108 may provide scan vectors to support predictions (made by the prediction application 132) according to a pre-defined schedule of monitoring periods, such as once per minute or once per hour. Alternatively, predictions may be made at irregular intervals (e.g., in response to certain process-based triggers, such as changes in measured pH levels and / or temperature), such that each monitoring period has a variable or uncertain duration. In some embodiments, the Raman analyzer 106 may send only one scan vector per monitoring period to the computer 110, or may send multiple scan vectors per monitoring period to the computer 110, depending on how many scan vectors the DL model 130 accepts as input for a single prediction. Multiple scan vectors (e.g., when aggregated or averaged) may, for example, improve the prediction accuracy of the DL model 130.

[0035] In some embodiments, the DL model 130 is not retrained / recalibrated after the initial training, or the training server 112 does so only infrequently (e.g., compared to traditional techniques or JITL). However, in other embodiments, the prediction application 132, another application in memory 128, retrains / recalibrates the local DL model 130 more frequently using JITL techniques (e.g., any of the techniques described in WO 2020 / 086635, which is incorporated herein by reference).

[0036] After receiving the Raman scan vectors, the prediction application 132 pre-processes the scan vectors (as discussed further below) to generate a pseudo-image and applies the pseudo-image as an input to the DL model 130. The DL model 130 then generates predictions based on the pseudo-image. In some embodiments, the DL model 130 also accepts other information (e.g., operating conditions, media profiles, process data, cell line information, protein information, metabolite information, etc.) as part of the input / feature set.

[0037] The database maintenance unit 136 may cause the analytical instrument 104 to periodically collect one or more actual analytical measurements at a frequency significantly less than the monitoring period of the Raman analyzer 106 (e.g., only once or twice a day, etc.). The measurements by the analytical instrument 104 may be destructive in some embodiments and may require permanent removal of the sample from the process in the bioreactor 102. At or about the time the database maintenance unit 136 causes the analytical instrument 104 to collect and provide an actual analytical measurement, the database maintenance unit 136 may also cause the Raman analyzer 106 to provide one or more Raman scan vectors. The database maintenance unit 136 may then cause the network interface 122 to transmit the Raman scan vectors and the corresponding actual analytical measurements via the network 114 to the training server 112 for storage as a new observational data set in the training database 138. The training database 138 may be updated according to any suitable timing, which may vary depending on the embodiment. In cases where the analytical instruments 104 output the actual analytical measurements within seconds of measuring the sample, for example, the training database 138 may be updated with new measurements almost immediately after the sample is taken. However, in certain other embodiments, the actual analytical measurements may be the result of minutes, hours, or days of processing by one or more analytical instruments 104, in which case the training database 138 is not updated until such processing is completed. In still other embodiments, new observational data sets may be added incrementally to the training database 138 as different ones of the analytical instruments 104 complete their respective measurements. In any of these embodiments, the training database 138 may provide a "dynamic library" of previous observations that the training server 112 can utilize to tune or retrain the DL models 130. However, in other embodiments, the database maintenance unit 136 is omitted, the training database 138 is not updated, and / or the DL models 130 are not tuned or retrained.

[0038] The predictive application 132 may predict parameters for various purposes depending on the embodiment and / or situation. For example, certain parameters may be monitored (i.e., predicted) as part of a quality control process to ensure that the process remains in compliance with relevant regulations. As another example, one or more parameters may be monitored / predicted to provide feedback in a closed-loop control system. For example, FIG. 2 illustrates a system 200 similar to system 100, but for controlling glucose concentration in a biopharmaceutical process (i.e., adding glucose to a predicted glucose concentration to match a desired set point within some acceptable tolerance). It will be appreciated that in other embodiments, system 200 may instead (or in addition) be used to control process parameters other than glucose level, or to control glucose level based on a prediction of one or more other process parameters (e.g., lactate level, pH, etc.). In FIG. 2, the same reference numbers are used to indicate corresponding components of FIG. 1.

[0039] 2, in the system 200, the memory 128 also stores a control unit 202. The control unit 202 is configured to control the glucose pump 204, i.e., to cause the glucose pump 204 to selectively introduce additional glucose into the biopharmaceutical process in the bioreactor 102. The control unit 202 may comprise, for example, software instructions executed by the processing unit 120 and / or suitable firmware and / or hardware. In some embodiments, the control unit 202 implements a model predictive control (MPC) technique that uses glucose concentration as an input in a closed-loop architecture. In embodiments in which the DL model 130 provides confidence limits or other confidence indicators with each prediction, the control unit 202 may also accept the confidence indicators as inputs. For example, the control unit 202 may only generate control instructions for the glucose pump 204 based on glucose concentration predictions that have a sufficiently high confidence index (e.g., only based on predictions associated with confidence limits that do not exceed a certain percentage or absolute measurement range, or only based on predictions associated with confidence scores above a certain minimum threshold score), or may increase and / or decrease the weighting of certain predictions based on their confidence index.

[0040] As further described below, the prediction application 132 converts the 1D spectral data (e.g., Raman scan vectors) into an image-like format that is a 2D matrix of values ​​(also referred to herein as a "pseudo-image"). For example, if the 1D spectral data is a sequence of at least j×k values ​​(e.g., an array of intensity values ​​where each position corresponds to a different wavenumber or a sequence of [wavenumber, intensity value] tuples, etc.), the prediction application 132 may convert the sequence into a 2D spectral data matrix having j rows and k columns, where each position in the 2D spectral data matrix corresponds to a different wavenumber. In particular, the prediction application 132 may place the first N (>1) intensity values ​​(or [wavenumber, intensity value] tuples) of the sequence in the first row (or first column) of the matrix, the second N intensity values ​​(or [wavenumber, intensity value] tuples) of the sequence in the second row (or second column) of the matrix, and so on.

[0041] The DL model 130 of FIG. 1 or FIG. 2 may be any deep learning model configured to process image data and thus capable of processing such pseudo-images. In some embodiments, the DL model 130 is (or includes) a convolutional neural network (CNN), which is a feed-forward neural network specialized for image processing. An example CNN 300 that can be used as (or part of) the DL model 130 is shown in FIG. 3. The CNN 300 includes an input layer, several convolutional layers, several pooling layers, a flattening layer, several fully connected (dense) layers, and an output layer. The prediction application 132 applies the pseudo-image (a 2D spectral data matrix) to the input layer, which is a passive layer that passes the pseudo-image to the first convolutional layer. The convolutional layer applies multiple filters to the pseudo-image through a convolution operation to extract features from the pseudo-image. The convolution operation may be defined by the following equation: G[m,n]=(f*h)[m,n]=Σ j | a Σ k | b h[j,k]f[mj,nk] (Equation 1) In Equation 1, f is the input (pseudo-image), h is the filter or kernel, m and n are the row and column indices (respectively) of the result matrix, and a and b are stride parameters (which may be assumed to be 1 for CNN 300).

[0042] The output of the convolutional layer may be fed into an activation function. The CNN 300 may implement activation functions such as sigmoid, tangent hyperbolic (tanh), and / or linear functions, but may instead use the rectified linear unit (Relu) to avoid the vanishing gradient problem. The Relu activation function is g(x) = Relu(x) = max(0,x) (Eq. 2) It can be defined as: The tangent hyperbolic function is

number

[0043] The CNN 300 may include a pooling layer after each convolutional layer (or each of some of the convolutional layers). Each pooling layer applies a pooling operation to the output of the previous convolutional layer. The pooling operation may be a maximum, average, minimum, or other statistical measure of the feature maps. Pooling layers increase the computational efficiency of the CNN 300 by reducing the size of the convolutional output while generally retaining the most relevant information. In some embodiments, the CNN 300 includes a max pooling layer and an average pooling layer.

[0044] A flattening layer of the CNN 300 may follow the last convolutional layer or the last pooling layer (e.g., when the last convolutional layer is followed by a pooling layer). The flattening layer converts the output of the last convolutional layer or pooling layer into a vector, which is then fed to a fully connected layer, a linear layer, and / or a softmax layer. A fully connected layer of the CNN 300 may follow the convolutional layer and the pooling layer. The fully connected layer is similar to the inner layer of a shallow neural network and performs high-level inference from the output of the convolutional layer and the pooling layer. The output layer of the CNN 300 may perform image classification applications and thus may be a softmax layer that identifies the class of the input pseudo-image. The CNN 300 may include a fully connected layer with a linear activation function as the output layer to solve regression problems.

[0045] As described above, the CNN 300 may include multiple convolutional, pooling, and dense layers (e.g., the activation functions of the convolutional and dense layers may be linear, tangent hyperbolic, or rectified linear units, and average and max pooling layers may be used). Once the CNN 300 is developed, various techniques may be used to optimize the model. In one embodiment, a cost function is employed. The cost function may be selected from, for example, mean absolute percentage error and mean squared error. In one embodiment, an optimization algorithm is employed. The CNN 300 may be optimized using any suitable optimization algorithm to learn the relationship between Raman (or other) spectra and desired metabolite levels (or other predicted properties), such as stochastic gradient descent, root mean square propagation (RMSProp), Adamax, Adagrad, Adadelta, etc. Once optimized, the CNN 300 may be tested against different datasets to evaluate / validate the model performance. If the model performance is not satisfactory, the number of layers, the activation function, and / or the optimization algorithm can be modified to achieve better model performance.

[0046] In some embodiments, DL model 130 includes multiple deep learning models (e.g., multiple CNNs similar to CNN 300), each trained and / or optimized to predict a different type of parameter. For example, predictive application 132 may apply a given Raman scan vector to a first CNN to predict glucose concentration, a second CNN to predict lactate concentration, a third CNN to predict osmolality, etc. Various CNNs may be developed using different numbers of layers and / or nodes, different activation functions, different training and / or optimization algorithms, different loss functions, etc.

[0047] 4 is an example data flow 400 that may occur in the system 100 of FIG. 1 to enable and perform analysis of pharmaceutical processes using a deep learning model such as DL model 130 (e.g., CNN 300). In the data flow 400, a historical data set 402 may reside in the training database 138. The historical data set 402 includes spectral data (e.g., Raman scan vectors or other 1D spectral data) and corresponding labels generated by a suitable device / system (e.g., a spectroscopic system similar or different to Raman analyzer 106 and Raman probe 108). The labels may be actual measurements of parameters of interest (e.g., metabolite levels) made by an analytical instrument (similar to instrument 104) at or near the same time that the spectral data is generated.

[0048] A computing device or system, such as the training server 112, then trains 404 a deep learning model (e.g., CNN 300) using the spectral data of the historical dataset 402 as features / inputs and the corresponding analytical measurements as labels to generate a trained deep learning model 406. In run-time operation, the deep learning model 406 operates on the spectral data 408 (e.g., Raman scans generated by the Raman analyzer 106 and Raman probe 108 or other 1D spectral data generated by different types of spectroscopic systems) to generate predicted outputs 410 (e.g., predicted metabolite concentrations).

[0049] 4, pre-processing of the spectral data occurs both during the training 404 stage (at any time before inputting Raman scan vectors or other spectral data into the model being trained) and when the deep learning model 406 is used during run-time to generate predicted outputs 410 (e.g., immediately after the Raman analyzer 106 generates a Raman scan vector). This pre-processing involves converting the spectral data from its original 1D format into a pseudo-image (i.e., a 2D spectral data matrix) so that the model can process the spectral data in essentially the same way it processes an image.

[0050] Each Raman (or NIR, etc.) spectroscopic measurement, when converted into a pseudo-image, can result in a relatively large input image with large x and y dimensions. To feed such image data directly into a machine learning model may require the model to have a large number of parameters, which can unnecessarily increase computation time. Therefore, before the predictive application 132 applies the Raman scan vectors (or other 1D spectral data) as model inputs, one or more steps of pre-processing and dimensionality reduction may be applied to the data.

[0051] 5 illustrates an example pre-processing 500 of 1D spectral data 502 (e.g., Raman scan vectors) that may be performed in the system 100 of FIG. 1 to prepare the 1D spectral data 502 for processing by the DL model 130, the CNN 300, or a deep learning model of the deep learning model 406. The pre-processing 500 may be performed both during training and during run-time operation, thereby ensuring that the model input has a consistent format at both stages. In some embodiments, the pre-processing 500 is performed by the predictive application 132 or the data cleaning software 134.

[0052] In the illustrated embodiment, preprocessing 500 includes pruning (504) the 1D spectral data 502. Pruning (504) may include removing spectral data points (e.g., spectral data points corresponding to particular wavenumbers of a Raman scan) that are known (e.g., through earlier experimentation) to have a lower correlation with the model output (i.e., lower predictive power). In some embodiments, pruning (504) includes removing spectral data points corresponding to one or more consecutive sequences of wavenumbers. For example, for a Raman scan vector having wavenumbers from 100 to 3425, pruning (504) may include removing (e.g., ignoring or not using) spectral data points corresponding to all wavenumbers outside the range of 450 to 1893. For example, the remaining range of 450 to 1893 may be particularly suitable for predicting metabolite concentrations.

[0053] In other embodiments, pruning (504) may additionally or alternatively include removing non-contiguous sequences of spectral data points. For example, in a Raman scan vector having wavenumbers from 100 to 3425, pruning (504) may include removing (e.g., ignoring or not using) all spectral data points corresponding to wavenumbers outside the range of 500 to 3199, and then iteratively (e.g., retain, remove, remove, retain, remove, remove, etc.) further removing X for every Y of the remaining data points (e.g., 2 for every 3 data points). Removing spectral data points corresponding to wavenumbers 100-499 may be beneficial because that range is known to experience interference from the Raman instrument. Removing spectral data points corresponding to wavenumbers 3200-3325 may be beneficial because that range is known to exhibit relatively high variability.

[0054] After pruning 504 the 1D spectral data 502, the remaining 1D spectral data is normalized 506. Normalizing 506 may include normalizing intensity values ​​across the remaining spectral (e.g., wavenumber) range of the pruned 1D spectral data. For example, normalizing 506 may include mapping the pruned 1D spectral data to a standard distribution with zero mean and standard deviation 1. As another example, normalizing 506 may include mapping minimum and maximum values ​​(e.g., intensity levels) of the 1D spectral data to -1 and +1, respectively.

[0055] The pruned and normalized 1D spectral data is then converted (reformed) from its original 1D format into a 2D matrix of appropriate size (508). In the example above where the Raman scan vector was pruned (504) to only wavenumbers between 450 and 1893 (resulting in a total of 1444 data points), the 2D spectral data matrix may be a 38 x 38 matrix. In another example where the Raman scan vector was pruned (504) to only wavenumbers between 500 and 3199, and then two of every three remaining wavenumbers were removed (resulting in a total of 900 spectral data points), the 2D spectral data matrix may be a 30 x 30 matrix.

[0056] After conversion 508, the prediction application 132 inputs the 2D spectral data matrix into the DL model 130. It is understood that in some embodiments, pre-processing 500 includes additional and / or different steps to those shown in FIG.

[0057] As noted above, the techniques described herein may eliminate the need for recalibration, or at least frequent recalibration. However, in some embodiments, the computer 110 or training server 112 occasionally recalibrates the DL model 130. FIG. 6 illustrates one such embodiment in an example data flow 600 that may occur in the system 100 of FIG. 1 or the system 200 of FIG. 2. In the data flow 600, a past data set 602 may be present in the training database 138. The past data set 602 includes 1D spectral data (e.g., Raman scan vectors) and corresponding levels generated by a suitable data / system (e.g., similar to the Raman analyzer 106 and Raman probe 108). The labels may be actual measurements of the parameters of interest (e.g., metabolite levels) made by an analytical instrument (similar to the instrument 104) at or near the same time that the 1D spectral data is generated.

[0058] A computing device or system, such as the training server 112, then trains 604 a deep learning model (e.g., CNN 300) using the 1D spectral data of the historical data set 602 as features / inputs and the corresponding analytical measurements as labels to generate a trained deep learning model 606. During run-time operation, the deep learning model 606 operates on the 1D spectral data 608 (e.g., Raman scan vectors generated by the Raman analyzer 106 and Raman probe 108) to generate predicted outputs 610 (e.g., predicted metabolite concentrations). Although not shown in FIG. 6, pre-processing of the 1D spectral data (e.g., similar to pre-processing 500) may occur both during the training 604 stage (at any time before inputting the Raman scan vectors or other 1D spectral data into the model being trained) and when the deep learning model 606 is used during run-time to generate the predicted outputs 610 (e.g., immediately after the Raman analyzer 106 generates the Raman scan vectors).

[0059] The data flow 600 may also cause the computer 110 or training server 112 to determine (612) whether an analytical measurement corresponding to the most recent Raman scan spectrum or other spectral data is available (e.g., from the analytical instrument 104). If so, the computer 110 or training server 112 uses the new measurement as labels (and the corresponding spectral data as model features / inputs) to further train (i.e., tune) the deep learning model 606. If no such measurement is available, the model is not further trained / tuned.

[0060] In some embodiments, the techniques described herein (e.g., pre-processing 500) are used in conjunction with JITL. Example method 700 of FIG. 7 illustrates one such embodiment. Method 700 may be performed, for example, by computer 110 (e.g., processing unit 120 executing instructions stored in memory 128) and / or training server 112. In method 700, at block 702, a new scan of a pharmaceutical process is obtained. The scan includes 1D spectral data (e.g., a sequence of intensity values ​​or [wavenumber, intensity] tuples ordered according to wavenumber) generated by a spectroscopic system (e.g., a Raman scan vector generated by Raman analyzer 106 using Raman probe 108) and may be a single raw scan, an aggregate of multiple scans, an average of multiple scans, etc.

[0061] At block 704, a database is queried that includes observational data sets (e.g., similar to training database 138). The observational data sets are associated with previous / past observations of a pharmaceutical process (e.g., the same types of pharmaceutical processes as those referenced above in connection with block 702). Each observational data set may include corresponding analytical measurements in addition to scans (e.g., Raman scan vectors or other 1D spectral data). Analytical measurements may be, for example, media component concentrations, media conditions (e.g., glucose, lactate, glutamate, glutamine, ammonia, amino acids, Na+, K+ and other nutrients or metabolites, pH, pCO2, pO2, osmolality, etc.), viable cell density, titer, critical quality attributes, and / or cell conditions.

[0062] Block 704 includes determining a query point based at least in part on the new 1D spectral data. Depending on the embodiment, the query point may be determined based on the raw 1D spectral data or after suitable pre-processing of the raw 1D spectral data (e.g., similar to pre-processing 500). In some embodiments, the query point is also determined based on a media profile associated with the pharmaceutical process (e.g., fluid type, specific nutrients, pH level, etc.) and / or other information, such as, for example, one or more operating conditions under which the pharmaceutical process is analyzed (e.g., metabolite concentration set points, etc.). Block 704 may include selecting, as training data, from among the observation data sets, observation data sets that satisfy one or more relevance criteria with respect to the query point. If the query point includes a Raman spectral scan vector, for example, block 704 may include comparing the Raman spectral scan vector to the spectral scan vectors associated with each previous observation represented in the training database.

[0063] In block 706, the deep learning model (e.g., DL model 130, CNN 300, or deep learning model 406) is recalibrated (retrained) in response to the query using the portion of the observational dataset selected in block 704. In block 708, a property or parameter of the pharmaceutical process is predicted by the recalibrated deep learning model operating on additional 1D spectral data (e.g., Raman scan vectors newly generated by Raman analyzer 106) after the additional 1D spectral data has been preprocessed (e.g., according to preprocessing 500).

[0064] 8-17 show experimental results for various parameters (VCD, viability, TCD, glucose concentration, lactate concentration, osmolality, glutamate concentration, glutamine concentration, potassium concentration, and sodium concentration, respectively) and an example implementation of a deep learning model (in these examples, a CNN model similar to CNN model 300). In the plots of FIGS. 8-17, each "x" symbol represents an actual measurement of the parameter / attribute being measured (e.g., generated by an analytical instrument similar to one of analytical instruments 104 of FIG. 1 or FIG. 2), while the solid line represents the predicted value of the parameter / attribute (as predicted by the CNN model). In each of FIGS. 8-17, the plots in the left column represent results obtained using a first pre-processing method, while the plots in the right column represent results obtained using a second pre-processing method. The "first method" is preprocessing 500 of FIG. 5, where the 1D spectral data (here Raman scan vectors) are pruned down to wavenumbers starting at 450 and ending at 1893, and the 2D spectral data matrix is ​​a 38×38 matrix. The "second method" is also preprocessing 500 of FIG. 5, but the 1D spectral data (again Raman scan vectors) are pruned down to only one out of every three wavenumbers in the range of 500 to 3199, and the 2D spectral data matrix is ​​a 30×30 matrix. In each of FIGS. 8-17, each row of the plot corresponds to a different formulation. In FIG. 8, for example, results for the first and second formulations are shown for both the first and second preprocessing methods, while results for the third and fourth pharmaceuticals are shown only for the second preprocessing method.

[0065] As can be seen in Figures 8-17, when using the first method of pretreatment, the predicted values ​​for VCD, viability, and glucose were generally in close agreement with the analytical measurements. However, the predicted values ​​for osmolality, glutamine, potassium, and sodium were less consistent. When the second method of pretreatment was applied, the predicted values ​​for all attributes were generally more consistent than seen using the first method. Depending on the metabolites being measured, it may be preferable to use one pretreatment method over another.

[0066] Additional considerations regarding the present disclosure are now provided.

[0067] The terms "polypeptide" or "protein" are used interchangeably throughout and refer to molecules comprising two or more amino acid residues linked together by peptide bonds. Polypeptides and proteins also include macromolecules having one or more deletions, insertions, and / or substitutions from the amino acid residues of the native sequence, i.e., polypeptides or proteins including molecules produced by naturally occurring non-recombinant cells or produced by genetically engineered or recombinant cells and having one or more deletions, insertions, and / or substitutions from the amino acid residues of the native protein's amino acid sequence. Polypeptides and proteins also include amino acid polymers in which one or more amino acids are chemical analogs of the corresponding naturally occurring amino acids and polymers. Polypeptides and proteins also include modifications including, but not limited to, glycosylation, lipid conjugation, sulfation, gamma-carboxylation of glutamic acid residues, hydroxylation, and ADP-ribosylation.

[0068] Polypeptides and proteins may be of scientific or commercial interest, including protein-based therapeutics. Proteins include, inter alia, secreted proteins, non-secreted proteins, intracellular proteins, or membrane-bound proteins. Polypeptides and proteins may be produced by recombinant animal cell lines using cell culture techniques and may be referred to as "recombinant proteins." The expressed protein may be produced intracellularly or secreted into the culture medium from which it can be harvested and / or collected. Proteins include proteins that exert a therapeutic effect by binding to targets, particularly those listed below, including targets derived from, related to, and modifications thereof.

[0069] The protein is an "antigen binding protein." Antigen binding protein refers to a protein or polypeptide that contains an antigen binding region or portion that has a strong affinity for another molecule (antigen) to which it binds. Antigen binding proteins include antibodies, peptibodies, antibody fragments, antibody derivatives, antibody analogs, fusion proteins (including single chain variable fragments (scFv) and dual chain (bivalent) scFv), muteins, xMAbs, and chimeric antigen receptors (CARs).

[0070] An scFv is a single chain antibody fragment that has the variable regions of the heavy and light chains of an antibody linked together. See U.S. Patent Nos. 7,741,465 and 6,319,494 and Eshhar, et al., Cancer Immunol. Immunotherapy (1997) 45:131-136. An scFv retains the ability of the parent antibody to specifically interact with a target antigen.

[0071] The term "antibody" includes reference to both glycosylated and non-glycosylated immunoglobulins of any isotype or subclass, or antigen-binding regions thereof that compete with the intact antibody for specific binding. Unless otherwise specified, antibodies include human, humanized, chimeric, multispecific, monoclonal, polyclonal, hetero-IgG, XmAbs, bispecific, and oligomers or antigen-binding fragments thereof. Antibodies include lgG1, lgG2, lgG3, or lgG4 types. Also included are proteins having antigen-binding fragments or regions, such as Fab, Fab', F(ab')2, Fv, diabodies, Fd, dAb, maxibodies, single-chain antibody molecules, single domain VHH, complementarity determining region (CDR) fragments, scFv, diabodies, triabodies, tetrabodies, and polypeptides that contain at least a portion of an immunoglobulin sufficient to confer specific antigen binding to a target polypeptide.

[0072] Also included are human, humanized, and other antigen binding proteins, such as human and humanized antibodies, that do not produce a significant adverse immune response when administered to humans.

[0073] Also included are peptibodies, which are polypeptides comprising one or more biologically active peptides linked together, optionally via a linker, by an Fc domain, see U.S. Patent Nos. 6,660,843, 7,138,370, and 7,511,012.

[0074] Proteins also include engineered receptors such as chimeric antigen receptors (CARs or CAR-Ts) and T cell receptors (TCRs). CARs typically incorporate an antigen-binding domain (such as an scFv) in conjunction with one or more costimulatory ("signaling") domains and one or more activation domains.

[0075] Also included are bispecific T cell engager (BiTE®) antibody constructs, which are recombinant protein constructs made from two flexibly linked antibody-derived binding domains (see WO 99 / 54440 and WO 2005 / 040220). One binding domain of the construct is specific for a selected tumor-associated surface antigen on the target cell; the second binding domain is specific for CD3, a subunit of the T cell receptor complex on T cells. BiTE® constructs may also include the ability to bind to a context-independent epitope at the N-terminus of the CD3s chain (WO 2008 / 119567) to more specifically activate T cells. Half-life extended BiTE® constructs include fusion of small bispecific antibody constructs, preferably to larger proteins that do not interfere with the therapeutic effect of the BiTE® antibody construct. Examples of such further development of bispecific T cell engagers include the bispecific Fc molecules described, for example, in US2014 / 0302037, US2014 / 0308285, WO2014 / 151910, and WO2015 / 048272. Another strategy is the use of human serum albumin (HAS) fused to the bispecific molecule, or simply the fusion of a human albumin binding peptide (see, for example, WO2013 / 128027, WO2014 / 140358). Another HLE BiTE® strategy involves fusing a first domain that binds to a target cell surface antigen, a second domain that binds to an extracellular epitope of the human and / or rhesus CD3e chain, and a third domain that is a specific Fc modality (WO 2017 / 134140).

[0076] In some embodiments, the protein may include a colony stimulating factor, such as granulocyte colony stimulating factor (G-CSF). Such G-CSF agents include, but are not limited to, Neupogen® (filgrastim) and Neulasta® (pegfilgrastim). Also included are Epogen® (epoetin alfa), Aranesp® (darbepoetin alfa), Dynepo® (epoetin delta), Mircera® (methoxypolyethylene glycol-epoetin beta), Hematide®, MRK-2578, INS-22, Retacrit® (epoetin zeta), Neorecormon® (epoetin beta), Silapo® (epoetin zeta), Binocrit® (epoetin alfa), epoetin alfa H. Also included are, but are not limited to, exal, Abseamed® (epoetin alfa), Ratioepo® (epoetin theta), Eporatio® (epoetin theta), Biopoin® (epoetin theta), epoetin alfa, epoetin beta, epoetin zeta, epoetin theta, and epoetin delta, epoetin omega, epoetin iota, tissue plasminogen activator, GLP-1 receptor agonists, and molecules or variants or analogs thereof and biosimilars of any of the foregoing.

[0077] In some embodiments, the proteins may include proteins that specifically bind to one or more CD proteins, HER receptor family proteins, cell adhesion molecules, growth factors, nerve growth factors, fibroblast growth factors, transforming growth factors (TGFs), insulin-like growth factors, osteoinductive factors, insulin and insulin-related proteins, coagulation and coagulation-related proteins, colony stimulating factors (CSFs), other blood and serum proteins, blood group antigens; receptors, receptor-related proteins, growth hormones, growth hormone receptors, T cell receptors; neurotrophic factors, neurotrophins, relaxins, interferons, interleukins, viral antigens, lipoproteins, integrins, rheumatoid factors, immunotoxins, surface membrane proteins, transport proteins, homing receptors, addressins, regulatory proteins, and immunoadhesins.

[0078] In some embodiments, the protein may include proteins that bind to one or more of the following, alone or in any combination: CD proteins, including but not limited to CD3, CD4, CD5, CD7, CD8, CD19, CD20, CD22, CD25, CD30, CD33, CD34, CD38, CD40, CD70, CD123, CD133, CD138, CD171, and CD174; HER receptor family proteins, including HER2, HER3, HER4; EGF receptor EGFRvIII; cytoplasmic reticulocytes (CRCs); and cytoplasmic reticulocytes (CRCs). Follicular adhesion molecules, such as LFA-1, Mol, p150, 95, VLA-4, ICAM-1, VCAM, and alpha v / beta 3 integrin, growth factors, including, but not limited to, vascular endothelial growth factor ("VEGF"), VEGFR2, growth hormone, thyroid stimulating hormone, follicle stimulating hormone, luteinizing hormone, growth hormone releasing factor, parathyroid hormone, Mullerian inhibitory substance, human macrophage inflammatory protein (MIP-1-alpha), erythropoietin (EPO), nerve growth factors such as NGF-beta, platelet-derived growth factor fibroblast growth factors (PDGFs), such as aFGF and bFGF, epidermal growth factor (EGF), Cripto, transforming growth factors (TGFs), such as TGF-α and TGF-β, including TGF-β1, TGF-β2, TGF-β3, TGF-β4, or TGF-β5, insulin-like growth factors-I and -II (IGF-I and IGF-II), des(1-3)-IGF-I (brain IGF-I), and bone morphogenetic factors, insulin, insulin A chain, insulin B chain, proinsulin, and insulin-like growth factor linker. Insulin and insulin-related proteins, including but not limited to synthase proteins, coagulation and coagulation-related proteins such as factor VIII, tissue factor, von Willebrand factor, among others, plasminogen activators such as protein C, alpha-1-antitrypsin, urokinase, and tissue plasminogen activator ("t-PA"), bombadin, thrombin, thrombopoietin, and thrombopoietin receptors, colony stimulating factors (CSFs), including M-CSF, GM-CSF, and G-CSF, among others, albumin, IgE,and other blood and serum proteins including, but not limited to, blood group antigens, e.g., receptors and receptor-associated proteins including the flk2 / flt3 receptor, obesity (OB) receptor, growth hormone receptor, and T cell receptor; (x) neurotrophic factors including, but not limited to, bone-derived neurotrophic factor (BDNF) and neurotrophin-3, -4, -5, or -6 (NT-3, NT-4, NT-5, or NT-6); (xi) interferons including relaxin A chain, relaxin B chain, and prorelaxin, e.g., interferons including interferon-alpha, -beta, and -gamma, interleukins (IL), e.g., IL-1 to IL-2; (xiv) viral antigens including, but not limited to, AIDS envelope virus antigens, lipoproteins, calcitonin, glucagon, atrial natriuretic factor, pulmonary surfactant, tumor necrosis factor-alpha and -beta, enkephalinase, BCMA, IgKappa, ROR-1, ERBB2, mesothelin, RANTES (Regulated Antigens of IL-10, IL-12, IL-15, IL-17, IL-23, IL-12 / IL-23, IL-2Ra, IL1-R1, IL-6 receptor, IL-4 receptor and / or IL-13RA2 which is a receptor for IL-13, or IL-1RAP which is an IL-17 receptor; on activation normal T-cell expressed and secreted), mouse gonadotropin-related peptide, Dnase, FR-alpha, inhibin, and activin, integrin, protein A or D, rheumatoid factor, immunotoxin, bone morphogenetic protein (BMP), superoxide dismutase, surface membrane protein, decay accelerating factor (DAF), AIDS envelope, transport protein, homing receptor, MIC (MIC-a, MIC-B), ULBP 1-6, EPCAM, addressin, regulatory protein, immunoadhesin, antigen-binding protein, somatropin, CTGF, CTLA4, eotaxin-1, MUC1, CEA, c-MET, claudin-18, GPC-3, EPHA2, FPA, LMP1, MG7, NY-ESO-1, PSCA, ganglioside GD2, ganglioside GM2, BAFF, ICOS, OPGL (RANKL), myostatin, Dickkopf-1 (DKK-1), Ang2, NGF,IGF-1 receptor, hepatocyte growth factor (HGF), TRAIL-R2, c-Kit, B7RP-1, PSMA, NKG2D-1, programmed cell death protein 1 and ligand, PD1 and PDL1, mannose receptor / hCGβ, Hepatitis C virus, mesothelin dsFv [PE38 conjugate, Legionella pneumophila pneumophila) (lly), IFN gamma, interferon gamma-inducible protein 10 (IP10), IFNAR, TALL-1, thymic stromal lymphopoietin (TSLP), proprotein convertase subtilisin / kexin type 9 (PCSK9), stem cell factor, Flt-3, calcitonin gene-related peptide (CGRP), OX40L, α4β7, platelet specific (platelet glycoprotein Iib / IIIb (PAC-1), transforming growth factor beta (TFGβ), zona pellucida sperm-binding protein 3 (ZP-3), TWEAK, TSLP, platelet-derived growth factor receptor alpha (PDGFRα), sclerostin, and biologically active fragments or variants of any of the foregoing.

[0079] In another embodiment, the protein of interest is selected from the group consisting of abciximab, adalimumab, adecatumumab, aflibercept, alemtuzumab, alirocumab, anakinra, atacicept, basiliximab, belimumab, bevacizumab, biosozumab, blinatumomab, brentuximab vedotin, brodalumab, cantuzumab mertansine, canakinumab, cetuximab, certolizumab pegol, conatumumab, daclizumab, denosumab, eculizumab, edrecolomab, efalizumab, epratuzumab, etanercept, evolocumab, galiximab, ganitumab, gemtuzumab, golimumab, ibritumomab tiuxetan, infliximab, ipilimumab, Mab, lerdelimumab, rumiliximab, ixekizumab (lxdkizumab), mapatumumab, motesanib diphosphate, muromonab-CD3, natalizumab, nesiritide, nimotuzumab, nivolumab, ocrelizumab, ofatumumab, omalizumab, oprelvekin, palivizumab, panitumumab, pembrolizumab, Includes pertuzumab, pexelizumab, ranibizumab, rilotumumab, rituximab, romiplostim, romosozumab, sargramostim, tocilizumab, tositumomab, trastuzumab, ustekinumab, vedolizumab, visilizumab, volociximab, zanolimumab, zalutumumab, and biosimilars of any of the foregoing.

[0080] Proteins encompass all of the above, and further include antibodies comprising one, two, three, four, five, or six of the complementarity determining regions (CDRs) of any of the above antibodies. Also included are variants that comprise regions that are 70% or more, particularly 80% or more, more particularly 90% or more, even more particularly 95% or more, particularly 97% or more, more particularly 98% or more, even more particularly 99% or more amino acid sequence identical to the reference amino acid sequence of the subject protein. Identity in this regard can be determined using a variety of well-known and readily available amino acid sequence analysis software. Preferred software includes those that implement the Smith-Waterman algorithm, which is believed to be a satisfactory solution to the problem of searching and aligning sequences. Other algorithms may also be used, especially where speed is an important consideration. Commonly used programs for DNA, RNA and polypeptide alignment and homology matching that can be used in this regard include FASTA, TFASTA, BLASTN, BLASTP, BLASTX, TBLASTN, PROSRCH, BLAZE and MPSRCH, the latter of which is an implementation of the Smith-Waterman algorithm for running on massively parallel processors produced by MasPar.

[0081] Some of the figures 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 illustrated. Furthermore, 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.

[0082] An embodiment of the present disclosure relates to a non-transitory computer-readable storage medium having computer code 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 sequence of instructions or computer code for performing the operations, methods, and techniques described herein. The medium and computer code may be specially designed and constructed for the purposes of the embodiments of the present disclosure, or they may be of the type well known and available to those skilled in the art of computer software technology. Examples of computer-readable storage media include, but are not limited to, magnetic media such as hard disks, floppy disks, and magnetic tapes, 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.

[0083] Examples of computer code include machine code, such as that produced by a compiler, and files containing relatively high-level code executed by a computer using an interpreter or compiler. For example, an embodiment of the present disclosure may be implemented using Java, C++, Python, or other object-oriented programming languages ​​and development tools. Further examples of computer code include encryption and compression codes. Furthermore, an embodiment of the present disclosure may be downloaded as a computer program product, which may be transferred from a remote computer (e.g., a server computer) to a requesting computer (e.g., a client computer or a different server computer) over a transmission channel. Another embodiment of the present disclosure may be implemented in hardwired circuitry in place of or in combination with machine-executable software instructions.

[0084] As used herein, the singular terms "a," "an," and "the" can include plural references unless the context clearly indicates otherwise.

[0085] As used herein, the terms "connect," "connected," and "connection" mean an operative coupling or linking. Connected components may be directly or indirectly coupled to each other, for example through a set of separate components.

[0086] As used herein, the terms "substantially", "substantially", and "about" are used to describe and explain small variations. When used in conjunction with an event or circumstance, these terms may refer to not only the exact occurrence of the event or circumstance, but also to the occurrence of an event or circumstance that is very close. For example, when used in conjunction with a numerical value, these terms may refer to a variation of the numerical value of ±10% or less, such as ±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" identical if the difference between the two numerical values ​​is ±10% or less of the average of the numerical values, such as ±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.

[0087] Additionally, amounts, ratios, and other numerical values ​​may be presented herein in a range format, with the understanding that such range format is used for convenience and brevity and includes numerical values ​​expressly specified as the limits of the range, but should be understood in a flexible manner to include all individual numerical values ​​or subranges subsumed within the range, as if each numerical value and subrange were expressly specified.

[0088] Although the present disclosure has been described and illustrated with reference to specific embodiments thereof, these descriptions and illustrations are not intended to limit the present disclosure. Those skilled in the art will understand 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. The illustrations may not necessarily be to scale. Due to manufacturing processes, tolerances, and / or other reasons, there may be differences between the technical expressions in the present disclosure and the actual devices. There may be other embodiments of the present disclosure that are not specifically illustrated. The present specification and drawings (other than those in the claims) should be regarded as illustrative rather than restrictive. Changes may be made to adapt a particular situation, material, composition of matter, 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. Although the techniques disclosed herein have been described with reference to particular operations being performed in a particular order, it will be understood that these operations may be combined, sub-divided, or re-ordered to form equivalent techniques without departing from the teachings of this disclosure. Thus, unless specifically indicated herein, the order and grouping of operations is not intended to be limiting of the disclosure.

Claims

1. 1. A computer-implemented method for monitoring and / or controlling a pharmaceutical process, comprising: acquiring, by one or more processors, one-dimensional (1D) spectral data generated by a spectroscopic system when scanning the pharmaceutical process; converting, by the one or more processors, the 1D spectral data into a two-dimensional (2D) spectral data matrix; predicting, by the one or more processors, a parameter of the pharmaceutical process, wherein predicting the parameter of the pharmaceutical process includes applying the 2D spectral data matrix to an input layer of a deep learning model. A method comprising:

2. 2. The computer-implemented method of claim 1, wherein the 1D spectral data comprises (i) a series of tuples each including an intensity value and a corresponding wavenumber, or (ii) a series of intensity values ​​where each location corresponds to a respective wavenumber.

3. 10. The computer-implemented method of claim 1, wherein the spectroscopy system is a Raman spectroscopy system, a near-infrared (NIR) spectroscopy system, a high performance liquid chromatography (HPLC) spectroscopy system, an ultra-performance liquid chromatography (UPLC) spectroscopy system, or a mass spectroscopy system.

4. 2. The computer-implemented method of claim 1, wherein the deep learning model is a convolutional neural network (CNN) model.

5. converting the 1D spectral data to the 2D spectral data matrix; pruning the 1D spectral data by removing a plurality of spectral data points; using the pruned 1D spectral data to fill the 2D spectral data matrix; and The computer-implemented method of claim 1 , comprising:

6. converting the 1D spectral data to the 2D spectral data matrix; normalizing the 1D spectral data before or after pruning the 1D spectral data. The computer-implemented method of claim 5 further comprising:

7. The computer-implemented method of claim 5 , wherein pruning the 1D spectral data comprises removing spectral data that is less correlated with the parameter.

8. The computer-implemented method of claim 5 , wherein pruning the 1D spectral data comprises removing spectral data points in one or more predetermined ranges of spectral data points.

9. removing spectral data points in the one or more predetermined ranges of spectral data points; removing spectral data points in one or more ranges of spectral data points that are known to have high variability; and Removing spectral data points in one or more ranges of spectral data points known to exhibit interference in a spectroscopic system The computer-implemented method of claim 8 , further comprising one or both of:

10. 6. The computer-implemented method of claim 5, wherein pruning the 1D spectral data comprises removing X of every Y spectral data points in a predetermined range of spectral data points, where X and Y are predetermined positive integers and Y is greater than X.

11. 11. The computer-implemented method of claim 10, wherein X is equal to 2 and Y is equal to 3.

12. controlling, by the one or more processors, at least one parameter of the drug process based at least in part on the predicted parameter of the drug process. The computer-implemented method of any one of claims 1 to 11, further comprising:

13. causing the one or more processors to present the predicted parameters to a user via a display. The computer-implemented method of any one of claims 1 to 11, further comprising:

14. The computer-implemented method of any one of claims 1 to 11, wherein the predicted parameter of the pharmaceutical process is a media component concentration, a media condition, a viable cell density, a titer, a critical quality attribute, or a cell condition.

15. The predicted parameters of the drug process are glucose, lactate, glutamate, glutamine, ammonia, amino acids, Na ＋ , or K ＋ The computer-implemented method of any one of claims 1 to 11, wherein the concentration of

16. The predicted parameters of the drug process include pH, pCO ２ , pO ２ 12. The computer-implemented method of claim 1, wherein the pressure is 0.01 or 0.15, or osmotic pressure.

17. Before acquiring the 1D spectral data, training the deep learning model using historical 1D spectral data generated by one or more spectroscopic systems and corresponding actual analytical measurements of pharmaceutical processes; The computer-implemented method of any one of claims 1 to 11, further comprising:

18. obtaining an actual analytical measurement of said pharmaceutical process with an analytical instrument; (i) additional 1D spectral data generated by the spectroscopy system when the actual analytical measurements were obtained, and (ii) training the deep learning model using the actual analytical measurements of the pharmaceutical process; The computer-implemented method of any one of claims 1 to 11, further comprising:

19. identifying, by the one or more processors, query points associated with scanning the pharmaceutical process by the spectroscopic system; querying, by the one or more processors, a database including a plurality of observational datasets related to past observations of pharmaceutical processes, each of the observational datasets including associated 1D spectral data and corresponding actual analytical measurements, wherein querying the database includes selecting, from the plurality of observational datasets, observational datasets that satisfy one or more relevance criteria with respect to the query points as training data; training, by the one or more processors, the deep learning model using the selected training data and the observation data set that meets the one or more relevance criteria with respect to the query points; The computer-implemented method of any one of claims 1 to 11, further comprising:

20. identifying the query points includes identifying the query points based at least in part on new 1D spectral data, the new 1D spectral data being generated by the spectroscopy system when scanning the pharmaceutical process; 20. The computer-implemented method of claim 19, wherein selecting as training data the observation data sets that satisfy one or more relevance criteria with respect to the query points comprises comparing the new 1D spectral data on which the query points were identified with 1D spectral data associated with the past observations of the pharmaceutical process.

21. identifying the query points, Identifying the query points based at least in part on one or both of (i) a media profile associated with the drug process and (ii) one or more operating conditions under which the drug process is analyzed.

20. The computer-implemented method of claim 19, comprising:

22. The computer-implemented method of any one of claims 1 to 11, wherein the pharmaceutical process is a cell culture process.

23. 23. A non-transitory computer readable medium storing instructions for monitoring and / or controlling a pharmaceutical process, the instructions, when executed by one or more processors, causing the one or more processors to perform a method according to any one of claims 1 to 17 or claims 19 to 22.

Images