Systems and methods for monitoring and controlling biotherapeutic processes through flow cytometry

Abstract

A non-contact system and method for real-time monitoring of biotherapeutic processes including: generating a hyperspectral image of cell body aggregates and cell media in a microfluidic flow cell device using a hyperspectral imaging device; determining transmission signals from the hyperspectral image; converting the transmission signals to absorbance signals; generating an attention map for quality attributes of the cell body aggregates and the cell media, based on the first cell absorbance signal and the second cell absorbance signal; determining, by one or more machine learning models, metabolite predictions for the cell body aggregates and the cell media; determining, by the one or more machine learning models, cell quality attributes and control material attributes; and providing cell culture device control parameters.

Description

Attorney Docket No.: 1475-126 PCTSYSTEMS AND METHODS FOR MONITORING AND CONTROLLING BIOTHERAPEUTIC PROCESSES THROUGH FLOW CYTOMETRYCROSS-REFERENCE TO RELATED APPLICATION(S)

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 733,081 filed on December 12, 2024, which is incorporated herein by reference in its entirety.TECHNICAL FIELD

[0002] The subject matter of the present disclosure relates generally to the monitoring and control of biotherapeutic processes through flow cytometry. More particularly, the subject matter of the present disclosure relates to optical flow cytometry that allows for the measurement of cell bodies and cell media in a fully contactless, non-destructive, and label- free manner through the use of a short-wave infrared (SWIR) hyperspectral camera.BACKGROUND

[0003] The bioprocessing industry relies on analytical tests to track intracellular health conditions, attributes, and improvements of biological agents during the manufacture of biotherapeutics. The current gold standard readings of intracellular activity and status are obtained using flow cytometry. Flow cytometry measures cell phenotypes and functions using targeted fluorescent stains. However, as a result of the fluorescent or antibody labeling, flow cytometry is a sample-destructive technique with low throughput. There are also other sensors that attempt to measure intracellular activity and status indirectly through the use of nearinfrared spectroscopy (NIR) and Raman Probes that are dipped inside the bioreactor. TheseAttorney Docket No.: 1475-126 PCT probes are highly sensitive to noise, probe fouling, temperature, and measurement drift. Thus, the use of current optical probes requires a laborious collection of sufficiently large datasets that can be used to calibrate optical probe signals to true metabolite concentrations. Further, most practitioners find it difficult to confidently adapt black box chemometric models that are difficult to troubleshoot in high-stakes applications such as biopharmaceutical manufacturing.

[0004] In contrast, short-wave infrared (SWIR) wave spectroscopy does not require a submerged probe nor sample fluorescent dye labeling prior to analysis. Furthermore, the use of a short wave infrared hyperspectral camera (SWIR-HSI) allows for fully contactless collection of thousands of absorption spectra using a single HSI image.

[0005] Accordingly, there remains a need for a contactless, non-destructive, and label-free optical flow cytometer capable of monitoring cell bodies directly and the surrounding cell media, thereby allowing for monitoring critical material attributes (CMAs) and monitoring and quantifying the resulting cell quality attributes (CQAs).SUMMARY

[0006] In accordance with aspects of the disclosure, a non-contact system for real-time monitoring of biotherapeutic processes comprises: a flow cell device configured to receive cell body aggregates and cell media from a cell culture device through an inline connector; a hyperspectral imaging system configured to capture a hyperspectral image of the cell body aggregates and the cell media; a processor; and a memory including instructions stored thereon which, when executed by the processor, cause the system to: generate the hyperspectral image of the cell body aggregates and the cell media; convert a first transmission signal of the cell body aggregate to a first cell absorbance signal from theAttorney Docket No.: 1475-126 PCT hyperspectral image; convert a second transmission signal of the cell media aggregate to a second cell absorbance signal from the hyperspectral image; generate an attention map and metabolite predictions for quality attributes of the cell body aggregates and the cell media based on the first cell absorbance signal and the second cell absorbance signal; determine, by one or more machine learning models, cell quality attributes of the cell body aggregates and control material attributes of the cell media, based on the first and second cell absorbance signals; and provide real-time bioreactor control parameters.

[0007] In an aspect of the present disclosure, the cell culture device may be a bioreactor or a shake flask.

[0008] In an aspect of the present disclosure, the cell body aggregates may be cancerous cells.

[0009] In an aspect of the present disclosure, the inline connector may be a closed loop inline connector and may be configured to preserve the sterility of the cell body aggregates and the cell media from the cell culture device.

[0010] In an aspect of the present disclosure, the hyperspectral imaging system may operate at least in a short-wave infrared wavelength range, and may include a reflective lens.

[0011] In an aspect of the present disclosure, the hyperspectral image may be at a 20x magnification.

[0012] In an aspect of the present disclosure, the instructions, when executed by the processor, may further cause the system to: convert the absorbance signal to a polynomial basis; pass the polynomial basis into a pretrained latent space model, wherein the latent space model forms a compressed latent space; generate metabolite predictions by a regression modelAttorney Docket No.: 1475-126 PCT through use of the compressed latent space; and determine coefficients from the latent space model and regression model.

[0013] In an aspect of the present disclosure, the generating of an attention map may further include taking all coefficients corresponding to a non- zero sparse regression from a latent space model model; averaging an absolute weight of each band across all latent space components; and normalizing the averaged absolute weight of each band across all latent space components.

[0014] In an aspect of the present disclosure, the attention map may be formed based on weights assigned to a trained sparse linear regression to each band of input signal.

[0015] In an aspect of the present disclosure, the instructions, when executed by the process, may further cause the system to perform real-time monitoring of cell body aggregate and cell media parameter levels.

[0016] In an aspect of the present disclosure, the metabolite predictions and quality attributes may be determined at least in part by providing the attention map to the one or more machine learning models.

[0017] In an aspect of the present disclosure, the bioreactor control parameters may be provided in real time.

[0018] In accordance with aspects of the disclosure, a processor-implemented method for real-time monitoring of biotherapeutic processes comprises: generating a hyperspectral image of cell body aggregates and cell media in a microfluidic flow cell device using a hyperspectral imaging device; determining transmission signals from the hyperspectral image; converting the transmission signals to absorbance signals; generate an attention map for quality attributes of the cell body aggregates and the cell media, based on the first cell absorbance signal andAttorney Docket No.: 1475-126 PCT the second cell absorbance signal; determine, by one or more machine learning models, metabolite predictions for the cell body aggregates and the cell media; determine, by one or more machine learning models, cell quality attributes and control material attributes; and provide real-time cell culture device control parameters.

[0019] In an aspect of the present disclosure, the processor-implemented method may further comprise pumping the cell body aggregates and the cell media into the microfluidic flow cell device from a cell culture device through an inline connector.

[0020] In an aspect of the present disclosure, the converting of transmission signals may include converting the transmission signals of the cell body aggregate to a cell absorbance signal; and converting the transmission signal of the cell media to a media absorbance signal.

[0021] In an aspect of the present disclosure, the processor-implemented method may further comprise converting the absorbance signal to a polynomial basis; passing the polynomial basis into a pretrained latent space model, and the latent space model forms a compressed latent space; generating metabolite predictions by a regression model through use of the compressed latent space; and determining coefficients from the latent space model and regression model.

[0022] In an aspect of the present disclosure, generating of the attention map may further include taking all coefficients corresponding to a non-zero sparse regression from a latent space model; averaging an absolute weight of each band across all latent space components; and normalizing the averaged absolute weight of each band across all latent space components.

[0023] In an aspect of the present disclosure, the cell culture device may be a bioreactor or a shake flask.Attorney Docket No.: 1475-126 PCT

[0024] In an aspect of the present disclosure, the cell body aggregates may be cancerous cells.

[0025] In an aspect of the present disclosure, the metabolite predictions and quality attributes may be determined at least in part by providing the attention map to the one or more machine learning models.

[0026] In an aspect of the present disclosure, the inline connector may be a closed loop inline connector and may be configured to preserve the sterility of the cell body aggregates and the cell media from the cell culture device.

[0027] In an aspect of the present disclosure, the hyperspectral imaging device may operate at least in a short-wave infrared wavelength range, and may include a reflective lens.

[0028] In an aspect of the present disclosure, the monitoring of biotherapeutic processes may occur in real time, and the cell culture device control parameters may be provided in real time.

[0029] Further details and aspects of exemplary embodiments of the present disclosure are described in more detail below with reference to the appended figures.BRIEF DESCRIPTION OF THE DRAWINGS

[0030] A better understanding of the features and advantages of the present disclosure will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the present disclosure are utilized, and the accompanying drawings of which:

[0031] FIG. 1 is a diagram of a system for sensing cell attributes, in accordance with examples of the present disclosure;Attorney Docket No.: 1475-126 PCT

[0032] FTG. 2 is a block diagram of a controller for the system of FIG. 1 , in accordance with aspects of the present disclosure;

[0033] FIG. 3 is an illustration of a cell culture device, a short-wave infrared hyperspectral imaging system, and an analytical model of the non-contact system for real-time monitoring of biotherapeutic processes, in accordance with aspects of the present disclosure;

[0034] FIG. 4 is a diagram of a Linear Regression of the analytical model of FIG. 3, in accordance with aspects of the present disclosure;

[0035] FIG. 5A is an illustration of an example embodiment of a hyperspectral image taken from the short-wave infrared hyperspectral imaging device of FIG. 3, in accordance with aspects of the present disclosure;

[0036] FIG. 5B is a graphical illustration of an example embodiment of a transmission signal of cell media and cell body aggregate of the analytical model of FIG. 3, in accordance with aspects of the present disclosure;

[0037] FIG. 6A is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for viable cell density (VCD) using training data, in accordance with aspects of the present disclosure;

[0038] FIG. 6B is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for viable cell density (VCD) using validation data, in accordance with aspects of the present disclosure;

[0039] FIG. 7A is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for glucose using training data, in accordance with aspects of the present disclosure;Attorney Docket No.: 1475-126 PCT

[0040] FTG. 7B is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for glucose using validation data, in accordance with aspects of the present disclosure;

[0041] FIG. 8A is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for lactate using training data, in accordance with aspects of the present disclosure;

[0042] FIG. 8B is a graphical illustration of LDA-Spare Linear Regression (L-SLR) correlation to ground truth for a best-performing quantification fold for lactate using validation data, in accordance with aspects of the present disclosure;

[0043] FIG. 9A is an illustration of an average attention map generated from the linear regression model of FIG. 4 for viable cell density (VCD), in accordance with aspects of the present disclosure;

[0044] FIG. 9B is an illustration of an average attention map generated from the linear regression model of FIG. 4 for glucose, in accordance with aspects of the present disclosure;

[0045] FIG. 9C is an illustration of an average attention map generated from the linear regression model of FIG. 4 for lactate, in accordance with aspects of the present disclosure; and

[0046] FIG. 10 is a flow diagram of a method for real-time monitoring of biotherapeutic processes attributes, in accordance with aspects of the present disclosure.DETAILED DESCRIPTION

[0047] The present disclosure relates generally to optical flow cytometry. More specifically, the present disclosure relates to a contactless, non-destructive, label-free opticalAttorney Docket No.: 1475-126 PCT flow cytometer capable of simultaneously monitoring cell bodies directly and surrounding cell media.

[0048] Although the present disclosure will be described in terms of specific examples, it will be readily apparent to those skilled in this art that various modifications, rearrangements, and substitutions may be made without departing from the spirit of the present disclosure.

[0049] For the purpose of promoting an understanding of the principles of the present disclosure, reference will now be made to exemplary embodiments illustrated in the drawings, and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the present disclosure is thereby intended. Any alterations and further modifications of the novel features illustrated herein, and any additional applications of the principles of the present disclosure as illustrated herein, which would occur to one skilled in the relevant art and having possession of this disclosure, are to be considered within the scope of the present disclosure.

[0050] FIG. 1 illustrates a system 100 for sensing cell attributes. System 100 is configured to sense the attributes of cells and cell media, for example, CQAs of cell body aggregates and CMAs of cell media. The system 100 generally includes a hyperspectral imaging system 400 configured to capture a hyperspectral image of cell body aggregates and cell media within a fluid, a flow cell 330, and a controller 200. The system 100 may be used as a contactless sensor for cell body aggregates and cell media from cell culture device using the hyperspectral imaging system 400 and machine learning methods for monitoring and analyzing cell body aggregates and cell media. The system 100 enables contactless and label free quantification of CMAs and CQAs using a Short-Wave Infrared (SWIR) hyperspectral imaging device 410 and enables fully transparent models that reveal how bands of nonlinear absorption spectra contribute to metaboliteAttorney Docket No.: 1475-126 PCT quantification. For example, by indicating contributions of respective spectral bands. The disclosed system 100 solves the technical problem of probe fouling and contamination by using a SWIR hyperspectral imaging system with a closed-loop sterile configuration with no exposure outside the cell culture device and microfluidic flow cell. The disclosed system 100 thus includes the benefit of providing a non-contact approach for monitoring cell body aggregates and cell media, as well as CQAs and CMAs. The system 100 is described in more detail below with reference to FIG. 3.

[0051] Referring now to FIG. 2, exemplary components in the controller 200 in accordance with aspects of the present disclosure include, for example, a database 210, one or more processors 220, at least one memory 230, and a network interface 240. In aspects, the controller 200 may include a graphical processing unit (GPU) 250, which may be used for processing machine learning models.

[0052] Database 210 can be located in storage. The term “storage” may refer to any device or material from which information may be capable of being accessed, reproduced, and / or held in an electromagnetic or optical form for access by a computer processor. Storage may be, for example, volatile memory such as RAM, non-volatile memory, which permanently holds digital data until purposely erased, such as flash memory, magnetic devices such as hard disk drives, and optical media such as a CD, DVD, Blu-ray disc, or the like. In various embodiments, data may be stored on the controller 200, including, for example, user preferences, historical data, and / or other data. The data can be stored in database 210 and sent via the system bus to the processor 220.

[0053] As will be described in more detail later herein, the processor 220 executes various processes based on instructions that can be stored in the server memory 230 and utilizing theAttorney Docket No.: 1475-126 PCT data from the database 210. The illustration of FIG. 2 is exemplary, and it will be understood by persons skilled in the art that other components may exist in a controller 200. Such other components are not illustrated in FIG. 2 for clarity of illustration.

[0054] FIG. 3 illustrates an exemplary non-contact system for in-line self-calibration and critical material parameters and quality attributes monitoring, that includes a bioreactor to flow-cell continuous non-destructive (close or open system) SWIR hyperspectral imaging system, in accordance with the present disclosure. This system boasts a small analyte volume of <500uL and a real-time timescale (e.g., minute-level measurements). The entire system enables a closed-loop feedback system for optimal smart manufacturing.

[0055] FIG. 3 illustrates the connection 350 between a cell culture device 300 and the hyperspectral imaging system 400. The cell culture device 300 may be either a shake flask, a bioreactor, or any other suitable device configured to culture, expand, or process cells in suspension or aggregate form. In aspects, the cell culture device 300 may include batch, fed- batch, or perfusion culture vessels, including single-use bioreactors, stirred-tank bioreactors, wave bioreactors, hollow-fiber systems, or other scale-up / scale-out culture platforms. In the present disclosure, the cell culture device 300 is a bioreactor 130 and contains cell body aggregates and cell medium. The connection 350 may be an in-line fluidic interface that transfers a portion of culture fluid from the cell culture device 300 to the microfluidic flow cell 330 and, in aspects, returns the fluid to the cell culture device 300, thereby forming a closed-loop sterile sampling path. The connection 350 may include one or more sterile connectors, aseptic fittings, valves, and / or tubing (e.g., SWIR-compatible, biocompatible tubing) selected to maintain sterility, minimize dead volume, and reduce shear or damage to cell body aggregates during transport. In further aspects, the connection 350 may beAttorney Docket No.: 1475-126 PCT configured for continuous, periodic, or on-demand sampling, and may include flowconditioning components such as bubble traps, filters, or mixing segments to provide consistent optical measurements in the flow cell 330.

[0056] The cell body aggregates are Chinese Hamster Ovary (CHO) cells, however, they may be any other suitable cell type. For example, the cell body aggregates may be cancer cells, tumor cells, or other mammalian, microbial, insect, plant, or stem-cell derived cultures, including but not limited to T cells, NK cells, iPSCs, hybridomas, or organoid cultures, depending on the manufacturing application. In aspects, the aggregates may contain viable cells, non- viable cells, debris, microcarriers, or secreted biomolecules within the surrounding media. The bioreactor 130 is connected to a feed pump 310. The feed pump supplies the bioreactor 130 with fresh cell medium, that is nutrient rich, to maintain and support the growth of the cell body aggregates. The feed pump also helps to maintain stable conditions within the bioreactor by adding necessary components such as buffers, pH adjusters, and / or chemicals. In aspects, the feed pump 310 may be controlled according to a predetermined feed schedule and / or based on real-time measurements derived from hyperspectral monitoring, such that the addition of nutrients or other supplements is adjusted to maintain target critical material attributes (CMAs) and cell quality attributes (CQAs). The feed pump 310 may include a peristaltic pump, syringe pump, diaphragm pump, or other low-shear pumping mechanism suitable for sterile cell culture operations, and may be configured to deliver one or more feeds (e g., carbon sources, amino acids, salts, growth factors, or other supplements).

[0057] In aspects, the in-line connection 350 enables sterile, contactless transfer of culture fluid between the cell culture device 300 and the microfluidic flow cell 330 while maintaining the culture in a closed system. By avoiding open sampling steps, the connection 350 reducesAttorney Docket No.: 1475-126 PCT contamination risk and eliminates the need for manual draw-and-analyzing procedures. The closed-loop configuration further reduces probe fouling and drift associated with immersed sensors, because optical measurements are performed through the flow cell 330 without introducing any component into the bioreactor 130. In some implementations, the connection 350 may be configured as a recirculating loop so that sampled fluid is returned to the cell culture device 300 after imaging, thereby minimizing sample loss and maintaining culture volume. In other implementations, the connection 350 may be configured for one-way diversion to waste or for intermittent sampling into the flow cell 330 according to a defined schedule. The rate and timing of sampling through the connection 350 may be selected to provide near-real-time monitoring of the culture while preserving cell viability and aggregate integrity during transport.

[0058] The microfluidic flow cell 330 is configured for enabling the imaging of the cell body aggregates and the cell media. The microfluidic flow cell 330 is a sample cell designed so that fluid samples may be continuously flowed through a beam path. New samples may be continually replenished such that continuous imaging is possible. The microfluidic flow cell 330 may be completely transparent and / or may include a transparent window. In aspects, the transparency (or transparent window) provides an optically clear beam path for SWIR illumination and detection, thereby enabling accurate transmission and absorbance measurements of fluids flowing through the flow cell 330. Transparent portions of the flow cell 330 reduce or eliminate spectral attenuation and parasitic absorption by the flow-cell walls, which improves signal-to-noise ratio, stabilizes baseline response, and enhances repeatability of spectral measurements across time and devices. The transparent configuration further allows the hyperspectral imaging device 410 to simultaneously image both cell bodyAttorney Docket No.: 1475-126 PCT aggregates and surrounding cell media within a common field of view, facilitating spatial segmentation of cell regions versus media regions and enabling separate cell-absorbance and media-absorbance signals to be derived from the same hyperspectral image. Additionally, optical access through the transparent flow cell 330 reduces imaging artifacts such as wall scattering, shadowing, and internal reflections, which can otherwise confound spectral analysis and machine learning outputs. Because the flow cell 330 provides optical access without requiring the imaging device 410 to contact the fluid, the system maintains a closed sterile boundary and avoids probe fouling or contamination while permitting high- magnification (e.g., 20x) imaging of cell aggregates. In further aspects, the flow cell 330 may include one or more SWIR-transparent materials or windows (e.g., quartz, sapphire, calcium fluoride, barium fluoride, or SWIR-transparent polymers) positioned along the optical path to provide spectral compatibility over a wavelength range of about 900 nm to about 2500 nm.

[0059] In aspects, the cell media and cell body aggregate samples are collected from the cell culture device 300 (e.g., bioreactor 130) via a closed loop inline connection 350. The microfluidic flow cell 330 is fluidically coupled to the cell culture device 300 by the closed- loop in-line connector 350 such that culture fluid may be transferred to the flow cell 330 for imaging and, in aspects, returned to the cell culture device 300 to maintain a closed sterile boundary.

[0060] In aspects, the system 100 may further include a pump (e.g., a peristaltic pump) configured to pump at a gentle pace (e.g., at 1ml / min speed) the fluid from the cell culture device 300 (e.g., bioreactor 130) to the microfluidic flow cell 330. The pump may be selected and controlled to provide low- shear transport of the culture fluid so as to preserve viability and structural integrity of cell body aggregates during transfer. In further aspects, the pumpAttorney Docket No.: 1475-126 PCT may operate continuously or intermittently and may be synchronized with image acquisition to provide time-resolved monitoring on a minute-scale or other suitable interval. The pump is configured to pump a fluid, which includes the cell body aggregates and the cell media, into the microfluidic flow cell 330. The pump may be connected to the microfluidic flow cell 330 using tubing. The tubing and connector 350 may include sterile, aseptic, and / or single-use fittings (e.g., luer locks, weldable lines, or valved connectors) and may be sized to minimize dead volume, prevent bubble formation, and reduce settling or clogging of aggregates within the sampling path. In some implementations, the connector 350 may include optional flowconditioning components such as bubble traps, inline filters, or mixing segments to present a consistent optical profile within the flow cell 330. Through this closed loop inline connector 350, the sterility of the cell body aggregates and the cell media within cell culture device 300 is not compromised while traveling to the microfluidic flow cell. Accordingly, the disclosed configuration enables non-destructive, real-time sampling without removal of culture to an open environment, thereby reducing contamination risk and eliminating probe fouling associated with immersed sensors. In taking a SW1R hyperspectral image of the cell body aggregates and cell media, the hyperspectral sensor can determine cell quality attributes (CQA) of the cell body aggregates and critical material attributes (CMA) of the cell media without breaking the sterility barrier of the bioreactor. In aspects, these determined CQAs and CMAs may be output to the controller 200 to support process monitoring, trending, and closed-loop control of culture conditions.

[0061] The SW1R hyperspectral imaging system 400 includes a SW1R hyperspectral imaging device 410, a stage, focusing markers, and a light source. The stage is configured to hold the microfluidic flow cell 330 steadily. In aspects, the stage may include translationalAttorney Docket No.: 1475-126 PCT and / or rotational adjustment components for aligning the flow cell 330 with an optical axis of the hyperspectral imaging device 410 and for maintaining a fixed working distance during repeated measurements. The focusing markers are located underneath the stage. The focusing markers may provide visual or algorithmic reference points that enable rapid autofocus and / or repeatable placement of the flow cell 330 at a predetermined focal plane, thereby improving run-to-run reproducibility and supporting in-line self-calibration. The focusing markers are located underneath the stage. The hyperspectral imaging device 410 may include a light source 340, such as a tungsten halogen light. In some aspects, the light source 340 may be positioned to provide uniform SWIR illumination through the transparent portion of the flow cell 330, and may include beam-shaping optics (e.g., diffusers, collimators, or apertures) to reduce hot-spots and enhance signal uniformity across the imaging field.

[0062] The SWIR hyperspectral imaging device 410 (e.g., a Short-Wave Infrared (SWIR) hyperspectral camera) is configured to capture a hyperspectral image of the cell body aggregates and cell media in the flow cell 330 and generate a signal based on the hyperspectral image 510. The hyperspectral image 510 may include a three-dimensional datacube comprising spatial information in two dimensions and spectral information across a plurality of SWIR wavelength bands. In aspects, the signal generated from the hyperspectral image 510 may include separate spectral signatures corresponding to (i) segmented regions containing cell body aggregates and (ii) regions containing surrounding media, thereby enabling parallel determination of CQAs and CMAs from a single in-line measurement.

[0063] The hyperspectral imaging device 410 further includes a lens. The lens is a mirror like reflective lens and may be set to about a 20x magnification, which allows for simultaneous viewing and imaging of both the cell body aggregates and cell media, n aspects,Attorney Docket No.: 1475-126 PCT the reflective lens provides high-magnification imaging with SWIR spectral compatibility by minimizing wavelength-dependent absorption and dispersion that can occur in transmissive optics. Normal glass like lenses often result in the loss of wavelength bands and are not compatible with 20x magnification of hyperspectral images. For example, glass or other transmissive lenses may attenuate or distort portions of the SWIR spectrum, thereby reducing spectral fidelity and introducing baseline artifacts into absorbance measurements. The reflective lens, however, is configured to reflect the wavelength bands, so that the wavelength does not get absorbed or lost. Accordingly, the reflective lens preserves spectral throughput over the SWIR range (e.g., about 900 nm to about 2500 nm), improves signal-to-noise ratio, and supports accurate conversion of transmission to absorbance for downstream modeling. The SWIR hyperspectral imaging device 410, and its reflective lens, is configured to zoom in on cells and cell aggregates for viewing and determining cell body aggregate and cell media characteristics. In aspects, the high-magnification SWIR imaging enables evaluation of aggregate morphology, size distribution, and spatial heterogeneity while concurrently measuring media composition, thereby providing a comprehensive optical representation of the culture state. Thus, the simultaneous viewing of the cell body aggregates and cell media enables a very precise and well-rounded understanding of the bioreactor environment.

[0064] Hyperspectral imaging (HSI) is a technique that analyzes a wide spectrum of light instead of just assigning primary colors (e.g., red, green, blue) to each pixel. The light striking each pixel is broken down into many different spectral bands in order to provide more information on what is imaged. Accordingly, each pixel in a hyperspectral image may be associated with a corresponding spectral signature, enabling simultaneous capture of spatial and chemical information from a scene or sample. In aspects, hyperspectral imaging providesAttorney Docket No.: 1475-126 PCT a spectral “fingerprint” of constituents within the imaged sample, allowing differentiation of cells, aggregates, and media based on wavelength-dependent absorption features. In aspects the hyperspectral imaging device 410 may use SWIR excitation.

[0065] The spectral signal includes SWIR spectra ranges from about 900nm to about 2500nm. The disclosed technology extracts the information from this range of spectra and finds a group of these wavelengths (i.e., signals) that are related to information or certain bands of signals of interest. In aspects, sample information may be obtained from the SWIR bands. In some aspects, the SWIR spectral range may be selected to include overtone and combination vibrational bands associated with water, sugars, organic acids, amino acids, and other metabolites relevant to cell-culture manufacturing. The system may identify one or more wavelength bands that correlate to particular CMAs or CQAs, and may weigh or prioritize such bands during model training, prediction, and / or generation of attention maps. By leveraging multiple SWIR bands simultaneously, the system improves robustness compared to single- wavelength sensing and enables accurate quantification even in optically complex, scattering culture environments.

[0066] Short Wave Infrared (900nm-2500nm) excitation strikes a healthy balance between sensitivity to water and water environment changes, including the addition of salts, n particular, SWIR wavelengths retain sensitivity to aqueous chemical composition while providing sufficient penetration through turbid or cell-dense samples, thereby enabling inline, non-destructive monitoring of culture fluid without dilution or labeling. Because lactic acid is a carboxylic acid, lactic acid forms a detectable signal in the SWIR range. These acids form a carboxylic acid cyclic dimer in solution through hydrogen bonding, which causes a spectral shift and broadening that is visible as a baseline shift in the spectra. Such spectralAttorney Docket No.: 1475-126 PCT shifts may be detected and quantified by the disclosed models, allowing lactate concentration to be inferred from in-line hyperspectral images. More generally, other metabolites or media components that undergo concentration-dependent hydrogen bonding or water- structure perturbations may likewise produce detectable changes in SWIR spectra, enabling broad CM A monitoring.

[0067] In an aspect of the present disclosure, water solutions were created. T o create water solutions, glucose and fructose were weighed using a balance (Mettler-Toledo, ME204) and placed in 15mL conical tubes. Next, the appropriate volume of water was added to reach the desired concentration. The tubes were then thoroughly mixed. Serum solutions were created with heat-inactivated horse serum using the same procedure as for water solutions. Since glucose is present in horse serum at unknown concentrations, the resulting solutions contained concentrations of glucose augmented by both fructose and additional glucose. Samples were randomized before image acquisition. 500 uL of each sample was loaded into a quartz cuvette (Hellma 110-1-40) and imaged using customized microfluidic flow cell with 1 wash of deionized water and 1 wash of imaging solution between image collections.

[0068] Referring to FIG. 3, system 100 includes a machine learning model 500. Machine learning model 500 may include a deep learning model or other suitable machine learning model. Machine learning model 500 is configured to analyze cells and media for output quality attributes and provide real-time bioreactor control parameter adjustments. In aspects, the machine learning model 500 receives hyperspectral image data and / or derived spectral signals and outputs predicted CMAs of the cell media and predicted CQAs of the cell body aggregates, such as metabolite concentrations, viable cell density, aggregate morphology metrics, or other culture-state indicators. In further aspects, predicted CMAs / CQAs may beAttorney Docket No.: 1475-126 PCT provided to controller 200 to generate control outputs, including adjustments to feed rates, perfusion rates, pH control, temperature control, dissolved oxygen control, or other bioreactor operating parameters.

[0069] Once the cell body aggregates and cell media are in the microfluidic flow cell 330, the SWIR hyperspectral imaging device 410, takes a hyperspectral image 510. The hyperspectral image 510 is a 20x photograph of the cell body aggregates 515 and the cell media 520 within the microfluidic flow cell 330. The 20x magnification allows concurrent imaging of cell body aggregates 515 and surrounding media 520 within a single frame, supporting pixel-level discrimination of cell regions versus media regions. The hyperspectral image(s) 510 are then processed. Processing may include background correction, normalization, and segmentation of cell body aggregate regions and media regions within the image(s) 510 to generate separate spectral profiles for cells and media. The transmission signals of both the cell body aggregates 515 and cell media 520 were converted to a cell absorbance signal and a media absorbance signal using equation 1 and smoothed using a gaussian filter (sigma = 2). A is absorbance signal, T is transmission signal.

[0071] For example, FIG. 5 illustrates the graphical illustration of the transmission signals and absorbance signals from the cell media and the transmission signals and absorbance signal from the cell aggregates. With reference to FIG. 3, 540 references the PAT Deep Learning Analytical Model for performing data analysis and model training for evaluating the hyperspectral image(s) 510 (see FIG. 4). In aspects, model 540 may include a training phase in which labeled hyperspectral data are associated with reference assay measurements (e.g., offline metabolite or cell-count assays) to learn relationships between SWIR spectral featuresAttorney Docket No.: 1475-126 PCT and target CMAs / CQAs, and a deployment phase in which new hyperspectral images 510 are analyzed in real time to generate predictions. In the present disclosure, a Partial Least Squares (PLS) with 16 components was used as the baseline model for contactless quantification of cell parameters. In aspects, the baseline PLS model provides a comparative analytical reference for evaluating the performance of the disclosed machine learning model(s), including interpretable sparse linear regression and / or deep learning models.

[0072] FIG. 4 is a diagram illustrating the architecture of the LDA-Square Linear Regression (L-SLR), which is utilized for both the cell body aggregate and the cell media, n aspects, any suitable regression model and / or latent-space shaping model may be substituted for, or used in combination with, the L-SLR framework described herein to generate metabolite predictions and other culture-attribute estimates from the SWIR absorbance signals. First, at 610, the SWIR absorbance signal is taken to a polynomial basis 612. In aspects, the polynomial basis expansion generates linear terms, higher-order terms, and crossterms from the SWIR absorbance signal to enhance separability of spectral features associated with different CM As and CQAs. Next, at 620, the SWIR absorbance signal in the polynomial basis 612 is then passed into pretrained Linear Discriminant Analysis (LDA) model 622. The pretrained LDA model 622 is trained by rounding ground truth labels in training dataset to the nearest integer for use as classification labels. In aspects, the rounded labels provide discrete classes that allow the LDA model 622 to learn projections that maximize separation between culture-state conditions while retaining spectral interpretability. Further, only the cross-terms of polynomial basis are used for quantification of fructose, while all polynomial terms are used for quantification of cell parameters in spent cell media. In aspects, this selective use of cross-terms versus all polynomial terms is based on the spectral sensitivity of the targetAttorney Docket No.: 1475-126 PCT analytes and is intended to improve prediction accuracy while preserving sparsity. For the spent cell media dataset, classification labels 624 were formed by assigning a unique integer to each distinct combination of VCD, glucose, and lactate. Next, LDA 622 forms a compressed latent space 630. The compressed latent space 630 generates LDA latent space signals. In aspects, the compressed latent space 630 includes a reduced number of components (e.g., a selected percentage of LDA components) that capture the most discriminative spectral variance relevant to the target CMAs / CQAs. Then, sparse linear regression (SLR) 640 using Orthogonal Matching Pursuit is trained to predict metabolite concentrations 650 via signals 635 from LDA compressed latent space 630, as shown in equation 2, where y is the predicted metabolite reading, 0 is the sparse weights, and z is the LDA latent space signal.

[0073] y = 0z (Eqn. 2)

[0074] All ground truth labels are normalized prior to fitting sparse regression model. Further, ground truth labels for cell media samples are kept unfiltered. In aspects, only data with cell viability greater than 50% and lactate levels less than 300 mg / dL may be used for model training from spent cell media dataset. Two flasks may be randomly selected for model training, and the remaining flasks are used as test data. 20 cross-validation folds were trained using this split. 50 % and 75 % respectively of the fructose samples from each glucose level were randomly drawn and reserved for training, while the remaining 50% and 25% of data respectively was reserved for model evaluation. 20 cross validation folds were trained using this split.

[0075] Because L-SLR is a fully linear model, it is possible to form attention maps 660 for each trained model based on the weights assigned to each band of input signal. In aspects, the attention maps 660 provide interpretable visualizations identifying which SWIRAttorney Docket No.: 1475-126 PCT wavelength bands or polynomial features contribute most strongly to each predicted metabolite or cell parameter. First, all coefficients corresponding to non-zero sparse regression components are taken from LDA model 622. The average absolute weight of each band across all latent space components is then taken and normalized. Cross terms are normalized separately from linear terms. Finally, all normalized coefficients less than 0.05 are set to zero. FIG. 9 illustrates exemplary attention maps 660.

[0076] In an aspect of the present disclosure, only data with cell viability greater than 80% was used as test data for models trained using spent cell media dataset, to minimize model extrapolation. All test data was used to evaluate other cell parameter models. Predicted cell parameter concentrations were compared to ground truth readings via r2, root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) metrics. These evaluation metrics quantify goodness-of-fit and prediction error across both metabolite models and cell -parameter models, enabling benchmarking of L-SLR performance relative to baseline approaches.

[0077] In other aspects, the prediction framework described with respect to FIG. 4 is not limited to sparse linear regression or to the particular latent-space construction shown. Rather, the metabolite predictions and / or cell-parameter predictions may be generated using any suitable regression model trained on the hyperspectral absorbance signals, polynomial-basis features, and / or latent-space features derived therefrom. Likewise, the latent-space shaping step may be performed using any suitable dimensionality-reduction, embedding, or featuretransformation technique configured to map the absorbance signals into a reduced or structured representation that preserves information relevant to the target Critical Material Attributes and Cell Quality Attributes. In such aspects, the regression model and latent-spaceAttorney Docket No.: 1475-126 PCT shaping model may be selected or combined based on desired predictive accuracy, robustness to scattering or matrix effects, computational efficiency for real-time operation, and / or interpretability. Further, regardless of the particular regression or latent-space approach employed, the system may be configured to generate an interpretability output indicating relative contributions of wavelength bands and / or band interactions to the resulting predictions.

[0078] FIG. 5A illustrates a sample hyperspectral image of cell body aggregates and cell media in accordance with the present disclosure and as described in FIG. 3. 510 references an exemplary image of cell body aggregates 515 and cell media 520 taken with the SWIR hyperspectral system 400 of FIG. 3. The hyperspectral image 510 includes a plurality of pixels, wherein each pixel is associated with a corresponding SWIR spectral signature across the wavelength range described herein (e.g., about 900 nm to about 2500 nm). The imaged field of view includes one or more regions containing cell body aggregates 515 and one or more regions containing surrounding cell media 520 within the microfluidic flow cell 330. The cell body aggregates 515 appear as optically distinct regions relative to the cell media 520 due to differences in scattering, absorption, and local composition, thereby enabling separation of cell-associated spectra from media-associated spectra from a single in-line measurement.

[0079] FIG. 5B illustrates the transmission signals from the cell media area and the transmission signals from cell body aggregates, as described in FIG. 3. The transmission signals are derived by selecting pixels corresponding to the media region and pixels corresponding to the aggregate region in the hyperspectral image 510, and then computing spectral transmission profiles for each region. The media-region transmission signalAttorney Docket No.: 1475-126 PCT represents wavelength- dependent attenuation attributable primarily to dissolved metabolites and media constituents, whereas the aggregate-region transmission signal represents combined attenuation attributable to aggregate-associated cellular constituents and local microenvironment effects. These region-specific transmission signals may be converted to corresponding absorbance signals (e.g., using Eqn. 1) to yield a cell absorbance signal associated with aggregates 515 and a media absorbance signal associated with media 520. Separation of these signals enables parallel quantification of CQAs from cell aggregates and CMAs from media, thereby providing a comprehensive optical snapshot of culture state without destructive sampling.

[0080] To evaluate model performance, all models were trained and evaluated on 20 random cross validation folds. Each cross-validation fold includes a distinct random split of the dataset into training and evaluation subsets, thereby minimizing bias associated with a single train / test partition and providing statistically robust performance estimates. Cross- validation further demonstrates that the disclosed models generalize across different culture flasks and culture conditions, supporting real-time deployment in manufacturing settings. Tables 1 through 3 summarize average viable cell density (VCD), glucose and lactate quantification performance across 20 cross validation folds when two cell culture flasks (25% of dataset) are randomly reserved for model training. Standard deviations across folds are also included. Table 1 illustrates the Quantification of Viable Cell Density in Spent Cell Media (25% of Data Used for Training)

[0081] Table 1.Train ValidAttorney Docket No.: 1475-126 PCTRMSE xMAE x RMSE xMAE x r2IO5IO5MAPE r2IO5IO5MAPEL-SLR 0.87±0.02 9.56±L36 7.38±L36 0.32±0.17 0.75±0.06 13.3±1.72 10.1±L53 0.34±0.11PLS (n=16) 0.95±0.02 3.OHO.98 2.41±0.79 0.07±0.04 0.36±0.39 12.0±2.51 8.80±2.36 O.3HO.1

[0082] Table 2 illustrates the Quantification of Glucose in Spent Cell Media (25% of DataUsed for Training)Table 2.Train Valid r2RMSE MAE MAPE r2RMSE MAE MAPEL-SLR 0.95±0.02 36.19±6.6529.18±4.490.19±0.04 0.87±0.04 53.65±7.67 41.55±6.780.29±0.07PLS (n=16) 0.99±0.01 12.06±4.769.39±3.61 0.05±0.02 0.72±0.31 58.86±26.5436.32±8.860.14±0.04

[0083] Table 3 illustrates the Quantification of Lactate in Spent Cell Media (25% of DataUsed for Training)Table 3.Train Valid r2RMSE MAE MAPE r2RMSE MAE MAPEL-SLR 0.96±0.01 15.62±2.1412.33±2.030.09±0.02 0.91±0.02 22.21±2.6917.59±2.470.13±0.02PLS (n=16) 0.99±0.01 7.23±3.79 5.68±2.68 0.04±0.02 0.71±0.1 31.14±4.2622.12±3.770.17±0.04

[0084] FIGS. 6, 7, and 8 visualize L-SLR correlation to ground truth for best performing quantification fold for VCD, glucose and lactate respectively, r2and RMSE are listed in the top left corner of each graph. VCD measurements are reported in cells / mL while glucose and lactate measurements are reported in mg / dL. FIG. 6 illustrates a correlation visualization between ground truth viable cell density (VCD) values and VCD values predicted by theAttorney Docket No.: 1475-126 PCT disclosed L-SLR model using spent cell media. FIG. 6A corresponds to the training subset for a best-performing cross-validation fold, and FIG. 6B corresponds to the validation subset for that fold. Each data point represents a sample for which a hyperspectral measurement was obtained, processed into a media-associated spectral signal, and analyzed by the L-SLR pipeline to generate a predicted VCD. The diagonal line represents ideal agreement between prediction and ground truth. The r2and RMSE values shown in the upper-left portion of each plot quantify goodness-of-fit and prediction error, respectively, with VCD reported in units of cells / mL.

[0085] FIG. 7 illustrates a correlation visualization between ground truth glucose concentrations and glucose concentrations predicted by the disclosed L-SLR model using spent cell media. FIG. 7A corresponds to training data for a best-performing cross-validation fold, and FIG. 7B corresponds to validation data for that fold. The plotted points show predicted glucose concentrations versus reference assay glucose concentrations, and the diagonal line denotes perfect prediction. The r2and RMSE values in the upper left of each plot demonstrate that the disclosed contactless SWIR hyperspectral workflow can quantify glucose in spent media with high correlation to ground truth. Glucose values are reported in mg / dL.

[0086] FIG. 8 illustrates a correlation visualization between ground truth lactate concentrations and lactate concentrations predicted by the disclosed L-SLR model using spent cell media. FIG. 8A corresponds to training data for a best-performing cross-validation fold, and FIG. 8B corresponds to validation data for that fold. Each point represents a spent-media sample imaged in the flow cell and analyzed via the L-SLR model to yield a predicted lactate concentration. The diagonal line indicates ideal agreement, and the r2and RMSE valuesAttorney Docket No.: 1475-126 PCT displayed in the upper left of each sub-figure quantify predictive performance. Lactate values are reported in mg / dL.

[0087] Across all folds, L-SLR exhibits significantly better correlation and less performance variation compared to PLS. In aspects, certain error metrics (e.g., MAE or MAPE) may remain similar between models because a small number of spurious outlier predictions can disproportionately affect aggregate accuracy statistics, even where overall correlation trends favor L-SLR. Both models achieve r2values less than 0.9 for VCD. The reduced VCD correlation relative to glucose and lactate may be attributable to the fact that spent, cell-free media is imaged, such that direct optical signatures of cell number are not present in the hyperspectral measurement. Since cell free permeate is imaged, perhaps hyperspectral images lack concrete signals that correlate to VCD. Most likely, the model is forced to correlate signals belonging to other cell parameters with VCD instead. Accordingly, the disclosed results indicate that media-based spectral features provide particularly strong predictive support for metabolite CMAs (e.g., glucose and lactate), while VCD prediction from spent media may rely on secondary correlations with metabolite or process-state signals.

[0088] Unlike Mid-Infrared (MIR) absorption, SWIR absorption reflects weak signals caused by overtones and combination vibrations. In aspects, SWIR spectra are dominated by higher-order vibrational features of common functional groups (e.g., O-H, C-H, N-H, and C=O), which produce broad and overlapping absorbance bands rather than narrow fundamental peaks typical of MIR. These signals are difficult to simulate and interpret, but model attention maps help the practitioner correlate high weights with known functional group harmonics in the literature. Accordingly, attention maps provide an interpretable linkage between spectral regions and predicted CMAs / CQAs, enabling practitioners toAttorney Docket No.: 1475-126 PCT validate that model predictions are driven by chemically meaningful features rather than noise or imaging artifacts. Attention maps formed via weights assigned by each trained L-SLR to each band of input signal are shown in FIG. 9. In aspects, the attention maps shown in FIGS. 9A-9B reflect normalized band-level contributions derived from the sparse regression coefficients and LDA weight vectors described above (see FIG. 4).

[0089] It is well known that hydrogen bonding within the solution matrix impacts SWIR absorption spectra. In particular, changes in hydrogen-bonding networks modulate O-H overtone intensity, peak location, and baseline behavior, alter SWIR absorbance in water-rich biological matrices. This phenomenon is reflected in the distinct differences between attention maps for models trained to identify fructose within glucose solutions vs horse serum solutions spiked with glucose (FIG. 9). For example, the model trained using glucose solutions pays more attention to absorption signals below 1400 nm, whereas the model trained using horse serum pays more attention to absorption signals between 1600- 1800nm. The former region is dominated by overtones related to O-H stretch, whereas the latter region is dominated by C- H stretch vibrations. Accordingly, the glucose- solution model appears to rely more heavily on water-matrix perturbations driven by fructose concentration, while the serum-solution model appears to rely more heavily on fructose-related C-H vibrational features and / or fructose interactions with organic constituents present in serum. This implies that the correlation between fructose concentration and SWIR absorption can be related to both changes in the hydrogen bonding matrix of water and interactions of fructose with organic compounds dissolved in the horse serum. These comparative attention-map shifts further demonstrate that the disclosed L-SLR framework can adapt to different solution matricesAttorney Docket No.: 1475-126 PCT while still maintaining spectral interpretability, which is valuable for deployment across varying media formulations in manufacturing.

[0090] FIG. 9 illustrates average attention map weights of bands across all cross-validation folds for L-SLR models trained on the spent cell media dataset. Specifically, FIG. 9A corresponds to the viable cell density (VCD) model, FIG. 9B corresponds to the glucose model, and FIG. 9C corresponds to the lactate model. In each of FIGS. 9A-9C, panel (I) depicts normalized weights assigned to individual SWIR wavelength bands, and panel (II) depicts normalized weights assigned to polynomial cross-terms between wavelength bands. In aspects, single-band weights (I) correspond to contributions from linear spectral features, whereas cross-term weights (II) correspond to polynomial interaction features that capture nonlinear relationships between pairs of wavelength bands.

[0091] Although FIGS. 9A-9C represent models trained to predict three distinct culture parameters, the attention-weight profiles share several common high-weight regions, indicating that multiple parameters co-vary with underlying media-matrix and metabolic spectral features. Notably, FIG. 9A (VCD) and FIG. 9C (lactate) exhibit overlapping high- weight regions, including a pronounced weighted band near approximately 2200 nm, whereas FIG. 9B (glucose) de-emphasizes this region. The -2200 nm region is consistent with overtone absorption associated with carboxylic-acid functional groups, which are characteristic of lactate, thereby supporting the chemical plausibility of the lactate model in FIG. 9C and the observed VCD-lactate coupling in FIG. 9A. Further, all three sub-figures (FIGS. 9A-9C) assign substantial weight to bands below approximately 1400 nm, suggesting that spectral changes associated with water hydrogen-bonding structure and bulk mediamatrix variation contribute to prediction of VCD, glucose, and lactate in this dataset.Attorney Docket No.: 1475-126 PCT

[0092] SWIR spectra nevertheless exhibit broad, overlapping overtone and combination peaks, and relationships between SWIR absorbance features and dissolved molecules may be nonlinear and matrix-dependent. Accordingly, similar SWIR spectra may correspond to different underlying solution matrix profiles, which in some cases can contribute to spurious cell-parameter predictions. In further aspects, interpretable models may be developed that enable both band identification and automated filtration of spurious metabolite predictions, for example by applying spectral-similarity metrics, latent-space clustering, confidence scoring, or residual-based outlier thresholds. Because LDA does not necessarily yield a smoothly parameterized latent space, identifying absorption spectra that overlap multiple metabolic profiles for automatic removal may be nontrivial; however, the disclosed L-SLR framework provides a computationally efficient and interpretable basis for such PAT deployments.

[0093] The L-SLR model does exhibit signs of overfitting when predicting fructose concentrations in glucose solutions. Perhaps this is caused by the similarity in structure between fructose and glucose. It is evident that traditional PLS is better able to find correlation between absorbance spectra and fructose concentrations. Still, the L-SLR framework is an efficient model to train, as evidenced by positive regression performance in cell culture flask dataset using only 25% of the data for training. In the future, this sensor framework may be used to quantify other valuable cell parameters such as pH, amino acids, and even antibody glycosylation.

[0094] In aspects, for the fructose-in-glucose dataset, overfitting may be attributable to structural and spectral similarity between fructose and glucose, and baseline PLS models may exhibit improved fructose correlation in that regime. In contrast, for spent cell media datasets,Attorney Docket No.: 1475-126 PCT the disclosed L-SLR framework demonstrates data-efficient performance using limited training data (e.g., about 25% of available flasks). In further aspects, the disclosed contactless SWIR hyperspectral sensor framework may be trained to quantify additional CMAs and CQAs, including pH, amino acids, and product-quality indicators such as antibody glycosylation, using corresponding hyperspectral datasets and ground-truth assays.

[0095] Referring to FIG. 10, a flow diagram for a method 700 for real-time monitoring of biotherapeutic processes attributes is shown. Although the blocks of FIG. 10 are shown in a particular order, the blocks need not all be performed in the illustrated order, and certain blocks can be performed in another order. F or example, FIG. 10 will be described below, with a controller 200 of FIG. 2 performing operations. In aspects, the operations of FIG. 10 may be performed all or in part by another device, for example, a server, a user device, and / or a computer system. These variations are contemplated to be within the scope of the present disclosure.

[0096] Initially, at block 702, the controller 200 causes the system 100 to cause the cell body aggregates and cell media to flow from the cell culture device to the microfluidic flow cell device through an inline connector. In aspects, the controller 200 controls fluid transfer through the closed loop in-line connector 350 shown in FIG. 3, such that culture fluid containing cell body aggregates and cell media is drawn from the cell culture device 300 (e.g., bioreactor 130) and delivered to the microfluidic flow cell device 330 for imaging. The culture fluid may be pumped through the connector 350 using a gentle low-shear pump (e.g., a peristaltic pump) as described with respect to FIG. 3, for example at an in-line sampling rate on the order of about 1 mL / min, so as to preserve aggregate integrity and viability during transport. The closed-loop configuration of connector 350 maintains a sterile boundaryAttorney Docket No.: 1475-126 PCT between the cell culture device 300 and the microfluidic flow cell 330, thereby preventing exposure of the culture to the external environment and avoiding contamination or probe fouling.

[0097] In further aspects, the controller 200 may cause fluid transfer continuously or intermittently and may synchronize sampling through connector 350 with hyperspectral image acquisition, thereby providing a representative, near-real-time sample volume within the microfluidic flow cell 330 (e.g., less than about 500 pL) for subsequent contactless SWIR hyperspectral monitoring as shown in FIG. 3. In implementations in which the connector 350 forms a recirculating loop, the sampled fluid may be returned to the cell culture device 300 after imaging to minimize sample loss and maintain culture volume.

[0098] At block 704, the controller 200 causes the system 100 to generate a hyperspectral image 510 of the cell body aggregates 515 and the cell media 520 in the microfluidic flow cell 330 device using a short-wave infrared hyperspectral imaging system 400. Once the culture fluid is present within the microfluidic flow cell 330, the controller 200 actuates the SWIR hyperspectral imaging system 400 (see FIG. 3), including the SWIR hyperspectral imaging device 410, to acquire hyperspectral image data of the sample. The hyperspectral imaging system 400 illuminates the flow cell 330 using the light source 340 (e.g., a tungsten halogen light source) and captures transmitted or otherwise detected SWIR spectral information over a plurality of wavelength bands (e.g., from about 900 nm to about 2500 nm). The microfluidic flow cell 330, being transparent or including a transparent window as described herein, provides an optically clear beam path that allows the imaging device 410 to simultaneously image both the cell body aggregates 515 and surrounding cell media 520 within a common field of view. In further aspects, the imaging device 410 may employ a reflective lens and aAttorney Docket No.: 1475-126 PCT magnification of about 20* to resolve aggregate regions and media regions with sufficient spatial detail for downstream separation of cell-associated versus media-associated spectral signals.

[0099] The resulting hyperspectral image 510 may comprise a spectral datacube associating each pixel in the field of view with a corresponding SWIR spectrum, thereby enabling subsequent processing to extract transmission and absorbance signals for the aggregates 515 and the media 520 as described with respect to FIG. 3. The controller 200 may trigger image acquisition continuously or at defined time intervals (e.g., minute-scale) to support real-time or near-real-time monitoring of culture attributes.

[0100] At block 706, the controller 200 causes the system 100 to determine transmission signals of the cell body aggregates 515 and transmission signals of the cell media from the hyperspectral image 510. The controller 200 initiates processing of the hyperspectral image 510 to extract wavelength-dependent transmission profiles for both regions corresponding to cell body aggregates 515 and (regions corresponding to surrounding cell media 520, as described with respect to FIG. 3 and FIG. 5B. To do so, the system 100 may identify or segment pixels associated with the cell body aggregates 515 and pixels associated with the cell media 520 within the hyperspectral image 510, for example based on spatial contrast, scattering differences, or other image-based separation techniques.

[0101] The system 100 then determines, for each of the aggregate and media regions, a transmission signal across the SWIR spectral bands captured by the imaging device 410. The resulting transmission signals represent wavelength-dependent attenuation of SWIR light through the cell body aggregates 515 and through the cell media 520, respectively, and serve as region- specific spectral inputs for subsequent conversion to absorbance and model-basedAttorney Docket No.: 1475-126 PCT quantification. In further aspects, transmission signals may be averaged over multiple pixels, multiple fields of view, or multiple timepoints to reduce noise and improve robustness for real-time monitoring.

[0102] At block 708, the controller 200 causes the system 100 to convert the transmission signals of the cell body aggregates to cell absorbance signals. The controller 200 further causes the system 100 to convert the transmission signals of the cell media 520 to media absorbance signals. The system 100 applies the absorbance conversion described herein (see Eqn. 1 and FIG. 5B) to each of the region- specific transmission signals determined at block 706. For the aggregate-associated transmission signal, the system 100 generates a first cell absorbance signal that reflects wavelength-dependent absorption attributable to the cell body aggregates 515 and their local microenvironment. For the media-associated transmission signal, the system 100 generates a second media absorbance signal that reflects wavelengthdependent absorption attributable primarily to dissolved metabolites and media constituents within the culture fluid. In aspects, conversion to absorbance linearizes relationships and facilitates comparison across samples and timepoints by reducing sensitivity to absolute illumination intensity. The resulting cell absorbance and media absorbance signals may be smoothed (e.g., using a Gaussian filter as described with respect to FIG. 3) to reduce high- frequency noise while preserving metabolite-related spectral features for downstream modeling.

[0103] At block 710, the controller 200 causes the system 100 to convert the absorbance signals to a polynomial basis 612 and pass into pretrained latent space model 622. In an aspect of the present disclosure, the latent space model is a Linear Discriminant Analysis (LDA) model. The system 100 transforms the cell absorbance signal and the media absorbance signalAttorney Docket No.: 1475-126 PCT into the polynomial basis 612 illustrated in FIG. 4, thereby generating linear terms, higher- order terms, and cross-terms that capture nonlinear spectral interactions between wavelength bands. The polynomial-basis features are then provided to the pretrained LDA model 622, which has been trained using classification labels derived from ground-truth culture measurements as described herein. The LDA model 622 projects the polynomial-basis features into a reduced- dimension, compressed latent space 630 that preserves discriminative spectral variance correlated with the target critical material attributes (CMAs) and cell quality attributes (CQAs). In further aspects, the same LDA projection may be applied to both aggregate-associated features and media-associated features, thereby enabling subsequent sparse regression on consistent latent components for real-time metabolite and cell-parameter quantification.

[0104] At block 712, the controller 200 causes the system 100 to generate metabolite predictions 650 via a regression model 640 through the use of a compressed latent space 630. In aspects of the present disclosure, the regression model is a sparse linear regression (SLR) 640. In aspects, the system 100 provides the LDA latent space signals 635 from the compressed latent space 630 (see FIG. 4) to the sparse linear regression (SLR) model 640. The SLR model 640 may be trained using Orthogonal Matching Pursuit to select a sparse subset of latent components that best correlate with ground-truth metabolite and / or cellparameter measurements. Using the sparse weights ® learned during training, the SLR model 640 computes predicted metabolite concentrations 650 according to Eqn. 2 (y = ©z), where z represents the latent space signal. In further aspects, the metabolite predictions 650 may include predicted critical material attributes (CMAs) of the cell media (e.g., glucose, lactate, fructose, or other metabolites) and / or predicted cell quality attributes (CQAs) of the cell bodyAttorney Docket No.: 1475-126 PCT aggregates (e g., viable cell density or other cell-state parameters), depending on whether the latent space signals are derived from the media absorbance signal or the cell absorbance signal. The controller 200 may execute the SLR inference in real time or near-real time as new hyperspectral images are acquired, thereby enabling continuous or periodic updating of metabolite predictions during the culture process.

[0105] At block 714, the controller 200 causes the system 100 to generate an attention map 660 for quality attributes of the cell body aggregates 515 and the cell media 520 through the absorbance signal based on the first cell absorbance signal and the second cell absorbance signal. In aspects, because the L-SLR framework is fully linear (see FIG. 4), the system 100 derives attention map weights from the trained LDA model 622 and the non-zero sparse regression coefficients of the SLR model 640. For each trained model, coefficients corresponding to the selected latent components are back-projected to identify contributions of individual SW1R wavelength bands and polynomial cross-terms to the metabolite predictions and / or quality-attribute estimates. The system 100 may compute an average absolute weight for each band across latent components, normalize the weights (with crossterms normalized separately from linear terms), and set weights below a threshold (e.g., 0.05) to zero, as described herein. The resulting attention map 660 thus provides a spatial and / or spectral visualization indicating which SWIR bands most strongly influence predicted cell quality attributes (CQAs) of the cell body aggregates and predicted critical material attributes (CMAs) of the cell media (see FIG. 9A-9C). In further aspects, the attention map 660 may be generated for each prediction timepoint and displayed to a user or provided to a controller to support interpretability, validation, and process decision-making.Attorney Docket No.: 1475-126 PCT

[0106] At block 716, the controller 200 causes the system 100 to determine, by machine learning model 500, metabolite predictions for the cell body aggregates 515 and the cell media 520. In aspects, the controller 200 provides to the machine learning model 500 the spectral features derived from the hyperspectral image data, including the first cell absorbance signal associated with the cell body aggregates and the second media absorbance signal associated with the cell media. The machine learning model 500 then performs inference to generate predicted metabolite concentrations and / or other culture- state parameters for both the aggregates and the media. In further aspects, the metabolite predictions produced at block 716 correspond to the metabolite predictions 650 generated via the sparse linear regression pathway described in blocks 710-712 and may be updated at each imaging timepoint to enable near-real-time tracking of culture conditions.

[0107] At block 718, the controller 200 causes the system 100 to determine, by machine learning model 500, cell quality attributes (CQA) of the cell body aggregates 515 and control material attributes (CMA) of the cell media 520. In aspects, based on the metabolite predictions and / or the attention map generated above, the machine learning model 500 determines one or more control -relevant CQAs of the cell body aggregates 515 and one or more control-relevant CMA of the cell media 520. The CQAs may include, for example, predicted viable cell density, aggregate growth phase indicators, or other aggregate-associated quality metrics, while the CMAs may include, for example, predicted concentrations of glucose, lactate, fructose, or other media constituents that influence culture performance. In further aspects, these determined cell quality attributes (CQA) and control material attributes (CMA) are used to assess deviation from target culture setpoints and to support generation of corrective control actions.Attorney Docket No.: 1475-126 PCT

[0108] At block 720, the controller 200 causes the system 100 to provide real-time cell culture control parameters. In aspects, the machine learning model 500 may provide the realtime cell culture control parameters. The controller 200 outputs real-time or near-real-time control parameters to the cell culture device 300 (e.g., bioreactor 130) based on the determined cell quality attributes (CQA) and control material attributes (CMA). The control parameters may include updated commands or setpoints for one or more operating variables of the cell culture device, such as nutrient feed rate, perfusion or sampling rate, buffer addition, pH adjustment, dissolved oxygen adjustment, temperature adjustment, agitation rate, or other process controls described herein. In further aspects, the machine learning model 500 may directly generate recommended control parameter adjustments, which the controller 200 then implements automatically or presents for operator approval, thereby enabling closed-loop or supervisory control of the biotherapeutic process.

[0109] Certain embodiments of the present disclosure may include some, all, or none of the above advantages and / or one or more other advantages readily apparent to those skilled in the art from the drawings, descriptions, and claims included herein. Moreover, while specific advantages have been enumerated above, the various embodiments of the present disclosure may include all, some, or none of the enumerated advantages and / or other advantages not specifically enumerated above.

[0110] The embodiments disclosed herein are examples of the disclosure and may be embodied in various forms. For instance, although certain embodiments herein are described as separate embodiments, each of the embodiments herein may be combined with one or more of the other embodiments herein. Specific structural and functional details disclosed herein are not to be interpreted as limiting, but as a basis for the claims and as a representative basis for teaching oneAttorney Docket No.: 1475-126 PCT skilled in the art to variously employ the present disclosure in virtually any appropriately detailed structure. Like reference numerals may refer to similar or identical elements throughout the description of the figures.[oni] The phrases “in an embodiment,” “in embodiments,” “in various embodiments,” “in some embodiments,” or “in other embodiments” may each refer to one or more of the same or different example embodiments provided in the present disclosure. A phrase in the form “A or B” means “(A), (B), or (A and B) .” A phrase in the form “at least one of A, B, or C” means “(A); (B); (C); (A and B); (A and C); (B and C); or (A, B, and C) ”

[0112] It should be understood that the foregoing description is only illustrative of the present disclosure. Various alternatives and modifications can be devised by those skilled in the art without departing from the disclosure. Accordingly, the present disclosure is intended to embrace all such alternatives, modifications, and variances. The embodiments described with reference to the attached drawing figures are presented only to demonstrate certain examples of the disclosure. Other elements, steps, methods, and techniques that are insubstantially different from those described above and / or in the appended claims are also intended to be within the scope of the disclosure.

Claims

Attorney Docket No.: 1475-126 PCTWHAT IS CLAIMED IS:

1. A non-contact system for monitoring biotherapeutic processes, comprising: a flow cell device configured to receive cell body aggregates and cell media from a cell culture device through an inline connector; a hyperspectral imaging system configured to capture a hyperspectral image of the cell body aggregates and the cell media; a processor; and a memory including instructions stored thereon which, when executed by the processor, cause the system to: generate the hyperspectral image of the cell body aggregates and the cell media; convert a first transmission signal of the cell body aggregate to a first cell absorbance signal from the hyperspectral image; convert a second transmission signal of the cell media aggregate to a second cell absorbance signal from the hyperspectral image; generate an attention map and metabolite predictions for quality attributes of the cell body aggregates and the cell media based on the first cell absorbance signal and the second cell absorbance signal; determine, by one or more machine learning models, cell quality attributes of the cell body aggregates and control material attributes of the cell media, based on the first and second cell absorbance signal; and provide bioreactor control parameters.

2. The system of claim 1, wherein the cell culture device may be a bioreactor or a shake flask.

3. The system of claim 1, wherein the cell body aggregates may be cancerous cells.

4. The system of claim 1, wherein the inline connector is a closed loop inline connector and is configured to preserve the sterility of the cell body aggregates and the cell media from the cellAttorney Docket No.: 1475-126 PCT culture device.

5. The system of claim 1, wherein the hyperspectral imaging system operates at least in a shortwave infrared wavelength range, and wherein the hyperspectral imaging device includes a reflective lens.

6. The system of claim 1, wherein the hyperspectral image is at a 20x magnification.

7. The system of claim 1, wherein the instructions, when executed by the processor, further cause the system to: convert the absorbance signal to a polynomial basis; pass the polynomial basis into a pretrained latent space model, wherein the latent space model forms a compressed latent space; generate metabolite predictions by a regression model through use of the compressed latent space; and determine coefficients from the latent space model and regression model.

8. The system of claim 1, wherein the generating of an attention map further includes: taking all coefficients corresponding to a non-zero sparse regression from a latent space model; averaging an absolute weight of each band across all latent space components; and normalizing the averaged absolute weight of each band across all latent space components.

9. The system of claim 8, wherein the attention map is formed based on weights assigned to a trained sparse linear regression to each band of input signal.

10. The system of claim 1, wherein the instructions, when executed by the processor, further cause the system to perform real-time monitoring of cell body aggregate and cell media parameter levels.Attorney Docket No.: 1475-126 PCT11. The system of claim 1, wherein the metabolite predictions and quality attributes are determined at least in part by providing the attention map to the one or more machine learning models.

12. The system of claim 1, wherein the bioreactor control parameters are provided in real time.

13. A processor-implemented method for monitoring biotherapeutic processes, the method comprising: generating a hyperspectral image of cell body aggregates and cell media in a microfluidic flow cell device using a hyperspectral imaging device; determining transmission signals from the hyperspectral image; converting the transmission signals to absorbance signals; generating an attention map for quality attributes of the cell body aggregates and the cell media, based on the first cell absorbance signal and the second cell absorbance signal; determining, by one or more machine learning models, metabolite predictions for the cell body aggregates and the cell media; determining, by the one or more machine learning models, cell quality attributes and control material attributes; and providing cell culture device control parameters.

14. The processor-implemented method of claim 13, further comprising pumping the cell body aggregates and the cell media into the microfluidic flow cell device from a cell culture device through an inline connector.

15. The processor-implemented method of claim 13, wherein the determining of transmission signals includes the transmission signal of the cell body aggregate and a transmission signal of the cell media from the hyperspectral image.

16. The processor-implemented method of claim 13, wherein the converting of transmissionAttorney Docket No.: 1475-126 PCT signals includes: converting the transmission signals of the cell body aggregate to a cell absorbance signal; and converting the transmission signal of the cell media to a media absorbance signal.

17. The processor-implemented method of claim 13, further comprising: converting the absorbance signal to a polynomial basis; passing the polynomial basis into a pretrained latent space model, wherein the latent space model forms a compressed latent space; generating metabolite predictions by a regression model through use of the compressed latent space; and determining coefficients from the latent space model and regression model.

18. The processor-implemented method of claim 13, wherein generating of the attention map further includes: taking all coefficients corresponding to a non-zero sparse regression from a latent space model; averaging an absolute weight of each band across all latent space components; and normalizing the averaged absolute weight of each band across all latent space components.

19. The processor-implemented method of claim 13, metabolite predictions and quality attributes are determined at least in part by providing the attention map to the one or more machine learning models.

20. The processor-implemented method of claim 13, wherein the inline connector is a closed loop inline connector and is configured to preserve the sterility of the cell body aggregates and the cell media from the cell culture device.Attorney Docket No.: 1475-126 PCT21 . The processor-implemented method of claim 13, wherein the hyperspectral imaging device operates at least in a short-wave infrared wavelength range, and. wherein the hyperspectral imaging device includes a reflective lens.

22. The processor-implemented method of claim 13, wherein the monitoring of biotherapeutic processes occurs in real time, and wherein the cell culture device control parameters are provided in real time.

Images