Scalable closed-loop BIO-manufacturing of tumor models or other volumetric biological samples

Abstract

Systems and methods are provided for automated incubation of large numbers of tumor organoids or other volumetric biological samples cultured from cells or other tissues. Label-free, non-destructive imaging of the samples is performed quickly and at high temporal resolution (e.g., once per day or more), allowing interventions (like the addition or subtraction of substances like water, glucose, buffers, or pharmaceuticals) to be applied to each sample based on the images, facilitating closed-loop per-sample control of sample growth. This can lead to improved incubation outcomes, like increased rates of generation of live samples and control of sample size or other sample properties. Image and other incubation data obtained via such automated incubation processes can be used to train machine learning models to plan interventions for the samples, further increasing incubation success rates and allowing for the incubation of rare and / or novel samples, for which reliable incubation protocols do not yet exist.

Description

SCALABLE CLOSED-LOOP BIO-MANUFACTURING OF TUMOR MODELS OR OTHER VOLUMETRIC BIOLOGICAL SAMPLESCROSS-REFERENCE TO RELATED APPLICATION

[0001] The present application claims priority to U.S. Provisional Patent Application No. 63 / 674,595 filed July 23, 2024, the entire contents of which are incorporated by reference into the present application.BACKGROUND

[0002] Tumor organoids, spheroids, three-dimensional multi-cellular cell or tissue samples, or other volumetric biological samples can be incubated or otherwise grown (e.g., in multi-well sample plates) to provide large numbers of volumetric biological samples that exhibit behaviors that mimic the behavior of ‘real’ tumors or other structures of interest. Such complex three-dimensional samples, which may incorporate multiple cell types / cell lines and which may exhibit structures and inter-cellular interactions that more closely recapitulate ‘real’ tumors (relative to, e.g., two-dimensional cell cultures), can provide significant benefits in the investigation of disease progression, evaluation of candidate therapies, or other scientific or therapeutic investigations. These benefits can also be obtained within the expense and variability that is associated with animal models, which may exhibit even more complex three- dimensional structure and interactive behavior with surrounding tissues (e.g., angiogenesis, inhibition of metastases) but which also present difficulties in mapping observed behaviors in animal physiology to human physiology.

[0003] However, conventional incubation of such three-dimensional tumor models or other organoids leads to highly variable results, with large numbers of ‘dead’ or otherwise failed samples and / or an unacceptably wide ranges of sizes or other properties amongst the ‘live’ samples. This variability in the generated samples may be related to a lack of feedback about the progress of the samples over time, resulting in the incubation being fully “open loop” and failing to account for the actual progress of the individual samples over time. While it is possible to stain and image the samples to provide such feedback, staining and / or imaging can perturb and / or destroy the samples. A subset of samples in a plate could be, e.g., stained and / or destructively imaged to provide feedback about the development of other samples in a plate. However, the inherent variability of behavior across biological samples makes it difficult to apply such information between samples of a plate. This uncertainly makes it difficult to translate such information into specific interventions (e.g., application of growth factors) tocontrol the incubation of the non-imaged samples in the plate.

[0004] These limitations also make highly difficult to consistently culture samples from limited cell sources (e.g., cells obtained from a biopsy of a particular patient). While ‘generic’ incubation parameters can be applied to novel samples, the variability of response across patients and cell types means that, for many individuals, no cells may be successfully cultured before the initial sample is exhausted. This limitation makes rare diseases or cancer types particularly difficult to investigate, since it is first necessary to use limited cell samples to develop cell culture protocols to enable further research.SUMMARY

[0005] In a first aspect, a computer-implemented method for manufacturing 3D biological structures that includes: (i) dispensing, by a laboratory system, a plurality of volumetric biological samples into culture medium in respective cell culture containers; (ii) transporting, by the laboratory system, the cell culture containers to an incubator; (iii) incubating the plurality of volumetric biological samples in the incubator; (iv) subsequent to incubating the plurality of volumetric biological samples in the incubator, using the laboratory system to non-destructively image the plurality of volumetric biological samples to generate first imaging data therefor, wherein the first imaging data represents at least one of physical or chemical information about the plurality of volumetric biological samples; and (v) storing the imaging data in a non-transitory computer-readable medium.

[0006] In another aspect, a non-transitory computer readable medium is provided having stored thereon program instructions executable by at least one processor to cause the at least one processor to perform any of the above methods.

[0007] In another aspect a system is provided that includes: (i) at least one processor; and (ii) a non-transitory computer-readable medium, having stored therein instructions executable by the at least one processor to cause the system to perform any of the above methods.

[0008] These as well as other aspects, advantages, and alternatives will become apparent to those of ordinary skill in the art by reading the following detailed description with reference where appropriate to the accompanying drawings. Further, it should be understood that the description provided in this summary section and elsewhere in this document is intended to illustrate the claimed subject matter by way of example and not by way of limitationBRIEF DESCRIPTION OF THE FIGURES

[0009] Figure 1 illustrates aspects of a system, according to an example embodiment.

[0010] Figure 2 illustrates a flowchart of an example machine learning model training and inference process, according to an example embodiment.

[0011] Figure 3 illustrates aspects of a system, according to an example embodiment.

[0012] Figure 4 illustrates aspects of a method, according to an example embodiment.

[0013] Figure 5 illustrates aspects of a system, according to an example embodiment.

[0014] Figure 6 illustrates aspects of a system, according to an example embodiment.

[0015] Figure 7 illustrates aspects of a system, according to an example embodiment.

[0016] Figure 8 illustrates aspects of a system, according to an example embodiment.

[0017] Figure 9A illustrates aspects of a method, according to an example embodiment.

[0018] Figure 9B illustrates aspects of a method, according to an example embodiment.

[0019] Figure 10 illustrates aspects of a method, according to an example embodiment.

[0020] Figure 11 illustrates aspects of a method, according to an example embodiment.

[0021] Figure 12 illustrates aspects of a system, according to an example embodiment.

[0022] Figure 13 illustrates a flowchart of an example method.DETAILED DESCRIPTION

[0023] The following detailed description describes various features and functions of the disclosed embodiments with reference to the accompanying figures. The illustrative embodiments described herein are not meant to be limiting. It may be readily understood that certain aspects of the disclosed embodiments can be arranged and combined in a wide variety of different configurations, all of which are contemplated herein.

[0024] To that end, example methods, devices, and systems are described herein. It should be understood that the words “example” and “exemplary” are used herein to mean “serving as an example, instance, or illustration.” Any embodiment or feature described herein as being an “example” or “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments or features unless stated as such. Thus, other embodiments can be utilized and other changes can be made without departing from the scope of the subject matter presented herein.

[0025] Accordingly, the example embodiments described herein are not meant to be limiting. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the figures, can be arranged, substituted, combined, separated, and designed in a wide variety of different configurations. For example, the separation of features into “client” and “server” components may occur in a number of ways.

[0026] Further, unless context suggests otherwise, the features illustrated in each of thefigures may be used in combination with one another. Thus, the figures should be generally viewed as component aspects of one or more overall embodiments, with the understanding that not all illustrated features are necessary for each embodiment.

[0027] Additionally, any enumeration of elements, blocks, or steps in this specification or the claims is for purposes of clarity. Thus, such enumeration should not be interpreted to require or imply that these elements, blocks, or steps adhere to a particular arrangement or are carried out in a particular order.

[0028] Unless clearly indicated otherwise herein, the term “or” is to be interpreted as the inclusive disjunction. For example, the phrase “A, B, or C” is true if any one or more of the arguments A, B, C are true, and is only false if all of A, B, and C are falseI. Overview

[0029] It is desirable to generate, in a short amount of time, at low cost, and with low amounts of manual human effort, large numbers of tumor organoids or other three- dimensionally complex biological cell and tissue samples, which may be referred to as “volumetric biological samples.” It is also desirable to generate such volumetric biological samples such that their size, geometry, cell composition, or other properties match a specification (e.g., to reduce variability across samples when using the sample to investigate an experimental intervention) and that exhibit reduced rates of sample failure or death or other undesirable sample end states. Increasing the number of available volumetric biological samples of this type allows more experimental investigations using the samples to be made in less time. Additionally, the ability to control the size or other properties of the samples allows sources of variability other than the controlled properties (e.g., applied pharmaceutical agents) to be more directly assessed without the confounding effects of variation in the controlled properties. Alternatively, the values of such properties could be controlled across a range of values to allow the interaction of such controlled properties with an experimental condition or intervention to be assessed, e.g., to directly investigate the effect of tumor organoid size on the efficacy of a candidate therapeutic.

[0030] However, it is difficult to culture large numbers of tumor organoids, complex three-dimensional cell or tissue samples, or other volumetric biological samples or to exert control over the size or other properties of such cultured samples. In particular, it has previously been very difficult to consistently produce volumetric biological samples from specific cells or sources (e.g., individual patients) and protocols available for existing sample lines cannot always be easily adapted to novel sample lines (e.g., cells from a new individual patient). Inpractice, tumor organoids are often cultured in an “open-loop” manner, with a large number of volumetric biological samples initially seeded in respective wells of a multi-well sample plate and then incubated together (optionally with other multi-well sample plates) in an incubator. Interventions (e.g., the addition of water, sugars or other nutrients, buffers, or other substances) are then provided in common to all of the samples of the plate according to a pre-determined schedule, with the particulars of the pre-determined schedule laboriously confirmed via repeated experiments for a particular sample type (e.g., a particular type of cancer cell). This leads to significant between-sample variability in outcome (e.g., with respect to size, viability, or other properties). Additionally, this makes it difficult to culture organoids or other volumetric biological samples of rare cell types or conditions (e.g., rare cancer types), since significant source amounts of the target cells are needed in order to generate enough samples to investigate the large space of possible intervention schedules in order to identify a schedule that results in significant numbers of patent samples.

[0031] The embodiments described herein overcome various shortcomings of the prior with respect to generating, at scale, large numbers of tumor organoids or other volumetric biological samples. These shortcomings include:• Generation of samples in small batches, resulting in correspondingly small amounts of data about the cell type or other biological system being cultured;• Lack of process monitoring, leading to lack of information about the growth and evolution of the samples;• Lack of quality control methods or measurements for the biological samples;• Lack of software or analytical methods that can predict sample growth;• A lack of automation, relying instead on manual effort, resulting in increased time and costs;• Low probability of culturing cells, using the same reagents, from different origins (e.g. different patients or different types or sub-types of cancer);• High inter-batch variability; and• Sensitivity to specific initial growth conditions and the idiosyncrasies of specific, often bespoke incubation equipment, leading to significant variability between incubation sites (e.g., between different labs attempting to implement the same incubation protocol of the same type of cell or tissue).

[0032] The embodiments described herein overcome these limitations by making extensive use of automation, software, and improved, fast, and non-destructive imaging tofacilitate closed-loop, image-guided sample incubation on a per-sample basis, using trained models or other methods to generate and deliver targeted per-sample interventions, thereby controlling the growth of volumetric biological samples in an individualized manner. This also facilitates standardization (since the sample seeding, applied interventions, and other aspects of sample incubation can be performed in an automated, instrumented, and sensor-verified consistent manner, in contrast to the variability inherent in manual sample preparation), leading to more consistent and controllable incubation of samples. Additionally, automation allows for the generation of a larger quantity of volumetric biological samples, which, in combination with imaging, allows a desired number and quality of samples to be more easily and consistently generated by generating an excess or otherwise increased number of samples and then selecting, from the samples, a subset that best satisfy specified quality standards or other constraints.

[0033] This can include using label-free, non-destructive, and short-duration microscopic sample imaging, intervention delivery, sample seeding, and sample handling in an automated laboratory context to generate repeated, per-sample image and other data at high temporal resolution. The embodiments described herein may be implemented as a complex “cyber-physical” system that integrates the imaging and automated laboratory systems into a multi-layer information architecture, allowing ‘raw’ image or other data at lower layers to inform more complex data analysis / synthesis, sample growth prediction, intervention planning, model training / updating, automated experimentation (e.g., via deep or other varieties of reinforcement learning), large scale sample incubation, novel sample investigation and optimization, and other operations at higher levels of the cyber-physical system stack. Data flow in such a cyber-physical stack can be bidirectional, with lower levels transmitting raw or processed image data, sample status or other sample information, the results of commands or other requests, records of past per-sample image, intervention, and / or incubation parameters, or other information and with higher levels transmitting commands, requests for imaging, status updates, or other information to lower levels.

[0034] For example, higher levels of such a cyber-physical information architecture can provide more general commands or information (e.g., commands to incubate a desired number of volumetric biological samples from a specified source of tissue sample(s) that represent a novel type of cancer cells) that are, via processing in intermediate levels (e.g., development of a plan of initial sample seeding and intervention and incubation parameter investigation, a request to retrain or fine-tune a machine learning model based on a specified set of data from lower level(s), a request for additional or updated imaging or other informationabout one or more volumetric biological samples), rendered into specific operations performed by imaging systems, sample handling systems, or other physical devices of one or more automated laboratory systems (e.g., to image a particular volumetric biological sample in a specified well of a specified multi-well sample plate according to specified imaging parameters, to provide a specified intervention to a specified well of a specified multi-well sample plate, to adjust a temperature, humidity, or other operating parameter of an incubator, to slice or otherwise prepare a specified input tissue or cell sample in a specified many to seed a specified well of a specified multi-well sample plate). Lower levels of such a cyber-physical information architecture can provide ‘raw’ data (e.g., unprocessed images of volumetric biological samples, temperatures, humidities, or other measured and / or commanded properties of an incubator, measured properties of reagents or growth media, confirmation of the performance of a specified intervention or method of sample seeding) that are, via processing in intermediate levels (e.g., image processing of images, generation of mode-switched images from a ‘raw’ source imaging modality to a simulated imaging modality, projection of image, intervention, tissue source, or other data into multi-dimensional vector embeddings or other high-level representations of volumetric biological samples, prediction of the trajectory of future growth of a volumetric biological sample), synthesized or otherwise transformed into high-level plans, objectives, or operations generated by high-level layers (e.g., a plan of investigative interventions to develop optimized incubation strategies for a novel cell or tissue type, a plan of interventions to generate a population of volumetric biological samples that satisfy a specified distribution).

[0035] Such per-sample, high temporal resolution data allows interventions (e.g., addition or subtraction of buffers, water, nutrients, culture medium, pharmaceuticals, or other substances, application of light, sound, heat, or other energies, adjustment of pH) to be determined in a per-sample manner. This contrasts with alternative imaging schemes like the use of stains and / or destructive imaging to generate image data for a single sacrificial sample of a multi -well sample plate in order to imperfectly infer the state of other volumetric biological samples on the plate. Additionally, the use of an automated laboratory system allows the determined per-sample interventions to be quickly applied to the corresponding wells of a multi -we 11 sample plate and for the label-free non-destructive imaging of each of the volumetric biological samples to be performed quickly, reducing the amount of time spent by the volumetric biological samples outside of the controlled environment of an incubator.

[0036] By facilitating repeated non-destructive and label-free imaging of many volumetric biological samples, the embodiments herein also make possible the consistentculturing of novel and / or rare samples. This is because the image data provides ongoing feedback to modify, over the course of incubation, interventions applied to the novel samples, reducing the chance that each sample dies or otherwise becomes non- viable. Additionally, by granting access to image and intervention data across time for large numbers of conventional samples, the embodiments described herein make it possible to train machine learning models to generate improved interventions, even in examples wherein a cell or sample type of interest is novel (e.g., wherein the model was trained using image and intervention data from cell or sample types other than the novel type).

[0037] For example, a deep reinforcement learning agent could be trained using large numbers of ‘conventional’ samples (for which reliable, pre-specified incubation protocols may have already been determined), resulting in an agent that ‘knows’ about the underlying behaviors of various cell types and how to control the incubation thereof by applying various interventions thereto based on non-destructive label-free images thereof. Such a trained agent could then be used to determine interventions for a small number of samples of a novel cell type . Since the agent has been pre-trained to have expertise in image-guided incubation of other types of samples, image-guided interventions generated by the agent for the novel samples can leverage the model’s ‘knowledge’ to predict interventions that are more likely to improve sample outcomes (e.g., reduce sample death or failure, improve sample consistency with desired size or other sample properties).

[0038] Such functionality could make possible research into rare diseases (e.g., rare cancer types) by making it possible to consistently incubate large numbers of samples for the rare diseases using small amounts of source material (e.g., biopsied tissue as a source of cells for incubation) or where the maximum available amount of source material may be limited (e.g. from a single patient or small groups of patients). This functionality could also facilitate research by reducing the amount of experimentation needed to develop incubation protocols for such novel sample types, since the model’s ‘knowledge’ about the behavior of other cell types can be readily applied to the novel sample type. Indeed, such functionality can enable large numbers of patent-specific samples to be consistently incubated from small amounts of source material from a single patient, allowing many different potential therapies to be investigated in vitro in order to objectively select an ‘optimal’ treatment for the patient. This can improve patient outcomes by, e.g., avoiding or reducing the need to investigate the relative efficacy of various different treatments for the patient in series and avoiding the provision of non-efficacious therapies that may be associated with side effects and / or undesirable delay before providing more efficacious treatment(s).

[0039] Additionally, the embodiments described herein can lead to improved sample incubation and subsequent use of the samples by providing, for each sample, a sample-specific history based on the many non-destructive label-free images taken thereof during incubation. Such high-resolution per-sample data about the incubation process could facilitate experimentation by allowing variability in the effects of various interventions or experimental conditions to be correlated with differences in the samples both with respect to their properties at the end of incubation (e.g., size, morphology) and during incubation (e.g., trajectory of growth, response to various applied interventions). Such data can also facilitate investigation into the underlying properties of the samples, e.g., to identify the presence of different cell or behavior sub-types within a set of samples and / or to quantify the influence of manufacturing conditions on the subsequent behavior of the samples (e.g. in response to drugs).

[0040] Using the per-sample non-destructive imaging data and / or other measurements to improve or otherwise control the incubation of volumetric biological samples can include optimizing or otherwise improving the cell culture matrix and / or media used to incubate volumetric biological samples (and / or to initially incubate seeds that are then transplanted into different matrix / medium for further incubation into volumetric biological samples of interest). The matrix and growth media for incubating the volumetric biological samples may be controlled in its chemical composition, which could be altered in time and / or between samples. Chemical interventions include the chemical species that originally comprise the growth media and chemical species that are later added to the growth media. The matrix and media can be different at the seed and growth stages. The matrix can consist of a cross-linked polymer derived from natural or synthetic sources. The matrix may be functionalized by adding chemical species to enhance the adhesion of cells, to cause degradation of matrix in response to specific chemical or physical cues (e.g., chemical species added to facilitate removal of sample material from the matrix), or to enhance the production of specific molecules from cells (e.g., collagen). To facilitate controlled incubation of samples, the systems described herein can be configured to accurately measure and / or control the media composition in which the volumetric biological samples are incubated. This composition may include a buffer to maintain pH and promote cell growth, a source of carbon for cell growth (such as Glucose), a source of nitrogen to provide amino acids that can aid protein synthesis, and inorganic salts to maintain osmotic pressure (such as sodium, potassium, calcium, magnesium, and phosphorus). Additional components may include those that support cellular functions such as vitamins, serum (e.g. fetal bovine serum) or serum -free alternatives (e.g. including growth factors, hormones, and other proteins) to stimulate cell growth and differentiation, or antibiotics orantifungals to prevent microbial contamination. Essential trace elements (like zinc, selenium, copper) may be added.

[0041] A design of experiments method (e.g., using one or more trained machine learning models) can be applied to specify the precise composition of the matrix and the media for growth. Such a method could include specifying combinations of components for testing in order to optimize or otherwise improved one or more attributes of the volumetric biological samples. Statistical techniques can be used to obtain the desired attributes by systematically varying the composition of media from amongst the controlled factors and observing their impact on sample attributes. The most influential factors and beneficial concentrations thereof can be identified, leading to efficient and improved methods to achieve the desired attributes of the volumetric biological samples. The matrix and growth media may be determined by screening experimental designs (which can identify key components) or optimization experimental designs (to fine-tune known compositions to achieve a specified outcome). The attributes for optimization may include controlled sample size, composition, or rate of proliferation.

[0042] A system as described herein may be used to optimize or otherwise improve the composition of a matrix and / or medium in a series of iterative steps. In a first step, a first matrix is used and one or more media compositions are used. The volumetric biological samples are grown using these, and assessed for physical and chemical attributes, biological function, and other characteristics that may be measured or calculated. In a second step, changes are made to the matrix and / or the media, after which further volumetric biological samples are grown and assessed. Such iterations may continue until matrix and / or medium compositions are found that result in desired growth conditions. At each iteration, the data collected may inform methods to predict, improve, or otherwise optimize the growth conditions, such as statistical methods or machine learning methods.

[0043] In some examples, volumetric biological samples could be incubated sequentially in two (or more) different culture media. For example, a first well of a multi-well sample plate could contain a first culture medium and could have disposed therein a seed (e.g., a specified number or amount of one or more specified types of cells, e.g., according to a spherical, shell, or other geometry and / or a slice or other contiguous portion of a tissue sample). After a period of incubation in the first well, the volumetric biological sample grown from the seed could then be transplanted from the first well into a second well (e.g., of a different multiwell sample plate) having disposed therein a second culture medium. Such a scenario could allow the volumetric biological sample incubated thereby to be improved in some manner (e.g.,with respect to likelihood of survival and / or correspondence with one or more constraints or criteria). For example, the first culture medium could be specified to enhance cell health and / or growth (e.g., to increase the likelihood that the seeded materials successfully grow into a volumetric biological sample) while the second culture medium could be specified to enhance correspondence with a specified experimental condition (e.g., to known particulars of a target tissue or other biological environment of interest), to enhance cell health and / or growth for larger volumetric biological samples (e.g., to support out-diffusion of cellular wastes, where such is less important for the growth of smaller samples), to prevent sample growth in undesired ways (e.g., to inhibit growth sufficient to prevent change in sample geometry from an initial spherical or other starting geometry, to inhibit cell mobility to prevent large-scale changes in staring sample geometry), or specified to address some other consideration.

[0044] Figure 1 depicts aspects of an example system 100 that includes and / or controls one or more automated laboratories as described herein. Note that an automated laboratory system or other type of system (e.g., a computing system configured to manage one or more automated laboratory systems) as described herein may include more or fewer elements (e.g., may lack an automated sample seeding system, may include additional sample imaging systems). The system 100 includes a seeding system 110 configured to seed samples in wells of multi-well sample plates (or “plates”), a robotic sample handling system 120 configured to move plates between elements of the system 100, a label-free non-destructive imaging system 130, a controller 140 configured to operate the elements of the system 100, an incubator 150 configured to control a temperature, humidity, or other environmental properties therein to incubate samples in wells of plates disposed therein, an intervention delivery system 160 configured to apply interventions to wells of plates, and an additional imaging system 170.

[0045] The robotic sample handling system 120 could include one or more robotic arms, gantries, conveyors, or other elements configured to deliver plates from one element of the system to another. The robotic sample handling system 120 could be configured to move only a single plate at a time, or could be configured to move multiple plates at a time. In some examples, the robotic sample handling system 120 and / or other components (e.g., 110, 130, 160, 170) could include queuing sub-systems (e.g., conveyor belts) to allow multiple plates to be operated on (e.g., imaged, subjected to interventions) without the intervention of single robotic arm or other plate-manipulating system of the robotic sample handling system 120. For example, such a robotic manipulator arm could move multiple plates from the incubator 150 to a conveyor of the imaging system 130, allowing the imaging system 130 to image all of the multiple plates without further intervention from the manipulator arm. The manipulator armcould then, when available, move imaged plates from the imaging system 130 back to the incubator 150 or to some other component of the system 100. Such an arrangement, which allows for ‘asynchronous’ use of manipulator arms or other plate -manipulating system of the robotic sample handling system 120, can increase throughput of the system 100 while reducing the number of complex manipulator arms or other plate-manipulating systems present in the robotic sample handling system 120.

[0046] The seeding system 110 can include robotic manipulators, robotic injectors (e.g., multi-well injectors), or other plate- or we 11 -manipulating systems to facilitate the seeding of cells or other tissue samples into wells of plates in order to begin the sample -incubating process. This can include injecting cell culture medium in order to control a location or geometry of cell(s) seeded into the well. For example, a first amount of culture medium could be deposited into a well, and then an amount of cells or other tissue sample material could be deposited thereon and / or therein. Additional cell culture medium could then optionally be added into the well, with the particulars of each deposition being specified in order to result in a desired initial state for the sample in the well (e.g., a desired total amount of culture medium, a desired distance of the sample from the surface of the culture medium, a desired geometry of the deposited cells). In some examples, multiple different types of cells could be deposited to seed a sample, e.g., a first type of cells in the center of a spherical geometry, and a second type of cells disposed in a shell around the first type of cells. Additive manufacturing techniques (e.g., three-dimensional controlled extrusion of cell-bearing fluid, cell culture medium, extracellular matrix components, and / or other substances) could be applied to seed samples having specified geometries or other properties in wells of a plate.

[0047] Cell samples used by the seeding system 110 to seed wells of a plate could be acquired from a variety of sources and pre-processed via a variety of methods. For example, cell samples could be acquired from a biopsy or other tissue sample and, partially digested, sorted (e.g., via centrifuge, via cell sorting), or otherwise processed to provide a source of cells that can be deposited into wells of a plate (e.g., onto or into a volume of cell culture medium, with a specified accuracy of, e.g., 100 microns or better in all three dimensions) in order to seed a sample of the cells in the well. The source of cells could be a single source (e.g., a source of sorted cells of a single types) or a specified mixture of sources and / or cell types. In some examples, a biopsy or other sample having a three-dimensional structure (e.g., a tumor) could be sliced and one or more of the individual slices (or portions thereof), with the two- dimensional and / or partial three-dimensional structure thereof relatively intact, could be deposited in a well of a plate to seed a sample therein. Pre-processing steps (e.g., sampledigestion, cell sorting, suspension of cells in a delivery medium) could be performed by components of the seeding system 110 and / or by other systems and the results of such processing provided to the seeding system 110.

[0048] The seed used to generate a volumetric biological sample may be in the form of cells or contiguous portions of tissue (e.g., slices), or a combination of cells and intact, contiguous tissue portions. The cells may be of one type or multiple types. The number of cells in the seed, and the number of each type of cells in the seed, may be controlled using various methods. For example, the cells may pass through a cell counter as they are deposited into a seed and / or sample container. The size of the seed and number of cells in the seed may be measured with optical imaging (e.g., scattering or fluorescence measurements). Imaging may additionally or alternatively be used to determine the type of cells in the seed. The use of the systems and methods described herein can lead to enhanced consistency of seeds generated thereby. For example, the number of each type of cells in a seed can be controlled to within a target size within a pre-determined tolerance or margin of error.

[0049] Such a seed may, in whole or in part, take the form of a tissue sample and / or a contiguous portion of such a tissue sample. Contiguous portions of a tissue sample used to seed a volumetric biological sample as described herein could be obtained by slicing, sectioning, cutting, chemically digesting, and / or mincing a larger tissue sample. Such a tissue sample- derived seed may be a single sample extracted from a larger tissue sample or it may be multiple tissue samples extracted from a larger tissue sample (e.g., with multiple different seeds, of respective different volumetric biological samples incubated therefrom, generated from the multiple extracted tissue samples). Multiple different sectioning or other sample -preparation techniques may be applied to a tissue sample or portion thereof to render a contiguous portion of the tissue sample that is then used, optionally in combination with other materials, to seed a volumetric tissue sample. For example, a tissue sample could be cut (e.g., to isolate a tumor or other portion thereof containing cell and / or tissue of interest) and then sliced to render a slice that is used to seed a volumetric tissue sample. Such a slice may be used as a seed without additional preparation, or some additional processing may be performed (e.g., additional cutting to isolate specific tissue or cells of interest of the initial slice) prior to using the processed slice to seed the incubation of a volumetric biological sample.

[0050] Such a contiguous portion of a tissue sample or other discrete tissue-derived source of sample material may be measured using optical imaging techniques before it, or portions of it (e.g., a slice), is seeded into a well or embedded in a gel matrix. Such measurements may determine the physical or chemical attributes of the seed, and maydetermine the number of cells, types of cells, size and shape of the cells, and spatial arrangement of the cells in the seed. The measured data and information may be stored for later use and may inform calculations or predictions of the volumetric biological sample that is grown from the seed.

[0051] Such a seeding process could, itself, result in the generation of imaging or other information that is later used to, e.g., predict the growth of volumetric biological samples generated therefrom, to generate a plan of interventions to such samples to obtain a desired pattern of growth, to train a model to perform growth prediction or intervention generation, or to perform some other operations. For example, an explanted tumor or other tissue sample could be imaged before and after slicing, after seeding the slices (whole, or dissociated) in growth medium in wells of a multi-well samples plate, and / or at other points in time, and such image information can then be associated with the volumetric biological samples(s) generated from such slices such that the associated image data can be used to predict alternate -modality image data (e.g., to predict stained-sample images from near-infrared spectroscopic tomography (NIRST) images or other label-free imagery), to predict a trajectory of future growth of a sample, to determine intervention(s) to apply to the sample, to determine a pattern of interventions to apply to a population of samples in order to learn about the incubation dynamics of the population, to update a model to generate such information, to evaluate whether a sample satisfies specified sample quality parameters, or to provide some other benefit or output.

[0052] NIRST, as an example of one type of imaging that could be employed as part of the embodiments described herein, combines the structural imaging capability of optical tomography (e.g., optical coherence tomography, diffractive, tomography) with the chemical specificity of near-infrared and mid-infrared spectroscopy. While the spectroscopic molecular detection capabilities of mid-IR spectroscopy are significant due to the strongly absorbing fundamental vibrational modes of molecules, the large absorption cross section of water and most organic materials precludes depth profiling beyond 1-10 micron. Near infrared spectra consist of overtones and combinations of these fundamental absorption modes. Thus, absorbance is -10-100 times weaker, allowing for 10-100-fold greater depth of penetration. A database of mid-IR frequencies important for cell culture is identified as optical biomarkers. Near-IR frequencies indicative of the mid-IR biomarkers, as well as those NIR spectral regions that are unaffected by cellular absorbance, are subsequently identified. The non-absorbing frequencies can then be used to build structural images of the samples. A variety of methods may be used to generate such imagery based on the phase retardation or scattering of light. Oneembodiment includes using optical coherence tomography (OCT) to generate such imagery with a continuous source being used as the “probe beam.” The sample is also illuminated with the selected NIR and / or mid-IR frequencies, modulated at frequencies sufficient to allow absorption and relaxation (~ms time scale). The absorption of these frequencies (which may be referred to as the “pump beam”) causes a transient rise in temperature. The pump and probe beams are focused into the sample where the transient absorption causes a thermally-modulated local expansion, resulting in the formation of a local microlens that alters the propagation of the probe laser and causes a change in the detected intensity or phase. The difference between the excited and resting structural images provides a measure of absorbance and, thus, chemical properties of the sample. This configuration allows for chemically sensitive optical imaging of thick sections, with up to 1 mm depth.

[0053] In some examples, whole-sample spectroscopy could be used to measure the overall composition of entire samples and / or to provide some other benefit. For example, quantitative similarity between the whole-sample spectra of different samples could be used to determine a degree of similarity between such samples without imposition of additional assumptions regarding the ‘meaning’ of the spectral information. For example, while such spectral information (optionally at higher resolution than the whole-sample level) could be used to estimate the concentration and distribution of various metabolites or other chemical species, and those estimates could then be compared in order to assess similarity between different samples, such predictions or other post-processing of the spectral information can abolish certain similarity-indicating information within the spectra and / or over-emphasize specific spectrally-derived predictions (e.g., specific chemical species) when evaluating the overall similarity between samples. Instead, the use of whole-sample spectra to assess the overall level of quantitative similarity between samples can reduce such bias, retain the entire set of spectral information for comparison, and also reduce the computational cost in performing such sample- to-sample similarity determinations. Such similarity determinations could be used, e.g., to select a subset of ‘most similar’ samples of a set of successfully incubated volumetric biological samples for retention and use in downstream applications (e.g., to reduce incubation- related variability in the measured outcome(s) of an experiment performed using a set of volumetric biological samples generated via the embodiments described herein). Additionally, such whole-sample spectra can be generated much more quickly (and potentially using less excitation light, reducing the amount of photobleaching or other unwanted damage to the sample) than higher-resolution spectral imaging.

[0054] The incubator 150 could include heaters, humidifiers, dehumidifiers, lightsources, or other elements to control a temperature, humidity, light level, or other environmental properties therein. The incubator 150 could include a single controlled- environment volume or multiple such volumes. Where multiple incubation volumes are available, the environmental conditions therein could differ, e.g., to facilitate the use of changes in such environmental conditions as interventions to control the growth of samples by controlling which volume the plates containing the samples are placed in.

[0055] The intervention delivery system 160 could include robotic injectors, sprayers, feedstock reservoirs, or other elements configured to allow specified interventions to be applied, on a per-well basis, to the wells of plates contained in the system 100. Such interventions can include providing / subtracting substances or energies, e.g., providing a pharmaceutical, cell culture medium, buffer, acid, base, cell product, virus, DNA, RNA, protein, water, solvent, or other substance, removing an amount of cell culture medium, heating, illuminating with a specified amplitude or spectrum of light, applying an electric field, or interacting with a sample in some other way. Additionally or alternatively, an intervention could include effecting a specified change in a property of the sample, e.g., causing the pH of the sample to match a specified value or causing the temperature of the sample to match a specified value. In some examples, an intervention could include changing a property of the incubation of the sample (e.g., changing a temperature, light level, humidity, pressure, gas composition, or other parameter of incubation); in such examples, providing such an intervention could include changing the operation of a controllable volume of the incubator 150 and / or controlling the robotic sample handling system 120 to place the plate containing the sample into a volume of the incubator 150 whose environmental properties match the specified intervention.

[0056] The imaging system 130 could include a variety of different systems for monochromatically, hyperspectrally, white light, or otherwise non-destructively imaging samples in wells of plates in a label-free manner. This could include imaging the samples with infrared, near infrared, visible, and / or ultraviolet light via bright field, confocal, stimulated Raman scattering microscopy (SRSM), near infrared spectroscopic tomography (NIRST), or other imaging methods. The imaging data generated could include images of the whole well or of portions of a well (e.g., a specified target volume, a slice, a point) at various spatial and / or spectral resolutions.

[0057] The imaging modalities used by the imaging system 130 could also exhibit reduced image capture times, to allow many wells to be imaged at high temporal resolution (e.g., each well of each plate in the system 100 being imaged at a rate of once or more per day).Desired chemical, spectral, or other image data could be obtained in such a reduced-time manner by employing multiple imaging modalities, which may have respective different strengths with respect to resolving power, spatial resolution, depth of field, spectral resolution, ability to detect geometry, or other properties. The image data output from the imaging system 130 for such two (or more) different imaging modalities could then be used (e.g., by the controller 140 applying a machine learning model thereto) to generate improved composite image data for a sample.

[0058] For example, bright-field imaging could be used to generate high spatial resolution morphologic and structural information about a sample while NIRST or some other spectrographic imaging modality is used to generate higher chemical or molecular detail but similar or lower spatial resolution spectroscopic information for the sample. A trained machine learning model could then be used to generate an output image based on the two (or more) input imaging modalities (e.g., a higher-resolution spectrographic image of the sample that exhibits the higher spatial resolution of the brightfield image and the higher spectral resolution of the NIRST image data). Such a model could also be trained to predict chemical distribution, optical scattering / absorption, or other image data generated by more time-intensive, more destructive, and / or label-based imaging techniques. For example, a model could be trained to predict, from two or more input images of a sample (e.g., a brightfield image and NIRST image data), a confocal or other type of volumetric image of a the sample as though it had been stained with one or more stains, a hyperspectral image of the sample as though it had been hyperspectrally imaged at a higher spatial and / or spectral resolution, a hyperspectral image of the sample as though it had been imaged using SRSM or some other more time-intensive and / or destructive hyperspectral imaging modality, or some other image data of interest.

[0059] Such an image predictive machine learning model could be trained using sets of training data obtained via some other means, e.g., a library of sets of corresponding brightfield, full spectral / spatial resolution, and reduced spectral / spatial resolution images of samples, or a library of sets of brightfield, hyperspectral, and stained or otherwise labeled images of samples. Additionally or alternatively, the system 100 could be used to generate such training datasets and / or to update such pre-trained models. For example, the system 100 could include an auxiliary imaging system 170 that is capable of generating image data that is improved, relative to the image data generated by the imaging system 130, with respect to spatial resolution, hyperspectral resolution, the ability to resolve chemical or other features of interest within the samples while also being worse with respect to per-sample image time, level of damage or perturbation to the sample, the additional of stains or other labeling substances to the sample,or some other factor that might limit its use in non-destructively imaging samples contained in the system 100 at a high temporal resolution and in a label-free manner. The controller 140 could operate the imaging systems 130, 170 and the robotic sample handling system 120 to occasionally image samples using both of the imaging systems 130, 170 (which may include using the robotic sample handling system 120 and / or intervention delivery system 160 to add stains or other labels thereto) to add training examples to a dataset used to train and / or update such a machine learning model.

[0060] This manner of operation could allow the model to be pre-trained to a high level of accuracy using larger numbers of images of conventional sample types. Such a pre-trained model could then be fine-tuned based on relatively fewer sets of images obtained of rare and / or difficult-to-incubate samples in order to improve the ability of the system 100 to incubate such rare samples.

[0061] Additionally or alternatively, multiple different imaging modalities could be used to reduce the time to generate image data by using one modality to target another. For example, a first modality could be used to quickly generate a first image of a sample that includes sufficient structural information to locate the sample within the well (e.g., brightfield imaging). The second imaging modality could be a slower, target-able three-dimensional imaging modality (e.g., NIRST, SRSM, a confocal modality) and the scan volume for the modality (e.g., a focal depth or slice, a set of imaged points within the well) could be set to match the determined location of the sample within the well. This could allow the time cost of slower imaging modalities to be reduced, since image data from such modalities is generally only needed for the volume of the sample within a well, and not for the surrounding cell culture medium or other non-sample volumes of the well.

[0062] In some examples, image data generated by the imaging system 130 could be used to determine when to use an auxiliary imaging system 170 that generates improved images (e.g., higher spatial or spectral resolution, or that provide improved information about chemical composition of a sample) by that does so more slowly or in some other manner that is less feasible to perform as frequently as with the imaging system 130. For example, a model that receives image data from the imaging system 130 and generates therefrom predictions of improved sample imagery could also output a confidence estimate in that output, and the auxiliary imaging system 170 could only be used to image the sample if the confidence output is below a specified level. In another example, a set of image data for a sample, taken over time, could be used together with newly-generated image for the sample by the imaging system 130 to determine whether to use the auxiliary imaging system 170 to generate higher-qualityimage data for the sample. This could include, e.g., applying the set of available image data to a deep reinforcement learning agent or some other trained model to predict a future pattern of growth of the sample, to determine an intervention to apply to the sample, to predict a state of the sample, or to predict some other information about the sample and obtaining, from the model, an indication that a level of certainty about the state of the sample is sufficiently low that image data from the auxiliary imaging system 170 is needed. Based on such an output, the auxiliary imaging system 170 could be used to image the sample.

[0063] Since the time needed to record a hyperspectral image from a sample is generally related to the number of wavelengths at which the sample is imaged, the number of wavelengths used to obtain a hyperspectral image by the imaging system 130 could be specified to be a smallest number that still provides sufficient information to accurately predict some information of interest about the sample (e.g., to predict a map of the chemical composition of the sample). For example, to predict the pattern of protein, lipid, and other chemical composition of a sample, the sample could be imaged (e.g., using NIRST, or using SRSM at a limited set of wavelengths) at a number of peaks in the infrared, e.g., at a number of peaks between 2800 and 2840 wavenumber. The specific number and location of the peaks could be selected via feature selection techniques based on higher-spectral-resolution images of training samples, e.g., based on SRSM images of samples taken at a regular spectral resolution regularly spanning a range of wavelengths. This could include, e.g., imaging two to five wavelengths in the CH stretching region of the Raman spectrum and a number of additional wavelengths in a ‘fingerprint’ region of the Raman spectrum around 1640 wavenumber (e.g., a number of wavelengths between 1600 and 1800 wavenumber).

[0064] The system 100 (or other systems described herein) could operate to periodically image samples in wells of plates contained by the system 100 and then determine one or more interventions to apply to the sample based on that image data (including the timing of each of the images in the image data and / or specific locations within each sample to be imaged, e.g., based on the output of a machine learning model to maximally improve an estimated internal state of a sample while minimizing the time, light exposure, or other costs of the imaging) and optionally other stored data for the samples (e.g., a record of past interventions and the timing thereof applied to the sample, information about the seeding of the sample, information about the identity, properties, or genetics of cells in the sample, information about a patient from whom the sample was generated). The intervention could be selected for a particular sample in order to increase the likelihood of the sample surviving, to cause the sample to match one or more specified properties (e.g., a specified size, geometry,density, chemical composition, or internal biological state), to cause the growth of the sample to correspond to a specified growth profile with respect to size, composition, etc., and / or to achieve some other end. As noted above, the availability of frequent, per-sample image data throughout the incubation of the samples enables the determination and provision of such per- sample interventions, improving the number of successfully cultured samples, providing additional information about the history of each sample, and allowing the specifics of the samples (e.g., size, geometry) to be more closely controlled, even in cases where the samples are cultured from novel or rare cell types.

[0065] The controller 140 could be, by way of example and without limitation, a computer (such as a desktop, notebook, tablet, or handheld computer, a server), an industrial controller system, all or a portion of (e.g., one or more servers, nodes, or other organizational units of) a cloud computing system, or some other type of computational substrate. It should be understood that controller 140 may represent a physical computing device such as a server, a particular physical hardware platform on which a machine learning and / or other applications operate in software, or other combinations of hardware and software that are configured to carry out computational functions as described herein, e.g., to implement, in combination with other aspects of the system 100 and / or aspects of additional automated laboratory systems, a cyber-physical system for culturing volumetric biological samples and for learning about known or novel volumetric biological samples (e.g., in the form of training one or more machine learning models to translate image information from one modality to another, to predict the trajectory of growth of a particular sample, to generate a set of interventions to apply to a set of samples in order to learn about the dynamics of the samples and / or to increase the number of the samples that survive to a designated incubation end state, and / or to perform some other task).

[0066] Controller 140 may comprise one or more general purpose processors - e.g., microprocessors - and / or one or more special purpose processors - e.g., digital signal processors (DSPs), graphics processing units (GPUs), floating point units (FPUs), network processors, tensor processing units (TPUs), or application-specific integrated circuits (ASICs). In some instances, special purpose processors may be capable of image processing (e.g., application of CNN kernels or other filters to images via, e.g., convolution), machine learning model training, execution, and / or inference, among other applications or functions. Controller 140 may include data storage that may include one or more volatile and / or non-volatile storage components, such as magnetic, optical, flash, or organic storage, and may be integrated in whole or in part with the controller 140 and / or may be accessible thereby. For example, aportion of such data storage may be implemented as cache or other on-chip memory of a graphics processing unit or tensor processing unit integrated circuit and / or as RAM or some other variety of storage that is collocated with a GPU or TPU, e.g., on a graphics card, tensor acceleration card, or other semi-discrete subsystem of the overall controller 140. Such storage could be used to store parameters that define a machine learning model (e.g., weights or other parameters of units of a multi-layer neural network or other multi-unit machine learning model). Such data storage may include removable and / or non-removable components.

[0067] Such a process could include projecting each image of the image data into a respective latent vector in a multidimensional latent vector space. The latent vectors can then be applied as inputs to the model instead of the images themselves. This provides a number of benefits. These benefits can include computational benefits like reduced storage costs to store records of such vectors instead of the full images, reduced memory and computational costs to execute a model based on such smaller latent vector inputs, and the ability to train smaller models, using smaller training datasets, to predict interventions or future sample growth based on such latent-vector inputs. These benefits also include the ability to easily use such trained models to accept image data from novel imaging modalities (i.e., imaging modalities not represented in the training dataset used to initially train the model) by projecting images generated from such new imaging modalities into the multidimensional latent vector space and / or to apply image data from different imaging modalities (e.g., brightfield imaging, SRSM hyperspectral images, NIRST hyperspectral images) to the model by projecting images from the different modalities into the same multidimensional latent vector space.

[0068] Figure 2 depicts aspects of a method 200 for using image data 201 about a sample to predict, for that sample, an output 225 that can be used to apply an intervention to the sample. The output 225 could be a predicted intervention and / or could be a prediction of the future growth of the sample, which can then be used in an iterative fashion to determine a final intervention to apply to the sample. The image data 201 could include image data from multiple image modalities (e.g., a brightfield image and an NIRST image) and / or could include image data predicted, using trained machine learning models (not shown), based on other image data (e.g., chemical composition data predicted from NIRST and brightfield image data, chemical composition data predicted from SRSM image data, higher-resolution hyperspectral image data generated from lower-resolution hyperspectral image data and / or brightfield image data).

[0069] The image data 201 is applied to a first trained machine learning model 210 to generate a latent vector 215 that represents the image data 201 in a multidimensional latentvector space. As shown, the image data 201 could be augmented with summary properties 203 (e.g., size, geometry, density) of the sample depicted in the image data 201; such summary properties 203 can be generated using image processing algorithms 222. Augmenting the input to the first model 210 can provide a variety of benefits. These can include allowing the first model 210, for a given level of accuracy, to be smaller (e.g., fewer parameters), to be less computationally expensive to execute or train, and / or to require fewer training examples to train. This is because the extracted summary properties 203 can be selected to be properties that are highly relevant to the health and biological state of the sample (e.g., size, geometry, density), allowing the first model 210 to avoid allocating internal model parameters or features to predicting such relevant parameters internally.

[0070] The latent vector 215 output from the first model 210 is applied, as part of a set of latent vectors 217 with one or more additional latent vectors representing additional prior image(s) taken of the sample at respective different prior points in time, to a second trained machine learning model 220. Applying the set of latent vectors 217 to the second model 220 can include applying a representation of the relative timing of generation of the image data represented by the latent vectors 217 to the second model 220, e.g., as an additional scalar value associated with each of the latent vectors 217. Optional additional inputs to the second model 220 can include a set of interventions 219 that have been and / or will be applied to the sample. Further optional inputs to the second model 220 can include information about the sample, e.g., about the methods used to seed the sample (e.g., cell types, geometry, amount or type of culture medium), about the genetic sequence of the sample, phenotypic information about a person from whom the sample was generated, information about a diagnosis of or treatments applied to the person from whom the sample was generated, or other information about the sample.

[0071] Depending on the method used to train the second model 220, the output 225 of the second model 220 could be an intervention to be applied to the sample (represented as, e.g., a timing of application of the intervention, a type of the intervention, and one or more values representing an amount or magnitude of the intervention to be applied). Additionally or alternatively, the second model 220 could output a prediction of future growth (or death) of the sample based on the provided inputs 217, 219. This could include a prediction of the state (e.g., alive, dead) of the sample at a specified future point in time (e.g., alive, dead), a prediction of the size, geometry, density, or other properties of the sample over time, or some other information about the future behavior or state of the sample. In such examples, the predicted future behavior or state of the sample could be compared to a desired future state (e.g., ‘alive’) or behavior (e.g., size trajectory relative to a desired pre-specified size trajectory) in order toevaluate 230 candidate interventions (indicated by the dashed-line intervention of the set of interventions 219). Based on this comparison, alternative candidate interventions can be determined and evaluated by applying the alternative candidate interventions as inputs (with the set of latent vectors 217 and remainder of the set of interventions 219) to the second model 220 to evaluate the effect of such alternative candidate interventions on the sample. Such a process can be performed iteratively to determine a particular candidate intervention to apply to the sample in order to increase the likelihood that the future behavior or state of the sample comports with the desired future behavior or state of the sample. Such evaluations 230 can include grid searches, gradient descents, or other methods.

[0072] The second model 220 can be trained on available training data to generate the desired output(s). If the second model 220 is trained to predict future behaviors and / or states of a sample, the model could be trained based on available states and / or trajectories of observed samples and their corresponding historical sets of latent image vectors (or other image information) and applied interventions. Additionally or alternatively, the second model 220 could be trained, based on previously observed sample image, intervention, and outcome data, to directly predict interventions to apply to the sample. For example, the second model 220 could be a reinforcement learning agent trained, in an offline or online manner, to increase the likelihood of samples surviving or to work toward some other criterion. For example, the second model 220 could be a deep reinforcement learning agent.

[0073] A reinforcement learning agent (e.g., a deep reinforcement learning agent) can provide a variety of benefits. Such an agent can act to ‘explore’ the space of possible interventions, ‘spending’ interventions and samples to increase the scope of scenarios observed by the agent and thus enriching the set of training data available to the agent in a manner that balances the ‘cost’ of such exploration with the benefits of ‘exploiting’ already-existing knowledge of the biological samples being cultured in order to increase the number of such samples that are successfully incubated and / or that are incubated in a manner that comports with pre-specified growth trajectories. This can reduce the number of samples needed in order to develop a sufficiently rich set of training data to train the agent to a desired level of accuracy for known cell or sample types.

[0074] Training such a model (e.g., a deep reinforcement learning model) using widely available sample types can be particularly valuable when applied to learn how to incubate rare samples (e.g., samples representing rare or novel conditions, or patient-specific samples). The embodiments described herein allow such a model to be trained using large amounts of ‘common’ samples imaged and incubated in the same manner (i.e., using the same system(s)),providing a high degree of predictive accuracy with respect to the ‘common’ samples while also learning about a the various mechanisms and behaviors of growth exhibited by biological samples when being incubated and when subjected to a wide range of possible intervention (e.g., when exposed to various different pharmaceuticals or other substances). The reduced persample cost in system bandwidth and reagents enabled by the embodiments described herein can also allow such a model to ‘sacrifice’ more samples to ‘explore’ the space of interventions and sample responses thereto.

[0075] The model can then leverage this broad knowledge about the behavior of biological samples in general to quickly adapt to the incubation of rare or novel samples, increasing the survival rate of such samples. Where the model is a reinforcement learning model, the model could also have an auxiliary input to throttle the amount of ‘exploration’ of the sample behavior space that the model engages in. Such a throttle input could be set to reduce the model’s ‘exploration’ when incubating samples whose cell or tissue sources are limited (e.g., sample cultured from cells obtained during a biopsy of a particular patient of interest), causing the model to perform more ‘cautiously’ and thereby increasing the likelihood that sufficient samples are successfully incubated. This likelihood of success can also be increased by training the model to receive, in additional to image and past intervention data (e.g., 217, 219), information about the genetic sequence, demographics, medical history, chemical composition, or other properties of the tissue or cell source used to seed the samples being incubated.

[0076] Indeed, the embodiments described herein allow large amounts of three- dimensional tumor organoids or other samples to be incubated an imaged in the same manner at decreased cost with respect to reagents, time, power, and hardware cost, in the process generating large datasets to train the various machine learning models described herein (e.g., image embedding models, growth prediction models, intervention prediction models). Some of the benefits of such a system (decreased cost with respect to reagents, time, power, and hardware cost) can be obtained even without the existence of the models described herein, allowing these benefits to sustain the operation of a system as described herein (using, e.g., conventional open-loop sample incubation techniques for known sample types) until sufficient training data has been generated to train the various models described herein.

[0077] Such a process could be extended by, e.g., first training an image embedding model, a growth prediction model, and / or a cross-image prediction model (i.e., a model that predicts some target image data based on one or more other modalities of image data and / or reduced-resolution image data) based on an initial set of training data. A reinforcement learningagent (which may be a type of machine learning model) could then be introduced that receives the outputs of such models (e.g., latent vectors representing image data) and outputs interventions to be applied to the samples being incubated. Such a reinforcement learning agent could act to ‘explore’ the space of possible interventions and sample growth behaviors, efficiently enriching the set of training data while also improving incubation success rates. Once such a reinforcement learning agent has developed a broad ‘knowledge’ of the dynamics of sample incubation for common samples or cell types, such a trained reinforcement learning agent can then be applied to improve the incubation of rare or novel samples (e.g., samples cultured from small amounts of tissue obtained from specific patients, in order to facilitate the in vitro evaluation of patient-specific therapies).

[0078] The embodiments described herein also allow samples to be incubated at multiple sites (e.g., each having a respective instance of the system 100 of Figure 1) and the data generated thereby aggregated in order to train and / or update models, allowing more training data to be generated while also allowing samples to be incubated at multiple different sites (e.g., different hospitals, proximate to respective populations of patients) having multiple different sizes or areas (e.g., to accommodate local space limitations and / or demand for sample incubation services). Computationally ‘expensive’ operations described herein (e.g., model training, model inference) could be performed by a central server or other computational system (e.g., a cloud computational system and / or a system collocated with one or more of the incubators providing data thereto). Such a central system could receive imaging and other data about samples from remote incubator systems (e.g., latent vectors representing image data, records of past interventions, information about the source, genome, or other compositional information about the samples) and apply that information to one or more trained machine learning models as described herein to determine one or more interventions to apply to one or more samples. The determined interventions could then be transmitted from the central system to the remote incubator, to be used to incubate samples thereby.

[0079] Additionally or alternatively, certain computational operations (e.g., model training, inference of deep reinforcement learning models) could be performed by the central system while other computational operations (e.g., inference of image generation models to determine imaging or chemical composition data from other image data, inference of image projection models to project image data into a multidimensional latent vector space) could be performed by a controller or other computational resource that is local to an incubator. For example, a local controller for a system (e.g., 100) as described herein could operate cell seeding, incubation, imaging, intervention, or other systems to generate image data for asample. The local controller could then apply the image data to a trained machine learning model to generate a latent vector therefor. The latent vector could then be transmitted to a central server that uses the latent vector (with, optionally, additional prior latent vectors or other information about the sample) to determine an intervention to apply to the sample. The intervention could then be transmitted to the local controller and delivered to the sample. Operating in this manner can lead to significantly reduced bandwidth of the communications between the local controller and the central server, since the latent vector(s) are significantly smaller than the image data from which they are generated.

[0080] The various hardware and software embodiments described herein could be integrated into a cyber-physical system, with the hardware and software devices, programs, or other elements or features organized into a multi-layer information architecture. Figure 3 depicts such a multi-layer information architecture 300, which includes a physical layer 310, an interface layer 320, a data layer 330, an orchestration layer 340, and a services layer 350. Each layer can communicate with its neighboring layers, passing information up from lower layers to upper layers via compression, analysis, synthesis, or other processing tasks and passing commands, incubation strategies, information requests, or other objectives down from upper layers to lower layers via decompression, replication, prediction, execution of commands, or other processing tasks.

[0081] Such an information architecture, and the organization of software and hardware resources that it represents, can facilitate the implementation of the embodiments described herein by allowing for processes at different level of abstraction, relative to the low- level raw data measured from and physical manipulations performed on the actual volumetric biological samples, to operate on corresponding types of input and output information. This can facilitate the initial development of such systems, e.g., allowing developers creating software at a particular level of the architecture to deal with appropriately simplified datatypes, assuming that lower levels have performed any necessary processing thereof. This can also facilitate the maintenance, upgrading, and / or expansion of such systems) as well as the expansion of the capabilities of the system through the addition of new hardware or software elements by allowing such new functionality to be designed to interact with existing functionality at the same or neighboring levels of the information architecture. For example, a new imaging system could be added by implementing the system and / or related image processing software (e.g., as an API or other software) to provide / receive inputs / outputs in a format or other manner similar to existing imaging systems and / or software related thereto. In another example, improved machine learning models for orchestrating the incubation ofpopulations of novel samples could be implemented by programming such models to receive similar inputs, and to generate similar outputs, to existing model(s) used to orchestrate such operations.

[0082] The physical layer 310 represents the various physical hardware of the system 300. This can include robotic sample handling arms or other equipment, incubators, imaging systems, plate handlers, plate lidders, plate de-lidders, sample slicers, microdissectors, sample digesters, pumps or other equipment for storing and delivering cell culture medium or other materials (e.g., pharmaceuticals, buffers, or other substances that can be delivers into sample wells to implement an intervention), sample strainers, or other physical hardware. The physical layer 310 could include instruments or other physical hardware depicted in Figures 1, 6, 7, 8, or 12.

[0083] The interface layer 320 interfaces with the physical hardware of the physical layer 310 and provides access to the abilities and information thereof to higher layers. The interface layer 320 acts to generate commands to operate the motors, pumps, heaters, light sources, robotic manipulators, cameras, shutters, electrodes, microtomes, actuators, and other outputs of the physical layer 310 and also to receive image data, measured physical parameters (e.g., incubator temperature, humidity), encoder outputs, or other information or signals output from sensors or other elements of the physical layer 310. The interface layer 320 can include logic to implement commands from higher levels (e.g., to generate an image of a specified sample) by operating the appropriate elements of the physical layer 310 (e.g., to operate a robotic arm and de-lidder to bring a multi-well sample plate containing the specified sample to an appropriate imaging apparatus, to operate the light sources, image sensor, and / or other elements of the imaging apparatus to generate the image, and then to operate a lidder and robotic arm to return the multi-well sample plate to an incubator). The interface layer 320 can also maintain a record of the status of elements of the physical layer 310 (e.g., the location of multi-well sample plates, the status of incubators, the location and status of robotic arms), maintaining a “digital twin” of the physical hardware in order to facilitate operation thereof.

[0084] The data layer 330 can act as a central repository for all of the data collected for each of the volumetric biological samples in the system. This can include storage and organization of sample-associated measurements, including but not limited to raw, postprocessed, and / or model-generated image data, records of delivered interventions, records of incubator conditions, recorded images, gene sequences, or other information related to the biological sources (cells, tissues) and methods used to seed the volumetric biological samples (or seeds or other precursor structures thereof) in their respective sample wells, or othermeasurements taken of the volumetric biological samples and / or records of manipulation thereof. The data stored by the data layer 330 can also include derived information, e.g., contrast-enhanced or otherwise processed images, style-transferred or otherwise-generated images that depict a sample in one modality (e.g., immune-stained) based on measured image data from another modality (e.g., NIRST or other label-free non-destructive imaging data), past predictions of the future growth trajectory of the sample, or embeddings of raw or derived image or other data. For example, images of a sample could be projected (e.g., by application to a trained machine learning model) to respective vectors in a representative multidimensional embedding space to represent the sample and / or such an embedding vector could be based on a combination or image and other data for the sample (e.g., a record of interventions applied to the sample, a cell type or genetic sequence of the sample).

[0085] Such per-sample records or other information objects provide quality assurance and a digital record of every manufactured tumor model (MTM) created as described herein. Such records may also be required for regulatory compliance and safety. Such per-MTM digital records can facilitate novel applications and research by providing an extensive, high-temporal- resolution record of the state of each MTM across its incubation and optionally after and before the incubation (e.g., images or other information about tumors, biopsies, or other samples from which an MTM was cultured and / or about the specific process used to generate the tissue slice or other cell sample used to culture the MTM). Such a record or other object representing an MTM could include representations of the MTM at a low level (e.g., individual images of the sample at specific points in time) or at high level (high-dimensional embedding vectors representing a cyber-physical summary of all or some specified portion of the available information about the MTM and / or related MTMs at one or more points in time). The record or other information object could include raw data measured from the MTM (e.g., images, temperatures, pH values) or related MTMs (e.g., MTMs cultured in the same multi well sample plate, in the same incubator, and / or at the same automated laboratory site, MTMs cultured from the same source or of the same cell type, MTMs exposed to the same or similar interventions) and / or raw data about methods applied to incubate the MTMs (e.g., timings and other information about interventions applied to the MTM, images of tissue slices or other image data related to the creation of the MTM from a biopsy, tissue sample, dissociated cells, or other source). Image data includes in the record or other information object could include images from multiple modalities, e.g., from a first modality that is relative quick to generate, recorded at a first, higher, temporal resolution and from a second modality that is relatively slower or otherwise more costly (e.g., with respect to incurring photobleaching or other damage or insultto the MTM) to generate and that thus is recorded only once or at some other reduced temporal resolution.

[0086] The record or other information object representing the MTM could include derived information about the MTM. Such derived information could include multidimensional vectors or other high-dimensional representations of the MTM generated from images, applied interventions, predicted growth trajectories, related MTMs, or other information about the MTM generated, e.g., as an intermediate or final output of a machine learning model trained to generate such representations and / or to perform some other task (e.g., to predict the future growth trajectory of the MTMs, to determine interventions to apply to the MTMs). Such derived information could include mode-translated images of the MTMs, e.g., images of the chemical composition of the MTMs generated, by trained mode-translating machine learning models or other algorithms, based on one or more NIRST images or other IR, visible, and / or ultraviolet images of the MTMs.

[0087] As noted elsewhere herein, such a record or other information object can be used throughout the incubation of an MTM to inform the interventions applied thereto. Such an information object may be gradually updated during incubation of the MTM (e.g., one or more summary multi-dimensional vector embeddings thereof could be updated with the addition of new images or other data about the MTM and / or about related MTMs) and may be used to determine interventions to apply to the MTM and / or to related MTMs. Such a record or other information object could include indications of confidence or certainty in its contents (e.g., confidence values, confidence intervals, likelihood values, or other measures of certainty in elements of a multidimensional vector representing the MTM and its history, in one or more predicted properties (e.g., size, chemical composition, cell type, geometry) of the MTM, in a predicted growth trajectory of the MTM, or in some other aspect of the information object) and such measures could be determined from the object itself, from objects representing related MTMs, from outputs (e.g., auxiliary confidence outputs) of models or other algorithms used to generate and / or update aspects of the object, or determined in some other manner. Such confidence or certainty measures could be used to trigger imaging or other measurements of an MTM and / or to increase the frequency of such measurements (e.g., in response to determining that a confidence for one or more aspects of an MTM have, with the passage of time, fallen below a specified level, a NIRST or other image of the MTM, or some other measurement of the MTM, could be made in order to provide information to increase the confidence in the one or more aspects).

[0088] Non-destructive imaging may be employed multiple times during the incubationperiod of a volumetric biological sample . Such imaging may be conducted on a pre-determined schedule, for example every sample may be imaged every day. Some samples may be imaged on some days and other samples may be imaged on different days. Some imaging modes may occur every day while other imaging modes may occur more frequently or less frequently. The imaging modes and schedule for each mode can be pre-determined or can be determined based on observations made during incubation (e.g., by one or more layers of the multi-layer information architecture 300). When processing a batch of samples, the specific volumetric biological samples to be imaged can be pre-determined or the set to be imaged can be determined based on observations made during incubation. A software system may interpret images, calculate sample attributes, or make predictions of physical or chemical attributes to determine the schedule for which samples to measure, what imaging mode(s) to use, and / or when to perform the measurements. Software may be used to determine which available measurements are most likely to improve the accuracy of predictions. Software may also be used to determine which measurements are most likely to aid in determining which chemical interventions to deploy, the schedule for deploying such interventions, or the effectiveness of such chemical interventions.

[0089] The data layer 330 can include trained machine learning models, filters, or other algorithms, models, programs, or other software to facilitate the generation of derived data stored in the data layer 330. Such software could generate estimations of the status (e.g., size, geometry, health) of a sample based on information measured from the sample (e.g., images of the sample, interventions applied to the sample, incubation conditions experienced by the sample) and / or based on information measured from other related samples (e.g., based on images or other raw or derived information of samples stored in other wells of the same multiwell sample plate, samples that have been incubated in the same incubator, samples that are of the same cell or tissue type samples that have been subjected to the same or similar interventions). Such derived information or other updates can be stored or otherwise reflected in the information object for the sample for which such estimates are made and / or for information objects of related samples (e.g., for samples on the same multi-well sample plate, samples incubated in the same incubator, samples of the same cell or tissue type, samples subjected to the same or similar interventions). The generation of such estimates could be accomplished using trained machine learning models. Such estimated information about samples could include estimates of confidence or noise in the estimates, e.g., an estimate of confidence in a derived image of a sample that has been generated based on an old image of the sample and more recent images or other data of one or more related but different samples.

[0090] The data layer 330 could be configured to organize the training or execution of models or other software thereof. For example, the data layer 330 could be configured to train and / or update of models used to estimate the statues of samples, to perform style-transfer, or other models used by the data layer 330 based on, e.g., having received sufficient additional training data to perform such training.

[0091] The data layer 330 could be configured to initiate the imaging of or other data gathering from samples (additionally or alternatively, the orchestration layer 340 could perform and / or organize such operations). For example, the data layer 330 could operate to periodically re-image the samples, e.g., in order to provide updated imagery and / or estimates of the size, geometry, health, or other information about the samples. This could include imaging a sample, selecting an imaging modality or imaging parameters to image a sample, or modifying a frequency with which a sample is imaged or otherwise measured based on an estimate of confidence in the accuracy of the available information for a sample with respect to the ability to estimate the current state of the sample.

[0092] For example, if a level of confidence with respect to the ability to estimate a size, geometry, health status, element of a representative multi-dimensional embedding, cell type, chemical cell contents, or other information about a particular sample drops below a specified level, the data layer 330 could operate to image the particular sample or take some other measure to obtain updated information about the particular sample sufficient to achieve the desired level of confidence with respect to the information of interest. The level of confidence and / or identity of the information whose confidence must meet that level may be pre-specified, e.g., by a command received from a higher layer (e.g., the orchestration layer 340), such that the data layer 330 information is proactively maintained at the specified level of confidence. Additionally or alternatively, a specified level of confidence and / or identity of the information may be received from a higher layer, prompting a determination of the relevant confidence and, if that confidence is below the specified level, further prompting imaging a sample or taking some other action to obtain relevant confidence -increasing information about a sample.

[0093] As noted above, such determinations (to image or otherwise measure a sample or to increase the frequency of such measurement) could be made based on the contents (e.g., confidence values stored in and / or derived from) of information objects for the samples. Indeed, various operations of some or all of the layers of the system 300 may trigger an update to the information objects for one or more samples, which may trigger such additional or increased frequency imaging or other measurements of the samples for which the objects areupdated and / or for related samples. For example, the operation of the orchestration layer 340 (described below) to predict the growth trajectory for one or more samples could result in a prediction with low confidence, triggering imaging or other measurement(s) of the one or more samples in order to permit generation of higher-confidence predicted growth trajectories based on newer and / or more complete image or other measurement data. In this way, operations up and down the layers of the 300 on information object(s) related to a particular sample can have the effect of ‘weaving’ additional information into the ‘digital thread’ relating to the sample, for example based on information from the ‘digital threads’ of related samples or sets of samples.

[0094] The orchestration layer 340 can act to organize the operation of the system 300 to incubate volumetric biological samples in order to achieve high-level objectives received from the services layer 350. This can include, based on sample information received from the data layer 330, using machine learning models or other software to predict the future growth trajectories of the samples in response to various putative interventions applied thereto and, based on such predictions, selecting interventions to apply to the samples in order to achieve a desired outcome, e.g., in order to obtain a population of volumetric biological samples that correspond to a specified distribution or other parameters. Additionally or alternatively, the orchestration layer 340 can execute a reinforcement learning model (e.g., a deep reinforcement learning model) to generate a set of interventions to apply to the samples in order to obtain useful information about the samples (e.g., about the growth properties, reaction to pharmaceuticals or other substances, or other properties of novel samples) while also balancing a drive to create an increased number of viable samples.

[0095] Information collected by the various imaging and other data-generating elements of the systems described herein (e.g., by one or more layers of the multi-layer information architecture 300 and / or systems associated therewith) may be analyzed to calculate physical attributes about target volumetric biological sample(s) (e.g., at the data layer 330). Information collected by the various imaging elements of the systems described herein may also be analyzed to calculate physical attributes, chemical attributes, or other information about the volumetric biological sample (s) based on physical information, chemical information, or other information about the sample(s) represented in the image or other information generated about the sample(s). Physical attributes and / or chemical attributes (absorption spectra, fluorescence spectra, or other optical properties of contents of a sample, predicted concentrations, distributions, or other information about one or more specified chemical species or groups of chemical species within a sample) may be calculated for a completevolumetric biological sample or a portion thereof. Portions of a sample for which such attributes may be determined could include a single cell, a portion of a cell, a single nucleus or portion of a nucleus, or a group of cells. Summary physical attributes and / or chemical attributes may also be calculated across a population of volumetric biological samples. Attributes may be calculated to provide additional quantitative information about samples or populations of volumetric biological samples for quality assurance purposes, to determine whether one or more volumetric biological samples meets one or more specified criteria (e.g., to be used for a downstream experiment or other application), to train or test machine learning models, or for other purposes.

[0096] Physical attributes may include size, geometry, location, volume, linear dimension, or other structural properties of individual cells, clusters of cells, or other contents of a volumetric biological sample. Physical attributes may include optical properties such as color, density, spectral reflectivity, spectral transmissivity, spectral absorptivity, or other information that may be measured using optical methods. Physical attributes may describe the location, geometry, shape, morphology or dimensions of any measured or inferred aspect of the volumetric biological sample (e.g., of individual cells, of portions of cells, of colonies or other populations of cells, of individual organelles, of portions of organelles). For example, physical attributes may include detailed information about the 2D or 3D shape of the volumetric biological sample, the length or shape of a cell wall, or the shape of a nucleus. Physical attributes may include spatial information or statistics about the distribution of features within a volumetric biological sample, for example the 2D or 3D arrangement of nuclei, the 2D or 3D arrangement of cell walls, or the 2D or 3D arrangement of chemical information or attributes such as the distribution of specific fluorescent markers.

[0097] Chemical attributes may include any aspect of the volumetric biological sample that indicates the types of chemicals composing the volumetric biological sample, e.g., as detected via non-destructive and / or destructive or otherwise sample-perturbing imaging (such as fluorescent imaging of fixed samples). Chemical attributes may include identity, concentration, or distribution, of one or more specific chemical species, sets of species, or chemical functional groups. These can include small molecules and metabolites, lipids, carbohydrates, nucleic acids, and proteins. Chemical attributes may also include aggregate measures of these species, including total protein concentration, protein conformation, types of nucleic acids or saturated and unsaturated lipids.

[0098] The physical and / or chemical information or attributes of a volumetric biological sample may be used to calculate or determine other attributes or properties of thesample, e.g., properties of growth or change of the sample over time. Examples of such volumetric biological sample attributes include the growth rate of the sample or part thereof, number of cells, identification of one or more constituent types of cells, a spatial arrangement of cells and / or cell types, number of living cells, number of dead cells, location of living cells, or location of dead cells as well as cellular function at one point in time and / or over time. Examples of cellular function include response to drugs, attributes under oxygenation conditions, or cellular processes like lipidogenesis.

[0099] A software or other system (e.g., combined cyber-physical system as described herein) may predict attributes of the biological sample (e.g., in the data layer 330 and / or orchestration layer 340 of the multi-layer information architecture 300). Such predictions can be used to determining whether volumetric biological samples meet or are likely to meet (e.g., at the end of an incubation period) one or more criteria (e.g., quality assurance criteria for delivery, for use in downstream experiments or other applications). Such predictions can also be used to determine chemical interventions to be applied to the sample(s). Examples of predicted attributes include sample size at a future time, shape at a future time, growth rate at a future time, types of cells at a future time, spatial arrangement of cells and / or cell types at a future time, identification of one or more constituent types of cells at a future time, number of cells of each of a plurality of different types at a future time, spectral signature at a future time, chemical composition at a future time, response to functional challenge at a future time, viability at a future time, growth rate at a future time, number of living cells at a future time, or number of dead cells at a future time.

[0100] The services layer 350 can receive inputs from users (e.g., from remote systems via the internet) to specify objectives of the system or other information. For example, the services layer 350 can interact with users to receive information about a number or other characteristics (e.g., a desired range or distribution of sizes, geometries, growth profiles or other characteristics) of a population of volumetric biological samples that the user desires to obtain. The services layer 350 can also receive, from the user, a specification of an identity of a type of tissue or cells to use to seed the desired samples, e.g., a novel tissue sample provided to a specified receptacle of a system of the physical layer 310, or a known sample kept ‘in stock’ within the system 300. The services layer 350 can also provide, to the users, records for each volumetric biological sample incubated for the user (e.g., a per-sample record of all the imagery or other data measured for a sample, the timing an identity of intervention applied to the sample, incubation parameters applied to incubate the sample, style-transferred images, cell size and geometry, growth profiles, or other derived information for the sample). The serviceslayer 350 can transmit other information about the volumetric biological samples and / or about their production, location, verification, or any other information, e.g., software outputs of the system.

[0101] The services layer 350 can also organize the operation of the other layers of the system 300, including organizing and segregating the flows of information and / or the allocation of automated laboratory systems or other resources, in order to allow particular portions of the system (e.g., individual automated laboratory sites located at, e.g., respective hospitals or other healthcare facilities) to be operated by or otherwise accrue benefits specifically to their owners or other specified individuals, groups of organizations. This can facilitate a variety of different methods of fabricating, maintaining, and organizing such a system (e.g., different business models for the financing, siting, and operation of individual automated laboratory systems operating within the overall system 300). The services layer 350 could operate to allow different users to make different types of requests (e.g., of different priorities, from different locations) and allocate corresponding portions of the overall system 300 to accomplish those requests (e.g., allocate requests from a particular user, which owns or is otherwise associated with a particular automated laboratory site, at an elevated priority to the associated particular automated laboratory site).

[0102] For example, the services layer 350 could allow a user that is associated with a particular automated laboratory or other portion of the physical layer 310 to obtain the benefits of such ownership or other association while still allowing them to benefit from, and contribute to, pooling of imagery or other data collected and analyzed by all of the aspects of the system 300. This could include the services layer 350 allowing a particular user, associated with a particular automated laboratory site (e.g., an automated laboratory of a hospital), to use the particular automated laboratory to incubate a population of volumetric biological samples (e.g., to incubate a population of samples from a specific patient or other individual). Additionally or alternatively, the services layer 350 could make the services of the particular automated laboratory available to other users (e.g., via the internet, through a webpage), allowing the benefits of such use (e.g., payment) to accrue to the particular user associated with the automated laboratory. Thus the services layer 350 facilitates information flow between different instances of the tumor growth system or different physical locations where tumors are being grown, analyzed, or tested (e.g., located at different hospitals or other physical sites, or even multiple locations at the same hospital). The services layer 350 may be used, for example, to help support or orchestrate the testing of pharmaceuticals that may be considered for treatment of cancer other disease.

[0103] The services layer 350 can also present such an interface via the internet (e.g., a webpage, an API for an e-commerce app) to allow users to make requests of the system 300 to incubate specified volumetric biological samples. This could include allowing the user to select the samples to be incubated from a list of samples or sample types that are natively available, e.g., that have already been incubated and stored by the system and / or for which the system already has a source of cells to incubate additional samples therefrom. The set of available samples could be user-specified, e.g., a user that owns or is otherwise associated with a particular automated laboratory site of the system 300 could have access to samples that are specific to that site (e.g., proprietary samples generated by a hospital, research institution, or other organization that houses or is otherwise associated with the particular automated laboratory site).

[0104] A user making requests could also include the user specifying properties of the incubated samples, e.g., size, number, geometry, chemical composition, growth trajectory, acceptable distributions or ranges of the preceding, or other specifications. A user making requests can also include the user specifying that the user will provide a biopsy, tissue sample, and / or cell sample to be incubated, optionally along with information about such sample(s) (e.g., cell type, origin, genotype, packaging or preparation method). Such a user can then send such samples to an automated laboratory site and / or a forwarding site. In such examples, the services layer 350 (e.g., the website ofthe services layer 350) could provide a shipping address, shipping label, or other information to facilitate the user sending the sample to the correct locations and also to facilitate identification of the sample when it arrives such that it can be correctly processed and associated with the user within the system 300. Note that the benefits of the embodiments described herein have been generally depicted with respect to the ability to incubate, in an automated manner and with increased consistency, lower cost, and other benefits, large numbers of volumetric biological samples, including volumetric biological samples cultured from unknown and / or novel cell and tissue types. Such samples can then be used to, e.g., assess the efficacy of a treatment or evaluate the behavior of the samples in response to some other environmental condition or intervention. However, the embodiments described herein can provide additional benefits. For example, the predictive models, reinforcement learning models, or other models or other software used to incubate a population of samples (in particular, to learn the growth behaviors and other properties of the samples being incubated) could, themselves, be used to evaluate the efficacy of a treatment, assess the effects of a pharmaceutical or other intervention on the samples, diagnose a disease (e.g., determine a type of a cancer represented in the samples), or provide some other benefit separatefrom any benefit provided by the incubated samples.

[0105] In some examples, a system as described herein can be interrogated (e.g., remotely) using a computer interface such as through a web browser or a stand-alone software application installed on a computer, tablet, or smart phone. The web browser or computer may be connected to the system using a wired connection, wireless connection, and / or over the internet. The computer interface may communicate with the laboratory equipment described here, a computer server that is connected the laboratory equipment, and / or a cloud computing system that stores data and information about tissue samples and / or performs some other operations related to the system and / or date generated thereby. Such a computer server or cloud computing system may store data and information about tissue samples grown at one site or at multiple sites. In this way, the computer server or cloud computing system can analyze data and create predictions or reports on tissue samples grown at different locations and at different times.

[0106] The user may send queries, report requests, database searches, or perform analyses using such a computer interface. These queries, requests, searches, and / or analyses may also be sent to a computer server, cloud computing system, or other system as described herein using an application programming interface (API) or other method for communicating information (e.g., commands) with such systems.

[0107] Such a computer interface can provide information to the user. The information may include raw data collected about individual MTMs or groups of MTMs or other types of volumetric biological samples as incubated, assessed, generated, or otherwise manipulated as described herein. The information may be organized by MTM type, intervention type, time or location of seeding, or time when certain information was collected. The information may include one or more attributes of the MTM, for example: types of cells in the tissue seed; number cells in the tissue seed; types of cells observed on a given day; number of each cell type observed on a given day; matrix composition; the tissue size on one or more days of growth; the MTM shape on one or more days of growth; brightfield image; spectral signature; chemical composition; type of intervention(s) applied thereto; when various interventions were applied; concentration of various interventions; growth conditions; viability; and / or growth rate. Such a computer interface may provide predictions for individual MTMs or groups of MTMs. The predictions may include: tissue size at a future time; type of cells at a future time; number of cells of each type at a future time; spectral signature at a future time; viability at a future time; growth rate at a future time; and / or MTM response to specific interventions or groups of interventions. Such a computer interface may offer suggestions for whatinterventions to apply or predictions of MTM behavior as a function of possible intervention types, concentrations, or deployment schedules.

[0108] One application of the systems and methods described herein is to obtain volumetric biological samples that meet given acceptance criteria. A volumetric biological sample that meets a given acceptance criterion or criteria can be used for subsequent studies or other downstream applications such as pharmaceuticals development and testing, toxicity studies, or basic research such as linking genomic information with biological function. The acceptance criteria may be linked to the presence or absence of physical or chemical attributes. The acceptance criteria may also be linked to a specified quantitative metric and to the variance of specific volumetric biological samples therefrom. For example, the metric for acceptance could be the number of cells of a given type plus or minus a given tolerance number of cells of that type. If the volumetric biological sample has a number of cells that falls within the tolerance range, then the volumetric biological sample is acceptable. Acceptance criteria may be physical attributes, chemical attributes, biological function(s), responses to interrogation(s), or predicted attributes of any type. Acceptance criteria can be assessed for individual biological samples or specifically determined for a number of samples.

[0109] While a variety of interventions or other perturbations of volumetric biological samples have been described herein for the purposes of enhancing growth of the samples, inducing the samples to comport with specified constraints, gaining information about the growth of such samples (e.g., to train a predictive machine learning model, to optimize incubation protocols for novel cell types), such perturbations can also be applied in order to explicitly evaluate a volumetric biological sample with respect to one or more experimental conditions or challenges of interest. Such “functional challenges” can include induced hypoxia, exposure to sub-lethal amounts of chemotherapy or other ‘poison’ substance, exposure to a ligand to excite one or more cellular receptors or metabolic pathways, or some other interrogation. Such a functional challenge could be applied at the end of incubation in order to enrich the data available to characterize each of the incubated samples. Additionally or alternatively, the timing of such functional challenges could be determined by a machine learning model or other controller in order to reduce the likelihood that the challenge kills the sample or otherwise prevents the sample from eventually comporting with one or more size, compositional, or other criteria. The duration, magnitude (e.g., concentration of an applied poison), or other particulars of an applied functional challenge could be pre-specified to be consistent across a set of samples (making the samples’ response thereto more directly comparable) while also reducing the likelihood that the challenge kills or otherwise rendersunusable any (or more than an acceptable number or amount) of the challenged samples. A set of such functional challenges (e.g., a hypoxic episode, a reduced-pH episode, and an episode of exposure to a chemotherapy drug or other poison) could be pre-specified for a set of volumetric biological samples such that each sample is exposed to the set of functional challenges. The timing of such a set of functional challenges could be pre-specified, or could be determined by a machine learning model or other controller as described herein in order to adequately evaluate the samples (e.g., subject to a constraint that the challenges are applied within a specified time period of the completion or incubation, or while the size of a sample is within a specified percentage of a final size) while still reducing the likelihood that the samples are killed, rendered unusable, or otherwise negatively affected with respect to some criterion or constraint.II. Example Embodiments

[0110] Provided herein are example manufacturing system(s) and aspects thereof that can provide consistent and repeatable production of three-dimensional (3D) tumor models of a variety of cancer types at scale. This agile manufacturing technology provides a turnkey solution to produce any solid tumor, including cancers from individual patients, cancers prevalent in underserved and minority communities, or rare cancers that cannot be addressed by current technologies.

[0111] 3D tumor models are physiologically superior to 2D cultures; however, they were previously only available for a few cancers and the previous methods for generating such tumor models included specialized protocols developed with trial-and-error testing and used uniquely skilled technicians to perform painstaking manual fabrication processes, among other deficits. Such previous approaches do not scale to the level required for widespread use and thus present an obstacle for the development of drugs, especially for personalized treatments or for rare cancers.

[0112] The manufacturing system(s) and aspects thereof described herein facilitates the scalable production of 3D tumor models with precise and consistent control over their quality (composition, physical and chemical and biological characteristics) that is not obtainable via prior methods.

[0113] The embodiments described herein provide on-demand, precision fabrication to produce manufactured tumor models (MTMs). Tumor model production can be accomplished as a digitally connected workflow. Such a manufacturing system can measure the physical and chemical attributes of every MTM as they grow. Software intelligence caninterpret these measurements and to predict, control, and verify every MTM. This software intelligence also allows the system to be rapidly reconfigured and optimized for different, or even completely novel, types of cancer.

[0114] Unlike state-of-the-art methods that lack rigorous quality control, the technology described herein provides quality assurance and a digital record of every MTM. Such a per-MTM digital records can facilitate novel applications and research by providing an extensive, high-temporal-resolution record of the state of each MTM across its incubation and optionally after and before the incubation (e.g., images or other information about tumors, biopsies, or other samples from which an MTM was cultured and / or about the specific process used to generate the tissue slice or other cell sample used to culture the MTM). Such a record or other object representing an MTM could include representations of the MTM at a low level (e.g., individual images of the sample at specific points in time) or at high level (high-dimensional embedding vectors representing a cyber-physical summary of all or some specified portion of the available information about the MTM and / or related MTMs at one or more points in time). The record or other information object could include raw data measured from the MTM (e.g., images, temperatures, pH values) or related MTMs (e.g., MTMs cultured in the same multi well sample plate, in the same incubator, and / or at the same automated laboratory site, MTMs cultured from the same source or of the same cell type, MTMs exposed to the same or similar interventions) and / or raw data about methods applied to incubate the MTMs (e.g., timings and other information about interventions applied to the MTM, images of tissue slices or other image data related to the creation of the MTM from a biopsy, tissue sample, dissociated cells, or other source). Image data includes in the record or other information object could include images from multiple modalities, e.g., from a first modality that is relative quick to generate, recorded at a first, higher, temporal resolution and from a second modality that is relatively slower or otherwise more costly (e.g., with respect to incurring photobleaching or other damage or insult to the MTM) to generate and that thus is recorded only once or at some other reduced temporal resolution.

[0115] The record or other information object representing the MTM could include derived information about the MTM. Such derived information could include multidimensional vectors or other high-dimensional representations of the MTM generated from images, applied interventions, predicted growth trajectories, related MTMs, or other information about the MTM generated, e.g., as an intermediate or final output of a machine learning model trained to generate such representations and / or to perform some other task (e.g., to predict the future growth trajectory of the MTMs, to determine interventions to applyto the MTMs). Such derived information could include mode-translated images of the MTMs, e.g., images of the chemical composition of the MTMs generated, by trained modetranslating machine learning models or other algorithms, based on one or more NIRST images or other IR, visible, and / or ultraviolet images of the MTMs.

[0116] As noted elsewhere herein, such a record or other information object can be used throughout the incubation of an MTM to inform the interventions applied thereto. Such an information object may be gradually updated during incubation of the MTM (e.g., one or more summary multi-dimensional vector embeddings thereof could be updated with the addition of new images or other data about the MTM and / or about related MTMs) and may be used to determine interventions to apply to the MTM and / or to related MTMs. Such a record or other information object could include indications of confidence or certainty in its contents (e.g., confidence values, confidence intervals, likelihood values, or other measures of certainty in elements of a multidimensional vector representing the MTM and its history, in one or more predicted properties (e.g., size, chemical composition, cell type, geometry) of the MTM, in a predicted growth trajectory of the MTM, or in some other aspect of the information object) and such measures could be determined from the object itself, from objects representing related MTMs, from outputs (e.g., auxiliary confidence outputs) of models or other algorithms used to generate and / or update aspects of the object, or determined in some other manner. Such confidence or certainty measures could be used to trigger imaging or other measurements of an MTM and / or to increase the frequency of such measurements (e.g., in response to determining that a confidence for one or more aspects of an MTM have, with the passage of time, fallen below a specified level, a NIRST or other image of the MTM, or some other measurement of the MTM, could be made in order to provide information to increase the confidence in the one or more aspects).

[0117] Figure 4 depicts aspects of the manufacturing technology described herein. The process begins with a specification of the end point for a desired MTM in terms of size, composition, and / or function. MTMs are seeded with precision placement of cells within an extracellular matrix (ECM) and then grown 10 - 28 days (about 4 weeks), depending on the cancer type and desired size. The growing MTMs are measured at regular intervals using label-free chemical and physical sensing to capture time series data about their size, shape, and chemical composition. This data may be interpreted by software intelligence based on machine learning (ML) to predict the future MTM state and optionally use such predictions to change the incubation of the MTMs in order to increase the probability that the MTMs survive and correspond to the pre-specific desired MTM specifications.

[0118] The sensing data and ML workflows can be used to rapidly identify improved (e.g., optimal) growth conditions through process parameters and interventions that promote or suppress growth. The improved growth conditions increase (e.g., maximize) the number of viable MTMs that satisfy the initial tumor specifications. In a final step for quality control, MTMs are verified for size and chemical composition prior to being transferred to a user.

[0119] Every physical tumor is accompanied by a digital record of the growth process and documentation that verifies MTM size and viability.

[0120] These embodiments can be implemented as a manufacturing service that can produce consistent and verified MTMs at an unprecedented rate of 100 or more per day. A “First” example system as described herein focuses on the most prevalent breast cancer subtype (estrogen receptor-positive). This includes precision seeding technology, in-situ sensing, ML models to predict the MTM end point, and a quality control system. A “Second” example system as described herein, which may be referred to as an ‘intermediate’ system, additionally focuses on producing MTMs from a rare breast cancer subtype (triple negative breast cancer), which is the deadliest subtype and especially prevalent among African- American women. This disclosure addresses a progression of four cell types of this disease, ranging from non-proliferative to metastatic. A “Third” examples system as described herein, which may be referred to as an ‘advanced’ system, additionally focuses on generalizability, focusing on six rare and patient-specific cancers. The First example system is capable of producing a total of 7,000 MTMs, rising to 75,000 for the Second example system, and 100,000 for the Third example system.

[0121] Simple 2D cultures remain the dominant model with their accessibility, cost, and ease of use, despite their well-known limitations. Animals, as an alternative, offer a more physiologically-representative model but are higher cost, laborious to develop, and less accessible for diverse tumor types.

[0122] In recent years, 3D cultures have emerged as a reasonably feasible model with greater physiologic reproducibility than 2D cultures and significantly more accessibility and ease of use compared to animal models. Despite these well-recognized benefits, 3D tumor models are very difficult to prepare compared to 2D cultures and impractical for most laboratories.

[0123] For example, there are several challenges associated even with the most popular and prototypical 3D tumor model: a spheroid of epithelial cells. Spheroids are typically fabricated in a two-step process: seeding tumor cells into a medium, followed by growth in an incubator. The seeding process is laborious and results in a single batch withhundreds of randomly distributed seeds of uneven sizes. Following a growth period of -15 days (about 2 weeks), the resulting spheroids have a random size distribution with size variability that can be as large as 10-fold.

[0124] To obtain models of a specific cancer type, growth conditions must be painstakingly developed with multiple iterative steps. These complex workflows and nonlinear biophysical processes accumulate to impart significant variance among the tumor models. In addition to excessive cost from an inherently wasteful process, the tumor model variance leads to weak statistical links between the tumor models and their intended application. In summary, tumor production today is diametrically opposed to modem industrial manufacturing that uses automation and standardized workflows to create scalable and repeatable products and to enable agile factories that can make different products at different times.

[0125] This disclosure uses breast cancer spheroids as a starting model and example to manufacture since they represent an intermediate complexity between 2D cultures and tissue structure. Spheroids reproduce the heterogeneous architecture and molecular gradients of signaling factors, nutrients, metabolites, interactions in spheroids, such as cell-to-cell and cell-to-extracellular matrix, also present easy-to-observe hallmarks for validation and judging the success of the manufacturing. Breast cancer tumor spheroids are also highly relevant for public health. Breast cancer is the second most common cancer among women in the United States (following skin cancer), with about 297,790 new cases and 43,170 deaths per year. Breast cancer also has significant health disparities as African American women die from breast cancer at a higher rate than White women. Men are also affected, though rarely, with about 2,800 new diagnoses and 530 deaths per year.

[0126] Breast cancer is recognized to consist of many subtypes. The most common are estrogen receptor-positive (ER+) tumors, which affect -70% of newly diagnosed patients. The First example system described herein provides manufacturing capability for ER+ MTMs. The process, powered by design of experiments (DOE), in situ sensing and ML workflows, allows us to optimally grow and to validate MTMs from ER+ tumors as well as different tumor types. The Second example system expands the First example system to produce ER+ breast cancer MTMs, providing industrial scale and quality benchmarks. The Second example system also produces four isogenic cell lines that are representative of tumor progression of the triple -negative breast cancer (TNBC) subtype, which is a rare (-10% of patients) but most lethal form of breast cancer. The Second example system provides a generalizable and agile manufacturing platform.

[0127] The high cost of developing new 3D cell culture protocols, combined with fewer resources to support rare diseases, presents significant barriers to making models for rare cancers or for tumors specific to underrepresented communities and patients. This challenge can be overcome with the ability of the system(s) described herein to rapidly optimize growth conditions in a systematic, automated manner. The Third example system expands the First and Second example systems to incubate rare and patient-specific cancers. The Third example system incubates samples of six cancers, three breast cancers from patient derived xenografts (PDX) and three of rare types, providing optimized process conditions to produce MTMs. The ability to find optimal conditions and develop a diversity of models at unprecedented speed is enabled by the integration of high content sensing and software intelligence of the system(s). The Third example system can be further expanded to generate individually precise tumors, organoids with multiple cell types, and other complex models.

[0128] The embodiments described herein include advancements in manufacturing automation, multi-modal chemical and physical sensing, and software intelligence. Manufacturing automation is the first technical pillar and comprises the integration of the hardware and software components. While scaffold-free methods can be used, this disclosure uses a scaffold-based process to provide a more physiologic environment, precise control over microenvironment properties and for future generalizability to include other cell types and tumors. MTMs are sensitive to their chemical and mechanical environment, seeding conditions, and the ECM.

[0129] State of the art practices for seeding are typically manual, leading to uncontrolled 3D variability in seed location, number of cells seeded, and the distribution of seeds within a volume. The embodiments herein use a precision seeding workflow that offers well-defined and reproducible seed location, number of seed cells, and tailored ECM. This disclosure considers a prototype system with capacity to seed and grow more than 100 tumor models per day. Robotic handling systems move the tumor models from seeding to the incubator and between incubator and measurement locations. This manufacturing automation is integrated with multi-modal in situ imaging to accumulate a digital record of growth and to verify tumor model size and viability at the end of production.

[0130] The second technical pillar is multi-modal chemical and physical sensing. Tumor models are typically characterized by their physical characteristics (size, shape, volume, number of cells), viability (live-dead cell fraction), molecular expression of specific biomarkers, and function. These measurements are usually achieved using (immuno)stains with confocal fluorescence microscopy; however, staining prohibits the fast in-situmeasurements required in an automated manufacturing process. The embodiments herein can use an imaging workflow that is specifically suited to monitoring and process control for MTMs using label-free sensing to acquire information rich measurements in a stable and non- perturbative manner. Brightfield imaging rapidly (-seconds) measures tumor model shape and localizes its position.

[0131] Optical microscopy generally lacks 3D imaging capability, precluding some analyses such as finding nuclei to count cells, and is also generally insensitive to chemical composition. Hyperspectral Stimulated Raman Scattering microscopy (SRSM) can be used for high spatial resolution (-250 nm) and depth sectioning that provides vibrational spectroscopic sensitivity to biochemical markers or cell growth phases, changes in lipids, and other metabolic pathways. However, this extensive data content makes SRSM too slow to measure every MTM more than a few times per growth period.

[0132] Hence, the embodiments herein can use Near Infrared Spectroscopic Tomography (NIRST) as a rapid, non-perturbative technique to record 3D structure using principles of optical coherence tomography (OCT) with chemical specificity from IR absorption spectroscopy. Additionally, the embodiments herein can use software that uses NIRST spectral data to predict more conventional measurements of chemical (SRSM) and molecular (fluorescent) biomarkers using deep neural style transfer. The optical label-free chemical and physical measurements are thus related to the typical characteristics of MTMs in a rapid, process-compatible manner.

[0133] The third technical pillar is software intelligence to predict and direct MTM growth. The embodiments herein can apply methods to predict the tumor model end point, and then apply models that optimize growth conditions. Growth conditions may include any measured or controlled process parameter, selectable input to the system, or intervention. Chemical or other interventions may be introduced in different quantities and at different times in the growth process; the quantity and time of intervention as well as the combination of different interventions presents a large number of possible permutations of growth conditions. The First example system provides prediction intelligence to predict MTM size and shape for MCF-7 cells. The training data set in some embodiments includes -6,000 MTMs with various growth conditions and -77 label-free measurements per MTM. The objective functions for the prediction model can include size, viability, and chemical composition as well as their response to functional tests. The prediction model architecture supports multimodal inputs and multi -objective functions trained to predict the final characteristics of a tumor at any state of the growth process with quantified uncertainty. TheSecond and Third example systems leverage the prediction models of the First example system to optimize or otherwise improve growth conditions. The prediction models learn how various conditions affect growth of various types of MTMs. The Second example system includes a Model Predictive Controller (MPC) architecture, which combined with the prediction model, optimizes tumor model growth conditions. The Third example system expands this into a generalizable software intelligence for rare and individual patient cancers and employs deep reinforcement learning methods that allow for automated discovery of optimal growth conditions, reducing the amount of training data required.

[0134] The embodiments described herein additionally validate the overall process for MTM manufacturing, MTMs of different cell types, and several component technologies that are individually useful and translatable. The overall manufacturing process is validated by demonstrating quantitative performance metrics for production rate, control over the MTM size (variance), and validation of the MTM chemical composition and function. Biological identity and functional validation will include genomic markers, immunofluore scent (IF) markers, and functional challenges that assess response to hypoxia, known drugs or molecular treatments (e.g., with estrogen for ER+ MCF-7 cells), and deuterated compounds to verify function. These measurements ensure the consistency of the MTM production process across all three example systems as new cancer types are introduced or production is scaled.

[0135] The component technologies can also be validated using specific, quantitative metrics. The automated seeding technology can be validated by meeting specifications for seeding rate and consistent control over seeded number of cells. The physical and chemical imaging technologies can be validated against standard samples with known spectral and structural properties. Fast SRSM and NIRST can also be validated for spatial fidelity and biochemical differentiation against labeled confocal fluorescence images to identify cells, quantify nuclei number and physical properties, concordance of live-dead zones in MTM, and accuracy in predicting biomarker expression. The sensing data can be used to develop the software intelligence, with the validation focused on the ability to predict and control MTM growth.

[0136] The system(s) and / or method(s) introduced by the embodiments herein transforms the availability of 3D tumor models by facilitating consistent and reliable production of tumors at scale, with the ability to optimize growth for any type of solid cancer, thereby making such samples accessible for low cost and with minimal local experience in sample culturing. This new capability can overcome obstacles to progress in research and inthe development of therapies, especially for rare diseases, for individuals and for those underrepresented in cancer medicine.

[0137] Table 1 below places such embodiments in the context of the current state of the art, the functionalities provided by the embodiments herein, and options for commercialization. The state of the art for growing in vitro tumors is to develop a specific protocol for each type of cancer cell via a manual, painstaking, and long process. Such a protocol does not readily translate to other cells from the same cancer type, nor does it translate readily to widespread use. Significant resources are required to develop protocols for each new type of cancer and education is needed to transfer a protocol from one laboratory to another.

[0138] The embodiments described herein optimize or otherwise improve MTM growth conditions rapidly to achieve scalable, repeatable, and verified manufacturing of 3D tumor models. This allows tumor model production to be scaled orders of magnitude beyond the state of the art while verifying the properties of every MTM produced.Table 1. Functionalities and comparison with state of the art

[0139] Table 2 below lists five aspects of the embodiments herein.Table 2. Example Aspects

[0140] First, manufacturing automation will integrate and automate seeding, growth, and qualification of tumor models. Previously, tumor model manufacturing was a largely manual process that offered little if any control over seed size and position within the matrix. The automation uses robotic systems that offer precise, repeatable, and rapid seeding, resulting in growth of more than 100 tumor models / day from a single system.

[0141] Second, fast, scalable, and optionally label-free methods for chemical and physical sensing are employed. The embodiments herein include NIRST, a sensing mode that is fast and compatible with the production process described herein. Sensing can be performed in-situ during growth, with each tumor being measured for size, shape, chemical composition, and optionally additional information in less than 1 minute. These fast measurements allow each tumor model to be measured at least daily during production with multiple imaging methods, resulting in more than ~77 measurements during a 15 -day growthperiod. In contrast to alternative methods, these label-free imaging methods are minimally disruptive and utilize contrast in label-free approaches. The embodiments herein apply software intelligence to interpret this data in terms of human-interpretable biochemical measurements (e.g., imagery of labeled samples). This fast, label-free imaging technology and its compatibility with a high throughput process is broadly applicable to other problems in cancer biology and pathology and can benefit many types of cancer research and medicine.

[0142] Third, closed loop manufacturing of MTMs is provided. The embodiments herein apply software methods to track and interpret time series measurement data of MTM growth, predict the tumor end point, and use these predictions to improve growth conditions by directed intervention(s) applied to the MTMs.

[0143] Fourth, measurements of individual MTMs during and at the end of their growth periods enable quality control methods to increase the likelihood that every tumor is viable and meets specified requirements. The embodiments herein include information technology tools such that each tumor model will be accompanied by a digital record of its growth history including morphology and chemical composition. Rigorous quality control of MTMs represents a significant advancement over the state of the art, which provides no objective measure of consistency and no likelihood that the tumor models meet quantitative criteria. Consistent tumor models and knowledge of the full manufacturing history will reduce uncertainty for cancer research and enable new types of precision medicine.

[0144] Fifth, the various components described herein are integrated into a scalable platform technology that can be used for most cancers. This approach and infrastructure allows any type of cancer to be quickly produced at scale, including rare cancer types or cancer types prevalent in underserved populations. The platform’s agility facilitates rapid discovery of optimal or otherwise improved growth conditions for nearly any type of solid cancer, resulting in a significantly improved ability to scale up production of MTMs as well as to provide the level of precision of producing MTMs using cancer cells from individual patients. Such a low-cost, on-demand tumor production service can lower barriers to access, enabling many different types of organizations to pursue cancer research and investigate new cancer diagnostics and therapies. The embodiments herein facilitate that goal, producing hundreds of consistent tumors from rare and patient-specific cancers.

[0145] The First example system of the embodiments herein produces MTMs from MCF-7, a familiar and widely known cell type. The Second example system produces MTMs from a rare breast cancer (TNBC) progression set, consisting of four isogenic cell lines, Ml (MCFlOa), M2 (MCF10aTlk.cl2), M3 (MCFlOCAlh), and M4 (MCFlOCAla.cll), whichare denoted collectively as M1-M4. The Third example system produces MTMs from three rare cancers and three PDX cancers from individual patients.

[0146] The three example systems exhibit increased production rates from one to the next, incubating 100 viable MTMs / day for the Third example system. MTM size variance also decreases from one example system to the next, with less than 10% size variance for the Third example system. The software intelligence predicts the MTM end state from sensor data and uses these predictions to improve growth conditions. The software of the Third example system also optimizes growth conditions and controls tumor model size with error lower than 10%.

[0147] The example system discussed above are summarized in Table 3 below:Table 3. Summary of Example Systems

[0148] 3D cultures, referred to as spheroids or organoids, are gaining increased usage.Several spheroid tumor models are available and factors that determine their growth andphysiology are becoming increasingly known. However, it is challenging to make organoids from a specified set of starting materials, and controlling their composition, growth, and diversity is an active area of study. Organ-on-a-chip models are another emerging area that allow for precise control over conditions and interactions between a tumor and the TME but require significantly more infrastructure and expertise to use. The extant challenge is to rapidly make a widely usable model, highly reproducible in meeting design specifications of composition and physical properties, scalable in numbers and convenient to use.

[0149] Imaging diagnostics of tumor model growth and viability are typically based on immunofluorescence measurements, however, label-free imaging has key advantages, offering information-rich non-destructive measurements through spectroscopy. The key challenges for label-free imaging are spatial and chemical resolution, light scattering limiting the depth of penetration, and throughput. Chemical imaging systems such as SRSM provide label-free methods with applications from live cell imaging to automated histopathology. The embodiments described herein facilitate the integration of SRSM with a high throughput manufacturing process. Compact, alignment-free, fiber laser-based SRSM can be implemented with balanced detection for shot-noise limited performance comparable to solid- state lasers

[0150] Multiple steps in an MTM production process as described herein can be connected via a “digital thread” that organizes process and product data. The embodiments herein create a digital thread for MTMs using computer vision, in-situ monitoring, and ML to interpret and control manufacturing pf 3D tumor models. The embodiments herein include the use of computer vision and ML algorithms to automatically inspect manufactured MTMs and to predict if an MTM meets dimensional specifications and desired other properties.

[0151] First Example System Embodiments

[0152] The embodiments herein include the use of commercial components for cell sorting, ECM dispensing, and liquid dispensing. The components are integrated into a seeding workflow as described below. The embodiments herein involve a 96-well microplate format for culture, which provides a balance between appropriate size and volume for handling and observation and sufficiently large sample numbers per plate. Other embodiments may use other formats. The ECM is dispensed into each well and a controlled number of cells are deposited to form the seed. ECM is deposited on top of the seed such that the seed is sandwiched between an upper layer and a lower layer of ECM. Initial seed sizes for breast cancer are 10, 100, and 1,000 cells per MTM. Other seed sizes are possible in other embodiments. The embodiments herein include configuring an incubator and other physicalequipment to grow the MTMs. Tumor models are grown through seeding cells in a PEG- based hydrogel that is functionalized with adhesive peptides that mimic the cell-ECM interactions that are formed with ECM components such as collagen and laminin. Through cell-ECM interaction, the cells anchor themselves in the ECM and undergo cell division to form an MTM. The MCF7 MTMs are grown in an incubator for a period of ~15 days (about 2 weeks), resulting in a mean MTM size of ~ 200 pm. The embodiments herein also include replenishing growth media and introducing chemical interventions from an automated liquid dispenser and performing preliminary tests with the different interventions. The resulting MTMs are used to refine the sensing technology.

[0153] The embodiments herein include fast, label-free sensing that is integrated with the manufacturing process to characterize MTMs using label-free imaging. The label-free sensing methods are brightfield microscopy, Stimulated Raman Scattering microscopy (SRSM) and near-infrared spectroscopic tomography (NIRST). For SRSM, the embodiments herein involve an optical configuration using the optical system design software tools Zemax and Code V. The optical system is well suited to multi-well plates and rapid location and measurement of MTMs. The optical configuration includes a plate loader integrated into a positioning stage, epi-scanning optics, relay optics, balanced detector, and a fiber-based laser. The performance of the fiber-based SRSM can be calibrated against an existing in-house benchtop system for spectral and spatial fidelity as well as assuring equal signal to noise ratio (SNR) in the same scan time (40 ps per pixel) or less. NIRST can be assembled from optical components as per the design described below. The configuration is optimized or otherwise configured to scan MTMs in multi-well plates and integrates a custom near-IR laser into a commercial spectral domain optical coherence tomography (OCT) module with a mid-IR excitation source. Software scans both with and without mid-IR excitation. Performance of the system can be validated for 3D imaging against SRSM, exceeding SRSM depth of penetration while providing a pixel size of ~10 pm. Spectral performance can be validated against mid-IR microscopy using microfabricated samples. Using the diversity of samples in the training data, data can be generated from both instruments that can obtain virtually stained images of MTMs from unstained NIRST data. Finally, a DL framework provides stain-less images of confocal optical microscopy data of chemical and molecular-stained MTMs from NIRST data using neural style transfer.

[0154] The First example system MTMs are grown under various growth conditions to create a dataset that can be used to refine the label-free sensing, train the prediction model, and control the manufacturing process. The MTM micro-environment can be controlled bynourishing the hydrogel cell-culture media along with growth factors and inhibitors that can affect the growth of the MTMs. The First example system focuses on 171 growth conditions, from which the embodiments herein systematically acquire data that spans the structural, chemical, and functional aspects of MTMs. The first set of measurements is done with non- perturbative methods of brightfield microscopy, NIRST, and SRSM. The second set uses a portion of the MTMs and performs destructive testing in the form of immunofluorescence, hematoxylin and eosin-stained, and IR chemical imaging data. Third, functional data is acquired that captures deviation in molecular characteristics or environmental response under conditions determined for each cell type. This data is used to develop a process sensing strategy and to train software intelligence models as well as style transfer software. NIRST data is interpreted using these algorithms to derive the information relevant to MTM growth that is ordinarily only available from multimodal label-free data, labeled data, and SRSM. The predicted images can be validated for accuracy and assessment of quality.

[0155] The embodiments herein include software that creates a unique information object for each MTM and that allows for storage, search, retrieval, and analysis of the MTM data. The information object contains the process data including growth conditions and the time series data collected about the MTM including image and spectroscopic data. The information object also contains features extracted from the image data such as MTM size and chemical composition. The embodiments herein also involve automation such that sensor data is automatically recorded and stored for each MTM. The software additionally provides basic analysis functions including automatic evaluation of MTM size and shape from brightfield imaging and chemical composition from chemical imaging. The software provides a digital certificate of conformance for every MTM, with tumor model size and viability verified by sensing data.

[0156] The embodiments herein also involve training an ensemble of ML models to predict quality characteristics of the tumor models defined as the size and viability of the models with quantified uncertainty. The ML models are trained using initial growth conditions, image data, and chemical information measured as described above. The objective function used to update the model during training can be the error between the predicted quality characteristics and the actual quality characteristics of the grown tumor model. The prediction models are trained on a training set of tumor models with various growth conditions. The performance of the model can be evaluated by the prediction error, the magnitude by which the prediction deviates from the actual size and viability of the tumor models in a production set. The prediction model can be expanded to result in the moresophisticated ML methods used in the Second and Third example systems. One result of the ML model development includes a list of useful sensing data. These are used to refine the measurement strategy to reduce the number of measurements by reducing redundant data streams, pixel density, and spectral content (adaptive sampling).

[0157] The embodiments herein also demonstrate production rate and repeatability by imaging and growing ~10 microplates (960 MTMs) as described below. The sizes of the MTMs are verified using brightfield microscopy, and the MTM viability is confirmed using spectroscopy and other validation techniques described below. The MTM size variance is < 30%.

[0158] The embodiments herein systematically acquire data that spans the structural, chemical, and functional aspects of tumor models. First, a baseline performance is established for MCF-7 cells cultured as per known experimental conditions. This baseline data includes structural characterization growing spheroids (brightfield, NIRST, SRSM, live-dead assay) over a 15 day period. RNA-seq data is also obtained under the growing conditions for 4 functional and molecular challenges. RNA-seq data is used to determine genes that characterize the cell lines under all conditions and those associated with changes for each functional challenge. These are used to characterize MTMs with real-time quantitative polymerase chain reaction (RT-qPCR) and single molecule fluorescence in situ hybridization (smFISH). Second, MTMs made under different conditions are characterized. MTMs grown under design of experiments’ determined condition are characterized for structural parameters (MTM size, shape, and number of cells from brightfield imaging and NIRST) and for chemical composition (SRSM, IR imaging and live-dead assay).

[0159] For chemical composition, NIRST data is compared to 3D culture of cells. The top 30 conditions are subjected to functional challenges. Structural and chemical composition are measured for each of these conditions for each challenge. The top 5 conditions from these experiments undergo functional challenges. In addition to structural and chemical measurements, Immunofluorescence and smRNA-FISH followed by confocal imaging and data analyses are conducted. These experiments are used to provide optimal (or otherwise improved) growth conditions. MTMs grown under optimal conditions are subject to each of the characterization methods above as well as RNA-seq. Finally, tumors in production are characterized by structural and chemical composition. A digital thread is populated from recorded data.

[0160] Second Example System Embodiments

[0161] The embodiments herein involve protocols for rapid, label-free sensing for M1-M4 cell types and their MTMs that allow completion of acquisition for training using brightfield, SRSM, and NIRST data. Activities include initial imaging, and assessment of SNR that provides data quality comparable to MCF-7 for the First example system. The embodiments herein also involve a signal processing strategy for SRSM that digitizes signals instead of lock-in detection, which allows digital filtering and estimation to increase SNR. A reduction in scan time from 40 ps per pixel is achieved, which is integrated into the acquisition software. NIRST data from the First example system is optionally used to determine optimal scan conditions and further refine parameters. Using the diversity of samples in the training data, data are generated from both instruments that can obtain virtually stained images of MTMs from unstained NIRST data.

[0162] The Second example system MTMs are grown under the growth conditions determined by the above approach to create a dataset that can be used to develop the label- free sensing, train the prediction model, and control the manufacturing process. The Second example system focuses on 423 growth conditions per cell type. Sensor and process data are collected as described above for each cell line.

[0163] The embodiments herein include training the ML architecture developed as described above on the expanded MTM dataset collected according to the methods described above. A model predictive controller (MPC) uses the prediction model to select optimal (or otherwise improved) growth conditions and interventions for each tumor model. The MPC controller uses the initial conditions of the MTMs and predictive model to simulate intervention strategies and select growth conditions such as seeding density and growth promoters or inhibitors to grow MTMs to a desired size. The performance of the controller is quantified as the error between the desired state of the tumor model and the final state of the grown tumor model.

[0164] The embodiments herein seed and grow ~10 microplates (960 MTMs) from each cell type as described below. The sizes of the MTMs are verified using brightfield microscopy, and the MTM viability is confirmed using spectroscopy and other validation techniques described below. The MTM size variance is < 20%.

[0165] The embodiments herein systematically acquire data that spans the structural, chemical, and functional aspects of tumor models. First, a baseline performance is be established for M1-M4 cells and cells from PDX models cultured as per known experimental conditions. This baseline data includes structural characterization growing spheroids (brightfield, NIRST, SRSM, live-dead assay) over a 15-day period. RNA-seq data is obtainedunder the growing conditions as well as for 3 functional and molecular challenges for each of the cell types. RNA-seq data is used to determine genes that characterize the cell lines under all conditions and those associated with changes for each functional challenge. These are used to characterize MTMs with real-time quantitative polymerase chain reaction (RTqPCR) and single molecule fluorescence in situ hybridization (smFISH). Second, MTMs made under different conditions are characterized that are grown under the design of experiments’ determined condition. Structural parameters (MTM size, shape and number of cells from brightfield imaging and NIRST) and chemical composition (SRSM, NIRST and live-dead assay) are characterized. For chemical composition, NIRST data is compared to SRSM data. The top 30 conditions are subjected to functional challenges. Structural and chemical composition is measured for each of these conditions for each challenge. The top 5 conditions from these experiments again undergo functional challenges.

[0166] In addition to structural and chemical measurements, confocal antibody imaging and smRNA-FISH is conducted. MTMs grown under optimal conditions are subjected to each of the characterization methods as well as RNA-seq. Additional characterization is conducted for PDX models and cells derived therefrom. RNA-seq is performed on 36 cultured cell samples from four PDX models, each grown in three different culture conditions in technical triplicates to select markers for growth conditions. From these studies, three rare PDX models are selected and RNA-seq conducted in triplicate for baseline data to optimize growth conditions. Finally, tumors in production are characterized for structural and chemical composition. A digital thread is populated from recorded data.

[0167] Third Example System Embodiments

[0168] The embodiments herein involve protocols for rapid, label-free sensing for MTMs generated by the Third example system by using brightfield, SRSM, and NIRST. Activities include initial imaging, assessment of SNR and data quality that provides data quality comparable to MTMs generated by the First and Second example systems. Data acquisition is further refined by fast, on-digitizer processing. A further reduction in scan time from 40 ps per pixel is achieved, which provides a cushion for imaging samples of increased complexity. For PDX models, data is checked to determine whether more than one cell type is present in the MTM and grows differentially. Cells are obtained for pilot projects and evaluations conducted for baseline and training data under different conditions. Evaluation of NIRST data is conducted to determine optimal scan conditions and further refine parameters, if necessary, for each cell type. Using the diversity of samples in the training data for the newcell types, data is generated from both instruments that can obtain virtually stained images of MTMs from unstained NIRST data.

[0169] Another task includes growing the patient-derived MTMs under various growth conditions to create a dataset that is usable to develop the label-free sensing, train the prediction model, and control the manufacturing process. The Third example system focuses on up to 423 growth conditions per cell type as described below. Sensor and process data is collected as described below.

[0170] A further task includes growing rare cancer MTMs under various growth conditions to create a dataset that is usable to develop the label-free sensing, train the prediction model, and control the manufacturing process. The Third example system focuses on up to 423 growth conditions per cell type as described below. Sensor and process data are collected as described below. The embodiments herein also employ deep reinforcement learning (RL) to find optimal or otherwise improved growth conditions in fewer experiments than in the training set. The RL architecture is trained with a simplified training set of fewer than 50 growth conditions. The RL model explores and improves MTM growth by selecting additional growth conditions to test such as seeding density or the addition of growth promoters or inhibitors. Optimal growth conditions are discovered and deployed by the RL model to control MTM growth. Model performance is calculated as the error between the desired state of the tumor model and the final state of the grown tumor model. The control strategy is deployed to optimize the production of cancers by the Third example system.

[0171] The embodiments herein also demonstrate the production rate and repeatability for each cell type as described below. The sizes of the MTMs are verified using brightfield microscopy, and the MTM viability is confirmed using spectroscopy and other validation techniques as described below. The MTM size variance is < 10%.

[0172] The embodiments herein systematically acquire data that spans the structural, chemical, and functional aspects of the six types of rare and patent-specific MTMs. MTMs grown under different conditions, using the design of experiments’ determined conditions for each cell type, are characterized. Structural parameters (MTM size, shape, and number of cells from brightfield imaging and NIRST) and chemical composition (SRSM, NIRST, and live-dead assay) are characterized. For chemical composition, NIRST data is compared to SRSM data. The top 30 conditions are subjected to functional challenges. Structural and chemical composition is measured for each of these conditions for each challenge. The top 5 conditions from these experiments undergo functional challenges.

[0173] In addition to structural and chemical measurements, confocal antibody imaging and smRNA-FISH are conducted. Each of the MTM types, grown under optimal conditions, is subjected to each of the characterization methods as well as RNA-seq. Finally, tumors in production are characterized for structural and chemical composition. A digital thread is populated from recorded data. Additionally, PDX models are assessed for their fidelity at predicting human responses. RNA-seq analysis and whole genome sequencing of 20 patient tumor samples at 60x coverage is performed and analyzed to determine the genomic characteristics of the tumors to further identify potential therapeutic strategies to be tested in vitro.Example Applications

[0174] The embodiments herein can be used to manufacture 3D tumors of a wide range of types, even of rare cancers. The technology makes MTMs widely accessible to anyone, making 3D tumor models as easy to use and cost-effective as 2D cultures were previously. The embodiments herein implement an MTM production system that results in ready availability of consistent, well-characterized tumor models to all laboratories, democratizing cancer research by providing 3D tumor models to anyone. This facile access also increases the resource efficiency of biomedical research, replacing manual and laborious processes with tumors inexpensively produced at scale. The standardization of 3D tumor models across laboratories can result in improved collaborations and repeatability of cancer research. The lasting improvements from the embodiments herein will be especially relevant for rare cancers, where diverse biological investigations are often stymied by limited resources and know-how needed to develop models each study ah initio.

[0175] A significant and lasting impact of the embodiments herein can be in permanently increasing the standards of reproducibility and reliability in research with tumor models. The combination of physically consistent MTMs and a digital record of their growth will greatly increase the quality of future studies on these counts. The embodiments herein will result in digital infrastructure allowing 3D tumor model data to be collected and shared across cancer types and laboratories, helping establish a new standard for data sharing, consistency in use and translational impact.Table 4. Technologies and Products

[0176] Embodiments herein provide a set of technologies that can be adapted for scalable production of MTMs for any solid tumor. The embodiments describe herein are also flexible in that it can be configured for production at different scales and in different environments. For example, the technology can be configured as an MTM “mega-factory,” as a Resource Core in institutional shared facilities or clinical settings, or scaled down for individual laboratories.

[0177] The embodiments herein allow a researcher to start from cells of any cancer type, rapidly identify growth conditions, and produce thousands of MTMs to support clinical research and drug discovery. The Third example system incubates samples of five breast cancer cell lines and six rare or patient-specific cancers, making them readily available to others. These and other MTMs can be provided as a research service.

[0178] The embodiments described herein can be used to sells 3D tumor models accessible as a service.Implementation Details

[0179] Figure 5 provides an overview of aspects of the example embodiments described herein, which are designed to make progress in individual areas and coordinate amongst them to achieve tangible milestones in each example systems, while mitigating risk as the technology matures. It starts from the specifications of the tumor to be manufactured, specific cell type(s) and their associated biomedical knowledge (Cancer Types). These activities progressively mature underlying component technologies and integrate them together for manufacturing automation, sensing, machine intelligence, and validation. These activities use these tools and parameters to achieve increased rate of incubation and quality of MTMs, while opening new opportunities for translation.

[0180] Cancer Types

[0181] The cancer types are chosen to present increasing complexity and challenge, starting from a common breast cancer model, to a rare subtype, to rare cancers, and finally to single patient cancers. The embodiments of the First example system utilize an MCF-7 cell line. This cell line is the prototypical model for the most common form of cancer in women - estrogen receptor (ER)-positive breast cancer. It comes with deep existing knowledge in the scientific community of the key physiologic and functional responses under various conditions, making it an ideal standard for validation of the First example system. The availability of high volume, standard samples of MCF-7 have a high likelihood of being of interest to many researchers also making it a desirable first cell type to produce high throughput MTMs.

[0182] The Second example system expands the First to provide a generalizable and agile manufacturing platform as well as the ability to address cancer across its progression lifecycle. A rare breast cancer subtype, MCF10AT, was selected for this purpose. The foundational cell line in this series is MCF10A (Ml), a non-tumorigenic cell line that arose from benign fibrocystic disease. Ml has been modified to produce isogenic lines with varying degrees of tumorigenicity, mainly through the expression of activated H-Ras. This sequence includes the nontumorigenic MCF10A (“Ml”), the pre-malignant MCF10ATlk.cl2 (“M2”), the early-stage malignant MCFlOCalh (“M3”), and the advanced, metastatic MCFlOCala.cll (“M4”) cell lines. These cell lines exhibit low expressions of ER, PR, and HER2, which makes them valuable models for researching TNBC progression. The Third example system expands on the Second and First to provide a generalizable platform for fabricating MTMs from individual breast cancers well as rare cancer types.

[0183] The embodiments herein explore culture conditions that optimize and control MTM growth. Table 5 below shows specific interventions for breast cancers, which are introduced through the cell culture media and controlled through a process- automated liquid dispensing system.Table 5. Example list of growth promoters and inhibitors

[0184] Reconstituted and decellularized ECMs such as Matrigel are commonly used to culture spheroids, which are expensive for high throughput MTM production and often suffer from batch to batch variability whose effects on cell growth cannot be known. To obtain additional control and reproducibility, the embodiments herein develop a well- characterized ECM that offers chemical and mechanical tunability. The composition and influence on cell growth are assessed combinatorically using the high throughput-sensing parts of the system(s) and optimized to yield typical results obtained with matrigel. Specifically, a PEG-based hydrogel made from cross-linking 4-arm polyethylene glycol)vinyl sulfone (4-arm PEG-V S) and PEG-dithiol is used. The hydrogel isfunctionalized using adhesive peptide ligands with the sequence GRCD. This sequence contains a cysteine residue to allow for covalent attachment to 4-arm PEG-VS. These adhesive peptides mimic the adhesive regions typically found in natural polymers like laminin.

[0185] This approach allows cell-ECM-like interactions without introducing major disruptions in the uniform hydrogel structure, such that the cells can anchor, proliferate, and form three-dimensional structures. The embodiments herein initially use laminin and collagen-derived peptides such as YIGSR, RGDS, and IKVAV peptides. This is then expanded over a wide range of peptides that mimic the interactions formed with other macromolecules such as entactin and fibronectin. Some embodiments also explore more complex peptides that are sensitive to Matrix metalloproteinases (MMPs).

[0186] MTM Seeding, Incubation, and Growth

[0187] Figure 6 shows the workflow to seed tumor models into 96 well microplates. The cell seeding unit prepares the empty microplate for cell seeding and seeds a defined number of cells with position control of ~ 0.1 mm. The system consists of a series of devices for individual operations - a plate carousel, a depth / level sensor for microplate quality check and ECM / media level measurement, an ECM / media dispenser, a controlled-cell number seeder, and a plate labeler - integrated with a rail-mounted robot arm. An empty microplate is retrieved from the carousel, de-lidded, and undergoes depth measurement of each well for verification. Next, chilled ECM is dispensed to a predefined depth.

[0188] After solidification, the single-cell seeder deposits the cells at defined XYZ locations. The seeder may dispense one or many cells to a given location, allowing for testing of different seed sizes and selecting of different seed sizes for different cancer types. After cell dispensing, additional ECM is dispensed to cover the seed, sandwiching the seed, and ensuring that the MTMs are exposed to consistent chemical and hydration conditions. The microplate is then lidded and labeled with a barcode and stored in the incubator. The entire process is contained within a biosafety level 2 enclosure.

[0189] Figure 7 shows the incubation and growth units, which include automated incubators with 1,000 microplate capacity, high-content imager, and chemical dispenser. The microplates move from the seeding unit into the incubator through a covered conveyor belt system. A robot retrieves microplates and delivers them to the Measurement and Intervention Unit for high content imaging, chemical interventions, and quality control. Microplates may be returned to the incubator without intervention or with intervention in the form of replenishing the media or additionally introducing chemical interventions that promote orsuppress tumor model growth. An automated liquid dispenser provides media replenishment and chemical intervention.

[0190] At the end of the growth period, each MTM is evaluated by NIRST imaging and graded (pass or fail) in a final quality check. Microplates with a sufficient number of viable MTMs are lidded and sealed before being removed from the enclosure for delivery. A digital record of the pass / fail status of each well and associated data of all passed MTMs is provided with the microplate.

[0191] Chemical and Physical Sensing of MTMs

[0192] MTMs are analyzed for two purposes: (a) Process sensing to ensure that the MTMs are growing and predicted to achieve design specifications; and (b) Validation of the final product. The embodiments herein use a battery of techniques that provide necessary data to make decisions on markers (sequencing), observation of specific pathways (immunofluorescence), response of individual ells and heterogeneity (FISH) as well as imaging to assess the growing tumor in the manufacturing process (brightfield microscopy, NIRST, SRSM), as shown in Figures 9A and 9B.

[0193] Figures 9A and 9B depict several characterization techniques for breast cancer tumor models and example data. Specifically, in Figure 9A, from left to right: live-dead assay, multiple fluorescence stains, sm RNA-FISH and heat map showing changes in gene expression from RNA-seq. Additionally, in Figure 9B, from left to right: 3D refractive index maps from optical tomography and SRS image with chemical contrast from spectra, chemical images at a plane and 3D data.

[0194] The embodiments herein consider ex-situ validation measurements in three categories: (a) measuring MTM integrity (e.g., 3D size, number of cells, live / dead cells), (b) assessing cell line-specific biomarkers to ensure cellular properties (e.g. formation of cell-cell junctions), and (c) ensuring functional characteristics of the model (e.g. mimicking known responses to drugs). The embodiments herein will use methods suited to the task, including imaging (staining with dyes or immunofluorescence), genomic biomarkers (expression measured by RT-qPCR or RNA-seq) or a combination (RNAfluorescence in situ hybridization or RNA-FISH), as shown in Figure 9A.

[0195] Table 6 below describes the methods to assess MTMs, technical details and results relevant to our systems.Table 6. Destructive techniques to characterize MTMs after formation

[0196] While Table 6 lists the general approach and protocols, specific molecular species measured change based on the cell types incubated by each example system. A set of conditions, using a design of experiments strategy is used to acquire data using these techniques to span the range of possible process conditions. This data is used to train the ML algorithms that determine the optimal seeding, growth, and sensing parameter of the process to grow each cell type. Next, important measurements and their ability to characterize and predict MTM states are determined, including those properties that can be determined by non-destructive imaging.

[0197] For in situ process sensing during tumor growth, the embodiments herein turn to label-free, high molecular and spatial content imaging. These techniques allow significantly more frequent monitoring, do not need reagents that can interfere with the growth process or lose signal (photobleach), and provide multiple molecular signals that can be integrated with the software intelligence of the manufacturing process.

[0198] First, Brightfield optical microscopy offers high resolution and speed to measure a 96-well microplate in under 4 mins. Brightfield imaging of full plates is performed approximately daily.

[0199] Second, SRSM obtains both chemical contrast and depth resolution. The embodiments herein will use a custom-built picosecond SRSM to record large hyperspectral data sets. SRSM can allow label-free nuclear segmentation (analogous to nuclei and actin- stained fluorescent images) as well as physicochemical parameters of MTMs via changes in the Raman scattering spectra. However, SRSM is limited in speed (~40 ps / voxel), wavelength range and the slow tuning rate of the optical parameter oscillator-based laser (minutes) as well as experiment setup time (focusing, alignment) in a conventional microscope. The embodiments herein use a conservative 40 ps pixel dwell time to generate an imaging schedule, leveraging the speed versus the information content (choice of number of bands, number of slices and MTM sampling fraction).

[0200] For process monitoring of MTM, the challenge is enabling a high throughput and alignment-free SRSM microscope. The First example system makes use of a fiber laserbased SRSM with a balanced detection scheme to overcome the higher noise of the laser and increase scan speed. The alignment-free operation of this laser system, the unprecedented wavelength tuning rate (spectral width of 2300 cm’1tunable in 5 ms), and the flexibility offered by an in-house imaging setup tailored to culture plates are able to meet the stringent requirements of the manufacturing pipeline with a much faster scan time of 10 ps / voxel. The Second example system provides enhanced spectral details for training, incorporating physics-informed, deep-leaming-based denoising to obtain scan times of 1-5 ps / voxel. Also, the embodiments herein implement an adaptive sampling strategy wherein fast brightfield images are used to guide SRSM imaging location; successively recorded SRSM bands then guide the imaging system to determine the next band and stay within the MTM during imaging. The SRSM data collected in training the ML for each cell type is used to optimize the number of bands and Z slices (spacing and number) to speed up scan time and decrease data overheads.

[0201] Third, Near-infrared spectroscopic tomography (NIRST) : While the two- photon capability of SRSM can provide depth sectioning, it also limits the size of a voxel to ~1 pm3, necessitating trade-offs to measure whole spheroids in the limited process time proposed herein. For a full, label-free, fast 3D reconstruction of the MTM, a custom imaging setup that couples a tunable low-coherence near-infrared source in a modified spectral domain optical coherence tomography (SDOCT, Thorlabs), as shown in Figure 8, is used.

[0202] Figure 8 depicts the NIRST setup to merge NIR SD-OCT mid-IR photothermal contrast.

[0203] In a tomographic reconstruction of the Z profile (A-Scan), the axial resolution depends on the bandwidth of the laser source, and the maximum depth-of-scan depends on the resolution of the spectrometer (minimum resolvable fringe spacing). Exploiting the high bandwidth of the near-IR light source (1320 nm, 110 nm bandwidth), grating resolution, and low NA focusing objective, an axial and lateral resolution of ~10pm and 5pm respectively is obtained using a fast (20 KHz, A Scan time) InGaS line camera (Sensors Unlimited). The embodiments herein undertake a full optical design for MTM scanning, fabricate the setup, test, and validate against gold standards following the methods for developing a scanning mid-IR imaging system. The scan time is 10 s per MTM or less.

[0204] Estimates of 3D size, cellular texture, size of the necrotic core, and cell migration assessment can be accomplished by this attenuation scattering intensity whereas polarization sensitivity will provide additional image texture and macromolecular orientation around the MTM.

[0205] To obtain enhanced chemical contrast and additional analysis of changes in the MTM microenvironment, the embodiments herein illuminate the sample with a modulated narrow band, tunable, high power spectral density Quantum Cascade Laser (QCL, Daylight Solutions). The resulting photothermal excitation is measured, at an equivalent pixel dwell time of 20 ms / band for 100 mid infrared spectral bands spanning 1000-1450 cm'1.Since the photothermal effect is proportional to the linear IR spectra, the embodiments herein also destructively record MTMs under different conditions with mid-IR spectroscopic imaging using established methods and a recently-developed QCL-imaging system during the training step. From this data, specific frequencies that allow sufficient penetration depth as well as chemical contrast are selected to be used for NIRST by the Third example system.

[0206] The embodiments herein link the strengths of two techniques-namely the deep tissue penetration of near-IR light with the molecular specificity of mid-IR light - through ML. Thus, single NIRST measurement offers rapid spatial and molecular coverage of entireMTMs at cell-level resolutions. Details of the experimental parameters of these techniques and the information content that they offer are summarized in Table 7 below. These results are obtained as per the schedule required for software intelligence. The design of experiments approach to training provides a diversity of measurements under different conditions for each modality.Table 7. Process-compatible technologies used

[0207] While several techniques are described herein, none of them alone can provide all the information needed to evaluate MTMs as they grow. Here, the embodiments herein have carefully curated complementary techniques to assess tissue and understand unique information content from each by making extensive measurements of MTMs in variousstages of growth and under a variety of conditions. By using contemporary ML methods, the strategy herein addresses the challenge of process monitoring with optimal imaging setups and their operating parameters. With sophisticated DL algorithms and the ability to handle large data and computation, a correlation is developed between image contents that describe the same physiology. Broadly termed “image -to-image translation”, this is a dynamic area in computer vision and across almost every imaging modality today.

[0208] DL methods provide powerful tools and sophisticated metrics of validation and assessment of quality to provide realistic reproduction of different image content from one measurement. The embodiments herein use generative adversarial networks (GANs), with CycleGAN capable of performing translations without the need for paired data.

[0209] While the embodiments herein follow methods developed for autofluorescence and infrared SRSM, the problem is relatively simpler in that a single cell type is analyzed with well-controlled variation using instruments with high SNR data. Further, the design of experiments approach to training provides a diversity of measurements under different conditions for each modality and there is validation available in each run by measurements of a subset of MTMs (e.g. 5.2% of samples undergo SRSM measurements).

[0210] The embodiments herein include a translation workflow with two generators: G1 and G2. The first model, Gl, is the virtual staining model that transforms NIRST images into high-resolution and high-content SRSM. Gl follows GAN formulation where its parameters are optimized by minimizing mean square error (MSE) loss and adversarial loss. The second model, G2, generates stained optical microscopy images (array of molecular confocal fluorescence and live-dead). Because the two domains images cannot be paired, the embodiments herein utilize cycle-GAN to train G2. The third generator, G3, maps images from the final domain to intermediate SRSM images as part of cycle-GAN methodology. For calculating the adversarial loss, multi-scale discriminators are used where there are two discriminators that have an identical or otherwise similar layout and they are applied to different image scales. Both generators can use a modified U-Net architecture. Since NIRST and virtual data are generated for every sample, accuracy measurements can continuously be obtained by comparing with the SRSM sampled subset. NIRST and virtual images, after passing the concordance test with SRSM, are used for ML predictions and process control.

[0211] Data acquisition for training and validation

[0212] Figures 9A and 9B show an overview of the data curation and validation strategy. First, for each cell type, the embodiments herein establish a baseline of properties (Figure 9A). RNA-seq analyses of baseline control cultures and functional challenged MTMsprovide the subset of genes (in addition to the ones that are already established by earlier studies) that are characteristic of the MTMs. RNA-seq and RT-qPCR provide quantitative data of the relative expression of genes at the population level. In addition, the embodiments herein perform single molecule RNA-FISH (smRNA-FISH) to analyze the expression and localization of the gene transcripts. These data quantify the subcellular localization and expression of each transcript at a single cell level and quantify the cellular heterogeneity within an MTM.

[0213] After establishing each cell type’s baseline data, the embodiments herein use design of experiments to seed and grow tumors in a variety of conditions and subject tumors grown under each set of parameters to functional challenges (Training Steps in Figure 9B). These experiments determine the optimal seeding conditions for MTMs and also provide the training data for ML workflows. There are three categories of measurements for each condition: (a) structure, composition and viability, (b) organizational fidelity and (c) functional consistency. The first two categories assess MTMs physical size and composition as well as key markers of tumor physiology and viability of cells within the tumor.

[0214] In addition to these traditional measures, the embodiments herein also perform several functional challenges to assure functional consistency. The first functional test assesses the key characteristic of the cell line. For example, MCF-7 is an ER+ line and is treated with 17-beta estradiol (E2), an estrogen derivative to study the upregulation of the estrogen responsive genes. The second functional challenge subjects the MTMs to hypoxic conditions by incubating them at 0.2% O2 for 24 hrs. Depending on the MTM, reactions span from apoptosis and cell cycle halt to cell growth and metastasis with specific pathways implicated in the response. Therefore, the markers of these pathways are used as a quality measure.

[0215] Finally, the third challenge subjects MTMs to deuterated water (D2O) and glucose (d-glucose) and measure with SRSM. Enzymatic assimilation from D2O or dglucose forms carbon-deuterium (C-D) bonds observable in the biologically silent zone (1800- 2800 cm-1) of the Raman spectrum. This facilitates the ability to specifically monitor biosynthesis within MTMs via SRSM or more extensive metabolite analysis in the future. Uptake rate, expression and an induced mild toxic response in cells are used to assess both functional response and heterogeneity in MTMs.

[0216] The Validation steps (Figure 9B) provide for extensive characterization of MTMs for each cell type grown under optimal conditions. Structure, composition, and viability are assessed by imaging measurements to yield MTM size, morphology, number ofcells, texture, size of necrotic core, fraction of live-dead cells, and growth curves. Organizational fidelity is assessed by 3D confocal immunofluorescence imaging using a panel of selected biomarkers. Since the embodiments herein focus on epithelial tumors, the distribution of baso-apical polarity (integrin-alpha 6 and ZO-1), EMT (E-cadherin and N- cadherin or Vimentin), Proliferation (Ki-67 or PCNA), and apoptosis (Caspase-3) markers are assessed for all MTMs.

[0217] Functional fidelity is assessed via a variety of tools. Functional challenges are assessed using relative changes from the baseline condition by gene transcripts for RT-qPCR and smRNA-FISH via molecular functional assessments. For PDX MTMs and those derived from rare cancers (Second and Third example systems), biomarkers are determined from baseline RNA-seq data. Below, some of the biomarkers from the baseline analysis for MCF-7 (First example system) and M1-M4 cells (Second example system) are listed:

[0218] MTMs without any functional challenge (control): genes characteristic of the cell line and the baseline expression level of genes to be studied for functional challenge are determined. The embodiments herein use ERa (ESRI) and PR (Progesterone) gene expression for MCF-7 cells (luminal sub type) and cytokeratin 5, 17 and EGFR for M1-M4 cells (basal-like sub type).

[0219] MTMs under metabolic challenge'. For cells under the deuterated metabolic challenges, cell metabolism and cell stress related pathway markers are monitored, including cell cycle regulators (CDKN1A), DNA damage (GADD45A, DDIT3), mitogenesis (EGR1), heme catabolism (HM0X1), (NFKB transcription regulator (NFKBIA), superoxide degradation (SOD2), cell proliferation kinases (MAPK1, INK) and protein translation regulation (eIF2a).

[0220] MTMs under hypoxia challenge'. Relative expression of genes involved in hypoxia responsive pathways, including angiogenesis (VEGFA), glucose metabolism (GLUT1), anaerobic glycolysis (LDHA), and HIF target IncRNAs (MALAT1 and NEAT1) are determined.

[0221] Molecular treatment challenge'. The First example system is configured to assess the relative expression of genes related to the ER / PR+ MCF-7 cell line includes ERa (ESRI), NRIP1 / RIP140 (ESR-interacting protein), PDZK1 (Cholesterol metabolism), GREB1 (ESR-regulated pathway), PGR (progesterone receptor) and cyclin DI (cell cycle). These genes are induced upon treating MCF-7 cells to estradiol as part of the estrogenic response. For quantifying the differential gene expression between nontumorigenic Ml and M2-M4 cells, the embodiments herein determine the expression of the following genes thatare induced in M2 (IL6, SERPINE2), M3 (CCND2, COX6B2) and M4 (PRSS21, IL36B) cells compared to Ml cells, based on preliminary RNA-seq experiments in 3D cultures.

[0222] Software Intelligence Design

[0223] Table 8 below summarizes software intelligence ML capabilities.Table 8. Machine learning technology

[0224] The First example system develops a model that can predict MTM size with error less than 20%. The prediction is made halfway through the growth period and error is calculated as the difference between the predicted and actual size and viability. The Second example system develops an ML-based controller that can optimize growth conditions for all five breast cancer types. Finally, the Third example system adds additional intelligence to automatically discover optimal growth conditions with fewer experiments and applies this to rare and patient specific cancers.

[0225] Each example system grows and measures MTMs for training and testing. In the First example system, the training data set for MCF-7 common breast tumor models includes three seed sizes (10, 100, and 1,000 cells) and seven intervention types - seven chemical interventions and one case of no chemical intervention (control). The interventions are introduced at the seed stage and remain constant for the entire growth period. The embodiments herein test five different concentrations of each intervention and 21 two- intervention combinations. This plan yields 171 tested conditions.

[0226] Each 96 well microplate has a distribution of three different seed sizes (32 MTMs each) and the same chemical intervention conditions. Thus, the First example systemtraining data set has 5,472 MTMs (32 models X 171 growth conditions). Each MTM is imaged ~ 77 times during growth: 15 times using brightfield; 15 times using NIRST; and on average of 47 times using SRS. The imaging data set includes approximately 421,000 images that can be used to train the ML models.

[0227] The Second example system has additional capabilities, including the ability to undertake chemical interventions during the tumor growth process. This capability allows for more precise control over and design of more sophisticated growth conditions to handle cells of varied potentials. The same seven interventions as in the First example system are tested. Additionally, the growth conditions are introduced at three different times (Days 4, 8, 12). The conditions remain constant after the intervention is introduced. This timed intervention results in 423 growth conditions with five cell types, leading to 67,680 different MTMs and a training data set of 5.2 million images.

[0228] The initial operation of the Third example system follows similar principles as above, although allowing for choosing the type and number interventions appropriate for the cancer types chosen. The Third example system is capable of identifying optimal growth conditions with fewer tests than the Second example system.

[0229] Predicting MTM Size and Viability

[0230] Figure 10 summarizes the prediction model workflow, which is organized into data inputs, ML model selection, and prediction. The data inputs include imaging data (brightfield, NIRST, SRSM), data from MTM validation studies, and data about cell seeding, growth conditions including interventions, and the final state of the MTM. The model is trained to predict the size and viability of the MTM at intermediate times and the final state of the MTM using the data from the initial conditions and intermediate measurements with quantified uncertainty.

[0231] The data inputs include defined features and discovered features. Defined features include the raw image and sensor data along with data about cancer type and growth conditions. Defined features include features derived from known physical, chemical, and biological characteristics extracted from image data such as MTM size, shape, number of cells, and spatial distribution chemical information. The defined features include image to image processing such as NIRST to SRS style transfer which increases imaging content without additional measurement time. Discovered features include data driven features extracted from the MTM process data to enhance the prediction performance of ML. The features are derived through dimensional reduction and data compression methods such as principal component analysis (PCA) or deep learning autoencoders. Defined features mayalso be converted into discovered features using attention-based modules which apply attention maps for adaptive feature refinement.

[0232] The training data is partitioned fortraining and testing purposes: one part trains the models with the other data being withheld to test model performance. The prediction model is trained as an ensemble of ML models selected from various architectures and hyperparameters. Ensemble methods can provide estimates of prediction uncertainty by analyzing the mean and variance of the predictions from multiple models.

[0233] For the individual models, the embodiments herein start with supervised architectures including neural networks, Gaussian process, and nearest neighbor models. Model architectures are adjusted to support multimodal inputs and multi -objective functions to process different data streams and optimize different final state metrics for the MTMs. Different model architectures are evaluated using the same training data to find the highest performing architecture. The models are then improved by testing different weight initialization and by training on different subsets of the training data in which the model is trained on each data stream. Redundant models may be removed via pruning.

[0234] The ML models are trained on high performance computing nodes having eight graphical processing units (GPUs), and the training time for the largest models may take no longer than several hours. Such ML models can be inferenced on lower performance computers integrated with a manufacturing production system.

[0235] The consolidated outputs of the model ensemble produce an MTM quality prediction. The quality prediction includes the tumor size, morphology, and performance on functional tests to ensure that the MTM is viable. The prediction model is trained to predict the end state of the MTM from any process data collected at seeding or any intermediate day. Model uncertainty is quantified from statistical analysis of the ensemble of predictions. The performance for the prediction model is calculated as the magnitude of error between the predicted quality metrics and the measured end state of the tumor model.

[0236] The First example system ML model has a prediction error of less than 20% on day 7 of the MTM growth. The prediction error also forms the objective functions used to train the prediction model. The embodiments herein also contemplate training the model with different quality metrics including the results of functional challenges which can reduce the error of the prediction model. Prediction error and uncertainty for the quality metrics informs how the prediction model can be improved through changes to data collection, feature extraction, or ensemble architecture.

[0237] Optimization and Discovery of Growth Conditions

[0238] The Second example system generates models for five breast cancer cell lines through the transfer learning of First example system models. In transfer learning the knowledge from a task is reused for a new problem allowing for higher performance. The model architectures can be re-used with modest adjustments and trained on new data to predict the growth of the Second example system MTMs in the presence of interventions. Transfer learning allows for the development of effective models for prediction and control of new tumor types with less training data than required for the First example system. Once a prediction model has been trained, the dominant data input streams that allow for accurate prediction can be determined to reduce the types of measurements for new models. Transfer strategies and general knowledge in architecture and hyperparameters also allow for the training of new models for rare cancer by the Third example system faster than the other example systems.

[0239] The Second example system includes a Model Predictive Control (MPC) model as shown in Figure 11 to select growth conditions that maximize the number of viable MTMs of a desired size. MPC directly uses the prediction model of the First example system. MPCs are well suited to control complex systems when provided with a sufficiently accurate prediction model. The ML architecture is trained and tested with the five cell lines incubated by the Second example system. The MPC controller interpolates within the training set and identifies growth conditions that are not represented in the training set. MPC performance is calculated as the error between the desired state of the tumor model and the final state of the grown tumor model. The Second example system controls MTM size with less than 20% error for five different target sizes. The embodiments herein contemplate the potential to update the predictive model with a continual learning process that will evolve the model with new training data from tumor models.

[0240] The Third example system applies deep RL to discover optimal growth conditions with greater data efficiency than MPC. In RL as shown in Figure 11, the model chooses growth conditions and interventions based on the previous system states and receives feedback based on how system states have changed. RL selects experiments balancing exploration and exploitation to find the optimal growth conditions without requiring a comprehensive DOE. These experiments may be conducted in an iterative fashion, with the RL algorithm learning at each iteration and using updated information to choose the next step. Implementation of RL for MTMs allows for the automated discovery of optimal growth conditions for new cancer types with fewer experiments than for MPC. The embodiments herein train the RL model with a simplified DOE, ~50 conditions depending on the numberof available interventions, and then allows the RL model to select follow-up experiments to explore and optimize over the design space. The Third example system determines the optimal growth conditions in fewer experiments than the Second example system and generates MTMs reaching the desired size with < 10% error for five different target sizes.

[0241] Manufacturing Automation Capabilities and Throughput

[0242] Figure 12 shows the lab automation system described herein for scalable manufacturing of tumor models in 96 well plates. This system is designed to support a production rate of 110 plates / day, or approximately 10,000 manufactured tumor models / day. The system is isolated in a biosafety level 2 enclosure and is 36’ by 8’ in size, allowing for complete automation of all the process steps without human intervention including seeding, incubation, sensing, and quality control.

[0243] The system of Figure 12 is divided into a seeding unit (left), an incubation unit(middle), and a growth unit (right). Each unit has a robotic system for manipulating the plates as well as plate -handling devices such as a delidder, sealer, and labeler.

[0244] Table 9 below summarizes the example systems with respect to MTM production rate and variability.Table 9. System Throughput Summary

[0245] The system throughput T is the number of 96-well microplates per day that can pass through the system. The First example system has a relatively low rate because it incorporates several manual processing steps, however the Second and Third example systems have higher rates as they implement fully robotic processes and thus increase the measurement speed. Tumor survival rate S = 50% is achieved by the First example system, increasing to 75% by the Third example system. The MTM population will have a Gaussiandistribution of sizes with standard deviation D = 50% for First-example system MTMs and improving to 25% for Third-example system MTMs. Improvements in S and D proceed as the optimization of growth conditions and control of the MTM manufacturing process is learned. The quality control step specifies that the size variance 0 is less than 30% for the First example system and less than 10% for the Third example system. The rate of success is R = S x erf (^=). Thus, the rate at which qualified MTMs are produced within the acceptable tolerance limit is P = 96 x T x R and the embodiments herein are able produce qualified MTMs at 11 MTM / day for the First example system, 47 MTMs / day for the Second, and 112 MTMs / day for the Third.

[0246] Rare and Patient-Specific Cancer Strategy

[0247] A major advantage of the approach herein is that any type of cancer can be optimized and produced at scale. This capability can open the door to MTMs from single individuals, including those that suffer from rare cancers. First, the embodiments herein consider individual patient cancers that have been grown successfully in mouse models. These PDX tumors form a natural resource to configure and evaluate the technology described herein since their behavior in an in vitro system is characterized, extensive molecular information is available, and they can be reproduced to some level. Thus, the embodiments herein select three genetically- and drug response-distinct tumors to use as test cases for this approach. These are selected to span the range of breast cancer disease. The embodiments herein start cryorecovery of PDX models (First example system), identify best growth conditions and perform in vitro validation of 4 therapeutic drugs (Second example system), and produce PDX models at scale (Third example system).III. Example Methods

[0248] Figure 6 is a flowchart of an example method 1300 for manufacturing 3D biological structures. The method 1300 includes dispensing, by a laboratory system, a plurality of volumetric biological samples into culture medium in respective cell culture containers (1310). The method 1300 additionally includes transporting, by the laboratory system, the cell culture containers to an incubator (1320). The method 1300 additionally includes incubating the plurality of volumetric biological samples in the incubator (1330). The method 1300 additionally includes, subsequent to incubating the plurality of volumetric biological samples in the incubator, using the laboratory system to non-destructively image the plurality of volumetric biological samples to generate first imaging data therefor, wherein the first imaging data represents at least one of physical or chemical information about the plurality ofvolumetric biological samples (1340). The method 1300 additionally includes storing the imaging data in a non-transitory computer-readable medium (1350). The method 1300 could include additional or alternative features.IV. Conclusion

[0249] It should be understood that arrangements described herein are for purposes of example only. As such, those skilled in the art will appreciate that other arrangements and other elements (e.g., machines, interfaces, operations, orders, and groupings of operations, etc.) can be used instead, and some elements may be omitted altogether according to the desired results. Further, many of the elements that are described are functional entities that may be implemented as discrete or distributed components or in conjunction with other components, in any suitable combination and location, or other structural elements described as independent structures may be combined.

[0250] While various aspects and implementations have been disclosed herein, other aspects and implementations will be apparent to those skilled in the art. The various aspects and implementations disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope being indicated by the following claims, along with the full scope of equivalents to which such claims are entitled. It is also to be understood that the terminology used herein is for the purpose of describing particular implementations only, and is not intended to be limiting.IV. Enumerated Example Embodiments

[0251] Embodiments of the present disclosure may thus relate to one of the enumerated example embodiments (EEEs) listed below. It will be appreciated that features indicated with respect to one EEE can be combined with other EEEs.

[0252] EEE 1 is a method for manufacturing 3D biological structures, the method including: (i) dispensing, by a laboratory system, a plurality of volumetric biological samples into culture medium in respective cell culture containers; (ii) transporting, by the laboratory system, the cell culture containers to an incubator; (iii) incubating the plurality of volumetric biological samples in the incubator; (iv) subsequent to incubating the plurality of volumetric biological samples in the incubator, using the laboratory system to non-destructively image the plurality of volumetric biological samples to generate first imaging data therefor, wherein the first imaging data represents at least one of physical or chemical information about the plurality of volumetric biological samples; and (v) storing the imaging data in a non-transitorycomputer-readable medium.

[0253] EEE 2 is the method of EEE 1, wherein the plurality of volumetric biological samples are disposed in respective wells of a sample plate, wherein incubating the plurality of volumetric biological samples in the incubator comprises incubating the sample plate in the incubator during a first period of time, wherein the first imaging data comprises first image data of a target volumetric biological sample in a target well, and wherein the method further comprises: (i) based on the first image data, determining a first intervention; (ii) providing, by the laboratory system, the first intervention to the target well; (iii) transporting, by the laboratory system after providing the first intervention, the sample plate to the incubator; and (iv) subsequently incubating the sample plate in the incubator during a second period of time.

[0254] EEE 3 is the method of EEE 2, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein a sample well of the sample plate other than the target well contains cells of a second cancer type that differs from the first cancer type.

[0255] EEE 4 is the method of any of EEEs 2-3, further comprising: (i) transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device; (ii) using the imaging device, non-destructively imaging the target well to generate second image data of the target volumetric biological sample in the target well; (iii) based on the second image data, determining a second intervention; (iv) providing, by the laboratory system, the second intervention to the target well; (v) transporting, by the laboratory system after providing the second intervention, the sample plate to the incubator; and (vi) subsequently incubating the sample plate in the incubator during a third period of time.

[0256] EEE 5 is the method of any of EEEs 2-4, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein providing the first intervention to the target well comprises providing a candidate treatment for the first cancer type to the target well.

[0257] EEE 6 is the method of any of EEEs 2-5, wherein the imaging device comprises a brightfield imager and a hyperspectral imager, and wherein non-destructively imaging the target well of the sample plate to generate the first image data comprises: (i) operating the brightfield imager to generate brightfield image data of the target volumetric biological sample ; (ii) based on the brightfield image data, determining a location of the target volumetric biological sample within the target well; and (iii) operating the hyperspectral imager to generate hyperspectral image data of the determined location of the target volumetric biological sample.

[0258] EEE 7 is the method of any of EEEs 2-6, wherein the imaging device comprises a hyperspectral imager, and wherein non-destructively imaging the target well of the sampleplate to generate the first image data comprises: (i) operating the hyperspectral imager to generate first hyperspectral image data of the target well; and (ii) applying the first hyperspectral image data to a first machine learning model to generate a first model output that includes at least one of: (a) second hyperspectral image data of the target well that represents the target well at more wavelengths than the first hyperspectral image data, or (b) a map of a chemical substance within the target well, wherein determining the first intervention based on the first image data comprises determining the first intervention based on the first model output.

[0259] EEE 8 is the method of any of EEEs 2-7, further comprising: (i) determining, based on the first image data, an imaging interval for the target volumetric biological sample; and (ii) at a time that is subsequent to the non-destructive imaging of the target well by a duration of the imaging interval, transporting the sample plate to the imaging device and using the imaging device to non-destructively image the target well to generate second image data of the target volumetric biological sample.

[0260] EEE 9 is the method of any of EEEs 2-8, wherein the imaging device is a first imaging device, and wherein the method further comprises: (i) determining, based on the first image data, that a second imaging device should be used to image the target well; and (ii) responsively using the second imaging device to non-destructively image the target well the sample plate to generate second image data of the target volumetric biological sample, wherein using the first imaging device to generate the first image data takes a first amount of time, and wherein using the second imaging device to generate the second image data takes a second amount of time that is greater than the first amount of time.

[0261] EEE 10 is the method of any of EEEs 2-9, further comprising: applying the first image data to a second machine learning model to generate a first latent vector, wherein the first latent vector represents the first image data in a multidimensional latent vector space, and wherein determining the first intervention based on the first image data comprises determining the first intervention based on the first latent vector.

[0262] EEE 11 is the method of EEE 10, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein the second machine learning model was trained using image data of volumetric biological samples that comprise a set of cancer types that does not include the first cancer type.

[0263] EEE 12 is the method of any of EEEs 10-11, wherein using the imaging device to non-destructively image the target well to generate first image data comprises operating a brightfield imager to generate brightfield image data of the target volumetric biological sample, wherein the method further comprises: based on the brightfield image data, determining amorphological feature of the target volumetric biological sample, wherein applying the first image data to the second machine learning model to generate the first latent vector comprises applying the brightfield image data and the morphological feature to the second machine learning model.

[0264] EEE 13 is the method of any of EEEs 10-12, wherein determining the first intervention based on the first image data comprises: (i) based on the first image data and information about a second intervention provided to the target well prior to generating the first image data, determining a first predicted growth trajectory of the target volumetric biological sample; and (ii) determining the first intervention based on the first predicted growth trajectory.

[0265] EEE 14 is the method of EEE 13, wherein determining the first intervention based on the first predicted growth trajectory comprises: (i) based on the first predicted growth trajectory, determining a candidate intervention; (ii) based on the first image data and information about the second intervention, determining a second predicted growth trajectory of the target volumetric biological sample assuming that the candidate intervention was provided to the target well; and (iii) determining the first intervention based on the second predicted growth trajectory.

[0266] EEE 15 is the method of any of EEEs 13-14, further comprising: (i) transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device; (ii) using the imaging device, non-destructively imaging the target well to generate second image data of the target volumetric biological sample; (iii) based on the second image data, information about the second intervention, and information about the first intervention, determining an updated predicted growth trajectory of the target volumetric biological sample; (iv) determining a third intervention based on the updated predicted growth trajectory; and (v) providing, by the laboratory system, the third intervention to the target well.

[0267] EEE 16 is the method of any of EEEs 10-14, wherein applying the first image data to the second machine learning model to generate the first latent vector is performed by a controller of the laboratory system, and wherein determining the first intervention based on the first predicted growth trajectory further comprises: (i) transmitting, from the controller to a server that is remote from the laboratory system, an indication of the first latent vector, wherein determining the first intervention based on the first latent vector is performed by the server; and (ii) transmitting, from the server to the controller, an indication of the first intervention.

[0268] EEE 17 is the method of EEE 16, wherein the server determining the first intervention based on the first latent vector comprises the server applying the first latent vector to a third machine learning model, and wherein the method further comprises: (i) receiving,from the controller of the laboratory system, a second latent vector that represents, in the multidimensional latent vector space, second image data of the target volumetric biological sample generated after providing the first intervention to the target well; (ii) receiving, from an additional controller of an additional laboratory system that is remote from the laboratory system, third and fourth latent vectors that represent, in the multidimensional latent vector space, third and fourth image data of a second target volumetric biological sample generated before and after, respectively, providing a third intervention to a second target well that contains the second target volumetric biological sample; and (iii) based on the first, second, third, and fourth latent vectors and information about the first and third interventions, training the third machine learning model to generate an updated machine learning model.

[0269] EEE 18 is the method of EEE 17, further comprising: (i) receiving, from the controller of the laboratory system, a fifth latent vector that represents, in the multidimensional latent vector space, fifth image data of as third target volumetric biological sample; (ii) applying, by the server, the fifth latent vector to the updated machine learning model to determine a fourth intervention; (iii) transmitting, from the server to the controller, an indication of the fourth intervention; and (iv) providing, by the laboratory system, the fourth intervention to a third target well that contains the third target volumetric biological sample, wherein the third target volumetric biological sample comprises cells of a second cancer type, and wherein the updated machine learning model was trained using image data of volumetric biological samples that comprise a set of cancer types that does not include the second cancer type.

[0270] EEE 19 is the method of any of EEEs 2-18, wherein determining the first intervention based on the first image data comprises applying the first image data or a representation thereof to a deep reinforcement learning agent, and wherein the method further comprises: (i) transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device; (ii) using the imaging device, non-destructively imaging the target well to generate sixth image data of the target volumetric biological sample in the target well; (iii) updating the deep reinforcement learning agent by applying the sixth image data or a representation thereof to the deep reinforcement learning agent to determine a fifth intervention; and (iv) providing, by the laboratory system, the fifth intervention to the target well.

[0271] EEE 20 is the method of EEE 19, wherein the first intervention and the fifth intervention are different interventions.

[0272] EEE 21 is the method of any of EEEs 19-20, wherein the target volumetricbiological sample comprises cells of a first cancer type, wherein the deep reinforcement learning agent was, prior to applying the first image data or a representation thereof thereto, trained using image and intervention data of volumetric biological samples that comprise a set of cancer types that does not include the first cancer type.

[0273] EEE 22 is the method of any of EEEs 2-21, further comprising: generating, for each well of the sample plate, a record of growth of a respective volumetric biological sample of the plurality of volumetric biological samples, wherein generating a record for the target volumetric biological sample comprises generating the record based on the first image data, a timing of imaging the target volumetric biological sample to generate the first image data, the first intervention, and a timing of providing the first intervention.

[0274] EEE 23 is the method of any preceding EEE, further comprising: (i) using the laboratory system to non-destructively image the plurality of volumetric biological samples a plurality of additional to generate additional imaging data therefor, wherein the additional imaging data represents at least one of physical or chemical information about the plurality of volumetric biological samples; and (ii) storing the additional imaging data in the non-transitory computer-readable medium.

[0275] EEE 24 is the method of any preceding EEE, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises seeding respective specified numbers of a single type of cell in each of the cell culture containers.

[0276] EEE 25 is the method of any of EEEs 1-23, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises seeding respective specified numbers of two or more types of cells in each of the cell culture containers.

[0277] EEE 26 is the method of EEE 25, wherein the two or more types of cells comprise a first cancer cell type and a second cancer cell type that is different from the first cancer cell type.

[0278] EEE 27 is the method of any of EEEs 1-23, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises disposing one or more contiguous portions of a tissue sample in one of the cell culture containers.

[0279] EEE 28 is the method of EEE 27, wherein disposing one or more contiguous portions of a tissue sample in one of the cell culture containers comprises disposing, in the culture medium in one of the cell culture containers, a slice extracted from the tissue sample.

[0280] EEE 29 is the method of any preceding EEE, further comprising: (i) retrieving the imaging data from the non-transitory computer-readable medium; and (ii) based on the retrieved imaging data, calculating, for at least one of the volumetric biological samples, a sample attribute that includes at least one of a size, a shape, a growth rate, a number of cells, an identification of one or more constituent types of cells, a spatial arrangement of cells and / or cell types, a chemical composition, a spectroscopic characteristic, a number of living cells, or a number of dead cells.

[0281] EEE 30 is the method of EEE 29, further comprising storing, in the non- transitory computer-readable medium, the calculated sample attribute.

[0282] EEE 31 is the method of any preceding EEE, further comprising: applying at least one of the imaging data or the calculated sample attribute to a trained machine learning model to generate, for at least one of the volumetric biological samples, a predicted attribute that includes at least one of a size at a future time, a shape at a future time, a growth rate at a future time, a spatial arrangement of cells and / or cell types at a future time, an identification of one or more constituent types of cells at a future time, a number of cells of each of a plurality of different types at a future time, a spectral signature at a future time, a chemical composition at a future time, a response to functional challenge at a future time, a viability at a future time, a growth rate at a future time, a number of living cells at a future time, or a number of dead cells at a future time.

[0283] EEE 32 is the method of any preceding EEE, further comprising, for a particular sample of the plurality of volumetric biological samples: (i) determining, based on the stored imaging data, one or more locations for further imaging within the particular sample; and (ii) operating a brightfield imaging system to non-destructively image the particular sample at the one or more locations.

[0284] EEE 33 is the method of any preceding EEE, further comprising, for a particular sample of the plurality of volumetric biological samples: (i) determining, based on the stored imaging data, one or more locations and target spectral content for further imaging within the particular sample; and (ii) operating a hyperspectral 3D imaging system to non-destructively imaging the target spectral content of the particular sample at the one or more locations.

[0285] EEE 34 is the method of any preceding EEE, further comprising, for a particular sample of the plurality of volumetric biological samples: (i) determining, based on the stored imaging data, target spectral content for further imaging from the particular sample; and (ii) operating a spectrometer to non-destructively imaging the target spectral content from the particular sample.

[0286] EEE 35 is the method of any preceding EEE, further comprising: (i) obtaining an acceptance criterion, wherein the acceptance criterion specifies at least one of a spatial property, a chemical property, or a functional property of a volumetric biological sample; (ii) determining whether each volumetric biological sample of the plurality of volumetric biological samples meets the acceptance criterion; and (iii) storing, in the non-transitory computer-readable medium, a respective record indicating whether the respective biological sample of the plurality of volumetric biological samples meets the acceptance criterion.

[0287] EEE 36 is the method of any preceding EEE, further comprising: (i) subsequent to generating the first imaging data, transporting, by the laboratory system, the plurality of volumetric biological samples to a data recording device to non-destructively image the plurality of volumetric biological samples to generate second data therefor, wherein the second data comprises at least one of imaging data or chemical data; and (ii) subsequent to generating the second data, incubating the plurality of volumetric biological samples in the incubator.

[0288] EEE 37 is the method of EEE 36, further comprising: based on the first imaging data, determining a data recording interval for at least one sample of the plurality of volumetric biological samples, wherein non-destructively imaging the plurality of volumetric biological samples to generate the second data occurs at a time that is subsequent to generating the first imaging data by the data recording interval.

[0289] EEE 38 is a non-transitory computer readable medium having stored thereon program instructions executable by at least one processor to cause the at least one processor to perform the method of any preceding EEE.

[0290] EEE 39 is a system including: (i) a controller comprising one or more processors; and (ii) a non-transitory computer readable medium having stored thereon program instructions executable by at least one processor to cause the at least one processor to perform the method of any of EEEs 1-37.

Claims

CLAIMSWe claim:

1. A method for manufacturing 3D biological structures, the method comprising: dispensing, by a laboratory system, a plurality of volumetric biological samples into culture medium in respective cell culture containers; transporting, by the laboratory system, the cell culture containers to an incubator; incubating the plurality of volumetric biological samples in the incubator; subsequent to incubating the plurality of volumetric biological samples in the incubator, using the laboratory system to non-destructively image the plurality of volumetric biological samples to generate first imaging data therefor, wherein the first imaging data represents at least one of physical or chemical information about the plurality of volumetric biological samples; and storing the imaging data in a non-transitory computer-readable medium.

2. The method of claim 1, wherein the plurality of volumetric biological samples are disposed in respective wells of a sample plate, wherein incubating the plurality of volumetric biological samples in the incubator comprises incubating the sample plate in the incubator during a first period of time, wherein the first imaging data comprises first image data of a target volumetric biological sample in a target well, and wherein the method further comprises: based on the first image data, determining a first intervention; providing, by the laboratory system, the first intervention to the target well; transporting, by the laboratory system after providing the first intervention, the sample plate to the incubator; and subsequently incubating the sample plate in the incubator during a second period of time.

3. The method of claim 2, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein a sample well of the sample plate other than the target well contains cells of a second cancer type that differs from the first cancer type.

4. The method of claim 2, further comprising: transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device;using the imaging device, non-destructively imaging the target well to generate second image data of the target volumetric biological sample in the target well; based on the second image data, determining a second intervention; providing, by the laboratory system, the second intervention to the target well; transporting, by the laboratory system after providing the second intervention, the sample plate to the incubator; and subsequently incubating the sample plate in the incubator during a third period of time.

5. The method of claim 2, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein providing the first intervention to the target well comprises providing a candidate treatment for the first cancer type to the target well.

6. The method of claim 2, wherein the imaging device comprises a brightfield imager and a hyperspectral imager, and wherein non-destructively imaging the target well of the sample plate to generate the first image data comprises: operating the brightfield imager to generate brightfield image data of the target volumetric biological sample; based on the brightfield image data, determining a location of the target volumetric biological sample within the target well; and operating the hyperspectral imager to generate hyperspectral image data of the determined location of the target volumetric biological sample.

7. The method of claim 2, wherein the imaging device comprises a hyperspectral imager, and wherein non-destructively imaging the target well of the sample plate to generate the first image data comprises: operating the hyperspectral imager to generate first hyperspectral image data of the target well; and applying the first hyperspectral image data to a first machine learning model to generate a first model output that includes at least one of: (i) second hyperspectral image data of the target well that represents the target well at more wavelengths than the first hyperspectral image data, or (ii) a map of a chemical substance within the target well, wherein determining the first intervention based on the first image data comprises determining the first intervention based on the first model output.

8. The method of claim 2, further comprising: determining, based on the first image data, an imaging interval for the target volumetric biological sample; andat a time that is subsequent to the non-destructive imaging of the target well by a duration of the imaging interval, transporting the sample plate to the imaging device and using the imaging device to non-destructively image the target well to generate second image data of the target volumetric biological sample.

9. The method of claim 2, wherein the imaging device is a first imaging device, and wherein the method further comprises: determining, based on the first image data, that a second imaging device should be used to image the target well; and responsively using the second imaging device to non-destructively image the target well the sample plate to generate second image data of the target volumetric biological sample, wherein using the first imaging device to generate the first image data takes a first amount of time, and wherein using the second imaging device to generate the second image data takes a second amount of time that is greater than the first amount of time.

10. The method of any of claims 2-9, further comprising: applying the first image data to a second machine learning model to generate a first latent vector, wherein the first latent vector represents the first image data in a multidimensional latent vector space, and wherein determining the first intervention based on the first image data comprises determining the first intervention based on the first latent vector.

11. The method of claim 10, wherein the target volumetric biological sample comprises cells of a first cancer type, and wherein the second machine learning model was trained using image data of volumetric biological samples that comprise a set of cancer types that does not include the first cancer type.

12. The method of claim 10, wherein using the imaging device to non- destructively image the target well to generate first image data comprises operating a brightfield imager to generate brightfield image data of the target volumetric biological sample, wherein the method further comprises: based on the brightfield image data, determining a morphological feature of the target volumetric biological sample, wherein applying the first image data to the second machine learning model to generate the first latent vector comprises applying the brightfield image data and the morphological feature to the second machine learning model.

13. The method of claim 10, wherein determining the first intervention based on the first image data comprises:based on the first image data and information about a second intervention provided to the target well prior to generating the first image data, determining a first predicted growth trajectory of the target volumetric biological sample; and determining the first intervention based on the first predicted growth trajectory.

14. The method of claim 13, wherein determining the first intervention based on the first predicted growth trajectory comprises: based on the first predicted growth trajectory, determining a candidate intervention; based on the first image data and information about the second intervention, determining a second predicted growth trajectory of the target volumetric biological sample assuming that the candidate intervention was provided to the target well; and determining the first intervention based on the second predicted growth trajectory.

15. The method of claim 13, further comprising: transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device; using the imaging device, non-destructively imaging the target well to generate second image data of the target volumetric biological sample; based on the second image data, information about the second intervention, and information about the first intervention, determining an updated predicted growth trajectory of the target volumetric biological sample; determining a third intervention based on the updated predicted growth trajectory; and providing, by the laboratory system, the third intervention to the target well.

16. The method of claim 10, wherein applying the first image data to the second machine learning model to generate the first latent vector is performed by a controller of the laboratory system, and wherein determining the first intervention based on the first predicted growth trajectory further comprises: transmitting, from the controller to a server that is remote from the laboratory system, an indication of the first latent vector, wherein determining the first intervention based on the first latent vector is performed by the server; and transmitting, from the server to the controller, an indication of the first intervention.

17. The method of claim 16, wherein the server determining the first intervention based on the first latent vector comprises the server applying the first latent vector to a third machine learning model, and wherein the method further comprises: receiving, from the controller of the laboratory system, a second latent vector that represents, in the multidimensional latent vector space, second image data of the targetvolumetric biological sample generated after providing the first intervention to the target well; receiving, from an additional controller of an additional laboratory system that is remote from the laboratory system, third and fourth latent vectors that represent, in the multidimensional latent vector space, third and fourth image data of a second target volumetric biological sample generated before and after, respectively, providing a third intervention to a second target well that contains the second target volumetric biological sample; and based on the first, second, third, and fourth latent vectors and information about the first and third interventions, training the third machine learning model to generate an updated machine learning model.

18. The method of claim 17, further comprising: receiving, from the controller of the laboratory system, a fifth latent vector that represents, in the multidimensional latent vector space, fifth image data of as third target volumetric biological sample; applying, by the server, the fifth latent vector to the updated machine learning model to determine a fourth intervention; transmitting, from the server to the controller, an indication of the fourth intervention; and providing, by the laboratory system, the fourth intervention to a third target well that contains the third target volumetric biological sample, wherein the third target volumetric biological sample comprises cells of a second cancer type, and wherein the updated machine learning model was trained using image data of volumetric biological samples that comprise a set of cancer types that does not include the second cancer type.

19. The method of any of claims 2-9, wherein determining the first intervention based on the first image data comprises applying the first image data or a representation thereof to a deep reinforcement learning agent, and wherein the method further comprises: transporting, by the laboratory system subsequent to the second period of time, the sample plate to the imaging device; using the imaging device, non-destructively imaging the target well to generate sixth image data of the target volumetric biological sample in the target well; updating the deep reinforcement learning agent by applying the sixth image data or a representation thereof to the deep reinforcement learning agent to determine a fifth intervention; andproviding, by the laboratory system, the fifth intervention to the target well.

20. The method of claim 19, wherein the first intervention and the fifth intervention are different interventions.

21. The method of claim 19, wherein the target volumetric biological sample comprises cells of a first cancer type, wherein the deep reinforcement learning agent was, prior to applying the first image data or a representation thereof thereto, trained using image and intervention data of volumetric biological samples that comprise a set of cancer types that does not include the first cancer type.

22. The method of any of claims 2-9, further comprising: generating, for each well of the sample plate, a record of growth of a respective volumetric biological sample of the plurality of volumetric biological samples, wherein generating a record for the target volumetric biological sample comprises generating the record based on the first image data, a timing of imaging the target volumetric biological sample to generate the first image data, the first intervention, and a timing of providing the first intervention.

23. The method of any of claims 1-9, further comprising: using the laboratory system to non-destructively image the plurality of volumetric biological samples a plurality of additional to generate additional imaging data therefor, wherein the additional imaging data represents at least one of physical or chemical information about the plurality of volumetric biological samples; and storing the additional imaging data in the non-transitory computer-readable medium.

24. The method of any of claims 1-9, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises seeding respective specified numbers of a single type of cell in each of the cell culture containers.

25. The method of any of claims 1-9, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises seeding respective specified numbers of two or more types of cells in each of the cell culture containers.

26. The method of claim 25, wherein the two or more types of cells comprise a first cancer cell type and a second cancer cell type that is different from the first cancer cell type-27. The method of any of claims 1-9, wherein dispensing the plurality of volumetric biological samples into culture medium in respective cell culture containers comprises disposing one or more contiguous portions of a tissue sample in one of the cell culture containers.

28. The method of claim 27, wherein disposing one or more contiguous portions of a tissue sample in one of the cell culture containers comprises disposing, in the culture medium in one of the cell culture containers, a slice extracted from the tissue sample.

29. The method of any of claims 1-9, further comprising: retrieving the imaging data from the non-transitory computer-readable medium; and based on the retrieved imaging data, calculating, for at least one of the volumetric biological samples, a sample attribute that includes at least one of a size, a shape, a growth rate, a number of cells, an identification of one or more constituent types of cells, a spatial arrangement of cells and / or cell types, a chemical composition, a spectroscopic characteristic, a number of living cells, or a number of dead cells.

30. The method of claim 29, further comprising storing, in the non-transitory computer-readable medium, the calculated sample attribute.

31. The method of any of claims 1-9, further comprising: applying at least one of the imaging data or the calculated sample attribute to a trained machine learning model to generate, for at least one of the volumetric biological samples, a predicted attribute that includes at least one of a size at a future time, a shape at a future time, a growth rate at a future time, a spatial arrangement of cells and / or cell types at a future time, an identification of one or more constituent types of cells at a future time, a number of cells of each of a plurality of different types at a future time, a spectral signature at a future time, a chemical composition at a future time, a response to functional challenge at a future time, a viability at a future time, a growth rate at a future time, a number of living cells at a future time, or a number of dead cells at a future time.

32. The method of any of claims 1-9, further comprising, for a particular sample of the plurality of volumetric biological samples: determining, based on the stored imaging data, one or more locations for further imaging within the particular sample; and operating a brightfield imaging system to non-destructively image the particular sample at the one or more locations.

33. The method of any of claims 1-9, further comprising, for a particular sample of the plurality of volumetric biological samples: determining, based on the stored imaging data, one or more locations and target spectral content for further imaging within the particular sample; and operating a hyperspectral 3D imaging system to non-destructively imaging the target spectral content of the particular sample at the one or more locations.

34. The method of any of claims 1-9, further comprising, for a particular sample of the plurality of volumetric biological samples: determining, based on the stored imaging data, target spectral content for further imaging from the particular sample; and operating a spectrometer to non-destructively imaging the target spectral content from the particular sample.

35. The method of any of claims 1-9, further comprising: obtaining an acceptance criterion, wherein the acceptance criterion specifies at least one of a spatial property, a chemical property, or a functional property of a volumetric biological sample; determining whether each volumetric biological sample of the plurality of volumetric biological samples meets the acceptance criterion; and storing, in the non-transitory computer-readable medium, a respective record indicating whether the respective biological sample of the plurality of volumetric biological samples meets the acceptance criterion.

36. The method of any of claims 1-9, further comprising: subsequent to generating the first imaging data, transporting, by the laboratory system, the plurality of volumetric biological samples to a data recording device to non- destructively image the plurality of volumetric biological samples to generate second data therefor, wherein the second data comprises at least one of imaging data or chemical data; and subsequent to generating the second data, incubating the plurality of volumetric biological samples in the incubator.

37. The method of claim 36, further comprising: based on the first imaging data, determining a data recording interval for at least one sample of the plurality of volumetric biological samples, wherein non-destructively imaging the plurality of volumetric biological samples to generate the second data occurs at a time that is subsequent to generating the first imaging data by the data recording interval.

38. A non-transitory computer readable medium having stored thereon program instructions executable by at least one processor to cause the at least one processor to perform the method of any preceding claim.

39. A system comprising: a controller comprising one or more processors; and a non-transitory computer readable medium having stored thereon program instructions executable by at least one processor to cause the at least one processor to perform the method of any of claims 1-37.

Images