Drug effect prediction system

Abstract

To provide a system capable of simply and efficiently predicting medication effect on nerves.SOLUTION: A medication effect prediction system comprises: a micro fluid device having a first culture space for culturing nerve cells and a second culture space for culturing neurites while separated from nerve cells; a storage part in which a learned model based on a plurality of teacher data obtained by associating learning input data based on images of neurites in the second culture extending from each of a plurality of nerve cells cultured in a state of individually adding to culture solution each of a plurality of existing substances and learning output data relating to information of areas of damage to nerve caused by the existing substance; and a determination part for outputting information resulting from a damage occurrence probability by nerve area by culturing nerve cells in a state of adding test substances to culture solution, and applying determination input data based on microscopic images of neurites in the second culture space to the learned model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a system for predicting the effects of drugs, and more particularly to a system for predicting the effects on nerves such as peripheral nerves and central nerves. [Background technology]

[0002] In drug discovery research, there is a phase where toxicity and safety are evaluated in addition to the phase where efficacy is confirmed. If peripheral nerve damage occurs as a result of drug administration, it is undesirable because it has a significant impact on the patient's quality of life (QOL). For this reason, in drug discovery research, it is important to understand in advance the effects (neurotoxicity) on nerves, including peripheral nerves.

[0003] A known method for evaluating neurotoxicity is to form neurites from nerve cells using a microfluidic device and observe drug stimuli and their responses (see, for example, Patent Document 1 below). Patent Document 1 discloses a microfluidic device having channels that can culture neurites (also called "axons") extending from the cell body. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6430680 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the channel in Patent Document 1 has dimensions of 100 μm to 150 μm in width and 100 μm to 200 μm in height. Therefore, the width and height of the channel are larger than the thickness of the neurites, and the dimensions of the width and height are similar. Therefore, when neurons are cultured in this channel, the neurites aggregate to form a bundle-like structure. When neurons are cultured in this manner, it is impossible to individually observe and analyze each neurite extending from the neurons.

[0006] Additionally, while a conventional method for culturing neurons in a petri dish is known, even in this case, the neurons and neurites tend to overlap during culturing. If attempts are made to culture neurons while avoiding this phenomenon, the density of neurites in the petri dish becomes low. Under these conditions, the function of the cells declines, and the cells become increasingly different from those in the living body. Furthermore, variations occur depending on the state of individual neurons. Therefore, there are concerns about using images of neurites cultured using this method to evaluate the effects of drugs.

[0007] Another possible method is to culture neurons at high density to form a neurite network structure. However, even in this case, the neurons and neurites overlap each other, making it difficult to identify the observation site. Furthermore, the pattern of this overlap is irregular with each test, making quantitative evaluation based on the observation images difficult.

[0008] Due to the above circumstances, at present, there is no established method for quantitatively evaluating the effects of drugs on nerves, including peripheral nerves, in vitro.

[0009] In view of the above problems, an object of the present invention is to provide a system that can easily and efficiently predict the effects of drugs on nerves. [Means for solving the problem]

[0010] The drug effect prediction system according to the present invention comprises: a microfluidic device having a first culture space in which neurons can be cultured, and a second culture space extending from the first culture space in which neurites extending from the neurons can be cultured separately from the neurons; a memory unit in which a trained model is recorded, the trained model being generated by machine learning based on a plurality of training data associating training input data based on a microscopic image of a plurality of neurites extending from each of a plurality of neurons cultured in the first culture space of a plurality of the microfluidic devices prepared for training in a state in which a plurality of known substances, each of which causes a nerve damage site to be a known site belonging to a group of candidate damage sites or no damage site, is individually added to the culture solution, with training output data relating to information on the nerve damage site caused by the known substances added to the first culture space in which the nerve cells corresponding to the microscopic image are cultured, or information indicating that no damage site is present; a data input receiving unit that receives input of input data for determination based on a microscopic image of the neurites extending from the neurons and positioned in the second culture space, in which the neurons are cultured in a state in which a test substance whose damage site to the neurons is unknown is added to a culture solution in the first culture space of the microfluidic device prepared for determination; The method is characterized by comprising a judgment unit that applies the judgment input data to the trained model and outputs information derived from the probability of damage occurring at each site belonging to the group of candidate nerve damage sites due to the introduction of the test substance.

[0011] The microfluidic device described above allows neurons to be cultured in the first culture space, and neurites extending from the neurons to be cultured in the second culture space. Therefore, multiple neurites are cultured so that each neurite is aligned in the width direction, rather than forming a bundle structure.

[0012] Therefore, by culturing neurons in this microfluidic device and observing the second culture space in particular, it becomes possible to observe the dense network structure of neurites at each site.

[0013] By using this microfluidic device to culture neurons in the presence of a substance (known substance) whose effect on neurons and the site of its influence are known, the effects of the introduction of the known substance are reflected in information on observation images (microscopic images) of neurites present in the second culture space. Furthermore, because this microfluidic device allows neurites to be cultured in a state separate from the neurons, the effects of the introduction of the known substance on the neurites are highly reproducible. In other words, when neurons are cultured in the microfluidic device with multiple known substances individually added to the culture medium, the microscopic images of the neurites present in the second culture space contain information on the effects of each known substance on the site of influence (damage site).

[0014] Therefore, it becomes possible to perform machine learning by preparing multiple pieces of training data in which data on the microscopic image of the region where neurites exist (within the second culture space) itself, or data obtained by performing predetermined image processing on this microscopic image, is used as training input data, and information on the site of damage caused by a known substance added to the culture space of the nerve cell corresponding to this microscopic image is used as training output data. Note that the known substances may include those that do not cause damage to nerves even when introduced. In this case, data based on the microscopic image of the neurites cultured with the addition of the known substance is used as training input data, and information indicating that no site of damage exists is used as training output data, and training data in which these are associated is prepared.

[0015] A known method can be used as a machine learning method based on training data including the training input data and the training output data. For example, a trained model can be generated by constructing a neural network having an input layer including a plurality of nodes configured with information about a plurality of feature quantities extracted from image data based on the microscopic image, an output layer including a plurality of nodes configured with information about each part belonging to a group of candidate lesion parts and information indicating that no lesion part exists, and one or more intermediate layers each including a plurality of nodes connecting the input layer and the output layer.

[0016] The drug effect prediction system of the present invention includes a memory unit in which such a trained model is recorded. Therefore, in a similar microfluidic device, neurons are cultured with the addition of a test substance, and data based on microscopic images of neurites located in the second culture space is input as input data for determination into the system, allowing the trained model to be applied. In other words, this input data for determination is identical in quality to the training input data used as training data to generate the trained model. Therefore, when applied to the trained model, highly reproducible determination results can be obtained. According to this system, the determination unit outputs information derived from the probability of injury for each site belonging to each candidate injury site group. Note that if the introduction of the test substance results in zero neuronal damage, all of the injury probabilities for each site will be zero.

[0017] As described above, according to the drug effect prediction system of the present invention, all that is required is to culture neurons in a microfluidic device having the above-described structure with a test substance added, and obtain a microscopic image of the neurites located in the second culture space separately from the neurons, and then automatically calculate information derived from the probability of damage to each nerve site (such as the value of the damage probability itself, or a score value based on the probability value) by arithmetic processing based on the data from this microscopic image.These results can then be used to determine the presence or absence of side effects from an early stage of drug discovery, enabling safer drug discovery.

[0018] The group of candidate lesion sites may include axonal lesions, cell body lesions, and myelin lesions.

[0019] The second culture space of the microfluidic device is preferably flat and has a width that allows the neurites extending from the nerve cells to be cultured, but has a height that prevents the nerve cells from invading.

[0020] The second culture space of the microfluidic device may have a width that is 10 times or more its height. Preferably, the second culture space has a flat shape and a width that is 7 times or more the diameter of the neurites. Preferably, the height of the second culture space is 5 μm to 80 μm.

[0021] The nerve cells may be derived from mouse DRG, rat DRG, human ES, or human iPS. [Effects of the Invention]

[0022] According to the present invention, a system is provided that can easily and efficiently predict the effects of drugs on nerves. [Brief explanation of the drawings]

[0023] [Figure 1] 1 is a block diagram showing a schematic configuration of one embodiment of a drug effect prediction system of the present invention. [Figure 2] FIG. 1 is an exploded perspective view of a microfluidic device prior to fabrication. [Figure 3] FIG. 1 is a plan view of a microfluidic device. [Figure 4] FIG. 4 is a cross-sectional view of the microfluidic device shown in FIG. 3 taken along line IV-IV. [Figure 5] FIG. 1 is a plan view schematically showing nerve cells cultured in a microfluidic device. [Figure 6] FIG. 6 is a cross-sectional view of the microfluidic device shown in FIG. 5 taken along the line VI-VI. [Figure 7A] This is a microscopic image of neurites when neurons are cultured in a culture medium containing suramin. [Figure 7B] This is a microscopic image of neurites when neurons are cultured in a culture medium containing oxaliplatin. [Figure 7C] This is a microscopic image of neurites when neurons are cultured in a culture medium containing vincristine. [Figure 7D]This is a microscopic image of neurites when neurons are cultured in a culture medium containing sucrose. [Figure 8] 1 is a diagram schematically showing training data used when generating a trained model. [Figure 9] FIG. 1 is a conceptual diagram of a trained model generated by a model generation unit. [Figure 10] This is a microscopic image of neurites when neurons are cultured in a culture medium containing paclitaxel. [Figure 11] 11 is a table showing the determination results when the data derived from the microscopic image shown in FIG. 10 is input into a drug effect prediction system. DETAILED DESCRIPTION OF THE INVENTION

[0024] An embodiment of a drug effect prediction system according to the present invention will now be described with reference to the drawings, in which: Figure 1 is a block diagram schematically showing the configuration of an embodiment of a drug effect prediction system.

[0025] As shown in Figure 1, the drug effect prediction system 10 includes a data input reception unit 11, a memory unit 13, a determination unit 15, and a display output unit 17. The memory unit 13 stores a trained model M1. Hereinafter, the "drug effect prediction system 10" will be abbreviated as "system 10" as appropriate.

[0026] [Overall system configuration] The system 10 has the function of calculating the possibility that a test substance, the site of nerve damage of which is unknown, will cause damage to a nerve site. The following describes how to use the system 10. The following explanation uses peripheral nerves as an example, but a similar discussion can be applied to the central nerve. Note that when the term "nerve" is used simply in this specification, it encompasses both peripheral and central nerves.

[0027] As shown in FIG. 1, an observation image of neurites 42 cultured in a culture medium to which a test substance 48 has been added is acquired by a microscope 50, and data based on this observation image (hereinafter referred to as "determination input data i42t") is input to a data input receiving unit 11. The data input receiving unit 11 is an interface for inputting image data. The data input receiving unit 11 is equipped with a function for converting the input image data into a data format usable within the system 10, as necessary.

[0028] 2, neurons 41 are cultured in a culture medium to which a test substance 48 has been added, and an image of a region including neurites 42 extending from the neurons 41 is captured by a microscope 50. This microscopic image itself, or data obtained by subjecting this microscopic image to image processing, is input as determination input data i42t into the system 10 via the data input receiving unit 11.

[0029] The determination unit 15 applies this determination input data i42t to the trained model M1 recorded in the storage unit 13. The trained model M1 is a model generated by machine learning based on multiple training data sets that associate previously acquired microscopic images of locations where other neurites 42 exist with information about the lesion site indicated by the microscopic images. That is, a neural network is constructed in the trained model M1, with input data derived from the microscopic images of the neurites 42 as the input side and information about the lesion site occurring in the neurites 42 as the output side.

[0030] Therefore, by applying the input judgment input data i42t to the trained model M1, it is possible to probabilistically output, from the information appearing in this judgment input data i42t, what type of damage may have occurred in the neurite 42. The judgment unit 15 outputs this judgment result to the display output unit 17, which then displays the result.

[0031] The judgment unit 15 is a calculation processing means, which applies the judgment input data i42t to the trained model M1, performs calculations using the trained model M1, and outputs the results, and is composed of, for example, a CPU or MPU. The memory unit 13 is a storage medium that stores predetermined information including the trained model M1, and is typically a non-volatile memory such as a flash memory, or a hard disk. The display output unit 17 is, for example, a monitor.

[0032] However, information relating to the determination result by the determination unit 15 may be transmitted to a remote location via a communication line such as the Internet (not shown).

[0033] [Microfluidic device configuration] Next, the configuration of the microfluidic device 30 used to generate the trained model M1 and acquire the judgment input data i42t will be described with reference to the drawings. Note that the following drawings are merely schematic illustrations, and the dimensional ratios in the drawings do not necessarily match the actual dimensional ratios. Furthermore, the dimensional ratios between the drawings do not necessarily match.

[0034] Fig. 2 is an exploded perspective view of a microfluidic device 30 before completion. The microfluidic device 30 has a first substrate 31 and a second substrate 32, which are manufactured by bonding them together. Fig. 2 corresponds to a perspective view showing both substrates (31, 32) immediately before bonding the second substrate 32 to the first substrate 31.

[0035] The material constituting the first substrate 31 and the second substrate 32 is a transparent thermoplastic resin. Examples of the thermoplastic resin include polymethyl methacrylate (PMMA), polycarbonate (PC), cycloolefin copolymer (COC), cycloolefin polymer (COP), and polystyrene (PS). Among these, medical-grade COP is particularly preferable. COP is a thermoplastic resin with high transparency, low autofluorescence, and low drug adsorption.

[0036] After the through holes (33a, 34a) and recesses 35 described below are formed in each substrate (31, 32) made of the above-mentioned material by, for example, injection molding, an ultraviolet irradiation process and a pressing process are performed to obtain the microfluidic device 30.

[0037] The microfluidic device 30 is formed by stacking and bonding one main surface 31b of the first substrate 31 onto one main surface 32a of the second substrate 32 so that they are in partial contact with each other. Here, the term "main surface" refers to a surface that is much larger in area than the other surfaces constituting the substrates (31, 32). Each of the substrates (31, 32) has two main surfaces that are arranged opposite each other.

[0038] One main surface 31a of the first substrate 31 is located on the opposite side to the second substrate 32 and has ports (33, 34). The other main surface 31b of the first substrate 31 is in partial contact with the second substrate 32. A recess 35 is formed in the main surface 31b of the first substrate 31.

[0039] In the following description, when the first substrate 31 and the second substrate 32 are bonded together, the XYZ coordinate system will be referred to as appropriate, in which a plane parallel to the main surfaces (31a, 31b) of the first substrate 31 and the main surfaces (32a, 32b) of the second substrate 32 is defined as the XY plane, and the direction perpendicular to this XY plane is defined as the Z direction.

[0040] In this specification, when expressing a direction, if a distinction is made between positive and negative directions, the direction is described with a positive or negative sign, such as "+X direction" and "-X direction." When expressing a direction without distinguishing between positive and negative directions, the direction is simply described as "X direction." In other words, in this specification, when simply referring to the "X direction," both the "+X direction" and the "-X direction" are included. The same applies to the Y direction and the Z direction. Note that the microfluidic device 30 is usually used with the Z direction as the up-down direction, and the -Z direction corresponds to the upward direction.

[0041] Fig. 3 is a plan view of a microfluidic device 30. Fig. 4 is a cross-sectional view of a microfluidic device 30 in which a first substrate 31 is bonded onto a second substrate 32, taken along line IV-IV in Fig. 3. Note that, in the cross-sectional view of Fig. 4, only the outlines of both substrates (31, 32) that appear on the cross section are shown to facilitate understanding of the drawing. The same applies to Fig. 6, which will be described later.

[0042] The first substrate 31 and the second substrate 32 are rectangular substrates each having a main surface of the same shape. The dimensions of the main surfaces of the first substrate 31 and the second substrate 32 are, for example, 10 mm×20 mm.

[0043] The thickness of the first substrate 31 is greater than the thickness of the second substrate 32. The thickness of the first substrate 31 is, for example, 0.2 mm to 10 mm, and is typically 3 mm. The thickness of the second substrate 32 is, for example, 0.1 mm to 2 mm, and is typically 1 mm. Note that the "thickness" here corresponds to the length in the Z direction.

[0044] 2, through holes (33a, 34a) extending from the first main surface 31a toward the second main surface 31b are formed in the first substrate 31. In the shape example shown in FIG. 2, these through holes (33a, 34a) are arranged side by side in the X direction.

[0045] Port 33, which is the open end of through-hole 33a, and port 34, which is the open end of through-hole 34a, are provided for at least one of the purposes of injecting liquid into microfluidic device 30 and discharging liquid from microfluidic device 30. Typically, port 33 is used as a liquid inlet, and port 34 is used as a liquid outlet.

[0046] By joining the first substrate 31 and the second substrate 32, the through-hole 33a functions as the first culture space 36 (see FIGS. 3 and 4). Here, an example is shown in which the through-hole 33a functions as the first culture space 36, but the through-hole 34a may also function as the first culture space 36. Furthermore, both the through-hole 33a and the through-hole 34a may also function as the first culture space 36.

[0047] The through holes (33a, 34a) are both circular holes extending in the Z direction. The diameter q1 (see FIG. 3) of the through holes (33a, 34a) is, for example, 0.05 mm to 5 mm, and typically 2 mm. The separation distance q2 (see FIG. 3) between the through holes 33a and 34a is, for example, 1 mm to 100 mm, and typically 7 mm.

[0048] 2, the first substrate 31 has a recess 35 on the second main surface 31b side. The recess 35 is formed to extend in the X direction. A cross section (a cross section parallel to the YZ plane) perpendicular to the extension direction (X direction) of the recess 35 is a flat rectangular shape.

[0049] The recess 35 communicates with the through-holes 33a and 34a. When the first substrate 31 and the second substrate 32 are joined together, the recess 35 functions as a hollow second culture space 37 sandwiched between the two substrates (31, 32).

[0050] The recess 35 has a slit shape extending in the X direction with a constant width and depth. The width of the recess 35 is preferably at least 10 times the depth. That is, when viewed in the extension direction, the second culture space 37 is configured so that the width q3 (see FIG. 3) is preferably at least 10 times the height q4 (see FIG. 4). The second culture space 37 is configured so that the width q3 is preferably at least 10 times, more preferably at least 20 times the height q4. Furthermore, the width q3 is preferably no more than 100 times the height q4. In conventional microfluidic devices made of silicone rubber, it was difficult for a flat flow channel whose width was at least 10 times its height to maintain its height due to the influence of gravity, etc.

[0051] As described above, the microfluidic device 30 includes a first culture space 36 and a second culture space 37 extending from the first culture space 36. The second culture space 37 is preferably configured so that its width q3 is at least 10 times its height q4 when viewed in the extension direction. With this configuration, when neurons 41 are cultured in the first culture space 36, neurites 42 extending from the neurons 41 can be positioned in the second culture space 37, which is a space separate from the neurons 41 (see FIGS. 5 and 6). In other words, the microfluidic device 30 allows neurons 41 and neurites 42 to be cultured in a separated state.

[0052] 5 and 6 are a plan view and a cross-sectional view schematically showing the state in which nerve cells 41 are cultured in the microfluidic device 30. Fig. 6 is a cross-sectional view taken along line VI-VI of the microfluidic device 30 shown in Fig. 5.

[0053] Neurons 41 are cultured in first culture space 36, and neurites 42 extending from neurons 41 are cultured in second culture space 37. By configuring microfluidic device 30 as described above, second culture space 37 can culture neurites 42 while inhibiting the entry of neurons 41 themselves.

[0054] Nerve cells 41 are classified by the presence or absence of myelin sheath, diameter, conduction velocity, etc., and the diameters of type A nerves with myelin sheaths are Aα (13 μm to 22 μm), Aβ (8 μm to 13 μm), Aγ (4 μm to 8 μm), and Aδ (1 μm to 4 μm). Even if they have myelin sheaths, autonomic nerves are called type B, and their diameters are 1 μm to 3 μm. Furthermore, type C nerves (C fibers) without myelin sheaths have a diameter of 0.2 μm to 1.0 μm.

[0055] Since the second culture space 37 is configured with a flat cross-sectional shape, multiple neurites 42 do not form a bundle structure, but can be cultured so that each neurite 42 is aligned in the width direction (Y direction), as shown in Figure 5.

[0056] When neurite 42 is composed of only axons (without myelin), the diameter of neurite 42 is 0.2 μm to 1.5 μm. When neurite 42 is composed of axons covered with myelin, the diameter of neurite 42 is 1 μm to 22 μm. By making the width q3 (see FIG. 3) of second culture space 37 at least seven times the diameter of neurite 42, multiple neurites 42 can grow without contacting the side walls of second culture space 37.

[0057] The width q3 of the second culture space 37 (see FIG. 3) is preferably 50 μm to 8000 μm, and more preferably 500 μm to 1000 μm. The width q3 of the second culture space 37 is typically 800 μm.

[0058] The height q4 of the second culture space 37 (see FIG. 4) is preferably 5 μm to 80 μm, and more preferably 20 μm to 80 μm. The height q4 of the second culture space 37 is typically 40 μm.

[0059] [Trained model] Next, we will explain how to create the trained model M1.

[0060] When a specific injury occurs to a peripheral nerve, the shape and structure of the neurites 42 change due to the injury. Some substances contained in drugs are known to cause or not cause injury to peripheral nerves (hereinafter referred to as "known substances"). As described above, the microfluidic device 30 can culture neurons 41 and neurites 42 in a separated state. That is, by culturing neurons 41 in the microfluidic device 30, more specifically, in the first culture space 36, after introducing a known substance, and observing the neurites 42 extending from the neurons 41 and positioned in the second culture space 37, it is possible to observe the neurites 42 in a state affected by the known substance.

[0061] To facilitate microscopic observation, the specimen may be stained with a specific chemical solution before observation. Examples of such staining methods include fluorescent antibody staining targeting β-tubulin III, a biomarker for nerve axons, and myelin basic protein (MBP), a biomarker for myelin.

[0062] 7A to 7D are images obtained by microscopically observing neurites 42 located in the second culture space 37 after culturing neurons 41 in the microfluidic device 30 using culture media containing different known substances. Note that, although rat DRG cells were used as the neurons 41 here, various types of cells, such as mouse DRG cells, human ES cells, or human iPS cells, can also be used.

[0063] Figure 7A shows a microscopic image of neurites 42 when neurons 41 were cultured in a culture medium containing suramin, a known substance. The exposure time was 24 hours. Suramin is known to cause myelin damage.

[0064] Figure 7B shows a microscopic image of neurites 42 when neurons 41 were cultured in a culture medium containing oxaliplatin, a known substance. The exposure time was 24 hours. Oxaliplatin is known to cause cell body damage.

[0065] Figure 7C shows a microscopic image of neurites 42 when neurons 41 were cultured in a culture medium containing vincristine, a known substance. The exposure time was 24 hours. Vincristine is known to cause axonal damage.

[0066] 7D is a microscopic image of neurites 42 when nerve cells 41 were cultured in a culture medium containing sucrose as a known substance. Sucrose is known to be non-toxic to peripheral nerves.

[0067] In this way, by culturing neurons 41 in microfluidic device 30 using culture solutions containing different known substances, it is possible to obtain multiple microscopic images of neurites 42 located in second culture space 37. Note that, from the viewpoint of improving learning accuracy, multiple microscopic images of neurites 42 cultured using culture solutions containing the same known substance may be obtained.

[0068] The above method provides a plurality of data strings each combining image data da derived from a microscopic image of neurites 42 cultured in a culture medium containing a known substance and information db (information on the damaged area or information indicating the absence of a damaged area) on peripheral nerves caused by the known substance, corresponding to the image data da. Machine learning is performed by introducing this data string into a model generation unit 61, as shown in FIG. 8. The model generation unit 61 is a computational processing means for executing machine learning through computational processing, such as a CPU or MPU. The model generation unit 61 may be incorporated into the drug effect prediction system 10 itself, or may be incorporated into a separate system.

[0069] The model generation unit 61 performs machine learning based on a plurality of training data sets, with image data da based on a microscopic image as learning input data and data db associated with the learning input data and containing information about peripheral nerve damage as learning output data. Note that "machine learning" here refers to learning performed solely by a machine (especially a computer) without the intervention of human thought.

[0070] A typical example of a learning method used in machine learning is a neural network. From the viewpoint of improving the accuracy of judgment, it is preferable to use a hierarchical neural network having one or more intermediate layers between the input layer and the output layer. However, other learning methods may also be used. In addition to neural networks, examples of usable learning methods include linear regression, decision trees, support vector regression, and ensemble methods. These may be used alone or in combination of two or more.

[0071] Machine learning can be performed using, for example, TensorFlow (registered trademark), Keras (registered trademark), Pytorch, Matlab (registered trademark), etc.

[0072] From the viewpoint of improving the accuracy of the judgment, the number of samples of the training data is preferably 10,000 or more, more preferably 100,000 or more, and particularly preferably 1,000,000 or more. From the viewpoint of improving the accuracy of the judgment, the number of learning times is preferably 100 or more, more preferably 500 or more, and particularly preferably 1,000 or more.

[0073] FIG. 9 is a conceptual diagram of a trained model M1 generated by the model generation unit 61. A network is constructed in which the input layer uses feature values ​​based on image data da, which is based on a microscopic image of a neurite 42, and the output layer uses data db related to damage to the neurite 42, connecting these values ​​via an intermediate layer. The damage-related data db includes four nodes: myelin sheath damage db1, cell body damage db2, axon damage db3, and no toxicity db4. In other words, in this embodiment, three types of potential damage sites are planned: myelin sheath, cell body, and axon. A general-purpose method used in machine learning based on image data is used to extract feature values ​​from the image data da.

[0074] Axonal damage db1 is a phenomenon in which damage occurs to the axon itself. Cell body damage db2 is a phenomenon in which damage occurs to the cell body, and the effects of this damage travel throughout the axon, causing damage to the entire axon. Myelin damage db3 is a phenomenon in which damage occurs to the myelin sheath that covers and protects the axon. In other words, various types of damage affect the linear shape of the neurites 42, the thickness of the neurites 42, the branching shape of the neurites 42, the spacing between the neurites 42, the distance between the branches of the neurites 42, and the uniformity, dispersion, and aggregation of the neural network. These effects are manifested in image data da derived from a microscopic image of the neurites 42. In other words, by extracting multiple features from this image data da, a trained model M1 is generated based on the relationship between each damaged area and each feature.

[0075] The trained model M1 constructed in this manner is stored in the storage unit 13 of the system 10.

[0076] In the above embodiment, the trained model M1 has three types of peripheral nerve damage sites and an output layer with four nodes, including cases where no damage occurs, but the number of types of damage sites may be four or more.

[0077] [verification] Paclitaxel, which is known to cause axonal damage in peripheral nerves, was used as an example of test substance 48, and nerve cells 41 were cultured in microfluidic device 30 using a culture medium containing test substance 48. Then, neurites 42 extending from nerve cells 41 and positioned in second culture space 37 were observed under microscope 50, and an observation image was obtained. The observation image is shown in FIG.

[0078] Image data based on the observed image was input to the system 10 from the data input receiving unit 11 as input data for determination i42t. The system 10 applied the input data for determination i42t to the trained model M1 in the determination unit 15. That is, each feature based on the input data for determination i42t was applied to the input layer of the trained model M1. As a result, the determination result shown in FIG. 11 was output from the display output unit 17. This confirms that the test substance 48 has a high probability of causing axonal damage. Considering that the test substance 48 was paclitaxel, which is known to cause axonal damage, this result was a correct determination result.

[0079] Using paclitaxel images as the input data for judgment i42t, 5,776 judgment tests were performed using the drug effect prediction system 10, and 1,226 tests (21.2%) correctly identified axonal damage. However, of the 2,027 judgment tests that were not identified as non-toxic, the probability of identifying axonal damage was 60.5% (1,226 / 2,027). This accuracy rate can be further improved by increasing the number of training data samples and the number of learning runs. [Explanation of symbols]

[0080] 10: Impact prediction system 11: Data entry reception section 13: Storage section 15: Judgment section 17: Display output section 30: Microfluidic devices 31:First board 31a: First main surface of first substrate 31b: second main surface of first substrate 32:Second board 32a: Main surface of second substrate 33: Port 33a: Through hole 34: Port 34a: Through hole 35: Recess 36:First culture space 37:Second culture space 41: Nerve cells 42:Neurites 48: Test substance 50: Microscope 61: Model generation unit

Claims

1. a microfluidic device having a first culture space in which neurons can be cultured, and a second culture space extending from the first culture space in which neurites extending from the neurons can be cultured separately from the neurons; a memory unit in which a trained model is recorded, the trained model being generated by machine learning based on a plurality of training data associating training input data based on a microscopic image of a plurality of neurites extending from each of a plurality of neurons cultured in the first culture space of a plurality of the microfluidic devices prepared for training in a state in which a plurality of known substances, each of which causes a nerve damage site to be a known site belonging to a group of candidate damage sites or no damage site, is individually added to the culture solution, with training output data relating to information on the nerve damage site caused by the known substances added to the first culture space in which the nerve cells corresponding to the microscopic image are cultured, or information indicating that no damage site is present; a data input receiving unit that receives input of input data for determination based on a microscopic image of the neurites extending from the neurons and positioned in the second culture space, in which the neurons are cultured in a state in which a test substance whose damage site to the neurons is unknown is added to a culture solution in the first culture space of the microfluidic device prepared for determination; a determination unit that applies the input data for determination to the trained model, and outputs information derived from the probability of damage occurrence for each site belonging to the group of candidate nerve damage sites due to the introduction of the test substance; A drug effect prediction system characterized in that the second culture space of the microfluidic device has a flat shape and a width that is more than 7 times the diameter of the neurite extending from the nerve cell.

2. 2. The drug effect prediction system according to claim 1, wherein the group of candidate lesion sites includes axonal lesions, cell body lesions, and myelin lesions.

3. The drug effect prediction system described in claim 1 or 2, characterized in that the second culture space of the microfluidic device has a flat shape and a width that allows the culture of the neurites extending from the nerve cells, but a height that prevents the nerve cells from invading.

4. The drug effect prediction system according to claim 1 or 2, wherein the second culture space of the microfluidic device has a width that is at least 10 times larger than its height.

5. 3. The drug effect prediction system according to claim 1, wherein the nerve cells are derived from mouse DRG, rat DRG, human ES, or human iPS cells.

Images