Sequence generation device and sequence generation method
The sequence generation device employs a learning model to select amino acids with similar properties and trains the model to enhance surrogate evaluation values, addressing the challenge of generating diverse sequences with high binding affinity and improving experimental success rates.
Patent Information
- Application Number
- PCT/JP2024/002663
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-29
- Filing Date
- 2024-01-29
- Publication Date
- 2024-12-05
AI Technical Summary
Existing methods face challenges in generating diverse amino acid sequences for antibodies while maintaining high binding affinity, often resulting in sequences with low experimental success rates due to significant changes in amino acid characteristics.
A sequence generation device and method that uses a learning model to select amino acids with properties closer to the original using a probability density function, incorporating mutations with a predetermined probability to generate sequences with higher binding affinity, and trains the model to improve surrogate evaluation values.
The approach generates diverse sequences with high binding affinity, improving the rate of passing experimental acceptance criteria by maintaining amino acid characteristics and optimizing sequence modifications.
Smart Images

Figure JP2024002663_05122024_PF_FP_ABST
Abstract
Description
Array generation device and array generation method
[0001] The present disclosure relates to a sequence generation device and a sequence generation method, and in particular to a device for generating a sequence composed of amino acids.
[0002] Antibodies are used as antibody drugs. Antibodies are composed of a total of about 1,500 sequences of 20 different amino acids. Therefore, the amino acid sequence that makes up a single antibody can have as many as 20 to the power of 1,500 possible amino acid combinations. It is unrealistic to actually create such a huge number of sequences and evaluate them experimentally.
[0003] To conduct experiments efficiently, it is possible to generate sequences on a computer, evaluate them through simulations, and then actually create and evaluate only promising sequences in experiments. However, there are many evaluation items for sequences (e.g., binding affinity, viscosity, solubility, immunogenicity), and simulations can only evaluate a portion of these items. Therefore, even if a promising sequence is highly evaluated in simulations, it often performs poorly in actual experiments and fails to meet the pass criteria.
[0004] Various methods have been proposed to increase the success rate of passing the test criteria in actual experiments.
[0005] In Patent Document 1, mutations are made at pre-identified positions in a sequence, and a sequence that improves the affinity score or stability score is selected. The mutations are exhaustive, and the evaluation based on the affinity score or stability score is a simple ranking.
[0006] In Patent Document 2, an amino acid sequence is generated using a Generative Adversarial Network (GAN) in which the threshold amount for the binding region is the same as that of a non-human antibody and the threshold amount for the heavy and light chain framework regions is the same as that of a human antibody, and a penalty is imposed for changing a hydrophobic amino acid to another hydrophobic amino acid.
[0007] Patent Document 3 uses a model that generates sequences with high binding affinity using LSTM (Long Short Term Memory), which is characterized by using the position and context of amino acids.
[0008] There are also known examples of using reinforcement learning. For example, the binding affinity of sequences is evaluated through simulation, and reinforcement learning is performed to output sequences that have a higher evaluation value. However, general reinforcement learning may end up generating sequences that have high binding affinity but similar amino acid structures. Since sequences with similar structures also have similar properties, even if a large number of sequences are generated, there is a possibility that they will all be eliminated in actual experiments.
[0009] To address this issue, Non-Patent Document 1 uses GFlowNets (Generative Flow Networks) to generate diverse sequences with high binding affinity, thereby avoiding total annihilation. GFlowNets is characterized by learning a Q function so that the total probability value up to a certain state in reinforcement learning search matches the total probability value from that state onward. Conventional reinforcement learning uses a monotonic Q function based on a fixed discount rate, which tends to lead to a single path with a high final reward. However, GFlowNets obtains more appropriate behavioral value in each state, enabling sampling from a variety of modes. The Q function thus learned calculates a Q value for each of the 20 next amino acid candidates based on the partial sequences generated up to that point. The amino acid candidate with the highest Q value is selected as the next amino acid and incorporated into the generated sequence.
[0010] If amino acids are always selected according to their Q values, it will be difficult to select amino acids that have a low Q function at that time. As a result, the possibility of constructing a better sequence using amino acids that were not selected may be overlooked. To solve this problem, GFlowNets has a function (mutation function) that randomly changes amino acids with a certain probability.
[0011] US Patent Application Publication No. 2020 / 0126640 US Patent Application Publication No. 2023 / 0005567 US Patent Application Publication No. 2022 / 0253669
[0012] Jain, M., et al., "Biological Sequence Design with GFlowNets", arXiv:2203.04115, DOI: 10.48550 / arXiv.2203.04115, [online], 2022, [Retrieved April 11, 2020], Internet <URL: https: / / arxiv.org / abs / 2203.04115>
[0013] However, conventional techniques have had the problem of difficulty in generating diverse sequences while maintaining high binding affinity.
[0014] For example, if amino acids are randomly changed as in Non-Patent Document 1, even a single amino acid change can significantly change the antibody's properties. In such cases, a sequence with low binding affinity is generated, resulting in a lower clearance rate in actual experiments. For this reason, amino acid changes must be made appropriately and not too drastic.
[0015] In view of the above circumstances, the present disclosure provides a sequence generation device and a sequence generation method that can generate diverse sequences while maintaining high binding affinity.
[0016] An example of a sequence generation device according to the present disclosure is a sequence generation device comprising: a sequence surrogate evaluation unit that calculates a surrogate evaluation value for a sequence composed of a plurality of amino acids; and a sequence generation unit that generates the sequence using a learning model, wherein the sequence generation unit: - changes some of the amino acids in the sequence generated by the learning model with a predetermined probability, and selects the changed amino acids using a probability density function that has a higher probability of selecting amino acids with properties closer to the amino acids before the change; - trains the learning model to generate sequences that exhibit higher surrogate evaluation values based on the surrogate evaluation values for the sequence after the amino acids have been changed, thereby generating a trained model; and - generates a generated sequence using the trained model.
[0017] An example of a sequence generation method according to the present disclosure is a sequence generation method comprising: a sequence surrogate evaluation step of calculating a surrogate evaluation value for a sequence composed of a plurality of amino acids; and a sequence generation step of generating the sequence using a learning model, wherein the sequence generation step includes: - a step of changing some of the amino acids in the sequence generated by the learning model with a predetermined probability, and selecting the changed amino acids using a probability density function such that amino acids with properties closer to the amino acids before the change are selected with a higher probability; - a step of training the learning model to generate sequences that exhibit higher surrogate evaluation values based on the surrogate evaluation values for the sequence after the amino acids have been changed, thereby generating a trained model; and - a step of generating a generated sequence using the trained model.
[0018] The sequence generation device and sequence generation method according to the present disclosure can generate diverse sequences while maintaining high binding affinity.
[0019] 1. A configuration example of a sequence generation device according to a first embodiment of the present disclosure. A diagram for explaining an overview of processing in the sequence generation device of FIG. 1. A flowchart showing the processing flow of the sequence generation device of FIG. 1. An example of a characteristic serving as a selection criterion. An example of a specific probability density function. A configuration example of a sequence generation device according to a second embodiment. A diagram for explaining an overview of processing in the sequence generation device of FIG. 6. A flowchart showing the processing flow of the sequence generation device of FIG. 6. A configuration example of a sequence generation device according to a third embodiment. A flowchart showing the processing flow of the sequence generation device of FIG. 10. A configuration example of a sequence generation device according to a fourth embodiment. A flowchart showing the processing flow of the sequence generation device of FIG. 11.
[0020] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. [Example 1] Fig. 1 shows an example configuration of a sequence generation device 10 according to Example 1 of the present disclosure. The sequence generation device 10 is a device that generates a sequence composed of a plurality of amino acids. The sequence is, for example, the sequence of amino acids in an antibody.
[0021] The sequence generation device 10 includes a sequence proxy evaluation unit 11 and a sequence generation unit 12. The sequence proxy evaluation unit 11 calculates a proxy evaluation value for a sequence using a function such as MLP (Multi Layer Perceptron). The sequence generation unit 12 generates a sequence using a learning model.
[0022] The sequence generation device 10 has a hardware configuration as a known computer, and includes, for example, a computing means and a storage means. The computing means includes, for example, a processor, and the storage means includes, for example, a storage medium such as a semiconductor memory device or a magnetic disk device. Some or all of the storage medium may be non-transitory storage media.
[0023] The sequence generating device 10 may also include input / output means, which include, for example, input devices such as a keyboard and a mouse, output devices such as a display and a printer, and communication devices such as a network interface.
[0024] The storage means of the sequence generating device 10 may store a program. The processor may execute the program, causing the computer to realize the functions described in this embodiment, thereby configuring the sequence generating device 10.
[0025] An overview of the processing in the sequence generation device 10 will be explained using Figure 2. The Q function is a learning model that can be generated by machine learning, and the one described in Non-Patent Document 1 can be used. The input to the Q function is a state representing the partial sequence (a sequence consisting of zero or more amino acids) generated up to that point, and in the example of Figure 2, it contains two amino acids. The state is empty (zero amino acids) at the start of generation and becomes longer as the generation process progresses.
[0026] In FIG. 2, according to the convention, each of the different types of amino acids (20 types in total) is represented by a different alphabetical letter, and this is the same in other figures and in the following description.
[0027] The output of the Q function is a Q value for each of the 20 amino acids. The sequence generation unit 12 selects the amino acid with the highest Q value according to the Q function at that time.
[0028] After selecting an amino acid, it is determined whether or not a mutation occurs. If a mutation occurs, a modified amino acid is randomly selected using a probability density function.
[0029] 3 is a flowchart showing the processing flow of the array generation device 10. The array generation device 10 can generate an array by executing the method according to this flowchart. The operation of the array generation device 10 will be described below with reference to the flowchart.
[0030] 3, steps S1 to S8 are Q-function learning processes, which generate a Q-function as a learned model. First, the sequence generator 12 selects amino acids one by one in order from the beginning of the sequence according to the Q-function (step S1).
[0031] Next, the sequence generation unit 12 determines whether or not to cause a mutation based on a predetermined probability (step S2).
[0032] If it is determined that a mutation is to occur, the sequence generator 12 generates a probability density function for selecting an amino acid after the change (step S3). Here, the probability density function differs depending on the amino acid before the change, and a probability density function is generated such that an amino acid closer in properties to the amino acid before the change is selected with a higher probability.
[0033] Figure 4 shows an example of characteristics that can be used as selection criteria. In this example, the characteristics include hydrophobicity. Eisenberg's hydrophobicity index can be used as a value to represent hydrophobicity. Using the hydrophobicity index, the similarity of the characteristics of amino acids can be expressed numerically. Hydrophobicity is one of the important indicators that affects various evaluation items of antibody drugs (solubility, aggregation, binding affinity, etc.), and selecting amino acids based on this can maintain a higher rate of meeting the pass criteria in experiments.
[0034] A specific example of a probability density function is shown in Figure 5. The probability density function can be, for example, a discretized Gaussian distribution, with the horizontal axis representing the hydrophobicity index. The probability density function in Figure 5 is a function in which the probability gradually decreases as the difference between the hydrophobicity index of the amino acid before the change and the hydrophobicity index of the amino acid after the change increases.
[0035] 5(a) shows the probability density function when the amino acid before the change is Y, where an amino acid with properties closer to Y is selected with a higher probability. FIG. 5(b) shows the probability density function when the amino acid before the change is T, where an amino acid with properties closer to T is selected with a higher probability.
[0036] Here, the probability density function is designed to allow mutation from any amino acid to any amino acid, thereby increasing sequence diversity. That is, regardless of the amino acid before the change, each amino acid is selected with a different probability. Note that if the same amino acid as the amino acid before the change is selected, the sequence will be the same as if no mutation had occurred.
[0037] The sequence generation unit 12 selects amino acids after the change using the probability density function generated in step S3 (step S4).
[0038] If it is decided in step S2 that no mutations will be generated, steps S3 and S4 are not performed, i.e., the amino acid selected in step S1 is used as is.
[0039] The sequence generator 12 adds the amino acid selected in step S1 or S4 to the end of the sequence, and determines whether or not the generation of a sequence of a predetermined length has been completed (step S5). If not, the process returns to step S1, i.e., the next amino acid is added.
[0040] When the sequence generation is completed, the sequence surrogate evaluation unit 11 acquires a surrogate evaluation value for the generated sequence (step S6). A method for acquiring a specific evaluation value based on an amino acid sequence can be appropriately designed by a person skilled in the art based on publicly known techniques, etc.
[0041] A specific example of the surrogate evaluation value can be expressed as a value representing the binding affinity between an antibody represented by a sequence and an antigen. By using such a surrogate evaluation value, a high surrogate evaluation value can be given to a sequence that has a high rate of meeting the acceptance criteria for binding affinity in experiments.
[0042] In addition, if one or more mutations have occurred (which is generally a high probability), this proxy evaluation value can be said to be a proxy evaluation value for the sequence after the amino acid has been changed.
[0043] Next, the array generation unit 12 acquires the proxy evaluation value from the array proxy evaluation unit 11, and based on this proxy evaluation value, learns and updates the Q function so as to generate an array showing a higher proxy evaluation value (step S7).
[0044] Next, the array generation unit 12 determines whether the Q function has been updated a predetermined number of times (step S8). If the predetermined number of times has not been reached, the process returns to step S1, that is, the next array is generated.
[0045] When the predetermined number of times is reached, the machine learning of the Q function is completed, and the Q function is generated as a trained model. The array generator 12 generates an array (generated array) using the generated Q function (step S9).
[0046] As described above, in Example 1, in the Q function learning process, the sequence generator 12 changes some amino acids in the sequence generated by the Q function with a predetermined probability, and a probability density function is used that selects amino acids with properties (e.g., hydrophobicity) closer to the amino acids before the change with a higher probability. This makes it possible to generate sequences with closer properties (e.g., sequences with high binding affinity), thereby improving the rate at which the pass criteria are met in actual experiments.
[0047] [Example 2] Example 2 uses a plurality of array proxy evaluators, each calculating a different proxy evaluation value, instead of the single array proxy evaluator in Example 1. Example 2 will be described below, but explanations of parts common to Example 1 may be omitted.
[0048] 6 shows an example of the configuration of an array generation device 20 according to Example 2. The array generation device 20 includes, as array proxy evaluators, a first array proxy evaluator 21 that calculates a first proxy evaluation value indicating a first characteristic of the array, and a second array proxy evaluator 22 that calculates a second proxy evaluation value indicating a second characteristic of the array. The array generation device 20 also includes the array generation unit 12 similar to that of Example 1 (FIG. 1).
[0049] An overview of the processing in the array generation device 20 will be described using Figure 7. There are various methods for calculating the proxy evaluation value of an array, and multiple methods can be combined. Multiple proxy evaluation values are calculated for the array, and the result of adding them up (added proxy evaluation value) is used as the overall proxy evaluation value.
[0050] In the example of Figure 7, the first property represents the binding affinity between the antibody represented by the sequence and the antigen, and the second property represents the solubility of the antibody represented by the sequence. The definition of solubility can be determined appropriately by those skilled in the art, but it may represent, for example, solubility in water, physiological saline, or any body fluid. Using such properties increases the likelihood of passing the experimental criteria.
[0051] The array generator 12 calculates an overall proxy evaluation value based on the first proxy evaluation value and the second proxy evaluation value. For example, a simple addition may be performed, but a more flexible evaluation becomes possible if the overall proxy evaluation value is calculated based on a weighted addition.
[0052] 8 is a flowchart showing the processing flow of the array generating device 20. Step S6 in the first embodiment (FIG. 3) is replaced by steps S6a to S6c in the second embodiment (FIG. 8). The other steps can be the same as those in the first embodiment. Only steps S6a to S6c will be described below.
[0053] After the sequence generation is completed, the first sequence surrogate evaluation unit 21 obtains a first surrogate evaluation value for the generated sequence (step S6a), and the second sequence surrogate evaluation unit 22 obtains a second surrogate evaluation value for the generated sequence (step S6b). As described above, an example of the first property related to the first surrogate evaluation value is binding affinity, and an example of the second property related to the second surrogate evaluation value is solubility.
[0054] Next, the array generating unit 12 adds the first proxy evaluation value and the second proxy evaluation value (for example, by weighted addition) to calculate an overall proxy evaluation value (step S6c).
[0055] As described above, in the second embodiment, an array can be generated that incorporates evaluations of actual experiments for a plurality of characteristics, and therefore the rate at which the pass criteria of the actual experiments are met can be further improved.
[0056] In the example of FIG. 7, two types of characteristics are used, but it is also possible to use three or more types of characteristics by providing three or more array proxy evaluators.
[0057] As a modification of Example 2, the probability density function may be changed based on other conditions. For example, if the solubility evaluation value of the sequence generated so far is small (e.g., if the solubility evaluation value in water or a predetermined body fluid is equal to or less than a predetermined threshold), the right half of the probability density function may be replaced with a probability value of zero. In other words, amino acids with higher hydrophobicity than the amino acids before the change may not be selected.
[0058] Example 3 is a modification of Example 1, in which evaluations of actual experiments are fed back to the simulation. Example 3 will be described below, but descriptions of parts common to Example 1 may be omitted.
[0059] 9 shows an example of the configuration of an array generation device 30 according to Example 3. The array generation device 30 includes an array substitute evaluation unit 11 and an array generation unit 12 similar to those in Example 1 (FIG. 1). Furthermore, the array generation device 30 includes an array output unit 13 including an output device, an evaluation true value input unit 14 including an input device, and an array accumulation unit 15 including a storage device.
[0060] 10 is a flowchart showing the processing flow of the array generating device 30. First, the array generating device 30 generates an array by the processing of the first embodiment or a similar processing.
[0061] Next, the array output unit 13 outputs the generated array (step S11). The output is performed to the outside of the array generating device 30, for example, by displaying it on a display device, printing it on a printing device, transmitting it via a communication network, etc.
[0062] 10, after step S11, a user of the sequence generation device 30 evaluates the output sequence through an experiment, obtains an actual evaluation value (true evaluation value) for the sequence, and inputs the obtained value to the sequence generation device 30. The input is performed from outside the sequence generation device 30, for example, by using a keyboard, a touch panel, or transmission via a communication network.
[0063] In response to this, the evaluation true value input unit 14 acquires the evaluation true value for the array output in step S11 (step S12). Thereafter, the array accumulation unit 15 stores a pair of the array output in step S11 and the evaluation true value acquired in step S12.
[0064] Next, the array proxy evaluator 11 acquires a pair of an array and an evaluation true value from the array storage unit 15, and updates itself by performing machine learning based on the pair (step S13). For example, the function that provides the proxy evaluation value is updated, and the array proxy evaluator 11 is updated accordingly.
[0065] Next, the array proxy evaluator 11 determines whether the array proxy evaluator 11 has been updated a predetermined number of times (step S14). If the predetermined number of times has not been reached, the process returns to steps S1 to S9, i.e., the next array is generated. If the predetermined number of times has been reached, the process of FIG. 10 ends.
[0066] As described above, in Example 3, the accuracy of the proxy evaluation value is improved by using the true evaluation value based on the actual experimental results, thereby reducing the number of cases where the generated array meets the pass criteria in the simulation but fails to do so in the actual experiment.
[0067] In the third embodiment, the same modifications as those in the first or second embodiment can be made.
[0068] Example 4 Example 4 is a modification of Example 3, combining it with Example 2, to optimize the weighting when adding two types of characteristics. In particular, an algorithm (a so-called multi-armed bandit algorithm) is used, in which multiple array generators with different weights are prepared and the array generator with the largest average weighted evaluation value is adopted. Example 4 will be described below, but descriptions of parts common to Examples 1 to 3 may be omitted.
[0069] 11 shows an example of the configuration of a sequence generation device 40 according to Example 4. The sequence generation device 40 includes a first sequence generation group 41, a second sequence generation group 42, and a sequence generation selection unit 43. Both the first sequence generation group 41 and the second sequence generation group 42 include a first sequence proxy evaluator 21, a second sequence proxy evaluator 22, and a sequence generation unit 12 similar to those in Example 2 ( FIG. 6 ).
[0070] 12 is a flowchart showing the processing flow of the array generating device 40. First, the array generating device 40 generates a plurality of arrays by the processing of the second embodiment or a similar processing.
[0071] Next, the first array generation group 41 calculates a proxy evaluation value (added proxy evaluation value) for each generated array based on a weighted addition of the first proxy evaluation value and the second proxy evaluation value, and then calculates the average of these added proxy evaluation values (step S21). The average is an example of a method for calculating a representative value from multiple added proxy evaluation values; the median or maximum value of the added proxy evaluation values may also be used. Here, the weighted addition is performed using a first ratio. As a specific example, the first proxy evaluation value multiplied by 0.7 and the second proxy evaluation value multiplied by 0.3 are added together.
[0072] Next, the second array generation group 42 similarly calculates a proxy evaluation value (added proxy evaluation value) for each generated array based on a weighted addition of the first proxy evaluation value and the second proxy evaluation value, and then calculates the average of these added proxy evaluation values (step S22). Here, the weighted addition is performed using a second ratio different from the first ratio used by the first array generation group 41. As a specific example, the first proxy evaluation value multiplied by 0.3 and the second proxy evaluation value multiplied by 0.7 are added together.
[0073] Next, the array generation selection unit 43 selects either the first array generation group 41 or the second array generation group 42 based on the average value of the added proxy evaluation values calculated by the first array generation group 41 and the average value of the added proxy evaluation values calculated by the second array generation group 42 (step S23). For example, it selects the one with the larger average value.
[0074] Next, the array output unit 13 outputs the array generated by the array generation group selected by the array generation selection unit 43 (step S11).
[0075] Next, the evaluation true value input unit 14 acquires an evaluation true value for the sequence output in step S11 (step S12). Since multiple properties are used in Example 4, multiple evaluation true values are acquired. For example, an evaluation true value related to binding affinity and an evaluation true value related to solubility are acquired.
[0076] Next, the first sequence generation group 41 and the second sequence generation group 42 acquire pairs of sequences and evaluation true values, and perform machine learning based on these to update themselves (step S13). For example, each first sequence surrogate evaluation unit 21 is updated based on the evaluation true value related to binding affinity, and each second sequence surrogate evaluation unit 22 is updated based on the evaluation true value related to solubility.
[0077] As described above, in Example 4, the weighting when adding two types of characteristics can be optimized, and learning is performed based on an array that obtains a better additive evaluation value, thereby further improving the rate at which the generated array meets the pass criteria in actual experiments.
[0078] 11 and 12, two array generation groups are used, but three or more array generation groups can also be used. In that case, in step S23, one array generation group having the largest average value is selected from the array generation group group including three or more array generation groups.
[0079] In the fourth embodiment, the same modifications as those in the first to third embodiments can be made.
[0080] Those skilled in the art can add, modify, or delete components as desired in the above-described first to fourth embodiments and their variations within the scope of the present disclosure. For example, the sequence generation device can be configured by multiple computers connected via a communication network.
[0081] 10, 20, 30, 40... Array generation device 11... Array proxy evaluation unit 12... Array generation unit 13... Array output unit 14... Evaluation true value input unit 15... Array storage unit 21... First array proxy evaluation unit 22... Second array proxy evaluation unit 41... First array generation group 42... Second array generation group 43... Array generation selection unit
Claims
1. A sequence generation device comprising: a sequence surrogate evaluation unit that calculates a surrogate evaluation value for a sequence composed of a plurality of amino acids; and a sequence generation unit that generates the sequence using a learning model, wherein the sequence generation unit: - changes some of the amino acids in the sequence generated by the learning model with a predetermined probability, and selects changed amino acids using a probability density function that has a higher probability of selecting amino acids whose characteristics are closer to the amino acids before the change; - trains the learning model to generate sequences that exhibit higher surrogate evaluation values based on the surrogate evaluation values for the sequence after the amino acids have been changed, thereby generating a trained model; and - generates a generated sequence using the trained model.
2. The sequence generating device according to claim 1, wherein the proxy evaluation value represents the binding affinity between the antibody represented by the sequence and an antigen.
3. The sequence generating apparatus of claim 1, wherein said characteristics include hydrophobicity.
4. A sequence generating device according to claim 1, wherein the probability density function is a function in which the probability gradually decreases as the difference between the hydrophobic index of the amino acid before the change and the hydrophobic index of the amino acid after the change increases.
5. The array generation device of claim 1, wherein the array proxy evaluation unit comprises: a first array proxy evaluation unit that calculates a first proxy evaluation value indicating a first characteristic of the array; and a second array proxy evaluation unit that calculates a second proxy evaluation value indicating a second characteristic of the array, and the array proxy evaluation unit calculates the proxy evaluation value based on the first proxy evaluation value and the second proxy evaluation value.
6. The sequence generating apparatus according to claim 5, wherein the first characteristic represents a binding affinity between the antibody represented by the sequence and an antigen, and the second characteristic represents a solubility of the antibody represented by the sequence.
7. The array generating device according to claim 5, wherein the array proxy evaluation unit calculates the proxy evaluation value based on a weighted addition of the first proxy evaluation value and the second proxy evaluation value.
8. The array generation device according to claim 1, further comprising an array output unit, an evaluation true value input unit, and an array storage unit, wherein the array output unit outputs the generated array, the evaluation true value input unit acquires an evaluation true value for the generated array, the array storage unit stores a pair of the generated array and the evaluation true value, and the array surrogate evaluator acquires the pair of the generated array and the evaluation true value from the array storage unit and performs machine learning based thereon.
9. The sequence generation device comprises a first sequence generation group, a second sequence generation group, and a sequence generation selection unit, each of the first sequence generation group and the second sequence generation group comprising: a first sequence proxy evaluation unit for calculating a first proxy evaluation value indicating a first characteristic of the sequence; and a second sequence proxy evaluation unit for calculating a second proxy evaluation value indicating a second characteristic of the sequence, the first sequence generation group calculates the proxy evaluation value by weighting the first proxy evaluation value and the second proxy evaluation value by a first ratio, the second sequence generation group calculates the proxy evaluation value by weighting the first proxy evaluation value and the second proxy evaluation value by a second ratio different from the first ratio, the sequence generation selection unit selects one from a sequence generation group group including the first sequence generation group and the second sequence generation group based on an average value of the multiple proxy evaluation values calculated by the first sequence generation group and an average value of the multiple proxy evaluation values calculated by the second sequence generation group, and the sequence output unit outputs the generated sequence generated by the one selected by the sequence generation selection unit. The array generating device according to claim 8.
10. A sequence generation method comprising: a sequence surrogate evaluation step of calculating a surrogate evaluation value for a sequence composed of a plurality of amino acids; and a sequence generation step of generating the sequence using a learning model, the sequence generation step including: - a step of changing some of the amino acids in the sequence generated by the learning model with a predetermined probability, and selecting amino acids after the change using a probability density function such that amino acids with characteristics closer to the amino acids before the change are selected with a higher probability; - a step of training the learning model to generate sequences exhibiting higher surrogate evaluation values based on the surrogate evaluation values for the sequence after the amino acids have been changed, thereby generating a trained model; and - a step of generating a generated sequence using the trained model.
Citation Information
Patent Citations
Information processing system, information processing method, program, and method for producing antigen-binding molecule or protein.
WO2020246617A1
Deep learning for de novo antibody affinity maturation (modification) and property improvement
WO2022026551A1
