Disturbance response and target cell state modeling

Abstract

One or more embodiments include predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models configured to accept as input one or more basal states associated with the one or more biological systems and the first plurality of perturbations; predicting a first plurality of cellular states associated with the first plurality of responses using one or more cellular state models configured to accept as input the first plurality of responses, wherein the one or more cellular state models have been trained using a second plurality of training samples comprising observations of a plurality of biological systems paired with a second plurality of cellular states; and based on the first plurality of cellular states, selecting one or more additional perturbations included in the first plurality of perturbations for experimental evaluation.

Description

[0001] Cross-references to related applications This application claims priority to U.S. Provisional Patent Application Serial No. 63 / 622,301, filed January 18, 2024, entitled "PERTURBATION RESPONSE AND TARGET CELL STATE MODELING". The subject matter of that related application is incorporated herein by reference. Technical Field

[0002] The embodiments of this disclosure generally relate to machine learning, and more specifically, to perturbation response and target cell state modeling. Background Technology

[0003] Biological systems exhibit a wide range of complex phenomena that influence and / or reflect broader properties such as health, aging, resilience, and / or response to interventions or treatments. In recent years, numerous technologies have been developed for acquiring and analyzing biological data to understand the underlying patterns or mechanisms of these phenomena and / or properties. For example, imaging techniques such as cryo-electron computed tomography (CryoET) can be used to generate high-resolution three-dimensional (3D) views of biological samples such as cells, tissues, organisms, or macromolecules. In another example, various “omics” disciplines can be used to study biomolecules that represent the structure, function, and dynamics of biological systems. Among these “omics” disciplines, genomics involves the study of the complete set of deoxyribonucleic acid (DNA) within a biological system (such as an organism); proteomics can be used to characterize proteins produced or modified by a biological system; transcriptomics involves the study of ribonucleic acid (RNA) molecules transcribed from the genome of a biological system; metabolomics can be used to study small molecule metabolites within a biological system; epigenomics involves the study of epigenetic modifications of the genetic material of a biological system; phenomics involves the study of observable characteristics or traits of a biological system; and metagenomics can be used to study genetic material recovered from environmental samples.

[0004] However, identifying or characterizing patterns or relationships between large biological datasets generated using different technologies can be challenging. More specifically, each type of biological data may provide a noisy and incomplete view of the state or function of a biological system. Meanwhile, unless these types of data are collected from a single biological sample using a single assay method, it can be difficult to correlate physiological, functional, imaging, and / or other types of biological data.

[0005] These difficulties in identifying patterns and correlations across large biological datasets can further complicate the development and evaluation of interventions that elicit specific responses in biological systems. For example, interventions may include a variety of genetic and / or chemical perturbations applied to cells, cell types, tissues, organs, and / or other biological systems. However, the responses of biological systems to these perturbations can vary depending on age, disease state, injury, genotype, phenotype, cell type (e.g., neurons versus fibroblasts), and / or other properties associated with the biological system. Furthermore, multiple combinations of perturbations are often required to have an effect on a biological system. With billions to trillions of potential combinations, it is impractical to experimentally explore the responses of various biological systems to different combinations of perturbations.

[0006] As mentioned above, what is needed in this field is more effective technology for analyzing and correlating data across different biological fields and interventions. Summary of the Invention

[0007] One or more embodiments include: predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, the one or more perturbation response models being configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbations as inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including a base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; predicting a first plurality of cell states associated with the first plurality of responses using one or more cell state models configured to accept the first plurality of responses as inputs, wherein the one or more cell state models have been trained using a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states; and selecting one or more additional perturbations included in the first plurality of perturbations for experimental evaluation based on the first plurality of cell states.

[0008] Compared to existing technologies, one advantage of the disclosed technology is its ability to predict the responses of different biological systems to various combinations of perturbations, and to predict the corresponding cell states. Therefore, the disclosed technology can be used to identify perturbations that may improve biological age, health status, phenotype, and / or other target cell states without requiring time-consuming and resource-intensive experiments to determine the effects of these perturbations on biological systems. Another advantage of the disclosed technology is its ability to focus limited computational and / or experimental resources on promising perturbations that are expected to generate target cell states within biological systems. Therefore, the disclosed technology can prioritize and validate interventions that improve health and rejuvenation in biological systems in a targeted manner, thereby enhancing the exploration and understanding of biological rejuvenation within a vast and complex perturbation search space. These advantages represent one or more technological improvements over existing methods. Attached Figure Description

[0009] To gain a more detailed understanding of the features described above in the various embodiments, reference can be made to various embodiments to describe in more specific the inventive concept briefly outlined above, some of which are illustrated in the accompanying drawings. However, it should be noted that the drawings illustrate only typical embodiments of the inventive concept and should therefore not be construed as limiting the scope in any way, and that other equally effective embodiments exist.

[0010] Figure 1 The illustration depicts a system configured to implement one or more aspects of various embodiments.

[0011] Figure 2 According to various embodiments Figure 1 A more detailed illustration of the training engine and execution engine is provided.

[0012] Figure 3 The illustrations depict various embodiments. Figure 1 How does the training engine in the program train a set of perturbation response models?

[0013] Figure 4 The illustrations depict various embodiments. Figure 1 How does the training engine in the model train a set of cell state models?

[0014] Figure 5 Flowcharts of method steps for performing perturbation response and target cell state modeling according to various embodiments are provided.

[0015] Figure 6 The illustration shows example architectures of disturbance response models according to various embodiments.

[0016] Figure 7The receiver operating characteristic (ROC) curves representing the performance of exemplary cell state models according to various embodiments are shown. Detailed Implementation

[0017] In the following description, numerous specific details are set forth to provide a more thorough understanding of the various embodiments. However, it will be apparent to those skilled in the art that the inventive concept can be practiced without one or more of these specific details.

[0018] As mentioned above, identifying or characterizing patterns or relationships between large biological datasets generated using different technologies can be challenging. More specifically, each type of biological data may provide a noisy and incomplete view of the state or function of a biological system. Furthermore, unless these types of data are collected from a single biological sample using a single assay method, it can be difficult to correlate physiological, functional, imaging, and / or other types of biological data.

[0019] These difficulties in identifying patterns and correlations across large biological datasets can further complicate the development and evaluation of interventions that elicit specific responses in biological systems. For example, interventions may include hundreds or thousands of genetic and / or chemical perturbations applied to cells, cell types, tissues, organs, and / or other biological systems. However, the responses of biological systems to these perturbations can vary depending on age, cell type, disease state, injury, genotype, phenotype, and / or other properties associated with the biological system. Therefore, experimentally evaluating the effects of various combinations of perturbations on all possible combinations of properties associated with the biological system would be impractical.

[0020] To improve the understanding of patterns and relationships among biological datasets, the disclosed techniques train and execute various machine learning models to implement perturbation response and target cell state modeling. The machine learning models include one or more perturbation response learning models trained to predict the response of one or more biological systems to genetic, chemical, and / or other types of perturbations based on inputs including vectors, codes, and / or representations of gene expression, basal state, covariates, and / or perturbations. For example, the one or more perturbation response models may include combined perturbation autoencoders (CPAs) and / or other types of deep learning models with encoder-decoder architectures. The encoder in the perturbation response model can be used to generate embeddings representing the input basal state, cell type, perturbation, and / or covariates associated with the biological system. The decoder in the perturbation response model can be used to decode these embeddings into gene expression, images, proteins, and / or other outputs representing the biological system's response to perturbations.

[0021] The machine learning model also includes one or more cell state models trained to predict health measures or attributes representing measurable properties indicative of the health status of the biological system, based on inputs including data collected from the biological system. For example, cell state learning models may include nonlinear regression models, deep learning models, and / or other types of machine learning models that predict biological age, health or disease status, cell type, phenotype, and / or other cell states corresponding to these representations, based on gene expression, images, proteins, and / or other data indicative of the biological system's response to perturbations.

[0022] After training, the perturbation response model and cell state model are integrated into the feedback loop of computer simulations and physical experiments to accelerate the discovery of interventions that promise to improve health and / or rejuvenate biological systems. More specifically, the one or more perturbation response models can be used to explore perturbation responses across cell types, disease conditions, other baseline states, covariates, modalities, perturbation types, combinations of perturbations, and / or other variables. The perturbation responses generated by the one or more perturbation response models can then be fed into the one or more cell state models. The predictions of cell states generated by the one or more cell state models based on the input perturbation responses can be used to identify subsets of perturbations that promise to rejuvenate various cell types and / or otherwise improve their health. The identified perturbations can then be validated through various experimental screenings, the results of which can be used to retrain one or two sets of machine learning models (e.g., by providing ground truth data for updating the parameters of one or two sets of machine learning models).

[0023] System Overview Figure 1 The diagram illustrates a computer system 100 configured to implement one or more aspects of various embodiments. In one embodiment, the computer system 100 includes a desktop computer, laptop computer, smartphone, personal digital assistant (PDA), tablet computer, or any other type of computing device configured to receive input, process data, and optionally display images, and is suitable for practicing one or more embodiments. The computer system 100 also includes, or alternatively includes, machines or processing nodes operating in a data center, cluster, or cloud computing environment that provide scalable computing resources (optionally as a service) over a network.

[0024] As shown in the figure, the computer system 100 includes, but is not limited to, a central processing unit (CPU) 102 and a system memory 104 coupled to a parallel processing subsystem 112 via a memory bridge 105 and a communication path 113. The memory bridge 105 is further coupled to an I / O (input / output) bridge 107 via a communication path 106, and the I / O bridge 107 is in turn coupled to a switch 116.

[0025] I / O bridge 107 is configured to receive user input from an optional input device 108, such as a keyboard or mouse, and forward the input to CPU 102 for processing via communication path 106 and memory bridge 105. In some embodiments, computer system 100 may be a server machine in a cloud computing environment. In such embodiments, computer system 100 may not have an input device 108. Instead, computer system 100 may receive equivalent input by receiving commands in the form of messages transmitted over a network and received via network adapter 118. In one embodiment, switch 116 is configured to provide connectivity between I / O bridge 107 and other components of computer system 100, such as network adapter 118 and various add-on cards 120 and 121. In one embodiment, I / O bridge 107 is coupled to system disk 114, which can be configured to store content, applications, and data for use by CPU 102 and parallel processing subsystem 112. In one embodiment, system disk 114 provides non-volatile storage for applications and data and may include fixed or removable hard disk drives, flash memory devices, and CD-ROMs (compact disc read-only memory), DVD-ROMs (digital versatile optical discs), Blu-ray, HD-DVDs (high-definition DVDs), or other magnetic, optical, or solid-state storage devices. In various embodiments, other components, such as universal serial buses or other port connections, optical disc drives, digital universal disc drives, film recording devices, etc., may also be connected to I / O bridge 107.

[0026] In various embodiments, memory bridge 105 may be a northbridge chip, and I / O bridge 107 may be a southbridge chip. Additionally, communication paths 106 and 113, as well as other communication paths within computer system 100, may be implemented using any technically suitable protocol, including but not limited to AGP (Accelerated Graphics Port), HyperTransport, or any other bus or point-to-point communication protocol known in the art. In some embodiments, the parallel processing subsystem 112 includes a graphics subsystem that transmits pixels to an optional display device 110, which may be any conventional cathode ray tube, liquid crystal display, light-emitting diode display, etc. In such embodiments, the parallel processing subsystem 112 includes circuitry optimized for graphics and video processing, including, for example, video output circuitry. This circuitry may be incorporated into one or more parallel processing units (“PPUs”), also referred to herein as parallel processors, which are included in the parallel processing subsystem 112. In other embodiments, the parallel processing subsystem 112 includes circuitry optimized for general and / or computational processing. Again, this circuitry may be incorporated into one or more PPUs included within the parallel processing subsystem 112, which are configured to perform such general and / or computational operations. In still other embodiments, one or more PPUs included within the parallel processing subsystem 112 may be configured to perform graphics processing, general processing, and computational processing operations. System memory 104 includes at least one device driver configured to manage the processing operations of one or more PPUs within parallel processing subsystem 112.

[0027] Parallel processing subsystem 112 can be with Figure 1 It can be integrated with one or more other components to form a single system. For example, the parallel processing subsystem 112 can be integrated with the CPU 102 and other interconnect circuitry on a single chip to form a system-on-a-chip (SoC).

[0028] In one embodiment, CPU 102 is the main processor of computer system 100, controlling and coordinating the operation of other system components. In one embodiment, CPU 102 issues commands to control the operation of PPUs. In some embodiments, communication path 113 is a PCI Express link, where a dedicated path is allocated to each PPU, as known in the art. Other communication paths may also be used. PPUs advantageously implement a highly parallel processing architecture. PPUs may be provided with any amount of local parallel processing memory (PP memory).

[0029] It should be understood that the system illustrated herein is illustrative, and variations and modifications are possible. First, the system's functionality can be distributed across multiple nodes in a distributed and / or cloud computing system. Second, the connection topology (including the number and arrangement of bridges, the number of CPUs 102, and the number of parallel processing subsystems 112) can be modified as needed. For example, in some embodiments, system memory 104 can be directly connected to CPU 102 instead of being connected to CPU 102 via memory bridge 105, and other devices will communicate with system memory 104 via memory bridge 105 and CPU 102. In another example, parallel processing subsystem 112 can be connected to I / O bridge 107 or directly to CPU 102 instead of being connected to memory bridge 105. In a third example, I / O bridge 107 and memory bridge 105 can be integrated into a single chip instead of existing as one or more discrete devices. Third, Figure 1 One or more of the components shown may be absent. For example, the direct connection of switch 116, network adapter 118, and add-on cards 120, 121 to I / O bridge 107 can be eliminated. Fourth, computer system 100 can be configured to perform any of the methods described herein.

[0030] In one or more embodiments, the computer system 100 is configured to execute a training engine 122 and an execution engine 124 residing in system memory 104. The training engine 122 and the execution engine 124 may be stored in system disk 114 and / or other storage devices and are loaded into system memory 104 when executed.

[0031] More specifically, training engine 122 and execution engine 124 include functions for training and executing various types of machine learning models to implement perturbation response and target cell state modeling. Training engine 122 trains one or more perturbation response models to predict the response of one or more biological systems (e.g., cell type, cell line, tissue, organ, organism, species, etc.) to genetic, chemical, and / or other types of perturbations based on inputs including vectors, encodings, and / or representations of gene expression, basal state, covariates, and / or perturbations associated with one or more biological systems. For example, the perturbation response models may include combinatorial perturbation autoencoders (CPA) and / or other types of deep learning models that generate embeddings representing basal state, cell type, perturbations, and / or covariates associated with the biological system. The one or more perturbation response models can also be used to decode combinations of these embeddings into gene expression, images (also referred to herein as “cell imaging data”), proteins, and / or other outputs representing the biological system’s response to perturbations. Therefore, the response predicted by the one or more perturbation response models can include measurable and / or quantifiable changes that occur in the biological system after the perturbation has been applied.

[0032] Training engine 122 also trains one or more cell state models to predict cell states associated with a biological system. For example, the one or more cell state models may include nonlinear multivariate regression models, deep learning models, and / or other types of machine learning models that use input gene expression, images, proteins, and / or other responses of the biological system to perturbations to predict the biological age, health or disease status, cell type, phenotype, and / or other cell states corresponding to the representation. Each predicted cell state may include measurable attributes that indicate and / or represent the health status of the corresponding biological system.

[0033] After training, execution engine 124 integrates the perturbation response model and cell state model into the feedback loop of computer simulation and physical experiment to accelerate the discovery of promising hypotheses or interventions. More specifically, the one or more perturbation response models can be used to explore perturbation responses across cell types, disease conditions, baseline states, covariates, modalities, perturbation types, perturbation combinations, and / or other variables. The perturbation responses generated by the one or more perturbation response models can then be input into the one or more cell state models, and the cell state predictions generated by the one or more cell state models based on the input perturbation responses can be used to identify subsets of perturbations that are expected to restore a youthful state to various cell types and / or otherwise improve the health status or biological age of various cell types. The identified perturbations can then be validated through various experimental screenings, the results of which can be used to retrain the one or more perturbation response models and / or the one or more cell state models. The operation of training engine 122 and execution engine 124 is described in further detail below.

[0034] Perturbation response and target cell state modeling Figure 2 According to various embodiments Figure 1 A more detailed illustration of the training engine 122 and execution engine 124 is provided. As described above, the training engine 122 and execution engine 124 operate to train and execute multiple sets of machine learning models to perform perturbation response and target cell state modeling. Figure 2 As shown, the machine learning model includes one or more perturbation response models 202 and one or more cell state models 204.

[0035] The perturbation response model 202 includes a deep learning model that generates latent values ​​210 representing cellular attributes 242 (e.g., basal state and covariates), perturbations 244, and / or other variables affecting the biological age, health status, and / or function of tissues, cell types, cell lines, and / or other types of biological systems. In some embodiments, the latent values ​​210 correspond to the embeddings of variables in a low-dimensional latent space. The perturbation response model 202 also uses various combinations of latent values ​​210 to generate predictions of a perturbation response 246 represented by combinations of variables.

[0036] More specifically, the perturbation response model 202 includes one or more encoders 206 that decompose the data representing the perturbation response 246 into different sets of latent values ​​210 corresponding to embeddings of cell type, perturbation, covariates, and / or other variables of interest. The perturbation response model 202 also includes one or more decoders 208 that reconstruct the perturbation response 246 based on combinations of embeddings representing the corresponding variables of interest. The decoders 208 can also (or alternatively) be trained to transform combinations of embeddings representing new combinations of these variables of interest into predictions of the perturbation response 246 associated with those variables of interest. For example, each of the perturbation response models 202 may include CPA, multimodal CPA, ChemCPA, PerturbNet, variational autoencoder (VAE), drug response VAE (Dr.VAE), Geneformer, scGPT, and / or other types of deep learning models that use one or more encoders 206 and one or more decoders 208 to generate latent values ​​210 associated with different perturbation responses 246 and transform combinations of the generated latent values ​​210 into predictions of new perturbation responses 246. Therefore, each of the perturbation response models 202 is trained to learn a latent space representation of data representing different variables of interest, wherein the latent values ​​210 in the latent space representation can be used to reconstruct individual variables of interest and / or combinations of variables of interest.

[0037] In one or more embodiments, the training engine 122 trains the perturbation response model 202 using one or more perturbation datasets 220. For example... Figure 2 As shown, the perturbation dataset 220 includes training base states 226, training covariates 228, training perturbations 224, and training responses 230 associated with one or more biological systems. The training base states 226 correspond to the states of the biological systems observed, measured, and / or inferred before the training perturbations 224 are applied to the biological systems (or, alternatively, in the absence of the training perturbations 224). For example, the training base states 226 may include information indicating the gene expression, appearance, protein, and / or other quantifiable states of the biological systems before they are exposed to the training perturbations 224.

[0038] Training perturbation 224 includes various types of interventions that can be applied to a biological system. For example, training perturbation 224 may include information indicating drugs, genetic perturbations, procedures, and / or other types of controlled events that can affect a biological system.

[0039] Training covariates 228 include variables that influence the outcome of applying training perturbation 224 to a biological system having a corresponding training baseline state 226. For example, training covariates 228 may include dose, time, and / or other variables associated with applying training perturbation 224 to the biological system. Training covariates 228 may also (or alternatively) include information indicating cell type, cell line, disease state, phenotype, species, patient, age, and / or other variables associated with the biological system to which training perturbation 224 has been applied.

[0040] Training response 230 includes the state of the biological system observed, measured, and / or inferred after training perturbation 224 has been applied to the biological system. For example, training response 230 may include information indicating gene expression, patterns, proteins, and / or other representations of the biological system after it has been exposed to genetic, chemical, and / or other types of perturbations.

[0041] During the training of the perturbation response model 202, the training engine 122 inputs observed data representing the training responses 230 of the biological system into the encoder 206. The encoder 206 generates latent values ​​210, which represent estimates of the training base states 226, training covariates 228, and training perturbations 224 associated with these training responses 230. The decoder 208 converts the generated latent values ​​210 into a training output 212, which includes a reconstruction of the training responses 230. The training engine 122 computes one or more losses 216 based on the latent values ​​210, the training output 212, and / or data from the perturbation dataset 220. The training engine 122 also uses training techniques (e.g., gradient descent and backpropagation) to update the parameters of the encoder 206 and decoder 208 within the perturbation response model 202 in a manner that reduces the losses 216.

[0042] Figure 3 The illustrations depict various embodiments. Figure 1 How does the training engine 122 in the model train a set of perturbation response models 202? For example... Figure 3 As shown, the training engine 122 inputs observation data 300 into one or more encoders 206 included in the perturbation response model 202. For example, the training engine 122 may input gene expression, images, protein sets and / or other observation data 300 from a given biological system into the encoder 206.

[0043] The training engine 122 uses encoder 206 to transform training samples, including observation data 300, into one or more latent values ​​210, which include a base state embedding 302, a covariate embedding 304, and / or a perturbation embedding 306. The base state embedding 302 represents the base state of the biological system (e.g., the base state in training base state 226), the covariate embedding 304 represents one or more training covariates 228 associated with the biological system, and the perturbation embedding 306 represents one or more training perturbations 224 applied to the biological system.

[0044] Training engine 122 merges the base state embedding 302, covariate embedding 304, and perturbation embedding 306 into a unified embedding 308. For example, training engine 122 may compute the unified embedding 308 as the sum of the base state embedding 302, covariate embedding 304, and perturbation embedding 306 and / or another aggregation form.

[0045] The training engine 122 also inputs the uniform embedding 308 into one or more decoders 208 included in the perturbation response model 202, and uses the decoders 208 to convert the uniform embedding 308 into reconstructed data 310 corresponding to the reconstruction results of the observation data 300. For example, the training engine 122 can use the decoders 208 to generate gene expression, protein data, images, and / or other reconstruction results of the input observation data 300.

[0046] The training engine 122 also uses the reconstruction loss 324, calculated between the observed data 300 and the reconstructed data 310, to train the encoder 206 and decoder 208. For example, the training engine 122 may calculate the reconstruction loss 324 as one or more Gaussian negative log-likelihoods, negative binomial losses, Kullback-Leibler (KL) divergences, and / or other error measures between the representation of the observed data 300 and the corresponding representation of the reconstructed data 310. The training engine 122 may also use a weighted combination of the calculated reconstruction loss 324 and the adversarial loss 322 to update the parameters of the encoder 206, which generates the basic state embedding 302, the covariate embedding 304, and the perturbation embedding 306, and the decoder 208, which transforms the uniform embedding 308 into the reconstructed data 310.

[0047] In some embodiments, the training engine 122 also uses a set of adversarial losses 322 to train the encoder 206 and decoder 208. More specifically, the training engine 122 may train one or more encoders 206 and one or more discriminator neural networks (not shown) in an adversarial manner to generate the base state embeddings 302. In this adversarial training process, a given encoder learns to generate a base state embedding 302 representing the base state of a biological system, and the one or more discriminator neural networks cannot use the base state embedding 302 to predict perturbation or covariate values. For example, the training engine 122 may perform adversarial training on the encoder 206 and discriminator neural networks by inputting other representations of gene expression and / or base state into the encoder and using the encoder to encode the input representations into the corresponding base state embeddings 302. To ensure that information about covariate and perturbation values ​​is removed from the base state embeddings 302, the training engine 122 alternates between two training steps, training the one or more discriminator neural networks and the encoders 206 and decoders 208 in the perturbation response model 202 separately. During the first training step, the training engine 122 trains the one or more discriminator neural networks using cross-entropy loss and / or other adversarial losses 322, which are calculated between the predictions of perturbations and covariates generated by the one or more discriminator neural networks from the base state embeddings 302 and the training perturbations 224 and training covariates 228 associated with the corresponding observed data 300. During the second training step, the training engine 122 trains the encoder 206 and decoder 208 using a combined loss, including a reconstruction loss calculated between the reconstructed data 310 and the corresponding observed data 300, and a cross-entropy loss.

[0048] Therefore, the adversarial loss 322 enables the encoder 206 to learn a latent space that decouples the fundamental states of the biological system from the perturbations and covariates associated with the biological system. Simultaneously, the reconstruction loss 324 is used to train the encoder 206 and decoder 208 to reconstruct a set of given observation data 300 from the corresponding fundamental state embeddings 302, covariate embeddings 304, and perturbation embeddings 306 in the latent space. After training the encoder 206 and decoder 208 in this manner, they can be used to predict the additional responses of the biological system to new combinations of perturbations, as discussed in further detail below.

[0049] Back to Figure 2As discussed herein, cell state model 204 includes multilayer perceptron (MLP), deep learning models, and / or other types of machine learning models that convert the perturbation response 246 output by perturbation response model 202 into cell states 248 representing the corresponding perturbed biological system. More specifically, cell state model 204 includes one or more classifiers and / or regression models that, given inputs including data representing the combined response of the biological system to perturbation 224, generate cell state predictions 236(1) to 236(N) (each referred to herein as cell state prediction 236) representing the cell states associated with the biological system. These cell state predictions 236 may include (but are not limited to) scores and / or other numerical values ​​representing biological age, health status, disease status, phenotype, cell type prevalence, and / or other measurable, observable, and / or quantifiable attributes related to the health or functioning of the biological system.

[0050] In one or more embodiments, the training engine 122 trains the cell state model 204 using one or more cell health datasets 222. Figure 2 As shown, the cell health dataset 222 includes training cell states 234 and training cell data 232 associated with one or more biological systems. Cell states 248 include various measurable "attributes of interest" associated with the health or functional operation of a biological system. For example, cell states 248 may include predictions of probability and / or magnitude related to biological age (e.g., transcriptomic age or chronological age) and / or health or disease status, disease progression, cell type, phenotype, and / or other attributes that can be used to characterize the overall state or health of a biological system (e.g., characterizing cellular energy metabolism by measuring mitochondrial function, and / or characterizing the biological function of cells such as fibroblasts by measuring fibroblast migration).

[0051] Training cell data 232 corresponds to observations of one or more biological systems. For example, training cell data 232 may include transcriptome data (also referred to herein as gene expression), images, protein expression, and / or other information indicating measurable properties of interest for the one or more biological systems.

[0052] During the training of cell state model 204, training engine 122 inputs training cell data 232 of one or more biological systems into cell state model 204. Training engine 122 executes cell state model 204 to transform the input training cell data 232 into cell state prediction results 236 for one or more biological systems. Training engine 122 uses individual cell state prediction results 236 and / or combinations of cell state prediction results 236 as training output 214 of cell state model 204. Training engine 122 calculates one or more losses 218 based on the training output 214 and the corresponding "ground truth" values ​​in cell health dataset 222 for training cell states 234. In addition, training engine 122 updates the parameters of cell state model 204 using training techniques (e.g., gradient descent and backpropagation) to reduce the losses 218.

[0053] Figure 4 The illustrations depict various embodiments. Figure 1 How does the training engine 122 in the model train a set of cell state models 204? Figure 4 As shown, the cell state model 204 includes a basic model 430 and several tissue-specific models 432(1) to 432(X) (each of which is referred to as tissue-specific model 432 in this paper).

[0054] The training engine 122 first trains a base model 430 to generate biological age prediction results 408, cell type prediction results 410, and health status prediction results 412 based on inputs including transcriptome data 400 (including but not limited to gene expression) and cell imaging data 402 from multiple tissue types. Transcriptome data 400 and cell imaging data 402 may be included in training cell data 232 from one or more cell health datasets 222.

[0055] Transcriptome data 400 includes observed gene expression from one or more biological systems. For example, transcriptome data 400 may include ribonucleic acid (RNA) transcripts from cells, cell lines, tissues, organisms, and / or other biological systems. These RNA transcripts may be generated via microarrays, RNA sequencing technology, and / or other technologies.

[0056] Cell imaging data 402 includes images of the one or more biological systems. For example, cell imaging data 402 may include bright-field images, stained cell images, and / or other types of images of the one or more biological systems acquired using optical and / or other microscopy techniques.

[0057] More specifically, training engine 122 inputs transcriptome data 400 and / or cell imaging data 402 from multiple tissue types into base model 430. Training engine 122 uses base model 430 to generate a biological age prediction result 408, which includes a numerical value representing the biological age of the corresponding biological system. Training engine 122 also uses (or alternatively uses) base model 430 to generate a cell type prediction result 410, which includes scores and / or other outputs representing the probability or prevalence of a given cell type in the biological system. Training engine 122 also uses (or alternatively uses) base model 430 to generate a health status prediction result 412, which includes scores and / or other outputs characterizing the probability, degree, progression, and / or other measures associated with a specific health or disease attribute in the biological system.

[0058] Training engine 122 calculates a first set of losses 414 between the biological age prediction 408, cell type prediction 410, and / or health status prediction 412 generated by the base model 430 and the corresponding training cell states 234 in the one or more cell health datasets. For example, training engine 122 may calculate the losses 414 as mean squared error (MSE), mean absolute error (MAE), Huber loss, quantile loss, and / or other error measures between the predictions generated by the base model 430 and the corresponding ground truth training cell states 234.

[0059] The training engine 122 also updates the parameters of the base model 430 by reducing the calculated loss 414. For example, the training engine 122 can train the base model 430 using the loss 414 until one or more conditions are met. These one or more conditions include (but are not limited to) the convergence of the parameters of the base model 430, the reduction of the loss 414 below a threshold, and / or reaching a certain number of training steps, iterations, batches, and / or epochs.

[0060] After training the base model 430, the training engine 122 uses the trained base model 430 as the starting point for training additional tissue-specific models 432. For example, the training engine 122 can initialize each tissue-specific model 432 using the parameters of the trained base model 430.

[0061] like Figure 4As shown, the training engine 122 uses a set of different tissue-specific data 418(1) to 418(X) (each of which is referred to herein as tissue-specific data 418) to train the corresponding tissue-specific model 432. In some embodiments, each set of tissue-specific data 418 includes transcriptome data 400, cell imaging data 402, and / or training cell states 234 for different tissue types (such as, but not limited to, kidney cells, endothelial cells, fibroblasts, nerve cells, muscle cells, and / or other types of tissues of interest).

[0062] In one or more embodiments, training engine 122 adjusts or fine-tunes each tissue-specific model 432 to generate a biological age prediction 408, a cell type prediction 410, and / or a health status prediction 412 for the corresponding tissue type. More specifically, training engine 122 inputs transcriptome data 400, cell imaging data 402, and / or other observed training cell data 232 included in the tissue-specific dataset into different tissue-specific models 432. Training engine 122 uses tissue-specific models 432 to generate predictions 416(1) to 416(X) of biological age, cell type, health status, and / or other attributes (each of which is referred to herein as prediction 416). For each tissue-specific model 432, the training engine 122 computes one or more losses 420(1) to 420(X) (each referred to herein as loss 420) as mean squared error (MSE), mean absolute error (MAE), Huber loss, quantile loss and / or other error measures between the prediction result 416 output by the tissue-specific model 432 and the corresponding ground truth training cell state 234 in the tissue-specific data 418 associated with the tissue-specific model 432.

[0063] Training engine 122 then updates the parameters of each tissue-specific model 432 in a manner that reduces the calculated loss 420. For example, training engine 122 may train the tissue-specific model 432 using the corresponding loss 420 until one or more conditions are met. These one or more conditions include (but are not limited to) convergence of the parameters of a given tissue-specific model 432, the loss 420 of the tissue-specific model 432 being reduced below a threshold, and / or reaching a certain number of training steps, iterations, batches, and / or cycles.

[0064] By training the base model 430 to generate biological age predictions 408, cell type predictions 410, and / or health status predictions 412 for various types of tissues and / or biological systems, the training engine 122 generates a multi-tissue model incorporating biological age, health status, cell type, and / or other trained cell states 234. The training engine 122 can then train additional tissue-specific models 432 to better understand aging, rejuvenation, health, and / or disease in the corresponding tissue types. As discussed in further detail below, the trained base model 430 and / or tissue-specific models 432 can then be used to predict cell states 248 associated with perturbation responses 246 generated by the perturbation response model 202 based on latent values ​​210 representing new combinations of base states, covariates, and perturbations in the biological system.

[0065] Back to Figure 2 The discussion continues, and in some embodiments, after the training engine 122 has trained a set of perturbation response models 202 and a set of cell state models 204, the execution engine 124 uses one or more trained perturbation response models 202 and trained cell state models 204 in the feedback loop of the computer simulation and physical experiment 250 to accelerate the discovery of promising hypotheses or interventions. More specifically, the execution engine 124 uses one or more perturbation response models 202 to transform different combinations of cell attributes 242 (e.g., cell type, cell line, disease status, baseline state, covariates, modality, etc.) and perturbations 244 into corresponding predicted perturbation responses 246. For example, the execution engine 124 may use perturbation response models associated with a specific modality, set of modalities, and / or biological system to transform different combinations of cell attributes 242 and perturbations 244 into perturbation responses 246. As used herein, “modality” refers to a data type such as (but not limited to) text data, image data, audio data, transcriptome data, and / or other data representing observations and / or characterizations of one or more cells.

[0066] To reduce the resource overhead associated with generating the perturbation response 246, the execution engine 124 may use expert knowledge, existing research, genome editing libraries, and / or other data sources related to perturbations in biological systems to determine and / or screen combinations of cell attributes 242 and perturbations 244. For example, the execution engine 124 may retrieve specific perturbations 244 and cell attributes 242 of cells to which applying these perturbations 244 results in cell health restoration and / or other outcomes of interest from data sources.

[0067] The execution engine 124 also uses one or more cell state models 204 to identify perturbation responses 246 that result in a target cell state 248 for the corresponding biological system. In some embodiments, the target cell state 248 represents a healthy, youthful, and / or other “desired” cell state of the biological system. For example, the target cell state 248 may include a target biological age, health score, cell type score, phenotype, and / or other values ​​associated with the output of the cell state model 204.

[0068] To determine the perturbation response 246 corresponding to the target cell state 248, the execution engine 124 inputs the perturbation response 246 into one or more cell state models 204. Using these cell state models, the execution engine 124 predicts the biological age, health status, cell type, and / or other cell states of the biological system based on the input perturbation response 246. The execution engine 124 also identifies perturbation responses 246 that have predicted cell states corresponding to the target cell state 248. For example, the execution engine 124 may identify a subset of perturbation responses 246 whose biological age, health score, cell type score, and / or other predicted cell states fall within or within the range of values ​​corresponding to the target cell state 248. In another example, the execution engine 124 may aggregate the biological age, health score, cell type score, and / or other predicted cell states generated by the cell state models 204 (e.g., using summation, weighted summation, averaging, etc.) to obtain an overall score for the corresponding perturbation responses 246. The execution engine 124 can sort the perturbation responses 246 according to the overall score and use this sorting to select the subset of perturbation responses 246 with the best overall score. In both examples, the perturbation responses 246 identified by the execution engine 124 may represent perturbations that may rejuvenate various types of cells and / or otherwise improve the health or biological age of various types of cells.

[0069] The execution engine 124 additionally uses one or more experiments 250 to verify the effects of the perturbation on the biological system. For example, the execution engine 124 may generate an output including a list of cell attributes 242 associated with the target cell state 248, a perturbation 244, and a perturbation response 246. This output can be used to conduct one or more physical experiments 250 to determine whether a perturbation 244 applied to a biological system having cell attribute 242 results in a predicted perturbation response 246. Experiment 250 may also be used (or alternatively) to determine whether the biological system's response to the perturbation 244 results in the target cell state 248.

[0070] Execution engine 124 may additionally provide the results of experiment 250 to training engine 122, and training engine 122 may use these results to retrain one or more perturbation response models 202 and / or one or more cell state models 204. For example, execution engine 124 may add the base states, covariates, perturbations, and observed responses to these perturbations from experiment 250 to one or more new perturbation datasets 220. Execution engine 124 may provide the new perturbation datasets 220 to training engine 122, and training engine 122 may use the new perturbation datasets 220 to retrain one or more existing perturbation response models 202, and / or train one or more new perturbation response models 202. In another example, execution engine 124 may populate one or more new cell health datasets 222 with training cell data 232, which includes observed responses to perturbation 244 from experiment 250 and training cell states 234, which include biological age, health score, cell type score, and / or other attributes associated with the observed responses to perturbation 244. Execution engine 124 may provide this new cell health dataset 222 to training engine 122, and training engine 122 may use this new cell health dataset 222 to retrain one or more existing cell state models 204, and / or train one or more new cell state models 204.

[0071] Training engine 122 and execution engine 124 can continue to train or retrain perturbation response model 202 and cell state model 204, using the trained or retrained perturbation response model 202 and cell state model 204 to predict perturbation responses 246 and cell states associated with various combinations of cell attributes 242 and perturbations 244, thereby identifying perturbation responses 246 corresponding to target cell states 248, and performing experiments 250 to verify the perturbation responses 246 and / or target cell states 248 associated with perturbations 244 of the corresponding biological system. In repeating this cycle, training engine 122 and execution engine 124 can generate more accurate perturbation response models 202 and cell state models 204 for various biological systems, and identify additional perturbation responses 246 and corresponding perturbations 244 that can lead to target cell states 248 in the biological system. Therefore, training engine 122 and execution engine 124 include the capability to improve the understanding of biological systems and aging within biological systems over time.

[0072] Figure 5 Flowcharts of method steps for performing perturbation response and target cell state modeling according to various embodiments are provided. Although the method steps are combined... Figures 1 to 4The system described herein is for informational purposes only; however, those skilled in the art should understand that any system configured to perform some or all of the method steps in any order falls within the scope of this disclosure.

[0073] As shown in the figure, in step 502, training engine 122 collects one or more perturbation datasets and one or more cell health datasets. For example, training engine 122 may generate perturbation datasets and / or cell health datasets based on experimental results, public datasets, and / or other datasets including the biological system's response to perturbations and / or cell states associated with the biological system.

[0074] In step 504, training engine 122 uses perturbations, covariates, and perturbation responses from the perturbation dataset to train a first set of machine learning models. For example, training engine 122 may train one or more encoders in this first set of machine learning models to transform observational data representing the biological system's response to perturbations into latent values ​​representing the base state, covariates, and perturbations. Training engine 122 may also train one or more decoders in the first set of machine learning models to transform combinations of latent values ​​into reconstructed responses of the biological system to perturbations.

[0075] In step 506, training engine 122 trains a second set of machine learning models using observational data, biological age, and health status from a cell health dataset. For example, training engine 122 may train one or more nonlinear multivariate regression models to predict the biological age, health score, and / or cell type score of various biological systems based on gene expression, images, and / or other representations of the biological system. This nonlinear multivariate regression model may include a base model trained to predict cell states in different tissue types. After training of this base model is complete, it can be used to initialize multiple tissue-specific models. Subsequently, each tissue-specific model can be fine-tuned using different tissue-specific datasets to generate cell state predictions for the corresponding tissue type.

[0076] In step 508, execution engine 124 predicts the responses of one or more biological systems to different combinations of perturbations by executing a first set of machine learning models. For example, execution engine 124 may use one or more encoders from the first set of machine learning models to generate embeddings of the basic states, covariates, and perturbations associated with the one or more biological systems. Execution engine 124 may combine the basic state embeddings, covariate embeddings, and perturbation embeddings (e.g., summing, averaging, aggregating, splicing, etc.) into a unified embedding representing the corresponding combinations of basic states, covariates, and perturbations in the one or more biological systems. Execution engine 124 may then use one or more decoders from the first set of machine learning models to convert the unified embedding into other representations of gene expression, proteins, images, and / or the responses of the one or more biological systems to perturbations.

[0077] In step 510, execution engine 124 predicts cell states associated with the response of the one or more biological systems to a perturbation by executing a second set of machine learning models. For example, execution engine 124 may input one or more gene expressions, images, and / or other data representing the predicted response of a biological system to a perturbation into a regression model included in the second set of machine learning models. Execution engine 124 can use the regression model to convert the data into a representation of the biological system's biological age, health score, cell type score, and / or other cell states.

[0078] In step 512, execution engine 124 determines one or more perturbations associated with one or more target cell states based on the predicted cell states. For example, execution engine 124 may aggregate the predicted biological age, health score, cell type score, and / or other cell state representations for each biological system into an overall score for that biological system. Execution engine 124 may rank the biological system's responses to various perturbations according to the corresponding overall scores. Execution engine 124 may then select a subset of perturbations that produce the highest or optimal overall score (e.g., a highest-ranking perturbation as a predefined number, or a variable number of perturbations where the overall score reaches or exceeds a threshold, etc.). In another example, execution engine 124 may identify a subset of perturbations capable of predicting the biological age, health score, cell type score, and / or other cell state representations corresponding to one or more predefined target cell states.

[0079] In step 514, execution engine 124 enables one or more experimental evaluations of one or more perturbations to the one or more biological systems to be performed. For example, execution engine 124 may generate an output including baseline state, covariates, perturbations, responses to perturbations, predicted cell states, and / or other attributes associated with the perturbations identified in step 512. This output can be used to conduct experiments to verify the responses of the one or more biological systems to perturbations, the cell states represented by the responses, and / or other predictions generated by the first set of machine learning models and / or the second set of machine learning models.

[0080] In step 516, training engine 122 and / or execution engine 124 determine whether to continue training and executing the machine learning model. For example, training engine 122 and / or execution engine 124 may continue training and executing the machine learning model within a feedback loop of computer simulations and physical experiments to better understand the life mechanisms and rejuvenation of the one or more biological systems. If training engine 122 and / or execution engine 124 determine that training and executing the machine learning model should not continue, no additional processing will be performed using these machine learning models.

[0081] If the training engine 122 and / or the execution engine 124 determine that the machine learning model should continue to be trained and executed, the execution engine 124 performs step 518, in which it determines the results of one or more experimental evaluations. For example, the execution engine 124 may determine perturbation response, cell state, and / or other results associated with experiments involving the application of the perturbations identified in step 512 to the one or more biological systems. In step 520, the execution engine 124 uses these results to update the perturbation dataset and / or cell health dataset. For example, the execution engine 124 may use these results to populate one or more new perturbation datasets and / or cell health datasets. The execution engine 124 may also (or alternatively) add these results to one or more existing perturbation datasets and / or cell health datasets.

[0082] After updating the perturbation dataset and / or cell health dataset using experimental results, training engine 122 and execution engine 124 repeat steps 504, 506, 508, 510, 512, 514, and 516 to train and / or retrain the first and second sets of machine learning models using the updated perturbation dataset and / or cell health dataset; identify additional perturbations that may improve cell health, biological age, and / or other cell states in one or more biological systems using the updated machine learning models; and experimentally verify the impact of the perturbations on the cell states of one or more biological systems. Thus, training engine 122 and execution engine 124 can continue the feedback loop during the feedback loop of the machine learning model used in computer simulations and physical experiments to enhance the understanding of life mechanisms and rejuvenation in one or more biological systems.

[0083] Exemplary Implementation The following content is presented as an example only and should not be construed as limiting the scope of the claims or the disclosed embodiments.

[0084] Exemplary Disturbance Response Model The following describes an exemplary perturbation response model that predicts the expression profile of cells after being perturbed by genetic or chemical intervention. In this example, the perturbation response model operates by transforming the input data into a low-dimensional, meaningful latent representation, which can then be used for various regression and / or classification tasks.

[0085] This exemplary perturbation response model can be applied to any perturbation dataset with transcriptome reads. A specific example dataset is a CRISPR-Cas9 gene knockout screening of patient-derived melanoma cells under three different culture conditions. Specifically, the Frangieh21 dataset includes 248 immune checkpoint response-related genes that have been knocked out in melanoma cells cultured under three different biological conditions. These conditions are for melanoma cells: 1) cultured alone; 2) stimulated with interferon-gamma (IFNg); and 3) co-cultured with tumor-infiltrating leukocytes, referred to herein as 1) none, 2) IFNg, and 3) co-culture conditions, respectively. These co-culture conditions are examples of covariates, defined as a data property that is not itself a perturbation but rather a data property that the perturbation response model should be aware of. A dataset may include multiple covariates (e.g., both cell type and culture conditions), but this dataset includes only one type of covariate. The Frangieh21 dataset also includes gene expression responses measured using single-cell RNA sequencing technology, resulting in gene expression profiles of 218,331 single cells, with each cell representing a separate data point.

[0086] The Frangieh21 dataset was split into training, validation, and test splits per perturbation and per covariate, with all cells belonging to a given perturbation under a given covariate assigned to the same split. The training split included all perturbations under the none and IFNg covariates, and 74 perturbations under the co-culture covariate. 87 perturbations under the co-culture covariate were reserved for the validation split, and another 87 perturbations under the co-culture covariate were reserved for the test split. This split was designed to simulate a scenario where perturbations under simpler biological conditions (none and IFNg) were measured and used to predict the effects of perturbations under more complex biological conditions (co-culture).

[0087] The Frangieh21 dataset was preprocessed as follows: 1. Subset-related genes: Select genes that are highly variable or differentially expressed. Using the `scanpy.pp.highly_variable_genes` function with the 'seurat_v3' filtering method, the top 4000 genes are retained. Alternatively, for each perturbation in each unique covariate set, the top 25 differentially expressed genes can be retained using the `scanpy.tl.rank_genes_groups` method with default parameters. Perturbation-affected genes are also selected.

[0088] 2. Normalization: To achieve comparability between cells, total count normalization was applied, which involved scaling the original counts based on the total count per cell and multiplying them by a scaling factor of 10000. Subsequently, the normalized expression values ​​were stabilized using a logarithmic transformation, with a pseudo-count of 1 added to avoid taking the logarithm of zero.

[0089] For the Frangieh21 example dataset, preprocessing produces a dataset with normalized gene expression values ​​for 9140 genes, which are the features of the dataset. The number of data points remains unchanged.

[0090] Figure 6 The illustration shows example architectures of disturbance response models according to various embodiments. More specifically, Figure 6 An example architecture of a perturbation response model is illustrated, which is trained to predict the expression profile of cells after being perturbed by genetic or chemical interventions, as described in this article regarding this example.

[0091] like Figure 6As shown, the perturbation response model includes two encoders 206(1) and 206(2) and a decoder 208(1). The input to encoder 206(1) includes a vector 602 of gene expression from randomly sampled control cells (or the average of multiple sampled control cells), i.e., x i These sampled control cells have been normalized and logarithmically transformed as described in the preprocessing section. The length is equal to the number of genes in the dataset, which is 9140 for the Frangieh19 dataset.

[0092] The input to encoder 206(1) also includes one or more covariates (cell type, treatment method, batch, etc.), which are represented as a vector 604 composed of one-hot codes of each covariate, where the length of vector 604 is equal to the number of unique covariates, i.e., c. i For example, in the Frangieh21 dataset, the covariate representation for the “none” condition will be similar to [1,0,0]; for the “IFNg” condition, the covariate representation will be similar to [0,1,0]; and for the “co-culture” condition, the covariate representation will be similar to [0,0,1].

[0093] The input to encoder 206(2) includes a perturbation response represented as a multi-hot coding vector 606, the length of which is equal to the number of unique perturbations, i.e., p. i For example, if a cell receives the first of five perturbations, its representation will be [1,0,0,0,0]. If a cell receives both perturbation 1 and perturbation 3, its representation will be [1,0,1,0,0]. For the Frangieh21 dataset, this vector 606 has a length of 248, which corresponds to the total number of unique perturbations excluding the control perturbation.

[0094] Encoder 206(1) is used to map vectors 602 and 604 to a first set of latent values ​​210(1) in the latent space associated with the perturbation response model. Encoder 206(2) is used to map vector 602 to a second set of latent values ​​210(2) in the latent space. Encoders 206(1) and 206(2) can be implemented as multilayer perceptrons (MLPs) with dropout to prevent overfitting and layer normalization to improve convergence and generalization. However, any type of encoder can be used, including but not limited to transformer neural networks. The encoder width is a hyperparameter ranging from 1024 to 8192 neurons, and is set to 2304 neurons for all encoders in the Frangieh21 dataset. The number of hidden encoder layers (excluding the input layer) is a hyperparameter ranging from 1 to 7 layers, and is set to 1 layer for all encoders in the Frangieh21 dataset. The ReLU activation function is used, and the dropout is a hyperparameter that ranges from 0.0 to 0.7, which is set to 0.3 for the Frangieh21 dataset.

[0095] Therefore, the operation of encoder 206(1) is represented as z x , i = f x (x i , c i ), or an encoding function that maps control cell gene expression and covariates to the latent space. The operation of encoder 206(2) can be represented as z p , i = f p (p i ), or an encoding function that generates perturbation data that perturbs a specific underlying representation.

[0096] Once in the latent space, the perturbation response model conditions the latent control cell representation with a perturbation representation by adding two sets of latent values ​​210(1) and 210(2): z i = z x , i + z p , i For the Frangieh21 dataset, a latent dimension of size 256 is used. Concatenation, cross-attention, and / or other conditionalization mechanisms can also be used to merge latent values ​​210(1) and 210(2).

[0097] The resulting latent representation z i The decoder 208(1) converts the data back to the gene expression space to predict the gene expression x of the disturbed cells using the following decoding function. i ':x i ' = g(zi The function is currently implemented in the same way as the encoder MLP, including the same number of layers and width, except that a softplus layer is added at the end to ensure non-negative output.

[0098] By operating in a latent space—where gene expression, covariates, and perturbations are first encoded into latent representations and then combined and transformed back into the original space—perturbation response models can capture nonlinear interactions within the data, thus providing an expressive and powerful approach to predicting the effects of perturbations on gene expression.

[0099] Training the perturbation response model involves iterating over perturbed cells in batches, with the batch size varying depending on the dataset and available computing power. For the Frangieh21 dataset, a batch size of 8000 cells was used.

[0100] For each perturbed cell, one or more control cells are sampled from a pool of control cells sharing the same covariates (cell type, treatment, batch, etc.) as the perturbed cell, and their gene expression is averaged. Taking the Frangieh21 dataset as an example, for a given perturbed cell, control cells are sampled from a pool of control cells sharing the same culture conditions (none, IFNg, co-culture). If 5 cells are sampled, 5 vectors of length 9140 (number of genes / features) are obtained, which are concatenated into a 5×9140 matrix. The average of this matrix is ​​calculated row-wise to obtain the average control vector of length 9140. The number of sampled control cells is an adjustable hyperparameter, ranging from 1 to 75 cells; for the Frangieh21 dataset, a single cell was sampled.

[0101] The exemplary perturbation response model is trained using the mean squared error (MSE) between the predicted and observed gene expression in all genes of the perturbed cells in the dataset as the loss function, which can be formally written as: in This represents the number of genes (9140 for the Frangieh21 dataset). For genes i Observational expression of disturbed cells For genes i The predicted expression is as follows. For a given batch of data, the MSE of each cell in the batch is calculated, and the average MSE of multiple cells in the given batch is taken to obtain the loss. This loss is fed back to the optimizer for a given gradient step.

[0102] The perturbation response model is trained using preset maximum and minimum epoch numbers, where each epoch refers to one complete iteration over the entire dataset. For the exemplary perturbation response model trained on the Frangieh21 dataset, the minimum number of epochs is set to 5, and the maximum number of epochs is set to 400. An early stopping mechanism is also employed, which monitors the validation loss (defined as the average MSE of all batches in the validation split set) for each epoch. If the validation loss does not decrease after 50 consecutive epochs, training of the perturbation response model is terminated before the maximum number of epochs is reached. Once training of the perturbation response model stops, model weights from the epoch with the minimum validation loss are loaded.

[0103] The perturbation response model was trained using the Adam optimizer implemented in PyTorch. The optimizer used default parameters, except for weight decay and the learning rate, both of which are hyperparameters. For the Frangieh21 dataset, the learning rate was set to 0.0000523 and the weight decay to 1.56 × 10⁻⁶. -07 This perturbation response model can use several different learning rate schedulers, including a cosine learning rate scheduler and a scheduler that reduces the learning rate when the validation loss tends to plateau. For the Frangieh21 dataset, the second scheduler was used, which reduces the learning rate by a factor of 5 when the validation loss has not decreased for 15 consecutive epochs. The entire perturbation response model 202 is trained end-to-end, with the MSE loss backpropagated through the entire network.

[0104] To deploy the trained perturbation-response model to generate counterfactual predictions, the perturbation and covariate (i.e., culture conditions) to be predicted are first identified. For each unique set of perturbations and covariates, a set of control cells with matching covariates (as described above) is identified, and these control cells and perturbations are used as input to the perturbation-response model. Using the Frangieh21 example dataset, to predict the effect of knocking out gene A under co-culture conditions, 1000 control cells are sampled from the co-culture conditions, and 1000 predicted perturbed cells are generated using the trained perturbation-response model. The default sampling size is 1000 cells per unique perturbation / covariate set.

[0105] The fidelity of the above perturbation response models was evaluated using several metrics, including cosine similarity and cosine similarity ranking. More specifically, the cosine similarity between the average predicted logarithmic change and the observed logarithmic change was defined as: in For genes i The observed logarithmic change, For genesi The predicted logarithmic change. The logarithmic change is calculated using the following formula: in The average expression (predicted or measured) in disturbed cells. The average expression of the control cells sampled for generating the prediction results. This is a pseudo-count used to avoid division by zero and is set to 0.1. A vector representing the logarithmic multiple change across all genes.

[0106] For the Frangieh21 dataset, all predictions include 1000 cells for each unique perturbation and covariate set. The number of observed cells per unique perturbation and covariate set varies depending on the perturbation.

[0107] A ranking-based metric was also used to measure the effectiveness of the predicted perturbations in ranking. Specifically, for a given observed perturbation, the prediction for that perturbation should be more similar to the predictions for other perturbations. The ranking metric is calculated per perturbation: in This is a vector whose length is equal to the number of genes in the dataset, representing a multiple of the predicted logarithmic change. This is a vector representing observations of the same length that have changed logarithmically. (Index) j Refers to disturbance. p To assess the total number of concentrated disturbances.

[0108] Predictions for the test split set, which includes 87 perturbations under co-culture conditions, have been generated and evaluated. For each predicted perturbation, a cosine similarity and cosine similarity ranking metric were predicted, and the average metric across the 87 perturbations is reported below. The “Random” line indicates the metric expected if the perturbation response model reports a completely random result. The “Perfect” line indicates the metric expected if the perturbation response model perfectly predicts the perturbation gene expression effect. As the results above show, the trained perturbation response model can predict expression changes caused by perturbations in a non-trivial manner.

[0109] Exemplary cell state model As described above, cell state models, given gene expression values ​​as input, output measurable attributes that indicate the health status of a biological system. These attributes include, but are not limited to: 1) predicting the actual age of a donor sample; 2) predicting the probability that a cell or sample originated from a diseased donor rather than a healthy donor; and 3) predicting donor-level biomarkers used to measure disease severity.

[0110] The exemplary cell state model can be trained using the Jerby-Arnon18 dataset. The Jerby-Arnon18 dataset is a single-cell RNA sequencing dataset of malignant melanoma cells from 22 patients, 12 of whom were treatment-naïve and 10 of whom were unresponsive to anti-programmed cell death protein 1 (PD-1) immunotherapy. This falls under the second type of target state attribute described above. It is assumed that the melanoma cells from treatment-naïve patients comprise a mixed population of immunogenic and immune-escape cells, while the melanoma cells from unresponsive patients are assumed to primarily comprise immune-escape cells. The goal is to build a classifier that can categorize cells as “untreated” or “unresponsive,” using the “unresponsive” label as a surrogate measure of immune escape. This results in a filtered scRNA-seq dataset containing 1101 treatment-naïve melanoma cells and 825 unresponsive melanoma cells. Each cell includes a gene expression count and a “untreated / unresponsive” label.

[0111] The dataset was randomly divided into 70% cells for training and 30% cells for testing. The split sets were stratified according to the "untreated / non-responsive" label to ensure class balance in each split set.

[0112] For the Jerby-Arnon18 dataset, the same preprocessing scheme described above for the exemplary perturbation response model was employed. During preprocessing, the dataset was subsetted into 6664 genes that showed significant differential expression between untreated and non-responsive cells in the training split set. Differential expression was defined as an adjusted p-value less than 0.01 using the Wilcoxon test implemented in scanpy.tl.rank_genes_group. This differential expression was performed only on cells from the training split set to avoid potential information leakage.

[0113] Cell state models can employ various architectures, but are typically classifiers or regression models. For the Jerby-Arnon18 dataset, a linear classifier with elastic net penalty was used to reduce overfitting. The input to this exemplary cell state model is a normalized gene expression profile of 6664 cells, and the output is a probability between 0 and 1, where 1 indicates that the cell is likely to be untreated and 0 indicates that the cell is likely to be unresponsive.

[0114] Formally, the cell state model maps gene expression vectors to probabilities: in The probability that the cell originated from a non-responsive donor. This is the normalized gene expression vector of the cell with a length of 6,664.

[0115] During the training of this exemplary cell state model, the LogisticRegressionCV class from scikit-learn is used to identify the optimal L1 and L2 penalty hyperparameters through 4-fold cross-validation on the training split set. Specifically, the training process iterates between 20 penalty intensities, the specific sequence of which is selected by the cell state model. The training process also iterates between 8 L1 / L2 ratios, specifically... The optimal cell state model has a penalty strength of 1.623 and an L1 / L2 ratio of 0.1. The saga solver is used with an error tolerance of 0.001 and a maximum period of 100.

[0116] After training, the optimal cell state model is used to predict the untreated / non-responsive labels in the test segmentation set based on gene expression data. Since a linear model was used for the Jerby-Arnon18 dataset, the target state inference can be expressed as: in It is a vector of length 6664 representing the model weights. It is a constant offset in the form of a vector of length 6664. Other cell state models may employ multilayer perceptron (MLP), random forest classifiers, and / or other types of machine learning architectures.

[0117] The accuracy of this cell state model is quantified using receiver operating characteristic (ROC) curves, which plot the relationship between the true positive rate and the false positive rate. The area under the ROC curve (AUC) represents the performance of the binary classifier.

[0118] Figure 7ROC curves representing the performance of exemplary cell state models according to various embodiments are shown. Figure 7 As shown, this cell state model demonstrates excellent performance in distinguishing between unresponsive melanoma cells and untreated melanoma cells.

[0119] Predicting melanoma cell immune escape using exemplary perturbation response models and exemplary cell state models Regulatory factors Predictive cells were generated for the Frangieh21 test set (which includes 87 perturbations under co-culture conditions), with 1000 cells generated for each perturbation. The non-response score for each predictive cell was then predicted using a cell state model. Only 2508 genes in the perturbation response model predictions overlapped with the target state model genes; therefore, the remaining 4156 genes were padded with zeros for all cells.

[0120] For each perturbation in the test set, the predicted non-response score of all cells with that perturbation was averaged to obtain the average predicted non-response score for all 87 perturbations. The non-response scores were ranked, and the table below lists the top 5 gene knockouts predicted to cause non-response to melanoma immunotherapy. The top three perturbations are associated with JAK1, JAK2, and IFN-γ receptors. In cancer, IFN-γ is secreted by cytotoxic T cells and binds to its receptor on melanoma cells, activating the JAK / STAT / IRF1 signaling pathway, which ultimately leads to PD-L1 expression. Clinical studies have shown that high levels of PD-L1 expression can predict response to anti-PD-1 immunotherapy, suggesting that blocking PD-L1 expression by knocking out IFN-γ receptors or certain components of the JAK / STAT / IRF1 signaling pathway may result in no response. The next two perturbations are CDKN2B and CDKN1A (also known as p16). Knocking out senescence regulators such as p16 / CDKN1A is known to reduce the efficacy of immunotherapy. CDKN2B is a known negative regulator of CDK4, and inhibiting CDK4 in a mouse melanoma model can improve the efficacy of immunotherapy. Therefore, all five top perturbations proposed by combining exemplary perturbation response models and exemplary cell state models are directly associated with immunotherapy resistance.

[0121] In summary, the disclosed techniques train and execute multiple sets of machine learning models to perform perturbation response and target cell state modeling. One or more perturbation response models are trained to predict the responses of one or more biological systems to genetic, chemical, and / or other types of perturbations. For example, the one or more perturbation response models may include combinatorial perturbation autoencoders (CPAs) and / or other types of deep learning models with encoder-decoder architectures. The encoders in the one or more perturbation response models can be used to generate embeddings representing the underlying state, cell type, perturbation, and / or covariates associated with the biological system. The decoders in the one or more perturbation response models can be used to decode combinations of embeddings into gene expression, images, proteins, and / or other outputs representing the biological system's response to perturbations.

[0122] One or more cell state models are trained to predict cell states corresponding to perturbation responses generated by a first set of machine learning models. For example, the one or more cell state models may include nonlinear regression models, deep learning models, and / or other types of machine learning models that translate gene expression, images, proteins, and / or other data indicative of a biological system's response to a perturbation into biological age, health or disease status, cell type, phenotype, and / or other corresponding cell states.

[0123] After training, the two sets of machine learning models are integrated into a feedback loop of computer simulations and physical experiments to accelerate the discovery of interventions that promise to improve the health and / or rejuvenate biological systems. More specifically, the one or more perturbation response models can be used to explore perturbation responses across cell types, disease conditions, baseline states, covariates, modalities, perturbation types, combinations of perturbations, and / or other variables. The perturbation responses generated by the one or more perturbation response models can then be fed into the one or more cell state models, and the predictions of cell states generated by the one or more cell state models based on the input perturbation responses can be used to identify subsets of perturbations that promise to rejuvenate various cell types and / or otherwise improve the health of various cells. The identified perturbations are then validated through various experimental screenings, the results of which can be used to retrain one or both sets of machine learning models.

[0124] Compared to existing technologies, one advantage of the disclosed technology is its ability to predict the responses of different biological systems to various combinations of perturbations, and to predict the corresponding cell states. Therefore, the disclosed technology can be used to identify perturbations that may improve biological age, health status, phenotype, and / or other target cell states without requiring time-consuming and resource-intensive experiments to determine the effects of these perturbations on biological systems. Another advantage of the disclosed technology is its ability to focus limited computational and / or experimental resources on promising perturbations that are expected to generate target cell states within biological systems. Therefore, the disclosed technology can prioritize and validate interventions that improve the health and rejuvenation of biological systems in a targeted manner, thereby enhancing the exploration and understanding of biological rejuvenation within a vast and complex perturbation search space. These advantages represent one or more technological improvements over existing methods.

[0125] Example Terms Clause 1: In various embodiments, a computer-implemented method includes: predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, the one or more perturbation response models being configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbations as inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including a base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; predicting a first plurality of cell states associated with the first plurality of responses using one or more cell state models configured to accept the first plurality of responses as inputs, wherein the one or more cell state models have been trained using a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states; and selecting one or more additional perturbations included in the first plurality of perturbations for experimental evaluation based on the first plurality of cell states.

[0126] Clause 2: The computer-implemented method as described in Clause 1 further includes: receiving one or more results associated with the experimental evaluation; generating a training dataset including the one or more results and the one or more perturbations; and retraining the one or more perturbation response models using the training dataset.

[0127] Clause 3: The computer-implemented method as described in Clause 1 or 2 further includes training the one or more perturbation response models using at least one of reconstruction loss or adversarial loss.

[0128] Clause 4: The computer-implemented method as described in any of Clauses 1 to 3 further includes: training the one or more cell state models using mean squared error.

[0129] Clause 5: A computer-implemented method as described in any one of Clauses 1 to 4, wherein predicting the first plurality of responses comprises: generating one or more embeddings corresponding to the one or more basic states and a first plurality of embeddings representing the first plurality of perturbations using the one or more perturbation response models; combining the one or more embeddings corresponding to the one or more basic states with the first plurality of embeddings representing the first plurality of perturbations to produce a plurality of uniform embeddings; and converting the plurality of uniform embeddings into the first plurality of responses.

[0130] Clause 6: A computer-implemented method as described in any of Clauses 1 to 5, wherein predicting the first plurality of cell states comprises: inputting one or more responses included in the first plurality of responses into a regression model included in the one or more cell state model; and executing the regression model to convert the one or more responses into one or more cell states of a corresponding biological system.

[0131] Clause 7: A computer-implemented method as described in any one of Clauses 1 to 6, wherein determining the one or more perturbations comprises: generating a plurality of scores for the first plurality of perturbations based on the first plurality of cell states; and determining the one or more perturbations based on a ranking of the first plurality of perturbations by the plurality of scores.

[0132] Clause 8: A computer-implemented method as described in any of Clauses 1 to 7, wherein the one or more perturbation response models are further configured to accept one or more covariates as input, the one or more covariates including information indicating at least one of cell line, disease state, or phenotype.

[0133] Clause 9: A computer-implemented method as described in any of Clauses 1 to 8, wherein the one or more basic states include information indicating the state of the biological system in the absence of the first plurality of perturbations, and the first plurality of responses include information indicating the states of the biological system when exposed to the first plurality of perturbations.

[0134] Clause 10: A computer-implemented method as described in any of Clauses 1 through 9, wherein the first plurality of perturbations includes at least one of a chemical perturbation or a genetic perturbation.

[0135] Clause 11: One or more non-transitory computer-readable media storing instructions, which, when executed by one or more processors, cause the one or more processors to perform the following steps: predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, the one or more perturbation response models being configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbations as inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including a base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; predicting a first plurality of cell states associated with the first plurality of responses using one or more cell state models configured to accept the first plurality of responses as inputs, wherein the one or more cell state models have been trained using a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states; and selecting one or more additional perturbations included in the first plurality of perturbations for experimental evaluation based on the first plurality of cell states.

[0136] Clause 12: One or more non-transitory computer-readable media as described in Clause 11, wherein the instructions further cause the one or more processors to perform the following steps: determining one or more results associated with the experimental evaluation; generating a training dataset including the one or more results and the one or more perturbations; and retraining the one or more perturbation response models or the one or more cell state models using the training dataset.

[0137] Clause 13: One or more non-transitory computer-readable media as described in Clause 11 or 12, wherein the instructions further cause the one or more processors to perform the step of training the one or more perturbation response models using reconstruction loss and adversarial loss.

[0138] Clause 14: One or more non-transitory computer-readable media as described in any of Clauses 11 to 13, wherein the instructions further cause the one or more processors to perform the step of training the one or more cell state models using multiple biological ages and multiple health conditions corresponding to the second plurality of cell states.

[0139] Clause 15: One or more non-transitory computer-readable media as described in any of Clauses 11 to 14, wherein training the one or more cell state models comprises: training a base model included in the one or more cell state models to accept the observations as input and produce the plurality of biological ages and the plurality of health conditions as outputs; initializing a plurality of tissue-specific models using the base model; and training the plurality of tissue-specific models using a plurality of tissue-specific datasets.

[0140] Clause 16: One or more non-transitory computer-readable media as described in any of Clauses 11 to 15, wherein the plurality of tissue-specific datasets are associated with at least one of kidney cells, endothelial cells, fibroblasts, nerve cells, or muscle cells.

[0141] Clause 17: One or more non-transitory computer-readable media as described in any of Clauses 11 to 16, wherein determining the one or more perturbations comprises: matching one or more cell states associated with the one or more perturbations with one or more target cell states of the one or more biological systems.

[0142] Clause 18: One or more non-transitory computer-readable media as described in any of Clauses 11 to 17, wherein the one or more cell state models include a multilayer perceptron.

[0143] Clause 19: One or more non-transitory computer-readable media as described in any of Clauses 11 to 18, wherein the one or more disturbance response models include a combined disturbance autoencoder.

[0144] Clause 20: A system includes: one or more memories storing instructions, and one or more processors coupled to the one or more memories, wherein, upon execution of the instructions, the one or more processors are configured to perform the following steps: predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, the one or more perturbation response models being configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbations as inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including a base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; predicting a first plurality of cell states associated with the first plurality of responses using one or more cell state models configured to accept the first plurality of responses as inputs, wherein the one or more cell state models have been trained using a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states; and selecting one or more additional perturbations included in the first plurality of perturbations for experimental evaluation based on the first plurality of cell states.

[0145] Any and all combinations of any claim element set forth in any claim and / or any element described in this application fall in any way within the intended scope of this invention and protection.

[0146] Various embodiments have been described for illustrative purposes and are not intended to be exhaustive or limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

[0147] Various aspects of this embodiment can be implemented as a system, method, or computer program product. Therefore, various aspects of this disclosure can take the form of a completely hardware embodiment, a completely software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, which are generally referred to herein as a “module,” “system,” or “computer.” Additionally, any hardware and / or software technologies, processes, functions, components, engines, modules, or systems described in this disclosure can be implemented as circuits or groups of circuits. Furthermore, various aspects of this disclosure can take the form of a computer program product embodied in one or more computer-readable media having computer-readable program code embodied thereon.

[0148] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. More specific examples of computer-readable storage media (not an exhaustive list) may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable optical disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the context of this document, a computer-readable storage medium can be any tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device.

[0149] Various aspects of this disclosure have been described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products, based on embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine. When executed via the processor of a computer or other programmable data processing apparatus, the instructions enable the implementation of the functions / actions specified in the blocks of the flowcharts and / or block diagrams. Such processors can be, but are not limited to, general-purpose processors, special-purpose processors, application-specific processors, or field-programmable gate arrays.

[0150] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing one or more specified logical functions. It should also be noted that in some alternative implementations, the functions recorded in the blocks may occur in an order other than that shown in the figures. For example, in practice, two blocks shown sequentially may execute substantially simultaneously, or, depending on the functions involved, the blocks may sometimes execute in reverse order. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system or a combination of dedicated hardware and computer instructions that performs a specific function or action.

[0151] While the foregoing describes embodiments of this disclosure, other and further embodiments of this disclosure may be designed without departing from its essential scope, which is defined by the appended claims.

Claims

1. A computer-implemented method, the computer-implemented method comprising: Predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, wherein the one or more perturbation response models are configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbation behavior inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including the base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; The first plurality of cell states associated with the first plurality of responses are predicted using one or more cell state models configured to accept the first plurality of responses as input, wherein the one or more cell state models have been trained with a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states. as well as Based on the first plurality of cell states, one or more additional perturbations included in the first plurality of perturbations are selected for experimental evaluation.

2. The computer-implemented method as described in claim 1, further comprising: Receive one or more results associated with the experimental evaluation; Generate a training dataset that includes one or more results and one or more perturbations; as well as The one or more perturbation response models are retrained using the training dataset.

3. The computer-implemented method as described in claim 1, further comprising: The one or more perturbation response models are trained using at least one of reconstruction loss or adversarial loss.

4. The computer-implemented method as described in claim 1, further comprising: The mean squared error is used to train the one or more cell state models.

5. The computer-implemented method of claim 1, wherein predicting the first plurality of responses comprises: The one or more perturbation response models are used to generate one or more embeddings corresponding to the one or more basic states and a first plurality of embeddings representing the first plurality of perturbations; The one or more embeddings corresponding to the one or more basic states are combined with the first plurality of embeddings representing the first plurality of perturbations to produce a plurality of unified embeddings; as well as The multiple unified embeddings are converted into the first multiple responses.

6. The computer-implemented method of claim 1, wherein predicting the first plurality of cell states comprises: One or more responses included in the first plurality of responses are input into the regression model included in the one or more cell state models; as well as The regression model is executed to convert the one or more responses into one or more cellular states of the corresponding biological system.

7. The computer-implemented method of claim 1, wherein determining the one or more disturbances comprises: Based on the first plurality of cell states, generate multiple scores for the first plurality of perturbations; as well as The one or more perturbations are determined based on the sorting of the first plurality of perturbations by the plurality of scores.

8. The computer-implemented method of claim 1, wherein the one or more perturbation response models are further configured to accept one or more covariates as input, the one or more covariates including information indicating at least one of cell line, disease state, or phenotype.

9. The computer-implemented method of claim 1, wherein the one or more basic states include information indicating the state of the biological system in the absence of the first plurality of perturbations, and the first plurality of responses include information indicating the states of the biological system when exposed to the first plurality of perturbations.

10. The computer-implemented method of claim 1, wherein the first plurality of perturbations includes at least one of a chemical perturbation or a genetic perturbation.

11. One or more non-transitory computer-readable media storing instructions, which, when executed by one or more processors, cause the one or more processors to perform the following steps: Predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, wherein the one or more perturbation response models are configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbation behavior inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including the base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; First plurality of cell states associated with the first plurality of responses are predicted using one or more cell state models configured to accept the first plurality of responses as input, wherein the one or more cell state models have been trained using a second plurality of training samples, the second plurality of training samples comprising observations of multiple biological systems paired with the second plurality of cell states; and Based on the first plurality of cell states, one or more additional perturbations included in the first plurality of perturbations are selected for experimental evaluation.

12. The one or more non-transitory computer-readable media of claim 11, wherein the instructions further cause the one or more processors to perform the following steps: Determine one or more results associated with the experimental evaluation; Generate a training dataset that includes one or more results and one or more perturbations; and The training dataset is used to retrain the one or more perturbation response models or the one or more cell state models.

13. The one or more non-transitory computer-readable media of claim 11, wherein the instructions further cause the one or more processors to perform the step of training the one or more perturbation response models using reconstruction loss and adversarial loss.

14. The one or more non-transitory computer-readable media of claim 11, wherein the instructions further cause the one or more processors to perform the step of: training the one or more cell state models using a plurality of biological ages and a plurality of health conditions corresponding to the second plurality of cell states.

15. One or more non-transitory computer-readable media as claimed in claim 14, wherein training the one or more cell state models comprises: Train the base model included in one or more cell state models to take the observations as input and produce the multiple biological ages and the multiple health conditions as output; The base model is used to initialize multiple tissue-specific models; as well as The multiple tissue-specific models were trained using multiple tissue-specific datasets.

16. One or more non-transitory computer-readable media as claimed in claim 15, wherein the plurality of tissue-specific datasets are associated with at least one of kidney cells, endothelial cells, fibroblasts, nerve cells, or muscle cells.

17. One or more non-transitory computer-readable media as claimed in claim 11, wherein determining the one or more disturbances comprises: Match one or more cell states associated with the one or more perturbations with one or more target cell states of the one or more biological systems.

18. One or more non-transitory computer-readable media as claimed in claim 11, wherein the one or more cell state models include a multilayer perceptron.

19. One or more non-transitory computer-readable media as claimed in claim 11, wherein the one or more disturbance response models comprise a combined disturbance autoencoder.

20. A system comprising: One or more memories for storing instructions, and One or more processors coupled to the one or more memories are configured to perform the following steps when executing the instructions: Predicting a first plurality of responses of one or more biological systems to a first plurality of perturbations using one or more perturbation response models, wherein the one or more perturbation response models are configured to accept one or more base states associated with the one or more biological systems and the first plurality of perturbation behavior inputs, wherein the one or more perturbation response models have been trained to reconstruct a first plurality of training samples, each of the first plurality of training samples including the base state of the biological system, one or more covariates associated with the biological system, and one or more perturbations applied to the biological system; The first plurality of cell states associated with the first plurality of responses are predicted using one or more cell state models configured to accept the first plurality of responses as input, wherein the one or more cell state models have been trained with a second plurality of training samples, the second plurality of training samples including observations of a plurality of biological systems paired with the second plurality of cell states. as well as Based on the first plurality of cell states, one or more additional perturbations included in the first plurality of perturbations are selected for experimental evaluation.

Images