Training program, training method, and information processing device

Abstract

The goal is to achieve continuous deformation of a plausible all-atom model. [Solution] The training program inputs a first image of the target compound into the encoder of an autoencoder having a latent space isometric with respect to the input space, inputs the latent variables output by the encoder and a typical compound model corresponding to a typical example of the 3D structure of the target compound into the decoder of the autoencoder, and updates the parameters of the encoder and decoder based on the reconstruction error between the first image and a second image reconstructed based on the output of the encoder.

Claims

1. A first image of the target compound is input to the encoder of an autoencoder having a latent space isometric with respect to the input space, and the latent variables output by the encoder and a typical compound model corresponding to a typical example of the three-dimensional structure of the target compound are input to the decoder of the autoencoder. The parameters of the encoder and the decoder are updated based on the reconstruction error between the second image, which is reconstructed based on the output of the encoder, and the first image. A training program characterized by having a computer perform a process.

2. The computer is further instructed to perform a process to generate similar compound models that have a three-dimensional structure similar to the aforementioned typical compound model. The updating process includes updating the parameters of the encoder and the decoder based on the distance between the three-dimensional model of the target compound output by the encoder of the autoencoder and the similar compound model. The training program according to feature 1.

3. The training program according to claim 2, characterized in that the second image is reconstructed based on the projection angle calculated based on the three-dimensional model and the first image.

4. The aforementioned generation process includes a process for generating the aforementioned analogous compound model based on MD (Molecular Dynamics) or AF (AlphaFold). The training program according to claim 2 or 3.

5. The input process includes further inputting the sequence corresponding to the typical compound model to the encoder of the autoencoder. The training program according to feature 1.

6. The training program according to claim 1, characterized in that the autoencoder is implemented by CryoTWIN.

7. The training program according to claim 1, characterized in that the latent space is formulated by a Gaussian mixture distribution.

8. The training program according to claim 1, characterized in that the first image is a first electron microscope image and the second image is a second electron microscope image.

9. A first image of the target compound is input to the encoder of an autoencoder having a latent space isometric with respect to the input space, and the latent variables output by the encoder and a typical compound model corresponding to a typical example of the three-dimensional structure of the target compound are input to the decoder of the autoencoder. The parameters of the encoder and the decoder are updated based on the reconstruction error between the second image, which is reconstructed based on the output of the encoder, and the first image. A training method characterized by having a computer perform the processing.

10. A first image of the target compound is input to the encoder of an autoencoder having a latent space isometric with respect to the input space, and the latent variables output by the encoder and a typical compound model corresponding to a typical example of the three-dimensional structure of the target compound are input to the decoder of the autoencoder. The parameters of the encoder and the decoder are updated based on the reconstruction error between the second image, which is reconstructed based on the output of the encoder, and the first image. An information processing apparatus characterized by having a control unit that performs processing.

Images