Outputs of MIDAS

In this section, we explain the outputs generated by MIDAS and show how to retrieve them.

Embedding Space Outputs

MIDAS generate reduction outputs including joint latents and modality-specific latents:

Type	Description
Modality-specific Latents	Latents corresponding to individual input modalities, consisting of biological information (denoted as c) and technical noise (denoted as u).
Joint Latents	The combined latent representation that integrates all modalities, along with the batch indices. This representation consists of biological information (denoted as c) and technical noise (denoted as u).

Feature Space Outputs

As a generative model, MIDAS can also generate outputs such as imputed data, batch-corrected data, and translated data.

Type	Description
Imputed Data	Data for all modalities imputed from the joint latent representation.
Batch-corrected Data	Imputed data with batch effects removed, generated by passing a standard noise latent instead of the real noise latent.
Translated Data	Data translation between modalities, where input from one or more modalities is used to generate the corresponding data in other modalities.

Note

By default, MIDAS performs sampling when generating batch-corrected data, based on the distribution settings. To apply sampling to other types of outputs, or to disable sampling entirely, modify the gen_real_data() in scmidas.model.predict().

Additionally, for debugging or other purposes, MIDAS can save the original input data:

Type	Description
Input	The original input data before being fed into the model.

Predicting Outputs

After training, you can use the predict method to generate and save predictions. Here’s the example:

model.predict(output_dir,
        joint_latent=True,
        mod_latent=True,
        impute=True,
        batch_correct=True,
        translate=True,
        input=True)

Where the parameters are:

• pred_dir: The directory where prediction results will be saved.

• joint_latent: Whether to calculate and save joint latent representations.

• mod_latent: Whether to calculate and save modality-specific latent representations.

• impute: Whether to perform data imputation, filling in missing or incomplete data.

• batch_correct: Whether to apply batch correction to the data to reduce batch effects.

• translate: Whether to perform modality translation.

• input : Whether to save the original input data.

Loading Predicted Outputs

To retrieve and load the predicted outputs, you can use the scmidas.utils.load_predicted() function. Here’s how to do it:

from scmidas.utils import load_predicted
load_predicted( output_dir,
                model.combs, # a list of modality combinations per batch
                joint_latent=True,
                mod_latend=True,
                impute=True,
                batch_correct=True,
                translate=True,
                input=True)