Multimodal fusion

mllabiome supports combining multiple data sources (modalities) into a single predictive experiment. The multimodal system follows the MISO (Multiple Input, Single Output) paradigm: several feature matrices are consumed as input, and a single target variable is predicted as output.

A multimodal experiment is configured by three orthogonal concerns:

• Fusion strategy controls how the modalities are combined (before or after model training).
• Modality configuration describes each secondary data source: its file, sample identifier, encoding options, and per-modality model overrides.
• NaN handling determines how missing values within a secondary modality are treated.

Each concern is independent. The subpages in this section describe the available options.

For a worked example with microbiome and metabolomics data, see the multimodal tutorial.