Modern biology can now measure many types of information from the same single cell. For example, scientists can examine gene expression using RNA sequencing, measure chromatin accessibility, quantify protein abundance, and even analyze spatial imaging data. While these technologies are powerful, combining the different data types into one clear picture of cell state remains challenging.

A new computational framework called APOLLO aims to address this problem. The work was led by researchers at institutions such as the Broad Institute and MIT, where computational biology and machine learning methods are actively developed. APOLLO stands for Autoencoder with a Partially Overlapping Latent space learned through Latent Optimization. Although the name sounds technical, the core idea is straightforward. The model learns which information is shared between data types and which information is unique to each type.

Overview of APOLLO for partially shared multi-modal embeddings and cross-modality prediction

a, As different experimental technologies capture different aspects of the cell state, we expect some cell-state information to be shared among different modalities and some information to be modality specific. b, APOLLO is an autoencoder-based approach that learns three latent spaces to disentangle information captured by each modality. APOLLO uses a two-step training procedure. In step 1, the decoders of each modality are trained so that the decoders can reconstruct the input data from the latent spaces. In step 2, modality-specific encoders are trained to enable inferring the latent space embedding for cells not used in training the model. c, Our model enables the prediction of missing modalities. The explicit learning of partial information sharing allows APOLLO to perform accurate cross-modality prediction. The example shows the predicted protein (CD16, CD3, CD4, CD8, lamin) fluorescence images of a cell in a held-out patient, given the cell’s chromatin and γH2AX protein stain. d, The information captured by the shared and modality-specific latent spaces learned by APOLLO is interpretable. The partially shared latent space can be interpreted by morphological features or genetic pathways.

Many existing integration methods combine multiple data types into a single mathematical representation. However, these approaches often blur the boundaries between modalities, making it difficult to understand which signals come from gene expression, which from chromatin structure, and which from protein measurements. APOLLO instead creates a partially shared internal representation. This means it preserves common biological signals while still keeping track of modality specific information.

The researchers tested APOLLO on both simulated datasets and real paired single cell datasets. These included SHARE-seq data that combines RNA sequencing with chromatin accessibility, CITE-seq data that pairs RNA sequencing with protein abundance, and multiplexed imaging datasets. Across these applications, APOLLO successfully integrated the data while maintaining interpretability.

One important feature of APOLLO is its ability to predict missing information. For example, if protein measurements are not available for a particular cell, the model can estimate them based on other data modalities. It can also help identify which specific modality or cellular compartment is associated with a particular phenotype, such as differences in protein localization between cells.

By distinguishing shared signals from modality specific ones, APOLLO provides a more complete and interpretable view of cell state. As single cell technologies continue to expand, frameworks like this may help researchers better understand how multiple layers of biology work together inside individual cells.

Availability – The code is available via Zenodo at https://doi.org/10.5281/zenodo.17841315 and in GitHub at https://github.com/uhlerlab/APOLLO/.