Fig. 1: Overview of the mvTCR.

The mvTCR model receives a gene expression vector, along with the CDR3-α and -β amino acid sequences. This input is transformed through separate encoders, and a mixture module combines different modalities to infer a joint representation (latent space) for downstream analysis. This joint representation is then fed into separate decoders to reconstruct the original gene expression and TCR sequences for each cell.