Multimodal VAE for paired single-cell transcriptome and TCR data, with a reference atlas of 3.35M immune cells for mapping new datasets.
No providers recorded yet. Browse all providers
A T cell's transcriptome says what it is doing now; its T cell receptor (TCR) records which antigen selected it and how far its clone has expanded. The two are related but neither substitutes for the other: cells with nearly identical expression profiles can belong to clones with different antigen histories and different fates under immunotherapy. Models that simply concatenate or fuse the two modalities tend to let whichever signal is stronger dominate the joint embedding, which erases exactly the modality-specific information a clonal analysis needs.
ImmuneLens, developed by Zhijian Duan, Shuhui Song and colleagues at the National Genomics Data Center of the China National Center for Bioinformation and the Beijing Institute of Genomics, Chinese Academy of Sciences, addresses this with explicit disentanglement. Each cell is encoded into three variational latents: a gene-expression-private space, a TCR-private space, and a shared space that holds what the two modalities jointly support. A statistical independence penalty (HSIC) pushes the shared space away from both private spaces, so neither modality can hide its own signal inside the joint representation.
The pretrained component is the gene-expression (GEX) encoder, trained as a multi-cohort immune reference atlas on about 3.35 million immune cells from colorectal cancer, non-small cell lung cancer and COVID-19 cohorts (GEO series GSE236581, GSE243013 and GSE158055). With its parameters frozen, that encoder maps an unseen external dataset onto the atlas and transfers cell-type labels, much as scVI-Census serves as a fixed reference for query mapping. The full GEX+TCR model is the framework built around that encoder and is applied to antigen-specificity and immunotherapy tasks. The work was posted to bioRxiv in September 2026 and has not yet been peer reviewed.
The GEX encoder combines a residual MLP over 2,000 highly variable genes with a Perceiver-style attention module (64 latents, 8 heads) over the 512 most highly expressed genes, fused through a GEGLU residual MLP. During atlas pretraining, a spectrally normalized batch discriminator attached through a gradient reversal layer removes cohort effects. The TCR branch uses TCR-BERT for CDR3β and CDR3α, fine-tuning only the last two CDR3β encoder layers and keeping the α encoder frozen; ESM-2 is supported as an alternative encoder. Expression is reconstructed by FiLM-conditioned shared and private decoders, and CDR3 sequences by two-layer autoregressive Transformer decoders. On atlas integration, ImmuneLens scored 0.572 overall against 0.535 for scVI, with ARI of 0.426 against 0.321. Mapped onto an external basal cell carcinoma dataset (GEO GSE123813), median annotation confidence ranged from 0.706 in NK/ILC cells to 0.877 in B-lineage cells.
For antigen specificity, fused embeddings fed per-donor classification heads on a 10x Genomics dextramer dataset from four donors; the fused model beat the best single-modality baseline in three donors, while expression alone did slightly better in the fourth. In 162 lung squamous cell carcinoma patients (151,935 CD8 T cells), the fused model reached a cross-validated AUROC of 0.826 against 0.805 for expression alone and 0.576 for TCR alone.
ImmuneLens is aimed at immunologists with paired single-cell RNA and V(D)J data. The reference atlas lets a new tumour or infection cohort be annotated against a harmonized multi-disease immune space. The fused representation separates functionally distinct clones within a transcriptionally homogeneous population: in a 60-patient lung adenocarcinoma neoadjuvant anti-PD-1 cohort, it split one exhaustion-like cluster into subclusters associated with response or non-response. The clone ranking nominates expanded, cytotoxic CD8 T cell clones as candidate response biomarkers.
ImmuneLens joins a line of paired GEX–TCR models including CoNGA, Tessa, mvTCR and MIST, and is distinguished by its explicit shared/private decomposition and its reference-atlas mode for mapping external data. Its limits are stated by the authors: it requires paired GEX and TCR data, which remain scarce; the antigen task rests on four donors and a small antigen panel; and the lung cancer findings are retrospective and need independent cohorts. Training and analysis code is public, but no trained atlas checkpoint is distributed, so users rebuild the reference from the public GEO cohorts before mapping their own data.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.