Trajectory-conditioned reconstruction of single-cell expression suggests regulatory programs

Wenjie Fan, Antonio Orvieto, Manfred Claassen

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: long paper (4–8 pages excluding references)

Keywords: scRNA-seq, single-cell foundation model, cell state dynamics, pseudotime, representation learning, masked reconstruction

TL;DR: scTNT conditions single-cell latent reconstruction on inferred cell history via a frozen reduced-layer scGPT autoencoder plus a trainable context adapter over ordered cell sequences.

Abstract: Foundation models for single-cell transcriptomics learn cell representations from millions of profiles, but are commonly pretrained on unordered cells and therefore do not explicitly condition on cell history. We introduce single-cell Transformer-iN-Transformer (scTNT), which conditions gene-expression reconstruction on inferred trajectories, represented here as ordered cell sequences. scTNT combines a frozen reduced-layer scGPT autoencoder with a trainable decoder-only transformer over sequences of latent cell embeddings and is trained by masked gene-expression reconstruction. On a CD8 T-cell exhaustion dataset with optimal transport-derived cell sequences, scTNT improves masked reconstruction relative to the scGPT baseline and outperforms alternative sequence backbones under controlled evaluations. We further propose a gradient-based gene-history attribution pipeline and apply TRRUST regulon enrichment to generate hypotheses about context-associated regulatory programs.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 69