Gene-Chronos

A pseudotime algorithm can order the cells of an embryo along a developmental axis, but what it returns is a rank on an arbitrary scale. Nothing in that ordering says "embryonic day 12.5", and two datasets ordered separately cannot share an axis. Developmental time inference asks the harder, supervised question: given one cell's transcriptome, what absolute biological time was it sampled at?

Gene-Chronos, from Wuhan University, answers that by attaching a small trained adapter to an existing single-cell foundation model rather than building a new one. All 18 transformer layers of a pretrained Geneformer backbone stay frozen, and its full sequence of contextual gene embeddings is kept rather than collapsed to a single classification token. Sixteen learnable temporal prompt tokens act as queries in a cross-attention module, pulling time-informative signal out of that sequence; the result is mean-pooled and passed to an MLP that regresses absolute time. Only 2.6M parameters — 0.8% of the 316M-parameter backbone — are ever updated.

The second ingredient is what the adapter is trained against. A plain regression loss fits the training atlas and drifts when the adapter meets a different embryo or protocol. A triplet-style temporal contrastive term pulls together cells close in true time and pushes apart cells far from it, organizing the latent space by developmental proximity rather than by whatever minimizes training error — which is what makes one fitted adapter usable on data it has never seen.

#Key Features

• Absolute time, not rank order: Predictions come back in the original units — embryonic days, or hours post-fertilization for zebrafish — so they are comparable across datasets rather than confined to one study's ordering.
• Frozen backbone, adapter-only training: Training touches 0.8% of the backbone's parameters and uses 0.4 GB of memory, against 19.5 GB for full fine-tuning of scGPT.
• Prompt-token cross-attention: Swapping the learned prompt queries for simple mean pooling drops Spearman correlation from 0.867 to 0.740, so the query-based aggregation, not the frozen features alone, carries the temporal signal.
• Temporal contrastive regularization: Dropping the contrastive term costs accuracy on its own (Spearman 0.867 to 0.832) and is what the paper credits for the cross-dataset transfer.
• Gene-level interpretability: The prompt tokens attend directly to gene positions, so attention weights rank which genes drive a prediction, recovering stage-dependent genes such as Hba-x, Il31ra and Lars2.

#Technical Details

The adapter uses a 256-dimensional hidden space and K = 16 prompt tokens, trained with an MSE loss on min–max normalized time plus the contrastive term at weight 0.1. Evaluation spans three small supervised benchmarks — Acinar, EBC and HFGC, 400 to 1,400 cells each, covering human pancreas, mouse beta cells and the human germline — and three atlases: MOSTA mouse spatiotemporal data (~770,000 cells, E9.5–E16.5), a human early-embryogenesis atlas harmonized from 12 public studies (~380,000 cells, E0–E29, scRNA-seq plus spatial), and ZESTA zebrafish (~62,000 cells, 3–24 hpf). Across the benchmarks Gene-Chronos raises average Spearman correlation from 0.799 to 0.850 and cuts normalized RMSE from 0.251 to 0.171 against baselines including TemporalVAE, Psupertime, PRESCIENT, CellRank 2, scBERT and fine-tuned Nicheformer. Transfer matters more than the benchmark table: an adapter fitted on mouse Embryo 1 and applied to the unseen Embryo 2 reaches Pearson R = 0.902 with RMSE 0.713 days, and one fitted on ten components of the human atlas and applied to two held-out datasets reaches R = 0.898 and ρ = 0.923. Ablation confirms the pretraining is load-bearing: random initialization drops Spearman to 0.617, and the same adapter over scFoundation or scGPT scores 0.765 and 0.687 against Geneformer's 0.867.

#Applications

The practical use is annotating unlabeled single-cell or spatial data with developmental age: staging cells from an embryo whose collection time is uncertain, checking whether a differentiation protocol or organoid has reached its intended stage, quantifying developmental delay between conditions, and placing cells from separate studies on one timeline. Query data needs no time annotations, so the cost falls on having a reference atlas that does. The attention-derived gene rankings give developmental and stem cell biologists a place to start following up a prediction at the bench.

#Impact

Published in Briefings in Bioinformatics in September 2026, Gene-Chronos is part of a shift from fine-tuning single-cell foundation models wholesale toward attaching task-specific adapters to frozen ones, putting atlas-scale work within reach of one modest GPU. Its limits are worth stating: the adapter must be fitted against a reference atlas carrying explicit time labels before it can be applied to anything, inference still requires a full forward pass through the 316M-parameter backbone, and the zebrafish transfer runs on a backbone pretrained on human transcriptomes alone. The code is MIT-licensed and the harmonized human embryo atlas CC BY 4.0, but no trained adapter checkpoint has been published, so reproducing the results means re-running the training step.