Research map / Architectures and efficiency / RNNs and normalization
xTrimoGene: An Efficient and Scalable Representation Learner for Single-Cell RNA-Seq Data
Jing Gong, Minsheng Hao, Xingyi Cheng, Xin Zeng, Chiming Liu, Jianzhu Ma, Xuegong Zhang, Taifeng Wang, Le Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use asymmetric encoder-decoder processing only non-zero unmasked genes in the encoder instead of full-length Transformer | xTrimoGene · pre-training compute (5M samples, 5 epochs) | total train FLOPs (resource vs Transformer): 100% -> 3.4% | Improved |
| use auto-discretization of expression values instead of round bin, up bin or equal bin (ablation) | xTrimoGene · cell clustering (ARI, NMI, HOMO, CP, SIL) | clustering metrics | Improved |
| use regression (MSE) masked objective instead of classification (ablation) | xTrimoGene pre-training · cell clustering | clustering metrics | Improved |
| supervise zero-valued positions during masking (ablation) | xTrimoGene pre-training · cell clustering | clustering metrics | Improved |
| raise masking ratio from 15% to about 30%, then further to 90% (ablation) | xTrimoGene pre-training · cell clustering | clustering metrics | Mixed |
| replace encoder-only Performer with asymmetric encoder-decoder (ablation) | xTrimoGene · masked value recovery and cell clustering | Pearson correlation / clustering | Improved |
| scale xTrimoGene from 3M to 10M to 100M parameters (ablation) | xTrimoGene · pre-training loss | loss | Improved |
| replace scBERT with xTrimoGene for cell type annotation | scBERT · Zheng68K | F1 score: 0.6695 -> 0.7354 | Improved |
| replace CellTypist with xTrimoGene on Zheng68K precision | CellTypist · Zheng68K | precision: 0.7454 -> 0.7335 | Hurt |
| add xTrimoGene embeddings to GEARS | GEARS · Norman perturb-seq top-20 DE genes | MSE: -14.8% MSE | Improved |
| use xTrimoGene embeddings in DeepDDS | DeepDDS · drug combination synergy prediction | ROC | Improved |
Closest papers by meaning
- sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models (ICML 2025 · similarity 0.87)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.86)
- CellPLM: Pre-training of Cell Language Model Beyond Single Cells (ICLR 2024 · similarity 0.86)
- CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.84)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.84)
- Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
