Loading…
Virtual-cell model that compresses a transcriptome into eight discrete tokens in a reasoning LLM's vocabulary, predicting module-level drug response.
No providers recorded yet. Browse all providers
When a drug nudges an aging immune cell back toward a youthful state, the change rarely arrives as a handful of strongly differentially expressed genes. In the 110-compound screen of primary human peripheral blood mononuclear cells (PBMCs) behind this work, most drug-induced differentially expressed genes fell outside the aging gene sets of interest entirely. A model graded on reconstructing that list can therefore score well and say nothing about whether the compound moved the phenotype. CellQ is built for the other question: given a baseline cell population and a candidate compound, does a named gene module shift in the beneficial direction?
CellQ (Cell Quantization) answers it by making cells directly speakable. A residual vector quantization (RVQ) autoencoder compresses each single-cell transcriptome into eight discrete codes, and those eight codebooks are inserted at fixed positions into the vocabulary of a Qwen3-4B-Thinking-2507 reasoning language model. Cell and text tokens then share one embedding table and one autoregressive objective, so no cross-modal projection network stands between measurement and reasoning. It was developed at Guangzhou National Laboratory with Westlake University, Shanghai Jiao Tong University and Guangzhou Medical University, and posted to bioRxiv in August 2026 as the virtual-cell component of a three-part self-auditing discovery agent.
It is a sibling of RVQ-Alpha from the same lab, not a rename of it: CellQ reuses that model's eight-codebook tokenizer and its alignment protocol of continued pretraining plus supervised fine-tuning, cited as prior work, but runs on Qwen3-4B-Thinking-2507 rather than Qwen3-4B, has no reinforcement-learning stage, and adds group-code aggregation and module-level prediction.
The RVQ autoencoder maps a cell's expression vector through a multilayer-perceptron encoder to a latent state and approximates it as a sum of eight codes from independent codebooks of size 32, with a symmetric decoder reconstructing expression. It trains as a denoising autoencoder under a composite objective whose codebook-utilisation term penalises collapsed entries. Pretraining used more than 10^8 human cells from public atlases across 105 dataset shards spanning diverse tissues, age groups, and healthy or disease contexts, with train and held-out splits stratified at the dataset level so that no dataset appeared in both. Alignment followed an Evidence-First Feature-Injection strategy: each prompt pairs the RVQ token tuple with the cell's most strongly expressed genes, and the completion lists gene-level evidence before its conclusion.
On cell-type classification over PBMC and intestine reference datasets, CellQ with group-code aggregation outperformed LangCell, Cell2Sentence and Cell-01. For perturbation prediction on the Tahoe-100M atlas, two test partitions hold out cell lines and compounds unseen during training; CellQ(group) achieved the highest macro-F1 under both, ahead of Xpert, PerturbNet, STATE, ChemCPA and Cell2Sentence. An ablation replacing the RVQ codes with bare language-model input performed substantially worse, isolating discrete tokenisation rather than the backbone as the source of the gain.
CellQ suits in-silico compound triage where the readout is distributed across many weakly moving genes instead of concentrated in one signature. In the preprint it sits inside an Analyzer–Planner–Auditor agent that screened 110 compounds in primary human PBMCs: retrieval shortlists candidates, CellQ returns a forced BETTER or WORSE call with a self-assessed confidence for each compound and module, and PACE — a curated panel of ten directionally scored, cell-type-resolved immune-aging gene sets scored with the UCell enrichment method — evaluates them. Compounds enter the prompt as names and mechanism descriptions rather than molecular structures, and because the phenotype is supplied as gene sets, the checkpoint carries over to any phenotype expressible that way.
CellQ argues that the productive interface between transcriptomes and language models is a short discrete code rather than a long gene list or a continuous embedding, and that evaluation should target coordinated module shifts rather than differential-expression recovery. Eight discrete codes necessarily discard fine-grained expression detail, and the preprint notes that characterising that information-loss boundary is open work. Validation covers a single phenotype, immune aging in PBMCs, over one revision round, and the preprint has not yet been peer reviewed. Code and trained weights are deposited on Zenodo, but the record is access-restricted during peer review and its identifier is absent from the preprint, so nothing is retrievable today. No license is stated for them, and no model card or data card accompanies the release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.