henhaobio
Sign in

bioRxivCell BiologyPreprint

SCALE: Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction

Shuizhou Chen, Lang Yu, Kedu Jin, Xinjie Mao, Songming ZhangiD, Xinyu Gu, Sheng XuiD, Lei BaiiD, Quan Qian, Qin CheniD, Qiang GaoiD, Siqi Sun, Zhangyang GaoiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Virtual cell models aim to enable in silico experimentation by predicting how cells respond to genetic, chemical, or cytokine perturbations from single-cell measurements. In practice, however, large-scale perturbation prediction remains constrained by three coupled bottlenecks: inefficient training and inference pipelines, unstable modeling in high-dimensional sparse expression space, and evaluation protocols that overemphasize reconstruction-like accuracy while underestimating biological fidelity. In this work we present a specialized large-scale foundation model SCALE for virtual cell perturbation prediction that addresses the above limitations jointly. First, we build a BioNeMo-based training and inference framework that substantially improves data throughput, distributed scalability, and deployment efficiency, yielding 12.51* speedup on pretrain and 1.29* on inference over the prior SOTA pipeline under matched system settings. Second, we formulate perturbation prediction as conditional transport and implement it with a set-aware flow architecture that couples LLaMA-based cellular encoding with endpoint-oriented supervision. This design yields more stable training and stronger recovery of perturbation effects. Third, we evaluate the model on Tahoe-100M using a rigorous cell-level protocol centered on biologically meaningful metrics rather than reconstruction alone. On this benchmark, our model improves PDCorr by 12.02% and DE Overlap by 10.66% over STATE. Together, these results suggest that advancing virtual cells requires not only better generative objectives, but also the co-design of scalable infrastructure, stable transport modeling, and biologically faithful evaluation.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (3 Sept 2026) · CC BY 4.0

A set-aware latent transport model predicts treated cell populations from unpaired controls, keeps perturbation-specific signal in CRISPR, drug and cytokine data, and guided a cytokine choice tested in three donors.

Single-cell assays measure separate control and treated cells, so cell-to-cell models depend on artificial pairings, while population-level models can lose perturbation-specific signal and drift toward an average response. SCALE trains on unpaired populations with endpoint supervision, and the paper tests whether it keeps target-specific responses.

Key results

  1. ShownAcross three benchmarks SCALE was strongest on Replogle and led several metrics on PBMC and Tahoe-100M, but no method won every dataset and metric.

    Replogle: SCALE led all seven metrics (PDCorr 0.7510 vs 0.4132 for STATE; DEOver 0.4040 vs 0.2294). PBMC: best PDCorr 0.6430, DEOver 0.4430, DEPrec 0.5280, but STATE had the lowest MSE and MAE. Tahoe-100M: best MSE 0.0003, PDCorr 0.8584, DirAgr 0.8092; STATE led DEOver and DEPrec. · 2.1 SCALE recovers perturbation-specific…

  2. ShownIn CRISPR data dominated by cell-line differences, SCALE kept gene-specific signal after cell-line and average-treatment effects were removed, and kept perturbation representations spread out rather than collapsed.

    Median residual PCC 0.33 vs 0.05 and Direction@100 0.88 vs 0.57 against STATE. Shared-response PCC 0.871 vs 0.082 and context-modulation PCC 0.332 vs 0.049 over 738 targets. SCALE was the only model above both mean baselines. · 2.2 SCALE recovers target-specific CRISPR…; C.1 Replogle cross-context response diagnostics; Shared response and context modulation

  3. ShownOn sci-Plex drug data, SCALE did better than STATE and a direct-addition baseline when the cell line, drug pair or dose was held out, including a 10 µM dose 100-fold above the largest training dose.

    Held-out drug pairs: Δ Pearson 0.70 to 0.93, gain over direct addition 0.23 to 0.46. Across 561 conditions at 10 µM, SCALE beat STATE on top-50 mean absolute error in 94% of comparisons, on Spearman in 72% and on direction accuracy in 68%. STATE had lower RMSE for four of five pairs. · 2.3 SCALE preserves drug-response specificity…

  4. ShownSCALE predicted cytokine responses for held-out cytokine conditions in partly measured donors, for held-out cell types and for held-out patients, with weaker performance in plasmablasts.

    PBMC atlas mean Δ Pearson 0.525, better than STATE in 85% of paired conditions. Cell-type holdouts ranged from 0.657 (CD4 memory T) to 0.306 (plasmablasts). Kang IFN-β patients: mean ΔPCC 0.932 vs 0.924 for STATE, SCALE better for six of eight patients. · 2.5 SCALE generalizes cytokine responses across…; Limitations and outlook

  5. SuggestedFour cytokines chosen from SCALE-CytoTI predicted scores (one per efficacy and inflammation quadrant) showed the predicted ordering of CD8 T-cell and NK-cell activation and inflammatory secretion in PBMCs from three donors.

    All three donors showed higher CD69 in CD8 T cells after IL-12 and IL-15; IL-32β stayed near PBS. IL-32β gave the largest log10 rises (IL-6 +1.14, TNF-α +1.01, CXCL8 +0.53, CCL2 +1.12), then IL-15 (+0.54, +0.59, +0.26, +0.64), then IL-12 (+0.32, +0.33, +0.10, +0.37). n = 3 donors, no significance testing. · 2.6 SCALE predictions of held-out cytokine…; SCALE-CytoTI separates predicted antitumour…

Methods, data and reproducibility

SCALE encodes unordered sets of control cells with a gene-attention encoder plus a DeepSets population summary, then uses a conditional DiT/JiT transport network to predict the treated endpoint in latent space, trained without cell pairing, and decodes to expression. Condition inputs (cell type, perturbation, dose, time) are combined by seed attention.

  • Replogle-Nadig CRISPRi screen (K562, HepG2, Jurkat, RPE1; 2,024 filtered perturbations) · human (cell lines) · n = 643,413 cells
  • Human Cytokine Dictionary PBMC atlas (90 cytokines, 16 cell types) · human · n = 12 donors
  • Tahoe-100M drug-response atlas · n = over 100 million profiles; 50 cancer cell lines; over 1,100 conditions
  • sci-Plex drug data (cell-line, drug-pair and dose holdouts) · n = 2,244 matched conditions in cell-line transfer; 561 in dose test
  • ZSCAPE developmental perturbation atlas · n = 28 genetic perturbations; 5 stages
  • Kang IFN-β cohort · human (patient cells) · n = 8 patients
  • Donor-matched PBMC cytokine experiment (readouts: CD69 flow cytometry and four secreted inflammatory molecules) · human · n = 3 donors, 5 conditions, 45 culture wells
Methods & tools
BioNeMo, Cell-Eval 0.6.6, STATE, CellFlow, Scanpy, DeepSets, DiT, JiT, CPA, scLAMBDA, BioLORD, LMDB
Reagents & environment
Cell-Eval version 0.6.6; training in BioNeMo; main model has 184M parameters (280M variant also tested).
Experimental conditions
BioNeMo system: epoch time 1298.43 s to 243.00 s; 27,818.20 cells/s vs 21,584.40 for STATE. Hardware not stated.
Reproducibility gaps
No code or data URLs appear in the text; no accessions are given for the datasets.
Limitations (6)
  • Results differ across datasets and metrics, and no single model wins everywhere; STATE remains stronger on some Tahoe differential-expression metrics. · stated by the authors
  • The PBMC split includes some training cytokines from every test donor, so it tests completion within partly measured donors, not entirely new donors; the Kang task does not show personalized clinical prediction. · stated by the authors
  • Endpoint-only training cannot reveal events between control and treatment, and developmental accuracy depends strongly on treated-cell numbers and stage. · stated by the authors
  • The cytokine experiment has three donors and four cytokines, tests only CD69 and four secreted molecules ex vivo, and does not test numerical accuracy of CytoTI, APC priming or in vivo toxicity. · stated by the authors
  • Table 2 combines results from different prediction pipelines and condition sets (public, reproduced and re-evaluated), so rows are not strictly like-for-like. · stated by the authors
  • The paper does not report whether other models would have selected the same cytokines in the prospective test, so the experiment does not show SCALE is better than alternatives at selection. · noted in this summary

Read this if you work on single-cell perturbation prediction and want a population-level model evaluated for perturbation specificity against mean baselines, plus a small prospective cytokine test.

Citations
2

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDibaeinia '26Zhang '26Weidener '26Adduri '26Chuai '26Heidari '26
71% semantic matchGenomics
Payam Dibaeinia, Sudarshan Babu, Mei K. Knudson+5 more·bioRxiv·Feb 9, 2026· 10 citations