SCALE:Scalable Conditional Atlas-Level Endpoint transport for virtual cell perturbation prediction
arXiv:2603.17380v1 Announce Type: new 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
arXiv:2603.17380v1 Announce Type: new 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.
Executive Summary
This article presents SCALE, a specialized large-scale foundation model for virtual cell perturbation prediction. SCALE addresses three coupled bottlenecks in virtual cell modeling: inefficient training and inference pipelines, unstable modeling in high-dimensional sparse expression space, and evaluation protocols that overemphasize reconstruction-like accuracy. The model utilizes a BioNeMo-based training and inference framework, conditional transport, and a set-aware flow architecture. SCALE demonstrates significant improvements in data throughput, distributed scalability, and deployment efficiency. The model also exhibits stronger recovery of perturbation effects and improves biologically meaningful metrics such as PDCorr and DE Overlap. These findings suggest that advancing virtual cells requires co-designing scalable infrastructure, stable transport modeling, and biologically faithful evaluation.
Key Points
- ▸ SCALE is a large-scale foundation model for virtual cell perturbation prediction
- ▸ The model addresses three bottlenecks in virtual cell modeling: training, modeling, and evaluation
- ▸ SCALE utilizes a BioNeMo-based training and inference framework and conditional transport
Merits
Strength in Addressing Bottlenecks
SCALE simultaneously addresses three major limitations in virtual cell modeling, providing a comprehensive solution.
Improved Scalability
The BioNeMo-based framework and conditional transport architecture significantly improve data throughput, distributed scalability, and deployment efficiency.
Biologically Faithful Evaluation
SCALE demonstrates stronger recovery of perturbation effects and improves biologically meaningful metrics such as PDCorr and DE Overlap.
Demerits
Limitation in Generalizability
SCALE's performance may be specific to the Tahoe-100M dataset and may not generalize to other virtual cell modeling tasks or datasets.
Complexity of Implementation
The BioNeMo-based framework and conditional transport architecture may be complex to implement and require significant computational resources.
Expert Commentary
This article represents a significant advancement in virtual cell modeling, addressing major limitations and providing a scalable, stable, and biologically faithful solution. The BioNeMo-based framework and conditional transport architecture demonstrate a clear understanding of the challenges and complexities of high-dimensional machine learning. However, the model's performance may be specific to the Tahoe-100M dataset, and its generalizability to other tasks and datasets remains to be seen. Nevertheless, the implications of SCALE are far-reaching and have the potential to revolutionize the field of virtual cell modeling.
Recommendations
- ✓ Future research should focus on developing more generalizable and robust virtual cell models that can be applied to diverse tasks and datasets.
- ✓ The development of more efficient and scalable machine learning architectures for high-dimensional sparse expression space is crucial for the advancement of virtual cell modeling.
