What is encoded?
How much predictive information is contributed by a molecular or gene representation, under controlled downstream conditions?
MVCBench / Multimodal virtual cells / 2026
What can we really predict about drug-induced cellular phenotypes?
A systematic benchmark of molecular and gene representations across transcriptomic, morphological, and multimodal response prediction.
MVCBench disentangles representation choice, readout modality, and evaluation setting—so stronger benchmark scores can become more useful design evidence for virtual-cell models.
Nature Communications · Accepted in Principle (2026)Preprint remains available while the journal version is pending.
BENCHMARK OVERVIEWFIGURE 01
~1.1M
Drug-response profiles across the benchmark resources
24
Representation methods 12 molecular · 12 gene
3
Progressive evaluation stages Molecular · Gene · Multimodal
2
Major phenotype readouts Transcriptomics · Morphology
01 / SCIENTIFIC MOTIVATION
A virtual cell needs more than a strong representation benchmark.
Drug-response prediction is increasingly built from powerful pretrained encoders. But a representation that helps predict a cell image need not help predict gene expression, and success on familiar compounds need not persist across experimental settings.
The question is not simply which encoder wins. It is what information the encoder transfers, to which phenotype, and how much of that gain remains under the evaluation protocol.
What is predicted?
Do representation rankings change between transcriptomic responses and high-content image-derived morphology?
What transfers?
Which conclusions survive changing compounds, cellular contexts, plates, and datasets rather than relying on a single split?
02 / BENCHMARK DESIGN
One benchmark. Three linked questions.
MVCBench proceeds from component-level evaluation to multimodal optimization. Its datasets span transcriptomic and morphological responses, but observations from the two readout families do not generally originate from the same cells or matched experimental replicates.
Cross-modal matching is primarily at the compound level; interpret joint-modality evidence accordingly.
A / Chemical focus
Hold the prediction setting fixed; vary the compound representation.
Compare molecular fingerprints and learned chemical encoders as inputs to drug-response prediction. Evaluate transcriptional and morphological readouts separately instead of assuming one universal model ordering.
ECFP4KPGTUniMol2Chemprop+ more
B / Biological focus
Ask whether pretrained cellular representations align with the response task.
Evaluate gene representations and single-cell foundation-model features under matched downstream conditions. Compare aggregate performance with targeted analyses such as highly expressed genes and dataset transfer.
STATEscGPTscFoundationGeneformer+ more
C / Multimodal modeling
Test joint optimization—not just the presence of two output heads.
Compare single-readout training, cross-modal auxiliary objectives, and joint multimodal training on eligible compound-matched data. Inspect optimization, loss weighting, and fusion design with their evaluation assumptions stated.
Single-taskAuxiliary objectivesJoint trainingFusion
Transcriptomic sourcesLINCS 2020 · CIGS · Tahoe-100M
Morphology sourcesCell Painting Gallery · BBBC datasets
Preprocessed releaseMVCBench datasets ↗
03–06 / EMPIRICAL EVIDENCE
The results are modality-dependent.
Interpret each result as evidence about its tested representations, datasets, readouts, and training protocol—not as a universal ordering of foundation models.
Representation gains are stronger for morphology.
Learned and geometry-aware molecular features show clearer advantages in image-derived phenotype prediction than in gene-expression response prediction.
See evidence ↓Task alignment matters for gene features.
Representations trained closer to perturbation response can outperform general-purpose alternatives in the tested downstream setting.
See evidence ↓Integration must be measured, not assumed.
Joint objectives improve the studied benchmarks, but the magnitude and interpretation depend on modality pairing and optimization choices.
See evidence ↓03 / MOLECULAR REPRESENTATIONS
A chemical encoder can be strong for images and ordinary for RNA.
Across the evaluated molecular representations, the benefit of learned chemical features is not uniform. For transcriptomics, improvements over classical fingerprints are often relatively modest. For morphology, several learned encoders provide more pronounced advantages.
This asymmetry motivates readout-specific benchmarking rather than selecting a single encoder from an aggregate score.
MOLECULAR INPUT → RNA RESPONSEFIGURE 02
MOLECULAR INPUT → IMAGE PHENOTYPEFIGURE 03
INTERPRETATION
Representation quality is task-conditional. Chemical information that helps recover structural phenotypes does not automatically explain high-dimensional transcriptional changes.
04 / GENE REPRESENTATIONS
Task relevance matters as much as pretraining scale.
MVCBench compares gene representation approaches, including single-cell foundation models, in a unified downstream evaluation. The reported results favor task-aligned features in several tested settings, with rankings changing by dataset and target subset.
This is a controlled comparison under shared downstream conditions, not a claim that every representation has been exhaustively optimized to its own best achievable performance.
GENE REPRESENTATION COMPARISONFIGURE 04
Alignment with the target task helps.
The results motivate pretraining and representation objectives that match the biological response readout.
A universal foundation-model ranking.
Relative performance is contingent on datasets, held-out conditions, feature extraction, and the common downstream training budget.
05 / MULTIMODAL VIRTUAL CELLS
Does joint training provide more than two separate predictors?
MVCBench tests whether sharing an optimization objective across transcriptomic and morphological targets improves prediction relative to the studied single-task and auxiliary-training alternatives.
The reported improvements are evidence for the evaluated joint-learning setup. Because modalities are primarily linked by compound rather than matched cells, they are not direct proof of cell-level biological complementarity.
MULTIMODAL OBJECTIVES AND EVALUATIONFIGURE 05
BASELINESingle-readout training
Independent objectives optimize transcriptional or morphological prediction.
COMPARISONCross-modal auxiliary training
One readout can be used as auxiliary supervision for the other.
EVALUATED APPROACHJoint multimodal optimization
Shared optimization is tested for benefits to both predicted response modalities.
06 / DESIGN PRINCIPLES
Build for the phenotype. Calibrate the comparison.
Design guidance should be conditioned on what was compared. MVCBench studies loss balancing, output-specific fusion, and the relationship between the two predicted response modalities.
OPTIMIZATION AND FUSION ANALYSESFIGURE 06
01
Evaluate readouts separately.
Advanced molecular encoders show different relative gains for gene-expression and image-derived morphology targets.
02
Balance losses explicitly.
In the studied settings, calibrated fixed-ratio loss weighting performs competitively or favorably against the tested adaptive schemes; this should not be generalized without further tests.
03
Match fusion to the target.
Reported fusion preferences depend on the predicted readout and benchmark conditions; there is no evidence here for one universally optimal fusion method.
07 / OPEN RESOURCES
Reproduce the comparison. Test your own representations.
The source repository provides benchmark scripts, example representation-extraction notebooks, and links to the processed data. The page reports the experimental findings; the code documents what was actually implemented.
01 / Source codeBenchmark repository ↗02 / DataProcessed datasets ↗03 / ExamplesEmbedding extraction ↗04 / ManuscriptbioRxiv preprint ↗
QUICK START / PYTHON 3.11+
git clone https://github.com/QSong-github/MVCBench.git
cd MVCBench
conda create -n mvcbench python=3.11
conda activate mvcbench
pip install -r requirements.txt
# Download benchmark data and configure the data root:
export VCBENCH_DATA_ROOT=/path/to/MVCBench_dataFull dataset mappings and benchmark commands are documented in the repository README.
PAPER & CITATION
MVCBench: A Multimodal Benchmark for Drug-induced Virtual Cell Phenotypes
University of Macau · University of Florida · Macao Polytechnic University · HKISI-CAS · NVIDIA · Wake Forest School of Medicine
Nature Communications · Accepted in Principle (2026) · DOI for final journal version pending
† Equal contribution · * Corresponding authors. Citation below points to the currently available bioRxiv version and notes the in-principle acceptance; it does not claim formal publication.
BIBTEX / PREPRINT RECORD
@article{li2026mvcbench,
title = {MVCBench: A Multimodal Benchmark for Drug-induced Virtual Cell Phenotypes},
author = {Li, Bo and Wang, Qing and Wang, Shihang and Zhang, Bob and Peng, Yuzhong and Zeng, Pinxian and Liu, Chengliang and Li, Mengran and Tang, Ziyang and Yao, Xiaojun and Deng, Chuxia and Song, Qianqian},
journal = {bioRxiv},
year = {2026},
doi = {10.64898/2026.04.22.720110},
note = {Accepted in principle at Nature Communications}
}