🧬 MVCBench

A Multimodal Benchmark for Drug-induced Virtual Cell Phenotypes

Which molecular and gene representations work, for which cellular readouts, and under which evaluation settings?

🧪 Molecular representations  ·  🧬 Gene representations  ·  🔬 Transcriptomics  ·  🖼️ Cell morphology

🌐 Project Page  •  📄 Read the Paper  •  🗂️ Dataset  •  ⚡ Quickstart  •  📝 Citation

---

🎯 Overview

MVCBench systematically studies how drug molecular and gene representations influence prediction of drug-induced cellular phenotypes across transcriptomic, morphological, and multimodal settings. Rather than declaring one encoder universally best, MVCBench asks which representation helps a particular readout and how well that advantage generalizes.

🧫 Drug-induced profiles	🧪 Molecular representations	🧬 Gene representations	🔬 Evaluation stages
~1.1 million	12	12	3

_{Figure 1. MVCBench integrates drug-perturbation resources and evaluates representation choices in a progressive benchmark, from individual response modalities to multimodal modeling. Click the figure to explore the project website.}

Note

What “multimodal” means here: The transcriptomic and morphological resources are not universally measured in the same cells, plates, or experimental replicates. When readouts come from different experiments, cross-modality correspondence may be established at the compound level. Do not interpret the total profile count as same-cell RNA–image pairs.

✨ Key Findings

	Research question	What MVCBench finds
🧪	Do stronger molecular encoders help all phenotypes equally?	Gains are substantially more pronounced for morphology prediction than for gene-expression prediction, where simpler fingerprints can remain competitive.
🧬	Are general-purpose gene foundation models always preferable?	Task alignment matters. In the evaluated transcriptomic settings, perturbation-oriented representations can outperform more general encoders.
🔗	Does multimodal training help?	Joint optimization improves results in the evaluated settings, but the size and interpretation of gains depend on the dataset, readout, and optimization setup.
🧭	What should we learn for future virtual cells?	Representation selection, generalization splits, loss balancing, and fusion strategy need modality-aware evaluation, not a single aggregate ranking.

📊 Show supporting results · Molecular representations (Figures 2–3)

_{Figure 2 · Drug molecular representations for gene-expression prediction.}

_{Figure 3 · Drug molecular representations for cellular morphology prediction.}

🧬 Show supporting results · Gene representations (Figure 4)

_{Figure 4 · Gene representation benchmarking across gene-expression prediction settings.}

🔗 Show supporting results · Multimodal modeling and design principles (Figures 5–6)

_{Figure 5 · Evaluation of multimodal virtual-cell construction.}

_{Figure 6 · Empirical analyses of multimodal design choices.}

🧭 Benchmark Design

The benchmark follows three complementary stages rather than collapsing different prediction questions into one score.

Stage	Focus	Scientific question	Training entry point
A · 🧪	Molecular representation	Which chemical representations support prediction of drug-induced RNA and morphology?	train_gene.py, train_image.py
B · 🧬	Gene representation	How do gene encoders affect transcriptomic response prediction?	train_gene.py
C · 🔗	Multimodal virtual cell	When does joint RNA–morphology learning help, and under which settings?	train_mvc.py

Evaluation dimensions: drug/compound, cellular context, and dataset-dependent generalization; see the paper and project website for the precise protocols and results.

🚀 Quickstart

1️⃣ Set up the environment

git clone https://github.com/QSong-github/MVCBench.git
cd MVCBench

conda create -n mvcbench python=3.11 -y
conda activate mvcbench
pip install -r requirements.txt

2️⃣ Prepare benchmark data and embeddings

Download the released data from 🤗 MVCBench on Hugging Face. The loaders expect prepared datasets and, for some settings, precomputed model embeddings. They do not automatically download every upstream resource.

# The default data root is ./data
export VCBENCH_DATA_ROOT=/path/to/MVCBench_data

Dataset paths and representation file mappings are defined in src/configs.py. Typical resource groups include LINCS2020, CIGS, Tahoe-100M, BBBC036, BBBC047, cpg0016, and molecular representations under Molecular_representations/.

3️⃣ Run a benchmark

🧬 A · Predict gene-expression responses

python train_gene.py \
  --dataset_name LINCS \
  --molecule_feature ECFP4 \
  --gene_encoder_type Default \
  --split_data_type smiles_split \
  --n_epochs 2 \
  --batch_size 1024

🖼️ A · Predict morphological responses

python train_image.py \
  --dataset_name cpg0016 \
  --molecule_feature ECFP4 \
  --image_encoder_type Default \
  --split_data_type smiles_split \
  --n_epochs 2 \
  --batch_size 1024

🔗 C · Train a multimodal virtual cell

python train_mvc.py \
  --dataset_name MVC_BBBC047 \
  --molecule_feature ECFP4 \
  --split_data_type smiles_split \
  --n_epochs 2 \
  --batch_size 1024

Tip

These commands use 2 epochs as a smoke-test configuration, not the training budget for reproducing published benchmark results. Supply the corresponding prepared datasets and embeddings before running. The training scripts use CUDA by default; check --device and available hardware as needed.

4️⃣ Inspect outputs

By default, training writes to results/, including model checkpoints (e.g., best_model.pt), per-sample evaluation tables (.csv), and predicted profiles (.h5) when profile export is enabled.

For benchmark sweeps, see the batch launchers:

• 🧬 scripts/run_gene_batch.sh
• 🖼️ scripts/run_image_batch.sh
• 🔗 scripts/run_mvc_batch.sh

🧰 Representation Zoo

MVCBench evaluates 12 molecular and 12 gene representation approaches in the paper. The reference tables below highlight linked implementations (not a complete model inventory); availability and reproducibility may depend on upstream projects.

🧪 Molecular representations · View model reference table

Model	Reference	Implementation
KPGT	Nature Communications · 2023
InfoAlign	ICLR · 2025
GeminiMol	Advanced Science · 2024
Ouroboros	Preprint
Mole-BERT	ICLR · 2023
ChemBERTa2	arXiv · 2022
MolT5	EMNLP · 2022
Chemprop	JCIM · 2024
MolCLR	Nature Machine Intelligence · 2022
UniMol	ICLR · 2023
UniMol2	NeurIPS · 2024

🧬 Gene representations · View model reference table

Model	Reference	Implementation
Geneformer	Nature · 2023	🤗 Hugging Face
tGPT	bioRxiv · 2022
UCE	bioRxiv · 2023
scBERT	Nature Machine Intelligence · 2022
CellPLM	ICLR · 2024
OpenBioMed	arXiv · 2023
scGPT	Nature Methods · 2024
scFoundation	Nature Methods · 2024
SCimilarity	Nature · 2025
Cell2Sentence	ICML
STATE	bioRxiv · 2025

🗃️ Datasets

Modality	Data sources	Access
🧬 Transcriptomics	CIGS · Tahoe-100M · LINCS 2020	CIGS · Tahoe · LINCS
🖼️ Cell morphology	Cell Painting Gallery (cpg0016, cpg0003)	AWS Open Data
🔗 Multimodal datasets	CDRP-BBBC047-Bray · CDRPBIO-BBBC036-Bray	Data resource reference
📦 Processed release	MVCBench benchmark data	🤗 Hugging Face

Cross-modal correspondence and experimental pairing should be interpreted at the level supported by each dataset, not assumed to mean same-cell image–RNA measurements.

🧩 Embedding Extraction

Representation	Notebook / code	Description
🧪 Molecular (e.g., UniMol2)	Get_Molecular_Embedding.ipynb	Generate compound/molecular representations for benchmark inputs.
🧬 Gene (e.g., STATE)	Get_STATE_Embedding.ipynb	Compute gene-expression/cellular representations with the corresponding upstream model.

For STATE, the upstream interface includes the following call (see the linked notebook for setup and actual parameters):

inferer.encode_adata(  # https://github.com/ArcInstitute/state
    input_file,
    output_file,
    emb_key=embed_key,
    dataset_name=dataset_name,
    gene_column=gene_column,
)

📖 Citation

If MVCBench supports your research, please cite the publicly available preprint until the final journal record is published.

@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},
  url     = {https://www.biorxiv.org/content/10.64898/2026.04.22.720110v1},
  note    = {Preprint; accepted in principle at Nature Communications}
}