Research map / Probabilistic and scientific modeling / Molecules and proteins
CellFlux: Simulating Cellular Morphology Changes via Flow Matching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CellFlux, a flow-matching model that transforms control-cell image distributions into perturbed-cell distributions within the same batch to simulate morphological responses to chemical and genetic perturbations.
Problem
Predicting cellular morphology changes under chemical and genetic perturbations from microscopy images, where paired control/treated cells cannot be observed (cell painting is destructive) and batch effects confound true perturbation effects.
Approach
Formulates morphology prediction as a distribution-to-distribution transformation: flow matching starts from same-batch control images rather than noise, with condition embedding, noise augmentation, and classifier-free guidance, so batch effects are preserved as shared context while the model learns the perturbation-specific difference.
Key results
- Overall FID of 18.7 on BBBC021, 33.0 on RxRx1, and 9.0 on JUMP, vs. best baseline IMPA at 33.7, 41.6, and 14.6.
- MoA accuracy of 71.1% on BBBC021 generated images vs. 63.7% for the best baseline (IMPA) and 72.4% for real perturbed images.
- Same-batch control conditioning improves conditional FID by 14% and MoA accuracy by 48% over conditioning on other-batch controls.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CellFlux instead of IMPA | cell morphology generation · BBBC021 | FID_o: 33.7 -> 18.7 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · RxRx1 | FID_o: 41.6 -> 33.0 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · JUMP | FID_o: 14.6 -> 9.0 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · RxRx1 | FID_c: 164.8 -> 163.5 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · BBBC021 | MoA Accuracy: 63.7 -> 71.1 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · BBBC021 OOD | FID_o: 44.5 -> 42.0 | Helped |
| Use CellFlux instead of IMPA | cell morphology generation · BBBC021 OOD | MoA Accuracy: 16.0 -> 43.2 | Helped |
| Use same-batch rather than other-batch control initialization | CellFlux · BBBC021 | FID_c: 66.3 -> 56.8 | Helped |
| Add perturbation condition (ablation) | CellFlux · BBBC021 | FID_o: 45.0 -> 18.7 | Helped |
| Add perturbation condition (ablation) | CellFlux · BBBC021 | FID_c: 113.0 -> 56.8 | Helped |
| Add classifier-free guidance (ablation) | CellFlux · BBBC021 | FID_o: 32.6 -> 18.7 | Helped |
| Add classifier-free guidance (ablation) | CellFlux · BBBC021 | FID_c: 92.4 -> 56.8 | Helped |
Surprising findings
- A CellFlux model trained on BBBC021 transfers perturbation effects zero-shot to RxRx1 and JUMP images despite domain shift.
- Conditioning on control images from other batches yields much lower MoA accuracy (48.2%) than same-batch conditioning (71.2%), showing batch-specific context matters.
Limitations
- Evaluated on a few cancer cell lines (MCF-7, HUVEC, U2OS/A549), which may limit biological representativeness.
- Biological validity of generated trajectory interpolations is unverified without ground-truth or experimental validation.
- Smaller improvement on RxRx1 attributed to limited images per perturbation type.
Experimental setup
Datasets: BBBC021 RxRx1 JUMP (CPJUMP1)
Benchmarks: BBBC021 RxRx1 JUMP
Metrics: FID KID MoA accuracy MoA macro-F1 MoA weighted-F1 CellProfiler nuclear size
Baselines: PhenDiff IMPA MorphoDiff
Builds on: flow matching rectified flow classifier-free guidance IMPA PhenDiff
Closest papers by meaning
- CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 2025 · similarity 0.85)
- MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays (NeurIPS 2026 · similarity 0.85)
- Multi-Modal and Multi-Attribute Generation of Single Cells with CFGen (ICLR 2025 · similarity 0.84)
- GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.84)
- Joint Velocity-Growth Flow Matching for Single-Cell Dynamics Modeling (NeurIPS 2025 · similarity 0.84)
- Bridging Protein Sequences and Microscopy Images with Unified Diffusion Models (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.