MorphoDiff: Cellular Morphology Painting with Diffusion Models
Authors: Zeinab Navidi, Jun Ma, Esteban Miglietta, Le Liu, Anne Carpenter, Beth Cimini, Benjamin Haibe-Kains, BO WANG
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We benchmarked Morpho Diff on three microscopy imaging datasets, evaluating image fidelity and biological properties of cells in in-distribution and unseen perturbations, demonstrating promising performance overall. |
| Researcher Affiliation | Academia | 1. Department of Computer Science, University of Toronto, Toronto, ON, Canada 2. Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada 3. Vector Institute, Toronto, ON, Canada 4. Imaging Platform, Broad Institute of Harvard and MIT, Cambridge, MA, USA 5. Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada 6. Structural Genomics Consortium, Toronto, ON, Canada 7. Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada 8. Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, ON, Canada 9. AI Hub, University Health Network, Toronto, ON, Canada |
| Pseudocode | No | The paper describes the model architecture and training process using text and mathematical equations, but it does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for generating chemical and genetic perturbation embeddings is available in the project s Git Hub. The code for data pre-processing, model training, evaluation, and result reproduction, along with the model weights are available on the project s Git Hub page. |
| Open Datasets | Yes | Three publicly available Cell Painting datasets were used for modelling and validation, representing a diverse range of perturbations and cell types, described as follows. [...] The images were directly downloaded from Recursion s website9. [...] Images from this dataset were directly downloaded from Broad Bioimage Benchmark Collection website10. [...] Images for this dataset were directly downloaded form the Cell Painting Gallery Git Hub page11. |
| Dataset Splits | No | The paper mentions using 'held-out compounds' and 'in-distribution drugs' for generalization, and discusses 'training images' and 'balanced images per perturbation' in Table 4. However, it does not provide specific train/validation/test dataset splits (e.g., percentages or exact counts) for the overall model training and evaluation. |
| Hardware Specification | Yes | All training and evaluation processes were conducted on NVIDIA T4 and NVIDIA A40 GPUs. |
| Software Dependencies | Yes | Publicly released parameters of Stable Diffusion v1.45 along with the accompanying training script from Huggingface were used as the base model6. |
| Experiment Setup | Yes | A batch size of 32 and a learning rate of 1 10 5 were used for all training runs. |