Summary

This repository contains code to train a U-Net to predict forces from protein distributions as described in Machine learning interpretable models of cell mechanics from protein images (arxiv).

The other two repositories for this paper are:

• Physical bottleneck analysis: https://github.com/schmittms/physical_bottleneck
• Green's function neural network: https://github.com/jcolen/gfnn

This code was used to train the networks Figures 1-4. The trained model weights can be downloaded here, and in load_trained_unet.ipynb we demonstrate how to load them. The raw data used for training can be downloaded from this link. Data pre-processing is described in the DataProcessing.ipynb notebook. This repository also contains a minimal working example notebook train_unet.ipynb which trains a U-Net on a small amount of data. This example dataset can be downloaded here. This notebook is for illustration purposes only; to train a network to the accuracy of those used in the paper, more data and longer training is needed.

Tip for applying our model to your data

We expect that our model is most useful as a pre-trained model which is further fine-tuned on some (potentially small) amount of data from your lab. This fine-tuning step will adapt the network to the new data so that factors such as microscope, substrate preparation, TFM hyperparameters etc. can be controlled for. As mentioned in the "Limitations of the study" section of the paper, our neural networks inherently rely on the data itself and their predictions depend on the data and experimental setup in subtle ways (see Figs. S7-S9). We found that using different microscopes, imaging fluorophores, substrate stiffnesses, etc. can affect the accuracy of predictions. We were able to account for some of these effects by normalization of protein fluorescence and force magnitude distributions, as described in DataProcessing.ipynb and the STAR Methods section. In particular, the choice of TFM regularization parameter proved to be an important factor, as it shifts both the magnitude and the typical "spread" of the measured forces (higher regularization parameter typically renders forces which are more ''smeared''). The dependence of force magnitude on this parameter is to some extent mitigated by normalization across each dataset, but the change in force spreads is difficult to account for (see Figs. S8-S9).

Data organization

All datasets are located in a master data directory. Each dataset consists of multiple cells. Each cell has its own folder where every frame of the time lapse is contained as a .npy file. This .npy file has shape [C, L, L] where L is the image size (images are square, typically L=992 or 1120), and there are C=7 or 8 channels. The channels correspond to [u_x, u_y, F_x, F_y, mask, forcemask, zyxin, actin] where u_x is the displacement in the x direction. Forces here are those directly measured from TFM, however before training and evaluation they are normalized according to the procedure in DataProcessing.ipynb and are presented in these normalized units in the main text.

data/
└───TractionData_16kPa/
│   │   cell_force_baselines_bydataset.csv
│   │   cell_force_baselines.csv
│   │   dataset.csv
│   │
│   └───dataset_A_cell_0
│   |   │   frame_0.npy
│   |   │   frame_1.npy
│   |   │   ...
│   |   │   frame_T.npy 
|   |
│   └───dataset_A_cell_1
│   |   │   frame_0.npy
│   |   │   frame_1.npy
│   |   │   ...
│   |   │   frame_T.npy
|   |
│   └───dataset_B_cell_0
│   |   │   frame_0.npy
│   |   │   frame_1.npy
│   |   │   ...
│   |   │   frame_T.npy 
|   |
│   └───dataset_B_cell_1
│   |   │   frame_0.npy
│   |   │   frame_1.npy
│   |   │   ...
│   |   │   frame_T.npy 
|  ...
...