biology microscopy text-to-image transformers

Huang Lab

CELL-E 2

Model description

CELL-E_2

CELL-E 2 is the second iteration of the original CELL-E model which utilizes an amino acid sequence and nucleus image to make predictions of subcellular protein localization with respect to the nucleus.

CELL-E 2 is novel bidirectional transformer that can generate images depicting protein subcellular localization from the amino acid sequences (and vice versa). CELL-E 2 not only captures the spatial complexity of protein localization and produce probability estimates of localization atop a nucleus image, but also being able to generate sequences from images, enabling de novo protein design. We trained on the Human Protein Atlas (HPA) and the OpenCell datasets.

CELL-E 2 utilizes pretrained amino acid embeddings from ESM-2. Localization is predicted as a binary image atop the provided nucleus. The logit values are weighted against these binary images to produce a heatmap of expected localization.

Spaces

We have two spaces available where you can run predictions on your own data!

Model variations

We have made several versions of CELL-E 2 available. The naming scheme follows the structure training set_hidden size where the hidden size is set to the embedding dimension of the pretrained ESM-2 model. We annotate the most useful models under Notes, however other models can be used if memory constraints are present. Since these models share similarities with BERT, the embeddings from any of these models may be benefical for downstream tasks.

HPA Models: HPA models are trained on the HPA dataset. They are best for general purpose predictions as they include a variety of cell types.

Model Size Notes
HPA_480 4.73 GB Best for Image Prediction
HPA_640 6.31 GB
HPA_1280 10.8 GB
HPA_2560 17.5 GB Best for Sequence Prediction

OpenCell Models: OpenCell models are trained on the OpenCell dataset. These only contain HEK cells and should ideally only be used for predictions on HEK cells. They perform well on image prediction but the generate heatmaps contain little information.

Model Size Notes
OpenCell_480 4.73 GB
OpenCell_640 6.31 GB
OpenCell_1280 10.8 GB
OpenCell_2560 17.5 GB Best for Sequence Prediction

Finetuned HPA Models: These models were used the HPA models as checkpoints, but then were finetuned on the OpenCell dataset. We found that they improve image generation capabilities, but did not necessary see an improvement in sequence prediction.

Model Size Notes
HPA_480 4.73 GB Best for Image Prediction
HPA_640 6.31 GB
HPA_1280 10.8 GB
HPA_2560 17.5 GB

To reduce download size, we removed the ESM-2 model from the checkpoint. This should be downloaded the first time the code is run, but is otherwise something to be aware of if loading into other projects.

How to use

The full codebase is available on GitHub. Download the model and make sure nuclues_vqgan.yaml, threshold_vqgan.yaml, config.yaml, and model.ckpt are present.

Here is how to use this model to do sequence prediction:

```python
configs = OmegaConf.load(configs/config.yaml);
model = instantiate_from_config(configs.model).to(device);
model.sample(text=sequence, condition=nucleus)

BibTeX entry and citation info

@inproceedings{
anonymous2023translating,
title={CELL-E 2: Translating Proteins to Pictures and Back with a Bidirectional Text-to-Image Transformer},
author={Emaad Khwaja, Yun S. Song, Aaron Agarunov, and Bo Huang},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=YSMLVffl5u}
}

Contact

We are an interdisciplinary lab based at UCSF. We are particularly seeking talents in optical engineering, machine learning, and cellular microscopy. Please reach out to Bo if you're interested in collaborating!