# Masked Autoencoders are Scalable Learners of Cellular Morphology

**Authors:** Oren Kraus, Kian Kenyon-Dean, Saber Saberian, Maryam Fallah, Peter McLean, Jess Leung, Vasudev Sharma, Ayla Khan, Jia Balakrishnan, Safiye Celik, Maciej Sypetkowski, Chi Vicky Cheng, Kristen Morse, Maureen Makes, Ben Mabey, Berton Earnshaw  
**arXiv:** [2309.16064](https://arxiv.org/abs/2309.16064)  
**Date:** 2023-09-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.16064v2/

## Key points

- Deep vision models are more effective at capturing biological signals compared to hand-crafted features
- Convolutional Neural Network (CNN) and Vision Transformer (ViT) based masked autoencoders outperform weakly supervised baselines
- ViT-L/8 model trained on over 3.5 billion unique crops achieved relative improvements of up to 28% in inferring known biological relationships
- Researchers adapted U-Nets for masked autoencoding (MU-Nets) with promising outcomes
- Study includes detailed figures illustrating StringDB recall and training FLOps, as well as recall across different cosine similarity percentiles for each database

## Summary

, , , , 

In the field of biological research, inferring biological relationships from cellular phenotypes in high-content microscopy screens presents both significant opportunities and challenges. Previous studies have demonstrated that deep vision models are more effective at capturing biological signals compared to hand-crafted features. This study delves into the scalability of self-supervised deep learning approaches when training larger models on extensive microscopy datasets. The results of this research show that both Convolutional Neural Network (CNN) and Vision Transformer (ViT) based masked autoencoders outperform weakly supervised baselines. At the highest scale examined in this study, a ViT-L/8 model trained on over 3.5 billion unique crops sampled from 93 million microscopy images achieved relative improvements of up to 28% over the best weakly supervised baseline in inferring known biological relationships sourced from public databases. The authors, including Oren Kraus, Kian Kenyon-Dean, Saber Saberian, Maryam Fallah, Peter McLean, Jess Leung, Vasudev Sharma, Ayla Khan, Jia Balakrishnan, Safiye Celik, Maciej Sypetkowski, Chi Vicky Cheng, Kristen Morse, Maureen Makes, Ben Mabey, and Berton Earnshaw conducted this study which was spotlighted at the NeurIPS 2023 Generative AI and Biology (GenBio) Workshop. Furthermore,
the researchers adapted U-Nets for masked autoencoding (MU-Nets) by training them to reconstruct masked sections of input images. The results for MU-Net-M showcase promising outcomes in line with previous studies. The study also includes detailed figures illustrating StringDB recall as a function of training FLOps and recall across different cosine similarity percentiles for each database. Overall, this work contributes valuable insights into leveraging self-supervised deep learning approaches for analyzing cellular morphology in high-content microscopy screens. The release of relevant code and select models associated with this research can be accessed at https://github.com/recursionpharma/maes_microscopy.