Machine Learning Model Predicts Protein Locations in Human Cells

The misplacement of proteins within cells is linked to diseases such as Alzheimer’s, cystic fibrosis, and cancer. Identifying the location of any of the approximately 70,000 proteins and their variants in a single human cell is a complex and resource-intensive task, as traditional experiments can only assess a few proteins at a time.

To address this challenge, researchers from MIT, Harvard, and the Broad Institute have developed a computational approach capable of predicting the location of any protein in any human cell line, including proteins and cell types not previously tested. This new method, described in a recent Nature Methods paper, advances beyond many existing techniques by providing single-cell localization, rather than averaged results across a cell population. For example, it can pinpoint a protein’s position in a specific cancer cell after treatment.

The technique, called PUPS (Prediction of Unseen Proteins' Subcellular localization), combines a protein language model with an image inpainting model. The protein language model analyzes the amino acid sequence and 3D structure of the protein, while the image inpainting model examines three stained images of a cell to gather information about its type, features, and stress state. The two models work together to generate an image highlighting the predicted protein location within a single cell.

During training, PUPS is tasked with both predicting the protein’s location and naming the compartment, such as the nucleus, to improve its understanding. This dual-task approach helps the model generalize to proteins and cell lines it has not encountered before. Users provide a protein sequence and three cellular images, one each for the nucleus, microtubules, and endoplasmic reticulum, and receive a prediction of the protein’s location.

Experimental validation showed that PUPS could accurately predict protein locations in new cell lines and proteins, outperforming baseline AI methods. The researchers aim to further develop PUPS to predict the localization of multiple proteins and understand protein-protein interactions in the future.