Lab Newsletter — September 18, 2026: Where Proteins Live

This week we’ve watched AI design antibodies and plan syntheses. Today we come home — to a question the lab has worked on since its beginning: where in the cell does a protein live? A kinase at the plasma membrane, in the nucleus, or in a mislocalized aggregate is, functionally, three different stories; the same amino-acid sequence can mean health or disease depending on its address. Mapping that address across the whole proteome is the mission of the Human Protein Atlas Cell Atlas — and today’s digest traces how deep learning learned to predict subcellular localization, from sequence and from images.

🧬 The founding move: localization from sequence alone

Classical predictors leaned on annotated homologues — useless for a truly novel protein. Almagro Armenteros et al. (Bioinformatics, 2017) set out from exactly that gap: “for novel proteins where no annotated homologues exist, and for predicting the effects of sequence variants, it is desirable to have methods for predicting protein properties from sequence information only.” Their DeepLoc answered with “deep neural networks to predict protein subcellular localization relying only on sequence information.” At its core is “a recurrent neural network that processes the entire protein sequence and an attention mechanism identifying protein regions important for the subcellular localization.” No database lookup — just the sequence, and a model that points to which residues decide the destination.

🗣️ The language of life replaces the alignment

DeepLoc still needed rich evolutionary features. Stärk et al. (Bioinformatics Advances, 2021) removed that cost by leaning on protein language models. They noted the bottleneck — machine learning “narrows the gap through predictions from expert-designed input features leveraging information from multiple sequence alignments (MSAs) that is resource expensive to generate” — then showed “embeddings from protein language models for competitive localization prediction without MSAs.” Their architecture, a “softmax weighted aggregation mechanism with linear complexity in sequence length referred to as light attention,” “significantly outperformed the state-of-the-art … by about 8 percentage points (Q10).” A pretrained language model, it turns out, already knows a great deal about where a protein goes.

🏷️ Many addresses at once, and why

Real proteins often live in more than one place, and a good predictor should say so — and explain itself. Thumuluri et al. (Nucleic Acids Research, 2022) delivered that with DeepLoc 2.0, “an update to the popular tool DeepLoc with multi-localization prediction and improvements in both performance and interpretability.” It “achieve[s] state-of-the-art performance … by using a pre-trained protein language model,” and, usefully, “uses sequence input rather than relying on slower protein profiles.” Two interpretability wins matter for biology: “an attention output along the sequence and highly accurate prediction of nine different types of protein sorting signals” — the model doesn’t just name the compartment, it points to the address label the cell itself reads.

🎮 The image side, at planetary scale: citizen science

Sequence is only half the story — the Cell Atlas is built from microscopy images, and annotating them is a mountain of work. Sullivan et al. (Nature Biotechnology, 2018) met it with a genuinely unusual pairing. Using “the publicly available data set from the Cell Atlas of the Human Protein Atlas,” they “integrated an image-classification task into a mainstream video game (EVE Online) as a mini-game, named Project Discovery.” The scale is staggering: “participation by 322,006 gamers over 1 year provided nearly 33 million classifications of subcellular localization patterns, including patterns that were not previously annotated by the HPA.” They paired the crowd with a machine — “an automated Localization Cellular Annotation Tool (Loc-CAT)” that “classifies proteins into 29 subcellular localization patterns and can deal efficiently with multi-localization proteins.” Humans and models, learning the cell together.

🏆 A community benchmark for the hardest patterns

To push the image models further, the HPA turned the problem into an open competition. Ouyang et al. (Nature Methods, 2019) — work from our own group — analyzed it: “pinpointing subcellular protein localizations from microscopy images is easy to the trained eye, but challenging to automate.” The response was enormous — “over 3 months, 2,172 teams participated,” wrestling with “training on highly imbalanced classes and predicting multiple labels per image.” The payoff was real progress: “the winning models far outperformed our previous effort at multi-label classification of protein localization patterns by ~20%,” and, importantly, “these models can be used as classifiers to annotate new images, feature extractors to measure pattern similarity or pretrained” networks — reusable infrastructure, not just a leaderboard. A benchmark that left the field better armed.

🧠 No labels at all: a self-supervised atlas of localization

The frontier is learning localization without the labels entirely. Kobayashi et al. (Nature Methods, 2022) reached it with cytoself, “a deep-learning approach for fully self-supervised protein localization profiling and clustering” whose “self-supervised training scheme … does not require preexisting knowledge, categories or annotations.” Trained “on images of 1,311 endogenously labeled proteins from the OpenCell database,” cytoself “reveals a highly resolved protein localization atlas that recapitulates major scales of cellular organization, from coarse classes, such as nuclear and cytoplasmic, to the subtle localization signatures of individual protein complexes.” A model that discovers the cell’s spatial vocabulary on its own — and clusters proteins into organelles and complexes it was never told about.

🧫 Why it’s our kind of problem

Read across the six and the arc is the lab’s own history and future in one line: from sequence-only predictors to language-model embeddings, and from hand-labeled images to community-scale annotation to fully self-supervised atlases. This is where much of the lab’s science begins — the Human Protein Atlas Cell Atlas — and it threads directly into what the lab builds now: the same imaging that feeds Agent-Lens and the REEF farm is exactly what these localization models consume, and open, reusable classifiers are the BioImage Model Zoo and BioEngine idea in miniature. The move toward self-supervised representations (cytoself, the language-model localizers) is the same bet the lab makes across single-cell and imaging foundation models. And above all, localization is a load-bearing layer for the virtual cell: you cannot simulate a cell you cannot spatially map. Knowing where every protein lives — and how that changes with state — is a prerequisite for modeling what the cell does. The address is part of the function.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.