The Problem: Every Model Speaks Its Own Embedding Language

Protein AI now has many useful models. Some read amino acid sequences. Some read natural-language descriptions. Some read structured biological annotations. Each model turns its input into embeddings, which are numerical representations of what the model thinks matters.

The problem is that those embedding spaces are not automatically compatible. A protein sequence model and a text model may both know something about a kinase, but they may encode that knowledge in different coordinates.

Vec2Vec asks whether those coordinates can be translated.

The Core Idea

The paper builds on a broader machine-learning idea called the Platonic Representation Hypothesis: models trained on related views of the same reality may learn representations with shared structure.

For proteins, the question becomes practical. Can we translate between a protein sequence embedding and a description embedding? Can we move between a protein language model and an Annotation Vocabulary model? Can the translation preserve enough identity to support search, annotation, and design?

If the answer is yes, researchers could connect models that were never designed to work together.

What The Paper Found

Protein representation spaces do share useful geometry, but not in a magical, fully unsupervised way.

The original Vec2Vec result from natural language suggested that unrelated embedding spaces could sometimes be aligned without paired examples. In proteins, geometry was shared, but per-protein identity required paired contrastive supervision. That is an important scientific distinction: the spaces rhyme, but they do not automatically index the same protein without guidance.

The strongest bridge was not free-text natural language. It was Annotation Vocabulary.

That makes biological sense. Annotation Vocabulary is a curated symbolic language for protein properties. It removes much of the noise in prose and gives the model a cleaner map of function. In the paper, sequence-to-Annotation Vocabulary translation was far stronger than sequence-to-natural-language description translation.

Why This Matters

Many teams already have embeddings from different systems. Recomputing everything with one universal model is expensive and often unnecessary.

A good representation translator could let researchers reuse existing embeddings, connect older models to newer ones, and build multimodal protein search tools that move between sequence, function, and text.

It also opens a path for frontier language models to contribute to protein modeling. If a language model can read a protein description or even a raw amino acid sequence, an adapter may translate that representation into a more protein-native space.

The paper showed large downstream gains from translating natural-language embeddings into a protein language model latent space. The broader message is that adapters can turn a weak biological representation into a stronger one without retraining the entire foundation model.

Synthyra's View

Synthyra products are built around practical biological workflows, not one model family. Vec2Vec-style alignment is valuable because it lets us connect specialized tools: sequence models, annotation models, interaction models, and user-facing language interfaces.

As with the other research directions, the product opportunity is not a simple port of the open paper. It is the extension of a research idea into better retrieval, better annotation, and easier model interoperability.

Limitations

Vec2Vec does not mean that all protein models know the same thing. Some representations preserve structure. Some preserve function. Some preserve taxonomy or composition. Translation can be asymmetric, and a small model cannot recover information it never encoded.

The practical use case is alignment and reuse, not universal model alchemy. When used carefully, that is still powerful.

---

This blog post summarizes work in the following paper:

Harnessing the Universal Geometry of Protein Embeddings
Logan Hallee, Jason P. Gleghorn
Manuscript draft, 2026