---
license: cc-by-4.0
tags:
- biology
- protein
- protein-language-model
- embeddings
- representation-learning
pretty_name: Vec2Vec protein embeddings of PLAT
size_categories:
- 100K<n<1M
---

# Vec2Vec protein embeddings of PLAT

Frozen, pooled embeddings of every protein in
[GleghornLab/plat_data](https://huggingface.co/datasets/GleghornLab/plat_data) from ten encoders that
read it three different ways: protein language models read the amino acid sequence, general text
models read the sequence or the curator description, and an annotation encoder reads the typed
annotation set. With every encoder embedding the same proteins, a translator between any two spaces
can be trained and scored without running a large model. These are the embeddings behind the Gleghorn
Lab's study of embedding translation between protein representation spaces.

| | |
|---|---|
| Proteins | 67,402 (PLAT train, valid and test) |
| Encoders | ESM2 8M, 35M, 150M, 650M and 3B; ModernBERT-base and -large; gpt-oss-20b and -120b; AT-Base |
| Vector | Mean then variance over tokens, concatenated: twice the encoder width, float32 |
| Format | One gzip-compressed PyTorch dictionary per encoder, keyed by the raw input |
| Licence | CC BY 4.0, inheriting the UniProt data |

## Homology leakage

The embeddings carry PLAT's splits unchanged. PLAT was clustered at 40% identity before the split, but a
later screen at 40% identity and 80% coverage of both sequences found a training neighbor for 12.57% of
test proteins, so it is not homology-disjoint; the
[PLAT card](https://huggingface.co/datasets/GleghornLab/plat_data) gives the counts.

## How it was built

1. Read every row of PLAT.
2. Ran each encoder frozen, in inference only, on the columns it reads, and truncated inputs instead
   of dropping them: sequences at 1,024 tokens, descriptions at 2,048 and annotation sets at 192.
3. Pooled each input's token embeddings into their mean and variance and concatenated the two,
   mean first.
4. Stored one vector per distinct input in a dictionary per encoder, so proteins with identical
   inputs share one entry, and saved it with `torch.save` under gzip.

## Columns

Each file is a dictionary from an input to its vector.

| Key | Encoders | Example |
|---|---|---|
| amino acid sequence, `str` | ESM2, ModernBERT, gpt-oss | `"MDQPQFSGAPRFLTRPKAFV..."` |
| description text, `str` | ModernBERT, gpt-oss | `"The following text describes a protein: ..."` |
| annotation ids, `tuple` of `int` | AT-Base | `(2459, 2633, 6640, ...)` |

A description or sequence is looked up exactly as it appears in PLAT, and an annotation set as the
tuple of its `annotation` list. For example, the ESM2-8 file holds 67,402 sequences with vectors of 640
values, ModernBERT-base holds 107,096 keys (67,402 sequences and 39,694 distinct descriptions) with 1,536, and AT-Base holds 53,371
distinct annotation sets with 1,024.

## Example rows

```python
import gzip, io, torch
from datasets import load_dataset
from huggingface_hub import hf_hub_download

path = hf_hub_download("GleghornLab/vec2vec_embeddings",
                       "embeddings/ESM2-650_False_mean_var.pth.gz", repo_type="dataset")
with gzip.open(path, "rb") as handle:
    table = torch.load(io.BytesIO(handle.read()), weights_only=True)

plat = load_dataset("GleghornLab/plat_data", split="test")
x = torch.stack([table[seq] for seq in plat["sequence"]])  # (3420, 2560)
```

Load with `weights_only=True`: the files are pickle containers, and nothing else is needed to read them.

## What a model trained on it may say

A translator trained on these vectors maps one encoder's pooled space to another's. Its scores say how
much of one space can be recovered from the other on PLAT's proteins, under mean and variance pooling and
these truncation lengths. They do not say how the full token-level representations relate, and an
encoder's pooled vector on a long protein covers only its first 1,024 tokens.

## Sources and licence

The inputs come from PLAT, built from [UniProtKB](https://www.uniprot.org) under
[CC BY 4.0](https://www.uniprot.org/help/license); cite UniProt when you use them. The encoders and
their licences:

| Encoder | Model | Reads | Licence |
|---|---|---|---|
| ESM2 8M to 3B | [facebook/esm2](https://huggingface.co/facebook/esm2_t33_650M_UR50D) family | sequence | MIT |
| ModernBERT-base, -large | [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base), [-large](https://huggingface.co/answerdotai/ModernBERT-large) | sequence, description | Apache 2.0 |
| gpt-oss-20b, -120b | [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b), [-120b](https://huggingface.co/openai/gpt-oss-120b) | sequence, description | Apache 2.0 |
| AT-Base | Annotation Transformer of the Gleghorn Lab, not public | annotation set | |

## Files

| File | Encoder | Keys |
|---|---|---|
| `embeddings/ESM2-8_False_mean_var.pth.gz` | ESM2 8M | sequence |
| `embeddings/ESM2-35_False_mean_var.pth.gz` | ESM2 35M | sequence |
| `embeddings/ESM2-150_False_mean_var.pth.gz` | ESM2 150M | sequence |
| `embeddings/ESM2-650_False_mean_var.pth.gz` | ESM2 650M | sequence |
| `embeddings/ESM2-3B_False_mean_var.pth.gz` | ESM2 3B | sequence |
| `embeddings/custom---answerdotai_ModernBERT-base_False_mean_var.pth.gz` | ModernBERT-base | sequence, description |
| `embeddings/custom---answerdotai_ModernBERT-large_False_mean_var.pth.gz` | ModernBERT-large | sequence, description |
| `embeddings/custom---openai_gpt-oss-20b_False_mean_var.pth.gz` | gpt-oss-20b | sequence, description |
| `embeddings/gpt-oss-120b_False_mean_var.pth.gz` | gpt-oss-120b | sequence, description |
| `embeddings/at_AT-Base_best_pad192_mean_var.pth.gz` | AT-Base | annotation set |