ESMC ProtHash

A protein language model that outputs amino acid sequence embeddings for use in clustering, classification, locality-sensitive hashing, and more. Distilled from the ESMC family of models, ProtHash produces contextual embeddings that align in vector space according to the sequences' underlying biological properties such as structure and function. Trained on the SwissProt dataset to mimic the activations of its ESMC teacher model, ProtHash embeddings have near-perfect similarity to ESMC embeddings but at a greatly reduced computational cost.

Key Features

• Blazing fast and efficient: ProtHash uses as few as 1.5% of its ESMC teacher's total parameters to achieve near-perfect cosine similarity between the two embedding spaces.

• Biologically-relevant: Biologically similar proteins will show up nearby in the embedding space enabling downstream tasks such as clustering, classification, and locality-sensitive hashing.

• Compatible with ESMC: ProtHash can output embeddings in its native or ESMC teacher's dimensionality - allowing it to serve as either a faster drop-in approximation to ESMC embeddings or a more efficient compressed representation.

• Quantization-ready: With quantization-aware post-training, ProtHash allows you to quantize the weights of the model while maintaining similarity to the teacher's embedding space.

Pretrained Models

Name	Context Length	Embedding Dimensions	Attention Heads (Q/KV)	Encoder Layers	Total Params	Teacher Model	Teacher Dimensions
andrewdalpino/ProtHash-384-Tiny	2048	384	16/4	4	5M	esmc_300m	960
andrewdalpino/ProtHash-384	2048	384	16/4	10	11M	esmc_300m	960
andrewdalpino/ProtHash-512-Tiny	2048	512	16/4	4	8.5M	esmc_600m	1152
andrewdalpino/ProtHash-512	2048	512	16/4	10	19M	esmc_600m	1152

Pretrained Example

First, you'll need the prothash and esm packages installed into your environment. We recommend using a virtual environment such as Python's venv module to prevent version conflicts with any system packages.

pip install prothash esm

Then, load the weights from HuggingFace Hub, tokenize a protein sequence, and pass it to the model. ProtHash adopts the ESM tokenizer as it's amino acids tokenization scheme which consists of a vocabulary of 33 amino acid and special tokens. The output will be an embedding vector that can be used in downstream tasks such as comparing to other protein sequence embeddings, clustering, and near-duplicate detection.

import torch

from esm.tokenization import EsmSequenceTokenizer

from prothash.model import ProtHash

tokenizer = EsmSequenceTokenizer()

model_name = "andrewdalpino/ProtHash-512-Tiny"

model = ProtHash.from_pretrained(model_name)

# Optionally quantize the weights.
model.quantize_weights()

sequence = input("Enter a sequence: ")

out = tokenizer(sequence, max_length=2048)

tokens = out["input_ids"]

# Input is a [1, T] tensor of token indices. 
x = torch.tensor(tokens, dtype=torch.int64).unsqueeze(0)

# Output the sequence embedding in native dimensionality.
y_embed_native = model.embed_native(x).squeeze(0)

# Output a drop-in replacement for the teacher's embeddings.
y_embed_teacher = model.embed_teacher(x).squeeze(0)

print(y_embed_native.shape)
print(y_embed_teacher.shape)

Training

Clone the project repo

We'll need the code from the project repository to train and/or fine-tune the model.

git clone https://github.com/andrewdalpino/ProtHash

Install Project Dependencies

Project dependencies are specified in the requirements.txt file. You can install them with pip using the following command from the project root. We recommend using a virtual environment such as venv to keep package dependencies on your system tidy.

python -m venv ./.venv

source ./.venv/bin/activate

pip install -r requirements.txt

Distilling

ProtHash is trained to mimic the activations of its ESMC teacher model. To begin distillation with the default arguments check the example below.

python train.py

You can change the default arguments like in the example below.

python train --teacher_name="esmc_300m" --max_steps=3500 --embedding_dimensions=256

Training Dashboard

We use TensorBoard to capture and display training events such as loss and gradient norm updates. To launch the dashboard server run the following command from the terminal.

tensorboard --logdir=./runs

Then navigate to the dashboard using your favorite web browser.

Training Arguments

Argument	Default	Type	Description
--teacher_name	'esmc_600m'	str	The teacher model name.
--num_dataset_processes	1	int	The number of CPU processes to use to preprocess the dataset.
--min_sequence_length	1	int	The minimum length of the input sequences.
--max_sequence_length	2048	int	The maximum length of the input sequences.
--quantization_aware_training	False	bool	Should we add fake quantized tensors to simulate quantized training?
--batch_size	4	int	The number of training images to pass through the network at a time.
--gradient_accumulation_steps	32	int	The number of batches to pass through the network before updating the model weights.
--max_steps	4200	int	The number of steps to train for.
--learning_rate	1e-4	float	The learning rate of the AdamW optimizer.
--max_gradient_norm	100.0	float	Clip gradients above this threshold norm before stepping.
--temperature	8.0	float	The smoothing parameter of the activations - higher temperature results in smoother activations.
--embeddings_dimensions	512	int	The dimensionality of the native embeddings.
--q_heads	16	int	The number of query heads used in the self-attention layers.
--kv_heads	4	int	The number of key and value heads used in the self-attention layers.
--hidden_ratio	2	(1, 2, 4)	The ratio of hidden neurons to embedding dimensions in the feed-forward layers of the network.
--num_encoder_layers	4	int	The number of layers within the body of the encoder.
--dropout	0.0	float	The proportion of activations to send to zero during training as regularization.
--activation_checkpointing	False	bool	Should we use activation checkpointing? This will drastically reduce memory utilization during training at the cost of recomputing the forward pass.
--eval_interval	100	int	Evaluate the model after this many epochs on the testing set.
--checkpoint_interval	100	int	Save the model checkpoint to disk every this many epochs.
--checkpoint_path	"./checkpoints/checkpoint.pt"	str	The path to the base checkpoint file on disk.
--resume	False	bool	Should we resume training from the last checkpoint?
--run_dir_path	"./runs"	str	The path to the TensorBoard run directory for this training session.
--device	"cpu"	str	The device to run the computation on.
--seed	None	int	The seed for the random number generator.

References

• The UniProt Consortium, UniProt: the Universal Protein Knowledgebase in 2025, Nucleic Acids Research, 2025, 53, D609–D617.
• T. Hayes, et al. Simulating 500 million years of evolution with a language model, 2024.
• B. Zhang, et al. Root Mean Square Layer Normalization. 33rd Conference on Neural Information Processing Systems, NeurIPS 2019.
• J. Ainslie, et al. GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints, Google Research, 2023.
• T. Kim, et al. Comparing Kullback-Leibler Divergence and Mean Squared Error Loss in Knowledge Distillation, 2021.