Esmplusplus Large

🚀 ESM++

ESM++ is a faithful implementation of ESMC that allows for batching and standard Huggingface compatibility without requiring the ESM Python package.

🚀 Quick Start

There was previously a bug with Huggingface weight tieing that caused the logits of ESM++ to differ from ESMC. That bug is now resolved.

ESM++ is a faithful implementation of ESMC (license) that allows for batching and standard Huggingface compatibility without requiring the ESM Python package. The large version corresponds to the 600 million parameter version of ESMC.

💻 Usage Examples

Basic Usage

from transformers import AutoModelForMaskedLM
model = AutoModelForMaskedLM.from_pretrained('Synthyra/ESMplusplus_large', trust_remote_code=True)
tokenizer = model.tokenizer

sequences = ['MPRTEIN', 'MSEQWENCE']
tokenized = tokenizer(sequences, padding=True, return_tensors='pt')

# tokenized['labels'] = tokenized['input_ids'].clone() # correctly mask input_ids and set unmasked instances of labels to -100 for MLM training

output = model(**tokenized) # get all hidden states with output_hidden_states=True
print(output.logits.shape) # language modeling logits, (batch_size, seq_len, vocab_size), (2, 11, 64)
print(output.last_hidden_state.shape) # last hidden state of the model, (batch_size, seq_len, hidden_size), (2, 11, 1152)
print(output.loss) # language modeling loss if you passed labels
#print(output.hidden_states) # all hidden states if you passed output_hidden_states=True (in tuple)

Advanced Usage

ESM++ also supports sequence and token level classification tasks like ESM2. Simply pass the number of labels during initialization.

from transformers import AutoModelForSequenceClassification, AutoModelForTokenClassification

model = AutoModelForSequenceClassification.from_pretrained('Synthyra/ESMplusplus_large', num_labels=2, trust_remote_code=True)
logits = model(**tokenized).logits
print(logits.shape) # (batch_size, num_labels), (2, 2)

ESM++ weights are fp32 by default. You can load them in fp16 or bf16 like this:

import torch
model = AutoModelForMaskedLM.from_pretrained('Synthyra/ESMplusplus_large', trust_remote_code=True, torch_dtype=torch.float16) # or torch.bfloat16

Embedding Entire Datasets

To embed a list of protein sequences fast, just call embed_dataset. Sequences are sorted to reduce padding tokens, so the initial progress bar estimation is usually much longer than the actual time it will take.

embedding_dict = model.embed_dataset(
    sequences=[
        'MALWMRLLPLLALLALWGPDPAAA', ... # list of protein sequences
    ],
    tokenizer=model.tokenizer,
    batch_size=2, # adjust for your GPU memory
    max_len=512, # adjust for your needs
    full_embeddings=False, # if True, no pooling is performed
    embed_dtype=torch.float32, # cast to what dtype you want
    pooling_types=['mean', 'cls'], # more than one pooling type will be concatenated together
    num_workers=0, # if you have many cpu cores, we find that num_workers = 4 is fast for large datasets
    sql=False, # if True, embeddings will be stored in SQLite database
    sql_db_path='embeddings.db',
    save=True, # if True, embeddings will be saved as a .pth file
    save_path='embeddings.pth',
)
# embedding_dict is a dictionary mapping sequences to their embeddings as tensors for .pth or numpy arrays for sql

model.embed_dataset()
Args:
    sequences: List of protein sequences
    batch_size: Batch size for processing
    max_len: Maximum sequence length
    full_embeddings: Whether to return full residue-wise (True) embeddings or pooled (False)
    pooling_type: Type of pooling ('mean' or 'cls')
    num_workers: Number of workers for data loading, 0 for the main process
    sql: Whether to store embeddings in SQLite database - will be stored in float32
    sql_db_path: Path to SQLite database
    
Returns:
    Dictionary mapping sequences to embeddings, or None if sql=True

Note:
    - If sql=True, embeddings can only be stored in float32
    - sql is ideal if you need to stream a very large dataset for training in real-time
    - save=True is ideal if you can store the entire embedding dictionary in RAM
    - sql will be used if it is True and save is True or False
    - If your sql database or .pth file is already present, they will be scanned first for already embedded sequences
    - Sequences will be truncated to max_len and sorted by length in descending order for faster processing

Fine-tuning with 🤗 peft

model = AutoModelForSequenceClassification.from_pretrained('Synthyra/ESMplusplus_large', num_labels=2, trust_remote_code=True)
# these modules handle ESM++ and ESM2 attention layers
target_modules = ["layernorm_qkv.1", "out_proj", "query", "key", "value", "dense"]

lora_config = LoraConfig(
    r=8, # choose lora parameters to your liking
    lora_alpha=16,
    lora_dropout=0.01,
    bias="none",
    target_modules=target_modules,
)

# Apply LoRA to the model
model = get_peft_model(model, lora_config)

# Unfreeze the classifier head
for param in model.classifier.parameters():
    param.requires_grad = True

For a more thourough example of fine-tuning, check out our example script here.

Returning Attention Maps

Usually F.scaled_dot_product_attention is used for the attention calculations, which is much faster than native PyTorch. However, it cannot return attention maps. ESM++ has the option to output_attentions, which will calculate attention manually. This is much slower, so do not use unless you need the attention maps.

output = model(**tokenized, output_attentions=True)
att = output.attentions
len(att) # 33, one for each layer, size (batch_size, num_heads, seq_len, seq_len) each

📚 Documentation

Comparison across floating-point precision and implementations

We measured the difference of the last hidden states of the fp32 weights vs. fp16 or bf16. We find that the fp16 is closer to the fp32 outputs, so we recommend loading in fp16. Please note that the ESM package also loads ESMC in fp32 but casts to bf16 by default, which has its share of advantages and disadvantages in inference / training - so load whichever you like for half precision.

Average MSE for FP16: 0.00000003

Average MSE for BF16: 0.00000122

We also measured the difference between the outputs of ESM++ vs. ESMC (both in bfloat16) on 1000 random sequences to ensure compliance with the ESM package.

Average MSE of last hidden state: 2.46e-09

You can load the weights from the ESM package instead of transformers by replacing .from_pretrained(...) to .from_pretrained_esm('esmc_600m')

Model Probes

We employ linear probing techniques on various PLMs and standard datasets, similar our previous paper, to assess the intrinsic correlation between pooled hidden states and valuable properties. ESMC (and thus ESM++) perform very well.

The plot below showcases performance normalized between the negative control (random vector embeddings) and the best performer. Classification task scores are averaged between MCC and F1 (or F1max for multilabel) and regression tasks are averaged between Spearman rho and R2.

Inference Speeds

We look at various ESM models and their throughput on an H100. Adding efficient batching between ESMC and ESM++ significantly improves the throughput, although ESM++ is also faster than ESMC for batch size one. ESM++ small is even faster than ESM2-35M with long sequences! The most gains will be seen with PyTorch > 2.5 on linux machines.

Citation

If you use any of this implementation or work please cite it (as well as the ESMC preprint).

@misc {ESMPlusPlus,
	author       = { Hallee, L. and Bichara, D. and Gleghorn, J, P. },
	title        = { ESMPlusPlus },
	year         = 2024,
	url          = { https://huggingface.co/Synthyra/ESMplusplus_small },
	doi          = { 10.57967/hf/3726 },
	publisher    = { Hugging Face }
}