Pytorch modules and utilities for modeling biological sequence data.

Here we will demonstrate the application of several tools we hope will help with modeling biological sequences.

Installation

$ pip install sequence-models
$ pip install git+https://github.com/microsoft/protein-sequence-models.git  # bleeding edge, current repo main branch

Loading pretrained models

Models require PyTorch. We tested on v1.9.0, v1.11.0,and 1.12. If you installed into a clean conda environment, you may also need to install pandas, scipy, and wget.

To load a model:

from sequence_models.pretrained import load_model_and_alphabet

model, collater = load_model_and_alphabet('carp_640M')

Available models are

  • carp_600k
  • carp_38M
  • carp_76M
  • carp_640M
  • mif
  • mifst
  • bigcarp_esm1bfinetune
  • bigcarp_esm1bfrozen
  • bigcarp_random

Convolutional autoencoding representations of proteins (CARP)

We make available pretrained CNN protein sequence masked language models of various sizes. All of these have a ByteNet encoder architecture and are pretrained on the March 2020 release of UniRef50 using the same masked language modeling task as in BERT and ESM-1b.

CARP is described in this preprint.

You can also download the weights manually from Zenodo.

To encode a batch of sequences:

seqs = [['MDREQ'], ['MGTRRLLP']]
x = collater(seqs)[0]  # (n, max_len)
rep = model(x)  # (n, max_len, d_model)

CARP also supports computing representations from arbitrary layers and the final logits.

rep = model(x, repr_layers=[0, 2, 32], logits=True)

Compute embeddings in bulk from FASTA

We provide a script that efficiently extracts embeddings in bulk from a FASTA file. A cuda device is optional and will be auto-detected. The following command extracts the final-layer embedding for a FASTA file from the CARP_640M model:

$ python scripts/extract.py carp_640M examples/some_proteins.fasta \
    examples/results/some_proteins_emb_carp_640M/ \
    --repr_layers 0 32 33 logits --include mean per_tok

Directory examples/results/some_proteins_emb_carp_640M/ now contains one .pt file per extracted embedding; use torch.load() to load them. scripts/extract.py has flags that determine what .pt files are included:

--repr-layers (default: final only) selects which layers to include embeddings from. 0 is the input embedding. logits is the per-token logits.

--include specifies what embeddings to save. You can use the following:

  • per_tok includes the full sequence, with an embedding per amino acid (seq_len x hidden_dim).
  • mean includes the embeddings averaged over the full sequence, per layer (only valid for representations).
  • logp computes the average log probability per sequence and stores it in a csv (only valid for logits).

scripts/extract.py also has --batchsize and --device flags. For example, to use GPU 2 on a multi-GPU machine, pass --device cuda:2. The default is to use a batchsize of 1 and cpu if cuda is not detected or cuda:0 if cuda is detected.

Masked Inverse Folding (MIF) and Masked Inverse Folding with Sequence Transfer (MIF-ST)

We make available pretrained masked inverse folding models with and without sequence pretraining transfer from CARP-640M.

MIF and MIF-ST are described in this preprint

You can also download the weights manually from Zenodo.

To encode a sequence with its structure:

from sequence_models.pdb_utils import parse_PDB, process_coords
coords, wt, _ = parse_PDB('examples/gb1_a60fb_unrelaxed_rank_1_model_5.pdb')
coords = {
        'N': coords[:, 0],
        'CA': coords[:, 1],
        'C': coords[:, 2]
    }
dist, omega, theta, phi = process_coords(coords)
batch = [[wt, torch.tensor(dist, dtype=torch.float),
          torch.tensor(omega, dtype=torch.float),
          torch.tensor(theta, dtype=torch.float), torch.tensor(phi, dtype=torch.float)]]
src, nodes, edges, connections, edge_mask = collater(batch)
# can use result='repr' or result='logits'. Default is 'repr'.
rep = model(src, nodes, edges, connections, edge_mask)  

Compute embeddings in bulk from csv

We provide a script that efficiently extracts embeddings in bulk from a csv file. A cuda device is optional and will be auto-detected. The following command extracts the final-layer embedding for a FASTA file from the mifst model:

$ python scripts/extract_mif.py mifst examples/gb1s.csv \
    examples/ \
    examples/results/some_proteins_mifst/ \
    repr --include mean per_tok

Directory examples/results/some_proteins_mifst/ now contains one .pt file per extracted embedding; use torch.load() to load them. scripts/extract_mif.py has flags that determine what .pt files are included:

The syntax is:

$ python script/extract_mif.py <model> <csv_fpath> <pdb_dir> <out_dir> <result> --include <pooling options>

The input csv should have columns for name, sequence, and pdb. The script looks in pdb_dir for the filenames in the pdb column.

The options for result are repr or logits.

--include specifies what embeddings to save. You can use the following:

  • per_tok includes the full sequence, with an embedding per amino acid (seq_len x hidden_dim).
  • mean includes the embeddings averaged over the full sequence, per layer (only valid for representations).
  • logp computes the average log probability per sequence and stores it in a csv (only valid for logits).

scripts/extract.py also has a --device flags. For example, to use GPU 2 on a multi-GPU machine, pass --device cuda:2. The default is to use cpu if cuda is not detected or cuda:0 if cuda is detected.

Biosynthetic gene cluster CARP (BiGCARP)

We make available pretrained CNN Pfam domain masked language models of BGCs. All of these have a ByteNet encoder architecture and are pretrained on antiSMASH using the same masked language modeling task as in BERT and ESM-1b.

BiGCARP is described in this preprint. Training code is available here.

You can also download the weights and datasets manually from Zenodo.

To encode a batch of sequences:

bgc = [['#;PF07690;PF06609;PF00083;PF00975;PF12697;PF00550;PF14765'],
       ['t3pks;PF07690;PF06609;PF00083;PF00975;PF12697;PF00550;PF14765;PF00698']]
model, collater = load_model_and_alphabet('bigcarp_esm1bfinetune')
x = collater(bgc)[0]
rep = model(x)

Sequence Datasets and Dataloaders

In sampler.py, you will find two Pytorch sampler classes: SortishSampler, a sampler to sort similarly length sample sequences into length-defined buckets; and ApproxBatchSampler, a batch sampler which grabs sequences from length-defined buckets until the batch has the set approximate max number of tokens or max number of tokens squared.

from sequence_models.samplers import SortishSampler, ApproxBatchSampler

# grab datasets
ds = dataset # your sequence dataset

# build dataloaders
len_ds = np.array([len(i[0]) for i in ds]) # list of lengths of the sequence in dataset (in order)
bucket_size = 1000 # number of length-defined buckets
max_tokens = 8000 # max number of tokens per batch
max_batch_size = 100 # max number of samples per batch
sortish_sampler = SortishSampler(len_ds, bucket_size)
batch_sampler = ApproxBatchSampler(sortish_sampler, max_tokens, max_batch_size, len_ds)
collater = collater # your collater function
dl = DataLoader(ds_train, collate_fn=collater, batch_sampler=batch_sampler, num_workers=16)

Pre-implemented Models

  • Struct2SeqDecoder (GNN)

The Struct2SeqDecoder model was adapted from Ingraham et al.. This model uses protein structural information encoded as a graph nodes and edges representing the structural information of each amino acid residue and their relations to each other, respectively.

If you already have node features, edge features, connections between nodes, encoded sequences (src), and edge mask (edge_mask); you can directly use the the Struct2SeqDecoder as demonstrated below:

from sequence_models.constants import trR_ALPHABET
from sequence_models.gnn import Struct2SeqDecoder

num_letters = len(trR_ALPHABET) # length of your amino acid alphabet  
node_features = 10 # number of node features
edge_features = 11 # number of edge features
hidden_dim =  128 # your choice of hidden layer dimension
num_decoder_layers = 3 # your choice of number of decoder layers to use
dropout = 0.1 # dropout used by decoder layer
use_mpnn = False # if True, use MPNN layer, else use Transformer layer for decoder 
direction = 'bidirectional' # direction of information flow/masking: forward, backward or bidirectional 

model = Struct2SeqDecoder(num_letters, node_features, edge_features, hidden_dim,
            num_decoder_layers, dropout, use_mpnn, direction)
out = model(nodes, edges, connections, src, edge_mask)

If you do not have prepared inputs, but have 2d maps representing the distance between residues (dist) and the dihedral angles between residues (omega, theta, and phi), you can use our preprocessing functions to generate nodes, edges, and connections as demonstrated below:

from sequence_models.gnn import get_node_features, get_k_neighbors, get_edge_features, \
    get_mask, replace_nan

# process features
node = get_node_features(omega, theta, phi) # generate nodes
dist = dist.fill_diagonal_(np.nan) # if the diagonal of dist tensor is not already filled with nans, it should 
                                    # to prevent selecting self when getting k nearest residues in the next step 
connections = get_k_neighbors(dist, n_connections) # get connections
edge = get_edge_features(dist, omega, theta, phi, connections) # generate edge
edge_mask = get_mask(edge) # get edge mask (in the scenario where there is missing edge features between neighbors)
edge = replace_nan(edge) # replace nans with 0s 
node = replace_nan(node) 

Alternatively, we have also prepared StructureCollater, a collater function found in collaters.py that also performs this task:

from sequence_models.collaters import StructureCollater

n_connections = 20 # number of connections per amino acid residue  
collater = StructureCollater(n_connections=n_connections)
ds = dataset # Dataset must return sequences, dists, omegas, thetas, phis 
dl = Dataloader(ds, collate_fn=collater)
  • ByteNet

The ByteNet model was adapted from Kalchbrenner et al.. ByteNet uses stacked convolutional encoder and decoder layers to preserve temporal resolution of sequential data.

from sequence_models.convolutional import ByteNet
from sequence_models.constants import trR_ALPHABET

n_tokens = len(trR_ALPHABET) # number of tokens in token dictionary
d_embedding = 128 # dimension of embedding
d_model = 128 # dimension to use within ByteNet model, //2 every layer
n_layers = 3 # number of layers of ByteNet block
kernel_size = 3 # the kernel width
r = ??? # used to calculate dilation factor
padding_idx = trR_ALPHABET.index('-') # location of padding token in ordered alphabet
causal = True # if True, chooses MaskedCausalConv1d() over MaskedConv1d()
dropout = 0.1 

x = torch.randn(32, 128) # input (n samples, len of seqs) 
input_mask = torch.ones(32, 128, 1) # mask (n samples, len of seqs, 1)
model = ByteNet(n_tokens, d_embedding, d_model, n_layers, kernel_size, r, 
            padding_idx=padding_idx, causal=causal, dropout=dropout)
out = model(x, input_mask) 

We have also an implemented versions of ByteNet to be able to use 2d inputs (ByteNet2d) and as a language model (ByteNetLM):

from sequence_models.convolutional import ByteNet2d, ByteNetLM

x = torch.randn(32, 128, 128, 64) # input (n samples, len of seqs, len of seqs, feature dimension)
input_mask = torch.ones # (n samples, len of seqs, len of seqs, 1), optional
model = ByteNet2d(d_in, d_model, n_layers, kernel_size, r, dropout=0.0)
out = model(x, input_mask)

x = torch.randn(32, 128) # input (n samples, len of seqs) 
input_mask = torch.ones(32, 128, 1) # mask (n samples, len of seqs, 1)
model = ByteNetLM(n_tokens, d_embedding, d_model, n_layers, kernel_size, r,
                    padding_idx=None, causal=False, dropout=0.0)
out = model(x, input_mask)
  • trRosetta The trRosetta model was implemented according to Yang et al.. In this model, multiple sequence alignments (MSAs) are used to predict distances between amino acid residues as well as their dihedral angles (omega, theta, phi). Predictions are in the format of bins. Omega, theta and phi angle are binned into 24, 24, and 12 bins, respectively with 15 degrees segments and one no-contact bin. Yang et al. has pretrained five models (model ids: 'a', 'b', 'c', 'd', 'e'). To run a single model:
from sequence_models.trRosetta_utils import trRosettaPreprocessing, parse_a3m
from sequence_models.trRosetta import trRosetta
from sequence_models.constants import trR_ALPHABET

msas = parse_a3m(path_to_msa) # load in msas in a3m format
alphabet = trR_ALPHABET # load your alphabet order
tr_preprocessing = trRosettaPreprocessing(alphabet) # setup preprocessor for msa
msas_processed = tr_preprocessing.process(msas)

n2d_layers = 61 # keep at 61 if you want to use pretrained version
model_id = 'a' # choose your pretrained model id
decoder = True # if True, return 2d structure maps, else returns hidden layer
p_dropout = 0.0
model = trRosetta(n2d_layers, model_id, decoder, p_dropout)
out = model(msas_processed) # returns dist_probs, theta_probs, phi_probs, omega_probs

To run an ensemble of models:

from sequence_models.trRosetta_utils import trRosettaPreprocessing, parse_a3m
from sequence_models.trRosetta import trRosetta, trRosettaEnsemble
from sequence_models.constants import trR_ALPHABET

msas = parse_a3m(path_to_msa) # load in msas in a3m format
alphabet = trR_ALPHABET # load your alphabet order
tr_preprocessing = trRosettaPreprocessing(alphabet) # setup preprocessor for msa
msas_processed = tr_preprocessing.process(msas)

n2d_layers = 61 # keep at 61 if you want to use pretrained version
model_ids = 'abcde' # choose your pretrained model id
decoder = True # if True, return 2d structure maps, else returns hidden layer
p_dropout = 0.0
base_model = trRosetta
model = trRosettaEnsemble(base_model, n2d_layers, model_ids)
out = model(msas_processed)

If you would like to convert bin prediction into actual values, use probs2value. Here is an example of converting distance bin predictions into values:

from sequence_models.trRosetta_utils import probs2value

dist_probs, theta_probs, phi_probs, omega_probs  = model(x) # structure predictions (batch, # of bins, len of seq, len of seq)
preperty = 'dist' # choose between 'dist', 'theta', 'phi', or 'omega'
mask = mask # your 2d mask (batch, len of seq, len of seq)
dist_values = probs2value(dist, property, mask):