Implementation of equivariant GVP-GNNs as described in Learning from Protein Structure with Geometric Vector Perceptrons by B Jing, S Eismann, P Suriana, RJL Townshend, and RO Dror.

UPDATE: Also includes equivariant GNNs with vector gating as described in Equivariant Graph Neural Networks for 3D Macromolecular Structure by B Jing, S Eismann, P Soni, and RO Dror.

Scripts for training / testing / sampling on protein design and training / testing on all ATOM3D tasks are provided.

Note: This implementation is in PyTorch Geometric. The original TensorFlow code, which is not maintained, can be found here.

• UNIX environment
• python==3.6.13
• torch==1.8.1
• torch_geometric==1.7.0
• torch_scatter==2.0.6
• torch_cluster==1.5.9
• tqdm==4.38.0
• numpy==1.19.4
• sklearn==0.24.1
• atom3d==0.2.1

While we have not tested with other versions, any reasonably recent versions of these requirements should work.

We provide classes in three modules:

• gvp: core GVP modules and GVP-GNN layers
• gvp.data: data pipelines for both general use and protein design
• gvp.models: implementations of MQA and CPD models
• gvp.atom3d: models and data pipelines for ATOM3D

The core modules in gvp are meant to be as general as possible, but you will likely have to modify gvp.data and gvp.models for your specific application, with the existing classes serving as examples.

Installation: Download this repository and run python setup.py develop or pip install . -e. Be sure to manually install torch_geometric first!

Tuple representation: All inputs and outputs with both scalar and vector channels are represented as a tuple of two tensors (s, V). Similarly, all dimensions should be specified as tuples (n_scalar, n_vector) where n_scalar and n_vector are the number of scalar and vector features, respectively. All V tensors must be shaped as [..., n_vector, 3], not [..., 3, n_vector].

Batching: We adopt the torch_geometric convention of absorbing the batch dimension into the node dimension and keeping track of batch index in a separate tensor.

Amino acids: Models view sequences as int tensors and are agnostic to aa-to-int mappings. Such mappings are specified as the letter_to_num attribute of gvp.data.ProteinGraphDataset. Currently, only the 20 standard amino acids are supported.

For all classes, see the docstrings for more detailed usage. If you have any questions, please contact [email protected].

The class gvp.GVP implements a Geometric Vector Perceptron.

import gvp

in_dims = scalars_in, vectors_in
out_dims = scalars_out, vectors_out
gvp_ = gvp.GVP(in_dims, out_dims)

To use vector gating, pass in vector_gate=True and the appropriate activations.

gvp_ = gvp.GVP(in_dims, out_dims,
            activations=(F.relu, None), vector_gate=True)

The classes gvp.Dropout and gvp.LayerNorm implement vector-channel dropout and layer norm, while using normal dropout and layer norm for scalar channels. Both expect inputs and return outputs of form (s, V), but will also behave like their scalar-valued counterparts if passed a single tensor.

dropout = gvp.Dropout(drop_rate=0.1)
layernorm = gvp.LayerNorm(out_dims)

The function gvp.randn returns tuples (s, V) drawn from a standard normal. Such tuples can be directly used in a forward pass.

x = gvp.randn(n=5, dims=in_dims)
# x = (s, V) with s.shape = [5, scalars_in] and V.shape = [5, vectors_in, 3]

out = gvp_(x)
out = drouput(out)
out = layernorm(out)

Finally, we provide utility functions for adding, concatenating, and indexing into such tuples.

y = gvp.randn(n=5, dims=in_dims)
z = gvp.tuple_sum(x, y)
z = gvp.tuple_cat(x, y, dim=-1) # concat along channel axis
z = gvp.tuple_cat(x, y, dim=-2) # concat along node / batch axis

node_mask = torch.rand(5) < 0.5
z = gvp.tuple_index(x, node_mask) # select half the nodes / batch at random

The class GVPConv is a torch_geometric.MessagePassing module which forms messages and aggregates them at the destination node, returning new node embeddings. The original embeddings are not updated.

nodes = gvp.randn(n=5, in_dims)
edges = gvp.randn(n=10, edge_dims) # 10 random edges
edge_index = torch.randint(0, 5, (2, 10), device=device)

conv = gvp.GVPConv(in_dims, out_dims, edge_dims)
out = conv(nodes, edge_index, edges)

The class GVPConvLayer is a nn.Module that forms messages using a GVPConv and updates the node embeddings as described in the paper. Because the updates are residual, the dimensionality of the embeddings are not changed.

layer = gvp.GVPConvLayer(node_dims, edge_dims)
nodes = layer(nodes, edge_index, edges)

The class also allows updates where incoming messages where src >= dst are computed using a different set of source embeddings, as in autoregressive models.

nodes_static = gvp.randn(n=5, in_dims)
layer = gvp.GVPConvLayer(node_dims, edge_dims, autoregressive=True)
nodes = layer(nodes, edge_index, edges, autoregressive_x=nodes_static)

Both GVPConv and GVPConvLayer accept arguments activations and vector_gate to use vector gating.

The class gvp.data.ProteinGraphDataset transforms protein backbone structures into featurized graphs. Following Ingraham, et al, NeurIPS 2019, we use a JSON/dictionary format to specify backbone structures:

[
    {
        "name": "NAME"
        "seq": "TQDCSFQHSP...",
        "coords": [[[74.46, 58.25, -21.65],...],...]
    }
    ...
]

For each structure, coords should be a num_residues x 4 x 3 nested list of the positions of the backbone N, C-alpha, C, and O atoms of each residue (in that order).

import gvp.data

# structures is a list or list-like as shown above
dataset = gvp.data.ProteinGraphDataset(structures)
# dataset[i] is featurized graph corresponding to structures[i]

The returned graphs are of type torch_geometric.data.Data with attributes

• x: alpha carbon coordinates
• seq: sequence converted to int tensor according to attribute self.letter_to_num
• name, edge_index
• node_s, node_v: node features as described in the paper with dims (6, 3)
• edge_s, edge_v: edge features as described in the paper with dims (32, 1)
• mask: false for nodes with any nan coordinates

The gvp.data.ProteinGraphDataset can be used with a torch.utils.data.DataLoader. We supply a class gvp.data.BatchSampler which will form batches based on the number of total nodes in a batch. Use of this sampler is optional.

node_counts = [len(s['seq']) for s in structures]
sampler = gvp.data.BatchSampler(node_counts, max_nodes=3000)
dataloader = torch.utils.data.DataLoader(dataset, batch_sampler=sampler)

The dataloader will return batched graphs of type torch_geometric.data.Batch with an additional batch attibute. The attributes of the Batch will then need to be formed into (s, V) tuples before passing into a GVP-GNN layer or network.

for batch in dataloader:
    batch = batch.to(device) # optional
    nodes = (batch.node_s, batch.node_v)
    edges = (batch.edge_s, batch.edge_v)
    
    out = layer(nodes, batch.edge_index, edges)

We provide two fully specified networks which take in protein graphs and output a scalar prediction for each graph (gvp.models.MQAModel) or a 20-dimensional feature vector for each node (gvp.models.CPDModel), corresponding to the two tasks in our paper. Note that if you are using the unmodified gvp.data.ProteinGraphDataset, node_in_dims and edge_in_dims must be (6, 3) and (32, 1), respectively.

import gvp.models

# batch, nodes, edges as formed above

mqa_model = gvp.models.MQAModel(node_in_dim, node_h_dim, 
                        edge_in_dim, edge_h_dim, seq_in=True)
out = mqa_model(nodes, batch.edge_index, edges,
                 seq=batch.seq, batch=batch.batch) # shape (n_graphs,)

cpd_model = gvp.models.CPDModel(node_in_dim, node_h_dim, 
                        edge_in_dim, edge_h_dim)
out = cpd_model(nodes, batch.edge_index, 
                 edges, batch.seq) # shape (n_nodes, 20)

We provide a script run_cpd.py to train, validate, and test a CPDModel as specified in the paper using the CATH 4.2 dataset and TS50 dataset. If you want to use a trained model on new structures, see the section "Sampling" below.

Run getCATH.sh in data/ to fetch the CATH 4.2 dataset. If you are interested in testing on the TS 50 test set, also run grep -Fv -f ts50remove.txt chain_set.jsonl > chain_set_ts50.jsonl to produce a training set without overlap with the TS 50 test set.

To train a model, simply run python run_cpd.py --train. To test a trained model on both the CATH 4.2 test set and the TS50 test set, run python run_cpd --test-r PATH for perplexity or with --test-p for perplexity. Run python run_cpd.py -h for more detailed options.

$ python run_cpd.py -h

usage: run_cpd.py [-h] [--models-dir PATH] [--num-workers N] [--max-nodes N] [--epochs N] [--cath-data PATH] [--cath-splits PATH] [--ts50 PATH] [--train] [--test-r PATH] [--test-p PATH] [--n-samples N]

optional arguments:
  -h, --help          show this help message and exit
  --models-dir PATH   directory to save trained models, default=./models/
  --num-workers N     number of threads for loading data, default=4
  --max-nodes N       max number of nodes per batch, default=3000
  --epochs N          training epochs, default=100
  --cath-data PATH    location of CATH dataset, default=./data/chain_set.jsonl
  --cath-splits PATH  location of CATH split file, default=./data/chain_set_splits.json
  --ts50 PATH         location of TS50 dataset, default=./data/ts50.json
  --train             train a model
  --test-r PATH       evaluate a trained model on recovery (without training)
  --test-p PATH       evaluate a trained model on perplexity (without training)
  --n-samples N       number of sequences to sample (if testing recovery), default=100

Confusion matrices: Note that the values are normalized such that each row (corresponding to true class) sums to 1000, with the actual number of residues in that class printed under the "Count" column.

To sample from a CPDModel, prepare a ProteinGraphDataset, but do NOT pass into a DataLoader. The sequences are not used, so placeholders can be used for the seq attributes of the original structures dicts.

protein = dataset[i]
nodes = (protein.node_s, protein.node_v)
edges = (protein.edge_s, protein.edge_v)
    
sample = model.sample(nodes, protein.edge_index,  # shape = (n_samples, n_nodes)
                      edges, n_samples=n_samples)

The output will be an int tensor, with mappings corresponding to those used when training the model.

We provide models and dataloaders for all ATOM3D tasks in gvp.atom3d, as well as a training and testing script in run_atom3d.py. This also supports loading pretrained weights for transfer learning experiments.

The GVP-GNNs for ATOM3D are supplied in gvp.atom3d and are named after each task: gvp.atom3d.MSPModel, gvp.atom3d.PPIModel, etc. All of these extend the base class gvp.atom3d.BaseModel. These classes take no arguments at initialization, take in a torch_geometric.data.Batch representation of a batch of structures, and return an output corresponding to the task. Details vary based on the exact task---see the docstrings.

psr_model = gvp.atom3d.PSRModel()

gvp.atom3d also includes data loaders to produce torch_geometric.data.Batch objects from an underlying atom3d.datasets.LMDBDataset. In the case of all tasks except PPI and RES, these are in the form of callable transform objects---gvp.atom3d.SMPTransform, gvp.atom3d.RSRTransform, etc---which should be passed into the constructor of a atom3d.datasets.LMDBDataset:

psr_dataset = atom3d.datasets.LMDBDataset(path_to_dataset,
                    transform=gvp.atom3d.PSRTransform())

On the other hand, gvp.atom3d.PPIDataset and gvp.atom3d.RESDataset take the place of / are wrappers around the atom3d.datasets.LMDBDataset:

ppi_dataset = gvp.atom3d.PPIDataset(path_to_dataset)
res_dataset = gvp.atom3d.RESDataset(path_to_dataset, path_to_split) # see docstring

All datasets must be then wrapped in a torch_geometric.data.DataLoader:

psr_dataloader = torch_geometric.data.DataLoader(psr_dataset, batch_size=batch_size)

The dataloaders can be directly iterated over to yield torch_geometric.data.Batch objects, which can then be passed into the models.

for batch in psr_dataloader:
    pred = psr_model(batch) # pred.shape = (batch_size,)

To run training / testing on ATOM3D, download the datasets as described here. Modify the function get_datasets in run_atom3d.py with the paths to the datasets. Then run:

$ python run_atom3d.py -h

usage: run_atom3d.py [-h] [--num-workers N] [--smp-idx IDX]
                     [--lba-split SPLIT] [--batch SIZE] [--train-time MINUTES]
                     [--val-time MINUTES] [--epochs N] [--test PATH]
                     [--lr RATE] [--load PATH]
                     TASK

positional arguments:
  TASK                  {PSR, RSR, PPI, RES, MSP, SMP, LBA, LEP}

optional arguments:
  -h, --help            show this help message and exit
  --num-workers N       number of threads for loading data, default=4
  --smp-idx IDX         label index for SMP, in range 0-19
  --lba-split SPLIT     identity cutoff for LBA, 30 (default) or 60
  --batch SIZE          batch size, default=8
  --train-time MINUTES  maximum time between evaluations on valset,
                        default=120 minutes
  --val-time MINUTES    maximum time per evaluation on valset, default=20
                        minutes
  --epochs N            training epochs, default=50
  --test PATH           evaluate a trained model
  --lr RATE             learning rate
  --load PATH           initialize first 2 GNN layers with pretrained weights

For example:

# train a model
python run_atom3d.py PSR

# train a model with pretrained weights
python run_atom3d.py PSR --load PATH

# evaluate a model
python run_atom3d.py PSR --test PATH

Portions of the input data pipeline were adapted from Ingraham, et al, NeurIPS 2019. We thank Pratham Soni for portions of the implementation in PyTorch.

@inproceedings{
    jing2021learning,
    title={Learning from Protein Structure with Geometric Vector Perceptrons},
    author={Bowen Jing and Stephan Eismann and Patricia Suriana and Raphael John Lamarre Townshend and Ron Dror},
    booktitle={International Conference on Learning Representations},
    year={2021},
    url={https://openreview.net/forum?id=1YLJDvSx6J4}
}

@article{jing2021equivariant,
  title={Equivariant Graph Neural Networks for 3D Macromolecular Structure},
  author={Jing, Bowen and Eismann, Stephan and Soni, Pratham N and Dror, Ron O},
  journal={arXiv preprint arXiv:2106.03843},
  year={2021}
}