Recent Releases of torchdrug

torchdrug - 0.2.1 Release

The new 0.2.1 release supports PyTorch 2.0 and Python 3.10. We fix quite a few bugs based on the suggestions from the community. Thanks everyone for contributing to this library.

Compatibility Changes

TorchDrug now supports Python versions from 3.7 to 3.10, and PyTorch versions from 1.8 to 2.0. There is no change in the minimal versions, so you can easily update your previous environment to use TorchDrug 0.2.1.
For PropertyPrediction , we change the its predict function to directly output original values rather than standardized values. This is more intuitive for new users. Note this change is not backward compatible. (#109, thanks to @kanojikajino)

Improvements

Add batch normalization and dropout in PropertyPrediction
Support custom edge feature function in GraphConstruction
Support ESM-2 models in EvolutionaryScaleModeling
Add full batch evaluation for KnowledgeGraphCompletion
Support dict, list and tuple in config dict
Add instructions for installation on Apple silicon (#176, thanks to @migalkin)
Reduce dependency to matplotlib-base (#141, thanks to @jamesmyatt)

Bug Fixes

Fix variable names in NeuralLogicProgramming (#126)
Fix interface for the new esm library (#133)
Fix the version of AlphaFoldDB to v2 (#137, thanks to @ShoufaChen)
Fix inconsistent output when using edge features in convolution layers (#53, #140)
Handle side cases in property optimization (#125, thanks to @jannisborn)
Fix a bug when using LR schedulers (#148, #152)
Fix a bug in graph construction (#158)
Fix a bug in layers.Set2Set (#185)
Avoid in-place RDKit operations in data.Molecule.from_molecule (#142)
Fix num_class in PropertyPrediction (#142)
Fix docker file for new CUDA image (#207, thanks to @cscandore)
Fix chain ID in data.Protein.from_molecule

Deprecations

Deprecate functional._size_to_index. Use torch.repeat_interleave instead.

- Python
Published by KiddoZhu almost 3 years ago

V0.2.0 is a major release with a new family member TorchProtein, a library for machine-learning-guided protein science. Aiming at simplifying the development of protein methods, TorchProtein encapsulates many complicated yet repetitive subroutines into functional modules, including widely-used datasets, flexible data processing operations, advanced encoding models, and diverse protein tasks.

Such comprehensive encapsulation enables users to develop protein machine learning solutions with one easy-to-use library. It avoids the embarrassment of gluing multiple libraries into a pipeline.

With TorchProtein, we can rapidly prototype machine learning solutions to various protein applications within 20 lines of codes, and conduct ablation studies by substituting different parts of the solution with off-the-shelf modules. Furthermore, we can easily adapt these modules to our own needs, and make systematic analyses by comparing the new results to a benchmark provided in the library.

Additionally, TorchProtein is designed to be accessible to everyone. For inexperienced users, like beginners or biological researchers, TorchProtein provides user-friendly APIs to simplify the development of protein machine learning solutions. Meanwhile, for professional users, TorchProtein also preserves enough flexibility to satisfy their demands, supported by features like modular design of the library and on-the-fly graph construction.

Main Features

Simplify Data Processing

It is challenging to transform raw bioinformatic protein datasets into tensor formats for machine learning. To reduce tedious operations, TorchProtein provides us with a data structure data.Protein and its batched extension data.PackedProtein to automate the data processing step.
- data.Protein and data.PackedProtein automatically gather protein data from various bio-sources and seamlessly switch between data formats like pdb files, RDKit objects and sequences. Please see the section data structures and operations for transforming from and to sequences and RDKit objects.
  
```python

construct a data.Protein instance from a pdb file

pdbfile = ... protein = data.Protein.frompdb(pdbfile, atomfeature="position", bondfeature="length", residuefeature="symbol") print(protein)

write a data.Protein instance back to a pdb file

newpdbfile = ... protein.topdb(newpdb_file) ```

bash Protein(num_atom=445, num_bond=916, num_residue=57)
- data.Protein and data.PackedProtein automatically pre-process all kinds of features of atoms, bonds and residues, by simply setting up several arguments.
  
```python pdbfile = ... protein = data.Protein.frompdb(pdbfile, atomfeature="position", bondfeature="length", residuefeature="symbol")

feature

print(protein.residuefeature.shape) print(protein.atomfeature.shape) print(protein.bond_feature.shape) ```

bash torch.Size([57, 21]) torch.Size([445, 3]) torch.Size([916, 1])
- data.Protein and data.PackedProtein automatically keeps track of numerous attributes associated with atoms, bonds, residues and the whole protein.
  - For example, reference offers a way to register new attributes as node, edge or graph property, and in this way, the new attributes would automatically go along with the node, edge or graph themself. More in-built attributes are listed in the section data structures and operations.
```python protein = ...

with protein.node(): protein.nodeid = torch.tensor([i for i in range(0, protein.numnode)]) with protein.edge(): protein.edgecost = torch.rand(protein.numedge) with protein.graph(): protein.graph_feature = torch.randn(128) ``- Even more, reference can be utilized to maintain the correspondence between two well related objects. For example, the mappingatom2residue` maintains relationship between atoms and residues, and enables indexing on either of them.

```python protein = ...
  
  create a mask indices for atoms in a glutamine (GLN)
  
  isglutamine = protein.residuetype[protein.atom2residue] == protein.residue2id["GLN"] maskindices = isglutamine.nonzero().squeeze(-1) print(mask_indices)
  
  map the masked atoms back to the glutamine residue
  
  residuetype = protein.residuetype[protein.atom2residue[maskindices]] print([protein.id2residue[r] for r in residuetype.tolist()]) ```
  
  bash tensor([ 26, 27, 28, 29, 30, 31, 32, 33, 34, 307, 308, 309, 310, 311, 312, 313, 314, 315, 384, 385, 386, 387, 388, 389, 390, 391, 392]) ['GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN', 'GLN']
It is useful to augment protein data by modifying protein graphs or constructing new ones. With the protein operations and the graph construction layers provided in TorchProtein,
- we can easily modify proteins on the fly by batching, slicing sequences, masking out side chains, etc. Please see the tutorials for more details on masking.
  
```python pdbfile = ... protein = data.Protein.frompdb(pdbfile, atomfeature="position", bondfeature="length", residuefeature="symbol")

batch

proteins = data.Protein.pack([protein, protein, protein])

slice sequences

use indexing to extract consecutive residues of a particular protein

tworesidues = protein[[0,2]] tworesidues.visualize() ```
- we can construct protein graphs on the fly with GPU acceleration, which offers users flexible choices rather than using fixed pre-processed graphs. Below is an example to build a graph with only alpha carbon atoms, please check tutorials for more cases, such as adding spatial / KNN / sequential edges.
  
```python protein = ...

transfer from CPU to GPU

protein = protein.cuda() print(protein)

build a graph with only alpha carbon (CA) atoms

nodelayers = [geometry.AlphaCarbonNode()] graphconstructionmodel = layers.GraphConstruction(nodelayers=node_layers)

originalprotein = data.Protein.pack([protein]) CAprotein = graphconstructionmodel(protein) print("Graph before:", originalprotein) print("Graph after:", CA_protein) ```

bash Protein(num_atom=445, num_bond=916, num_residue=57, device='cuda:0') Graph before: PackedProtein(batch_size=1, num_atoms=[2639], num_bonds=[5368], num_residues=[350]) Graph after: PackedProtein(batch_size=1, num_atoms=[350], num_bonds=[0], num_residues=[350])

Easy to Prototype Solutions

With TorchProtein, common protein tasks can be finished within 20 lines of codes, such as sequence-based protein property prediction task. Below is an example and more examples of different popular protein tasks and models can be found in Protein Tasks, Models and Tutorials.

```python import torch from torchdrug import datasets, transforms, models, tasks, core

truncatetransform = transforms.TruncateProtein(maxlength=200, random=False) proteinviewtransform = transforms.ProteinView(view="residue") transform = transforms.Compose([truncatetransform, proteinview_transform])

dataset = datasets.BetaLactamase("~/protein-datasets/", residueonly=True, transform=transform) trainset, validset, testset = dataset.split()

model = models.ProteinCNN(inputdim=21, hiddendims=[1024, 1024], kernel_size=5, padding=2, readout="max")

task = tasks.PropertyPrediction(model, task=dataset.tasks, criterion="mse", metric=("mae", "rmse", "spearmanr"), normalization=False, nummlplayer=2)

optimizer = torch.optim.Adam(task.parameters(), lr=1e-4) solver = core.Engine(task, trainset, validset, testset, optimizer, gpus=[0], batchsize=64) solver.train(num_epoch=10) solver.evaluate("valid") ```

bash mean absolute error [scaled_effect1]: 0.249482 root mean squared error [scaled_effect1]: 0.304326 spearmanr [scaled_effect1]: 0.44572

Compatible with Existing Molecular Models in TorchDrug

TorchProtein follows the scientific fact that proteins are macromolecules. The core data structures data.Protein and data.PackedProtein inherit from data.Molecule and data.PackedMolecule respectively. Therefore, we can apply any existing molecule model in TorchDrug to proteins

```python import torch from torchdrug import layers, datasets, transforms, models, tasks, core from torchdrug.layers import geometry

truncatetransform = transforms.TruncateProtein(maxlength=200, random=False) proteinviewtransform = transforms.ProteinView(view="residue") transform = transforms.Compose([truncatetransform, proteinview_transform])

dataset = datasets.EnzymeCommission("~/protein-datasets/", transform=transform) trainset, validset, test_set = dataset.split()

model = models.GIN(inputdim=21, hiddendims=[256, 256, 256, 256], batchnorm=True, shortcut=True, concat_hidden=True)

graphconstructionmodel = layers.GraphConstruction( nodelayers=[geometry.AlphaCarbonNode()], edgelayers=[geometry.SpatialEdge(radius=10.0, mindistance=5), geometry.KNNEdge(k=10, mindistance=5), geometry.SequentialEdge(maxdistance=2)], edgefeature="residue_type" )

task = tasks.MultipleBinaryClassification(model, graphconstructionmodel=graphconstructionmodel, nummlplayer=3, task=list(range(len(dataset.tasks))), criterion="bce", metric=("auprc@micro", "f1_max"))

optimizer = torch.optim.Adam(task.parameters(), lr=1e-4) solver = core.Engine(task, trainset, validset, testset, optimizer, gpus=[0], batchsize=4) solver.train(num_epoch=10) solver.evaluate("valid") ```

bash auprc@micro: 0.187884 f1_max: 0.231008
In Protein-Ligand Interaction (PLI) prediction task, we can utilize a molecular encoder module to extract the representations of molecules. Please check tutorial 2 for more details.

```python trainset, validset, test_set = ...

protein encoder

model = models.ProteinCNN(inputdim=21, hiddendims=[1024, 1024], kernel_size=5, padding=2, readout="max")

molecule encoder

model2 = models.GIN(inputdim=66, hiddendims=[256, 256, 256, 256], batchnorm=True, shortcut=True, concat_hidden=True)

task = tasks.InteractionPrediction(model, model2=model2, task=dataset.tasks, criterion="mse", metric=("mae", "rmse", "spearmanr"), normalization=False, nummlplayer=2)

optimizer = torch.optim.Adam(task.parameters(), lr=1e-4) solver = core.Engine(task, trainset, validset, testset, optimizer, gpus=[0], batchsize=16) solver.train(num_epoch=5) solver.evaluate("valid") ```

bash mean absolute error [scaled_effect1]: 0.249482 root mean squared error [scaled_effect1]: 0.304326 spearmanr [scaled_effect1]: 0.44572

Support From the Developer (@DeepGraphLearning/torchdrug-maintainers)

There is always an active supporting team to answer questions and provide helps. Feedbacks of use experience and contributions for development are welcomed.

New Modules

Data Structures and Operations

`data.Protein`

Representative attributes:
- data.Protein.edge_list: list of edges and each edge is represented by a tuple (nodein, nodeout, bond_type)
- data.Protein.atom_type: atom types
- data.Protein.bond_type: bond types
- data.Protein.residue_type: residue types
- data.Protein.view: default view for this protein. Can be “atom” or “residue”
- data.Protein.atom_name: atom names in each residue
- data.Protein.atom2residue: atom id to residue id mapping
- data.Protein.is_hetero_atom: hetero atom indicator
- data.Protein.occupancy: protein occupancy
- data.Protein.b_factor: temperature factors
- data.Protein.residue_number: residue numbers
- data.Protein.insertion_code: insertion codes
- data.Protein.chain_id: chain ids
Representative Methods:
- data.Protein.from_molecule: create a protein from an RDKit object.
- data.Protein.from_sequence: create a protein from a sequence.
- data.Protein.from_sequence_fast: a faster version of creating a protein from a sequence.
- data.Protein.from_pdb: create a protein from a PDB file.
- data.Protein.to_molecule: return an RDKit object of this protein.
- data.Protein.to_sequence: return a sequence of this protein.
- data.Protein.to_pdb: write this protein to a pdb file.
- data.Protein.split: split this protein graph into multiple disconnected protein graphs.
- data.Protein.pack: batch a list of data.Protein into data.PackedProtein.
- data.Protein.repeat: repeat this protein.
- data.Protein.residue2atom: map residue id to atom ids.
- data.Protein.residue_mask: return a masked protein based on the specified residues.
- data.Protein.subresidue: return a subgraph based on the specified residues.
- data.Protein.residue2graph: residue id to protein id mapping.
- data.Protein.node_mask: return a masked protein based on the specified nodes.
- data.Protein.edge_mask: return a masked protein based on the specified edges.
- data.Protein.compact: remove isolated nodes and compact node ids.

`data.PackedProtein`

Representative attributes:
- data.PackedProtein.edge_list: list of edges and each edge is represented by a tuple (nodein, nodeout, bond_type)
- data.PackedProtein.atom_type: atom types
- data.PackedProtein.bond_type: bond types
- data.PackedProtein.residue_type: residue types
- data.PackedProtein.view: default view for this protein. Can be “atom” or “residue”
- data.PackedProtein.num_nodes: number of nodes in each protein graph
- data.PackedProtein.num_edges: number of edges in each protein graph
- data.PackedProtein.num_residues: number of residues in each protein graph
- data.PackedProtein.offsets: node id offsets in different proteins
Representative methods:
- data.PackedProtein.node_mask: return a masked packed protein based on the specified nodes.
- data.PackedProtein.edge_mask: return a masked packed protein based on the specified edges.
- data.PackedProtein.residue_mask: return a masked packed protein based on the specified residues.
- data.PackedProtein.graph_mask: return a masked packed protein based on the specified protein graphs.
- data.PackedProtein.from_molecule: create a protein from a list of RDKit objects.
- data.PackedProtein.from_sequence: create a protein from a list of sequences.
- data.PackedProtein.from_sequence_fast: a faster version of creating a protein from a list of sequences.
- data.PackedProtein.from_pdb: create a protein from a list of PDB files.
- data.PackedProtein.to_molecule: return a list of RDKit objects of this packed protein.
- data.PackedProtein.to_sequence: return a list of sequences of this packed protein.
- data.PackedProtein.to_pdb: write this packed protein to a list of pdb files.
- data.PackedProtein.merge: merge multiple packed proteins into a single packed protein.
- data.PackedProtein.repeat: repeat this packed protein.
- data.PackedProtein.repeat_interleave: repeat this packed protein, behaving similarly to torch.repeat_interleave_.
- data.PackedProtein.residue2graph: residue id to graph id mapping.

Models

GearNet: Geometry Aware Relational Graph Neural Network.
ESM: Evolutionary Scale Modeling (ESM).
ProteinCNN: protein shallow CNN.
ProteinResNet: protein ResNet.
ProteinLSTM: protein LSTM.
ProteinBERT: protein BERT.
Statistic: the statistic feature engineering for protein sequence.
Physicochemical: the physicochemical feature engineering for protein sequence.

Protein Tasks

Sequence-based Protein Property Prediction:

tasks.PropertyPrediction predicts some property of each protein, such as Beta-lactamase activity, stability and solubility for proteins.
tasks.NodePropertyPrediction predicts some property of each residue in proteins, such as the secondary structure (coil, strand or helix) of each residue.
tasks.ContactPrediction predicts whether any pair of residues contact or not in the folded structure.
tasks.InteractionPrediction predicts the binding affinity of two interacting proteins or of a protein and a ligand, i.e. performing PPI affinity prediction or PLI affinity prediction.

Structure-based Protein Property Prediction:

tasks.MultipleBinaryClassification predicts whether a protein owns several specific functions or not with binary labels.

Pre-trained Protein Structure Representations:

Self-Supervised Protein Structure Pre-training: acquires informative protein representations from massive unlabeled protein structures, such as tasks.EdgePrediction, tasks.AttributeMasking, tasks.ContextPrediction, tasks.DistancePrediction, tasks.AnglePrediction, tasks.DihedralPrediction .
Fine-tuning on Downstream Task: fine-tunes the pre-trained protein encoder on downstream tasks, such as any property prediction task mentioned above.

Protein Datasets

Protein Property Prediction Datasets

BetaLactamase : protein sequences with activity labels
Fluorescence: protein sequences with fitness labels
Stability: protein sequences with stability labels
Solubility: protein sequences with solubility labels
BinaryLocalization: protein sequences with membrane-bound or soluble labels
SubcellularLocalization: protein sequences with natural cell location labels
EnzymeCommission: protein sequences and 3D structures with EC number labels for catalysis in biochemical reactions
GeneOntology: protein sequences and 3D structures with GO term labels, including molecular function (MF), biological process (BP) and cellular component (CC)
AlphaFoldDB: protein sequences and 3D structures predicted by AlphaFold

Protein Structure Prediction Datasets

Fold: protein sequences and 3D structures with fold labels determined by the global structural topology
SecondaryStructure: protein sequences and 3D structures with secondary structure labels determined by the local structures
ProteinNet: protein sequences and 3D structures for the contact prediction task

Protein-Protein Interaction Prediction Datasets

HumanPPI: protein sequences with binary interaction labels for human proteins
YeastPPI: protein sequences with binary interaction labels for yeast proteins
PPIAffinity: protein sequences with binding affinity values measured by $p{Kd}$

Protein Ligand Interaction Prediction Datasets

BindingDB: protein sequences and molecule graphs with binding affinity between pairs of protein and ligand
PDBBind: protein sequences and molecule graphs with binding affinity between pairs of protein and ligand

Data Transform Modules

TruncateProtein: truncate over long protein sequences into a fixed length
ProteinView: convert proteins to a specific view

Graph Construction Layers

SubsequenceNode: take a protein subsequence of a specific length
SubspaceNode: extract a subgraph by only keeping neighboring nodes in a spatial ball for each centered node
RandomEdgeMask: mask out some edges randomly from the protein graph

Tutorials

To help users gain a comprehensive understanding of TorchProtein, we recommend some user-friendly tutorials for its basic usage and examples to various protein-related tasks. These tutorials may also serve as boilerplate codes for users to develop their own applications.

Tutorial 1 - Protein Data Structure for basic usage of TorchProtein, like how to represent proteins and what operations are feasible.
Tutorial 2 - Sequence-based Protein Property Prediction for 5 types of protein sequence understanding tasks provided in TorchProtein.
Tutorial 3 - Structure-based Protein Property Prediction for property prediction tasks based on protein structure representations.
Tutorial 4 - Pre-trained Protein Structure Representations for self-supervised pre-training of protein structure encoders and its fine-tuning on downstream tasks.

Bug Fixes

Fix an error in the decorator @utils.cached (#118)
Fix an index error in data.Graph.split() (#115)
Fix setting attribute node_feature , edge_feature and graph_feature (#116)
Fix incorrect node feature shape for the synthon dataset USPTO50k (#116)
Fix a compatible issue when adding node/edge/graph reference and changing node/edge to atom/bond (#116, #117)

- Python
Published by KiddoZhu almost 4 years ago

torchdrug - 0.1.3 Release

TorchDrug 0.1.3 release introduces new features like W&B intergration and index reference. It also provides new functions and metrics for common development need. Note 0.1.3 has some compatibility changes and be careful when you update your TorchDrug from an older version.

W&B Integration
Index Reference
New Functions
New Metrics
Improvements
Bug Fixes
Compatibility Changes

W&B Integration

Tracking experiment progress is one of the most important demand from ML researchers and developers. For TorchDrug users, we provide a native integration of W&B platform. By adding only one argument in core.Engine, TorchDrug will automatically copy every hyperparameter and training log to your W&B database (thanks to @manangoel99).

python solver = core.Engine(task, train_set, valid_set, test_set, optimizer, logger="wandb")

Now you can track your training and validation performance in your browser, and compare them across different experiments.

Wandb demo

Index Reference

Maintaining node and edge attributes could be painful when one applies a lot of transformations to a graph. TorchDrug aims to eliminate such tedious steps by registering custom attributes. This update extends the capacity of custom attributes to index reference. That means, we allow attributes to refer to indexes of nodes, edges or graphs, and they will be automatically maintained in any graph operation.

To use index reference, simply add a context manager when we define the attributes.

python with graph.edge(), graph.edge_reference(): graph.inv_edge_index = torch.tensor(inv_edge_index)

Foor more details on index reference, please take a look at our notes. Typical use cases include

A pointer to the inverse edge of each edge.
A pointer to the parent node of each node in a tree.
A pointer to the incoming tree edge of each node in a DFS.

Let us know if you find more interesting usage of index reference!

New Functions

Message passing over line graphs is getting more and more popular in the recent years. This version provides data.Graph.line_graph to efficiently construct line graphs on GPUs. It supports both a single graph or a batch of graphs.

We are constantly focusing on better batching of irregular structures, and the variadic functions in TorchDrug are an efficient way to process batch of variadic-sized tensors without padding. This update introduces 3 new variadic functions.

variadic_meshgrid generates a meshgrid from two variadic tensors. Useful for implementing pairwise operations.
variadic_to_padded converts a variadic tensor to a padded tensor.
padded_to_variadic converts a padded tensor to a variadic tensor.

New Metrics

New metrics include accuracy, matthews_corrcoef, pearsonr, spearmanr. All the metrics are the same as their counterparts in scipy, but they are implemented in PyTorch and support auto differentiation.

Improvements

Add data.Graph.to (#70, thanks to @cthoyt)
Extend tasks.SynthonCompletion for arbitrary atom features (#62)
Speed up lazy data loading (#58, thanks to @wconnell)
Speed up rspmm cuda kernels
Add docker support
Add more documentation for data.Graph and data.Molecule

Bug Fixes

Fix computation of output dimension in several GNNs (#92, thanks to @kanojikajino)
Fix data.PackedGraph.__getitem__ when the batch is empty (#83, thanks to @jannisborn)
Fix patched modules for PyTorch>=1.6.0 (#77)
Fix make_configurable for torch.utils.data (#85)
Fix multi_slice_mask, variadic_max for multi-dimensional input
Fix variadic_topk for input containing infinite values

Compatibility Changes

TorchDrug now supports Python 3.7/3.8/3.9. Starting from this version, TorchDrug requires a minimal PyTorch version of 1.8.0 and a minimal RDKit version of 2020.09.

Argument node_feature and edge_feature are renamed to atom_feature and bond_feature in data.Molecule.from_smiles and data.Molecule.from_molecule. The old interface is still supported with deprecated warnings.

- Python
Published by KiddoZhu about 4 years ago

torchdrug - 0.1.2 Release

0.1.2 Release Notes

The recent 0.1.2 release of TorchDrug is an update on Colab tutorials, data structures, functions, datasets and bug fixes. We are grateful to see growing interests and involvement from the community, especially on the retrosynthesis task. Welcome more in the future!

Colab Tutorials
New Data Structures
New Functions
New Datasets
Bug Fixes

Colab Tutorials

To familiarize users with the logic and capacity of TorchDrug, we compile a full set of Colab tutorials, covering from basic usage to different drug discovery tasks. All the tutorials are fully interactive and may serve as boilerplate code for your own applications.

Basic Usage and Pipeline shows the manipulation of data structures like data.Graph and data.Molecule, as well as the training and evaluation pipelines for property prediction models.
Pretrained Molecular Representations demonstrates the steps for self-supervised pretraining of a molecular representation model and finetuning it on downstream tasks.
De novo Molecule Design illustrates the routine of training generative models for molecule generation and finetuning them with reinforcement learning for property optimization. Two popular models, GCPN and GraphAF, are covered in the tutorial.
Retrosynthesis shows how to use the state-of-the-art model, G2Gs, to predict a set reactants for synthesizing a target molecule.
Knowledge Graph Reasoning goes through the steps of training and evaluating models for knowledge graph completion, including both knowledge graph embeddings and neural inductive logic programming.

New Data Structures

A new data structure data.Dictionary that stores key-value mapping of PyTorch tensors on either CPUs or GPUs. It enjoys O(n) memory consumption and O(1) query time, and supports parallelism over batch of queries. This API provides a great opportunity for implementing sparse lookup tables or set operations in a PyTorchic style.
A new method data.Graph.match to efficiently retrieve all edges of specific patterns on either CPUs or GPUs. It scales linearly w.r.t. the number of patterns plus the number of retrieved edges, regardless the size of the graph. Typical usage of this method includes querying the existence of edges, generating random walks or even extracting ego graphs.

New Functions

Batching irregular structures, such as graphs, sets or sequences with different sizes, is a common demand in drug discovery. Instead of clumsy padding-based implementation, TorchDrug provides a family of functions that efficiently manipulate batch of variadic-sized tensors without padding. The update contains the following new variadic functions.

variadic_arange returns a 1-D tensor that contains integer intervals of variadic sizes.
variadic_softmax computes softmax over categories with variadic sizes.
variadic_sort sorts elements in sets with variadic sizes.
variadic_randperm returns random permutations for sets with variadic sizes, where the i-th permutation contains integers from 0 to size[i] - 1.
variadic_sample draws samples with replacement from sets with variadic sizes.

New Datasets

PCQM4M: A large-scale molecule property prediction dataset, originally used in OGB-LSC (thanks to @OPAYA )

Bug Fixes

Fix import of sascorer in plogp evaluation (#18, #31)
Fix atoms with stereo bonds in retrosynthesis (#42, #43)
Fix lazy construction for molecule datasets (#30, thanks to @DaShenZi721 )
Fix ChEMBLFiltered dataset (#36)
Fix ZINC2m dataset (#33)
Fix USPTO50k dataset (#32)
Fix bugs in core.Configurable (#26)
Fix/improve documentation (#16, #28, #41)
Fix installation on macOS (#29)

- Python
Published by KatarinaYuan over 4 years ago

Recent Releases of torchdrug

torchdrug - 0.2.1 Release

Compatibility Changes

Improvements

Bug Fixes

Deprecations

torchdrug - 0.2.0 Release

Main Features

Simplify Data Processing

construct a data.Protein instance from a pdb file

write a data.Protein instance back to a pdb file

feature

create a mask indices for atoms in a glutamine (GLN)

map the masked atoms back to the glutamine residue

batch

slice sequences

use indexing to extract consecutive residues of a particular protein

transfer from CPU to GPU

build a graph with only alpha carbon (CA) atoms

Easy to Prototype Solutions

Compatible with Existing Molecular Models in TorchDrug

protein encoder

molecule encoder

Support From the Developer (@DeepGraphLearning/torchdrug-maintainers)

New Modules

Data Structures and Operations

data.Protein

data.PackedProtein

Models

Protein Tasks

Sequence-based Protein Property Prediction:

Structure-based Protein Property Prediction:

Pre-trained Protein Structure Representations:

Protein Datasets

Protein Property Prediction Datasets

Protein Structure Prediction Datasets

Protein-Protein Interaction Prediction Datasets

Protein Ligand Interaction Prediction Datasets

Data Transform Modules

Graph Construction Layers

Tutorials

Bug Fixes

torchdrug - 0.1.3 Release

W&B Integration

Index Reference

New Functions

New Metrics

Improvements

Bug Fixes

Compatibility Changes

torchdrug - 0.1.2 Release

0.1.2 Release Notes

Colab Tutorials

New Data Structures

New Functions

New Datasets

Bug Fixes

`data.Protein`

`data.PackedProtein`