Skip to content

v2.0.0

Choose a tag to compare

View all tags
@nscorley nscorley released this 29 Nov 19:42
· 30 commits to production since this release
v2.0.0
5811631

Performance

  • Parser 2-3x faster: Significant optimizations to structure parsing, especially for symmetric assemblies
  • Cache loading 3-5x faster: Improved pickle/gzip cache handling with 2-level directory sharding for better filesystem performance
  • Vectorized annotations: add_pn_unit_iid_annotation() now uses boolean masks instead of expensive subarray operations (10-100x speedup on symmetric assemblies)

Breaking Changes

Dataset Module Restructuring

The dataset module has been restructured to align with TorchVision/TorchAudio and HuggingFace conventions, using a dataset/loader pattern:

  • Removed dataset.dataset nesting: Datasets are now flat; access data directly from the dataset object
  • MetadataRowParser deprecated: The StructuralDatasetWrapper + dataset_parser pattern is replaced with a loader parameter directly on datasets (backwards-compatible but deprecated)

Migration example:

# Old (deprecated)
from atomworks.ml.datasets import StructuralDatasetWrapper, PandasDataset
from atomworks.ml.datasets.parsers import PNUnitsDFParser

dataset = StructuralDatasetWrapper(
    dataset=PandasDataset(data="df.parquet"),
    dataset_parser=PNUnitsDFParser(...)
)

# New
from atomworks.ml.datasets import PandasDataset
from atomworks.ml.datasets.loaders import create_base_loader

dataset = PandasDataset(
    data="df.parquet",
    loader=create_base_loader(
        example_id_colname="example_id",
        path_colname="path",
    )
)

Parser Changes

  • CCD mirror path validation: ccd_mirror_path now raises FileNotFoundError if the path doesn't exist. Pass None explicitly to use Biotite's bundled CCD
  • build_assembly="_spoof" removed: Use "all" instead (raises deprecation warning)
  • convert_mse_to_met default changed: Now True by default (was False)
  • STANDARD_PARSER_ARGS renamed: Was DEFAULT_PARSE_KWARGS; now uses tuples instead of lists for hashability

Environment Changes

  • Removed automatic .env loading: dotenv is no longer auto-loaded on import. Call load_dotenv() explicitly if needed: 
    from dotenv import load_dotenv
load_dotenv()

Removed Exports

  • monkey_patch_atomarray removed from top-level exports. Use from atomworks.biotite_patch import monkey_patch_biotite instead

Added

New Modules

  • atomworks.ml.conditions - Unified conditioning management for model training
  • atomworks.ml.preprocessing.msa - MSA preprocessing (organize, filter, generate)
  • atomworks.ml.executables - External executable management (hbplus, hhfilter, mmseqs2, x3dna)
  • atomworks.ml.transforms.design_task - Design task transforms
  • atomworks.ml.transforms.mask_generator - Mask generation for training
  • atomworks.ml.utils.condition - Condition utilities
  • atomworks.io.utils.compression - Compression utilities (zstd support)

New Dataset Classes

  • FileDataset - Each file is one example (extracted from old monolithic datasets.py)
  • PandasDataset - DataFrame-backed dataset with loader support

New Loader Functions

  • create_base_loader() - Standard CIF loading
  • create_loader_with_query_pn_units() - Loading with PN unit queries
  • create_loader_with_interfaces_and_pn_units_to_score() - Interface scoring loader

New Constants

  • PROTEIN_BACKBONE_ATOM_NAMES - Backbone atoms including OXT
  • RNA_BACKBONE_ATOM_NAMES - Sugar-phosphate + 2' hydroxyl atoms
  • DNA_BACKBONE_ATOM_NAMES - Sugar-phosphate atoms
  • NUCLEIC_ACID_BACKBONE_ATOM_NAMES - Union of RNA+DNA backbones
  • MASKED - Token code for masked positions
  • MSAFileExtension enum - Supported MSA file formats
  • Expanded METAL_ELEMENTS - Now includes lanthanides and actinides

New Features

  • AtomArrayPlus support in parser - Extended atom array with additional metadata
  • Spawn multiprocessing support for data loading
  • zstd compression support for MSA files
  • Atom37 encoding with atomization support
  • JSON-level atom selection for bonds argument

Fixed

  • Residue starts bug with dependent functions
  • SASA calculation for empty amino acid arrays
  • Null handling in A3M files
  • Design tasks with zero frequency now handled gracefully instead of erroring
  • Non-uniform shard sizes handling
  • Pickling during data loading with spawn multiprocessing

Changed

  • Loaders module restructured from loaders.py to loaders/ subpackage (imports still work via __init__.py)
  • Parser cache structure now uses 2-level sharding (old caches automatically regenerated)

Deprecated

  • atomworks.ml.datasets.parsers module - Use loaders instead
  • StructuralDatasetWrapper - Use loader parameter on datasets directly

See CHANGELOG.md for full history.

Assets 2
Loading