v3.2.3

Release Notes - ESM 3.2.3

• bugfix: added chain ids to MolecularComplex
• bugfix: use biotite for mmcif conversion in MolecularComplex
• bugfix: updated sequence filtering in esm/utils/msa/filter_sequences.py for better handling of edge cases

v3.2.2.post2

All-Atom Molecular Complex Support

• Added fold_all_atom() method to handle structure prediction for complexes containing proteins, nucleic acids (DNA/RNA), and ligands
• Introduced MolecularComplex and MolecularComplexResult classes for representing multi-molecule systems
• New StructurePredictionInput builder for constructing complex molecular inputs

MSA Improvements

• Refactored MSA handling with new dedicated MSA class in esm.utils.msa
• Added FastMSA for optimized MSA operations
• Implemented sequence filtering utilities including greedy selection and hhfilter integration
• Support for removing insertions from sequences

SDK Enhancements

• Added SequentialDataclass base class for improved dataclass operations
• Enhanced retry mechanism with better error handling and callback support
• Fixed pLDDT expansion for chain breaks in protein structures
• Added mean_hidden_state support for ESMC forward passes
• Improved tensor serialization and handling throughout the SDK

Bug Fixes

• Fixed pLDDT array length mismatch when chain breaks are present
• Corrected logits tensor handling to avoid unnecessary byte conversions
• Fixed retry decorator to properly handle exceptions and return values
• Updated mean embedding/hidden state handling in generation trimming

Documentation Updates

• Updated Forge API URLs in tutorial notebooks (removed /console suffix)
• Standardized token prompt messages across notebooks

v3.2.2

Release Notes

Documentation

• Added comprehensive Available Models section to README with tables for:
  • ESM3 Family models (flagship, published, and experimental variants)
  • ESM C models with specifications
• Updated notebook examples with improved import organization

API & SDK Enhancements

• Experimental Guided Generation with Constraints: New ESM3GuidedDecodingWithConstraints class supporting:
  • Constraint-based protein generation using Modified Differential Method of Multipliers
  • Multiple constraint types (GREATER_EQUAL, LESS_EQUAL, EQUAL)
  • Trajectory visualization for constrained optimization
• Enhanced LogitsConfig and LogitsOutput with mean embedding support
• Improved PDB export functionality to match expected behavior
• Added confidence scores to protein complex chains

Bug Fixes & Improvements

• Fixed Flash Attention import handling for better compatibility
• Improved error handling in Forge client with proper exception propagation
• Enhanced potential_sequence_of_concern parameter handling (now optional)
• Removed unnecessary DSSP annotation from default protein chain conversion
• Improved oxygen inference in PDB string generation

Examples & Tutorials

• Updated guided generation tutorial with constraint-based examples
• Added visualization for constraint optimization trajectories
• Improved code examples with better formatting and organization

v3.2.1.post1

Changelog v3.2.1.post1

Changes

• Documentation: Updated README table of contents for better navigation
• Code Quality: Removed unused imports to clean up codebase
• SDK Updates: Various improvements and bug fixes to the ESM SDK
• Tutorials: Updated cookbook tutorials with latest examples
• Dependencies: Updated pixi.lock with latest dependency versions

Bug Fixes

• Cleaned up unused imports in forge.py
• Minor fixes and improvements across various SDK modules
• Fix a formatting error leading to SDK errors for .post1 release.

Documentation

• Improved README structure and table of contents
• Updated tutorial notebooks with current best practices

This is a maintenance release focused on code quality improvements, documentation updates, and minor SDK enhancements.

v3.2.0

Updates:

• Upgraded biotite version so we're not restricted to numpy < 2.0. this required removing the to_npz / to_npz_string functions (an internal biotite format they no longer support)
• Added MSA support
• Added a globularity function so it's easy to calculate
• Fixed a bug where Inf SASA was encoded as -1, will now be encoded as 1000.

v3.1.6

• gfp notebook updated to use a max num_steps of 20 instead of 229
• updated docstring for batch executor

v3.1.5

We had a few changes:

• extended the pyproject.toml to be pixi-compliant for folks using pixi
• bugfix to the gfp_design.ipynb in the tutorial
• updated cookbook/snippets/esmc.py to include examples of how to call esmc via forge
• updated the default for num_steps in GenerationConfig

v3.1.3

Main changes:

1. Tutorials have been reorganized to be completed in a step-by-step fashion. We added a new tutorial for using ESM C to get protein embeddings.
2. Layer number specification is required when requesting hidden_states of ESM C 6B.
3. SASA track in forward_and_sample API is fixed.

v3.1.2

In this release, we made two changes:

1. Brought Flash Attention and Triton Rotary embedding to the ESM C model
2. Updated forge client to enable the return of ESM C's binary serialized hidden states. (note this the return of ESM3 hidden states are not supported).

v3.1.1

In this release, we made three changes:

1. Updated ESM C tokenizer to no longer depend on ESM3. So now you can use ESM C without having to accept the ESM3 license.
2. Updated the transformer class so it returns hidden states.
3. Pined transformers version to be <4.47.0 to prevent a tokenizer issue. If you are currently seeing tokenizer issues, try reinstalling the new release.