Roadmap
Additional predictors from maintainers
Quarter 2, 2026
We have the following predictors planned to be wrapped:
ProteinMPNN (as a zero shot predictor)
RequiresStructureMixin
Does not require wildtype, eg. can do mutant marginal on variable length structures.
Type 3 dependencies (sub environment)
NOMELT (as a zero shot predictor)
Temperatute specific scores
Type 3 dependencies
ESM3 (Zero shot prediction and embedder)
Structure aware if available
For now, the annotation data mode will not be considered. This will require an additional attribute of the ProteinSequence class (in addition to sequence, structure, msa data types already supported).
Contributions from the community
While we will maintain the codebase and address bugs identified by the community, the usefulness of the tool will ultimately require contributions from the community a la higgingface scikit-learn, etc. When / If we add (2) community contributed models we will start developping the next major version (v2)
Major update v2
Undetermined
The component specification and software engineering exercise conducted in AIDE for different types of predictors would also be helpful for the variable and dispirate generator methods available, eg. methods for producing new sequences. These broadly categorize into:
Unconditional generators
Conditional generators (eg. infilling, homolog aware, structure or other property conditioning)
ProteinMPNN
NOMELT
Tranception
Score optimizers (black box or maybe gradient aware), eg. BADASS (already included in the package) which use a scoring function (AIDE predictor) to bias generation.