v1.2.0

What's Changed

• use chebifier graph instead of chebai graph by @sfluegel05 in #14
• Global Cache per smiles per model by @aditya0by0 in #15
• Dynamic imports - for model-specific packages by @aditya0by0 in #12
• Fix - Api Models by @aditya0by0 in #9
• Simplify getting started by @cthoyt in #16
• Ensemble update by @sfluegel05 in #19
• gentle smiles parsing for lookup by @sfluegel05 in #20

New Contributors

• @cthoyt made their first contribution in #16

Full Changelog: v1.1.1...v1.2.0

v1.1.1

What's Changed

• Move inconsistency resolution and chebi graph building to chebifier by @sfluegel05 in #8
• This release also solves some problems related to the installation (1.1.0 required manually installing wandb which is not needed, nor installed automatically, the installed c3p version was not windows-compatible)

Full Changelog: v1.1.0...v1.1.1

v1.1.0

What's Changed

• Support for 3CP
• Support for chemlog-extra
• A ChEBI lookup function
• Improved error handling
• self-downloadable models via Hugging Face #2

v1.0.0

Initial release. This contains a basic ensemble implementation for three model types:

• Electra (from chebai)
• ResGated GCNs (from chebai-graph)
• ChemLog Peptides (from chemlog-peptides)

Aggregation is Weighted Majority Voting (WMV) based on

• binary classification (either positive or negative, according to 0.5 threshold)) -> only using this would constitute majority voting
• confidence (distance of prediction from threshold)
• model weight (use case: set a higher weight for ChemLog since a rule-based classification is more trustworthy than an ML model)
• trust (based on f1-score on validation set)

v1.0.0.dev

Initial release. This contains a basic ensemble implementation for three model types:

• Electra (from chebai)
• ResGated GCNs (from chebai-graph)
• ChemLog Peptides (from chemlog-peptides)

Aggregation is Weighted Majority Voting (WMV) based on

• binary classification (either positive or negative, according to 0.5 threshold)) -> only using this would constitute majority voting
• confidence (distance of prediction from threshold)
• model weight (use case: set a higher weight for ChemLog since a rule-based classification is more trustworthy than an ML model)
• trust (based on f1-score on validation set)