This documentation is in Beta mode! Please give feedback by suggesting/commenting in embedded documents or via the Feedback Form!
Note: Typically the inner workings of RFdiffusion will choose the recommended model checkpoint file to use on its own.
Many groups and researchers have created their own model checkpoint files, this documentation page is only for the model weights that are part of the RFdiffusion distribution.
Trained only on monomer tasks (unconditional design and motif-scaffolding). This model is automatically chosen for thse tasks and should only be used for these tasks (i.e. not binder design).
Not recommended for use.
This file comes from the same training run as 'Base_ckpt.pt' and was only included for reproducability purposes.
Trained on complexes (binder given target), this model is useful (and automatically chosen) for de novo binder design.
Trained on complexes with the addition of fold specification (secondary structure and block adjacency information - see the 2023 Nature paper for more information.) Used when the user knows a priori the global fold of what they are designing, such as a TIM barrell.
Specifically trained for scaffolding very small motifs, such as enzyme active sites. This is one of the few models you may need to specifically point RFdiffusion to in order to appropriately model enzyme active sites. For an example, see the enzyme design example in the GitHub repository.
Allows for conditioning on structure in the absence of a sequence and conditioning on sequence in the absence of a structure. Can be used to design binders to flexible peptides. For an example of its usage, see Vazquez-Torres et al., 2023.
Similar to InpaintSeq_ckpt.pt, with the addition of fold information (secondary structure and block adjacency). For an example of its usage in designing binders to intrinsically disordered proteins (IDPs) while specifying the secondary structure of the IDP, see Liu et al., 2024.
This is similar to Complex_base_ckpt.pt, in that it is trained on complexes, however it generates more beta-structure binders.
RFdiffusion was trained from an intermediate version of RoseTTAFold, these weights were provided for reproducability purposes only. They will not work for model generation.