Please use this workflow to perform a conditional generation experiment, making use of the flexibility provided by AbBFN2. Here, we provide the options to condition AbBFN2 on full sequences (referred to as sequence labelling in the manuscript), partial sequences (inpainting), partial sequences and metadata (sequence design), metadata only (conditional de novo generation), or any combination of these tasks.
Please note that while the model explicitly models 45 different data modes, here we provide control of a subset of these, although the output from the web app will include all modelled attributes. For a full description of the data modes, please refer to the AbBFN2 manuscript. For full control over conditional generation as described in the manuscript as well as the ability to generate a larger set of samples, please refer to the code.
Disclaimer: As discussed in the manuscript, the flexibility of AbBFN2 requires careful consideration of the exact combination of conditioning information for effective generation. For instance, conditioning on a kappa light chain locus V-gene together with a lambda locus J-gene family is unlikely to yield samples of high quality. Such paradoxical combinations can also exist in more subtle ways. Due to the space of possible conditioning information, we have only tested a small subset of such combinations.