roccomoretti
commented
1 month ago In contig specifications, when you specify a range for generated structures, this is a range of potential lengths, which is randomly sampled.
For example, if you have 'contigmap.contigs=[A10-25/27-32]', RFdiffusion will randomly sample from the range of 27-32 for the number of residues to add. Sometime it will pick 30 aa (yielding a 46 aa long protein), sometimes it will pick 32 (yielding a 48 aa long protein) and sometimes it will pick 28 (yielding a 44 aa long protein).
If you run RFdiffusion multiple times, you should get different lengths for each run.
When comparing the input and output values, it seems that the output values represent the central residue number of each contig from the input.
For example, for the input range 27-32, the output is 30-30. This is because 30 is the central residue number within the range 27-32.
This suggests that RFdiffusion, when processing contigs internally, replaces each contig with its corresponding central residue number.
However, the precise reason for why this processing is done remains unclear.