As noted here in discussion with @timcyu here, in the simulated RBD dataset most of the escape values at non-escape sites for epitopes 1 and 2 are ~0.5 rather than 0. This may cause similarity regularization to zero out epitopes as they all have some overlap.
I examined this, and found it's because the actual "true" values fed into the simulation are like this, when they probably should not be.
So I adjusted the "true" simulation values to be zero at non-escape sites, which more or less addresses issue. This could be useful for future testing of regularizations.
As noted here in discussion with @timcyu here, in the simulated RBD dataset most of the escape values at non-escape sites for epitopes 1 and 2 are ~0.5 rather than 0. This may cause similarity regularization to zero out epitopes as they all have some overlap.
I examined this, and found it's because the actual "true" values fed into the simulation are like this, when they probably should not be.
So I adjusted the "true" simulation values to be zero at non-escape sites, which more or less addresses issue. This could be useful for future testing of regularizations.