IC-1477TAPe performs poorly compared to a simple CNN for protein fitness prediction

Andrew Kirjner, Jason Yim, Raman Samusevich, Shahar Bracha, Tommi S. Jaakkola, Regina Barzilay, Ila R Fiete

SourceImproving protein optimization with smoothed fitness landscapes

The authors note that TAPe, a popular transformer-based protein language model, showed poor performance as a fitness evaluator relative to a simpler 1D CNN architecture. This observation, which they state matches the findings of Dallago et al. (2021), led them to adopt the CNN as their in-silico evaluator for all experiments. No specific performance numbers for TAPe are reported; the comparison is qualitative and serves as methodological justification.

Evidence

observational

Caveat

No quantitative comparison is provided; the claim is a single qualitative sentence used to justify the authors' choice of evaluator architecture. The comparison is not controlled for model size or training data.

Model

TAPe

Datasets

GFP [eval]

Extraction

automatic-extraction