Note: all proteins binding the target (or its homologs) were removed from the training set.
Sep 2, 2020 • 8 tweets • 5 min read
Excited to share an update to our work on evolutionary-scale modeling (ESM)! Over the past year, we rewrote our paper with better pretraining and downstream models, leading to state-of-the-art results across multiple benchmarks. (1/8) biorxiv.org/content/10.110…
Last year, we showed that Transformer language models learn intrinsic properties of proteins from sequences. But on quantitative benchmarks, these models did not improve over alignment-based methods, as shown by @roshan_m_rao, et al in TAPE.😵(2/8)