Fig. 1
From: Medium-sized protein language models perform well at transfer learning on realistic datasets
Schematic view of the transfer learning via feature extraction approach used throughout this work. A) Protein representation (embeddings matrix) extraction using the pLMs. B) Embeddings matrix compression using various methods including mean pooling, max pooling, BOS, PCAs, and iDCT. C) Transfer learning to predict downstream tasks, such as secondary structure (SS), fitness, and physico-chemical properties (PCPs).
