Deepest Layer Suboptimal for Protein Language Models

anshulkundaje · x · 2026-08-17

A study systematically analyzing 13 Protein Language Models (PLMs) across 15 downstream tasks reveals that using the last layer embedding is often suboptimal. The deepest layer performs best only 17.9% of the time, with many protein-level tasks peaking in intermediate layers. Authors trained linear and k-NN probes on every layer to map information distribution, showing a consistent split: residue-level tasks improve monotonically with depth, whereas protein-level tasks often peak earlier.

Original post →

More from Research

Research channel →