Post

Five years ago, we released FLIP. The core question was: can ML models for protein fitness prediction generalize in the ways that actually matter for protein engineering, i.e. low data, extrapolation to more mutations, out-of-distribution sequences?

FLIP spawned fast development of several different benchmarking efforts across protein design, engineering, and variant effect assesment. The answer in 2021 was: sometimes, but simpler models hold up surprisingly well.

We were interested in how things had changed 5 years after our first release. So, we built FLIP2 on select datasets from great labs across the world, many of which have gracefully agreed to make their data freely available.

The answer in 2026 is largely the same. Simple ridge regression on one-hot sequences, optionally supplemented with zero-shot pLM likelihoods, often matches or outperforms fine-tuned protein language models.

Especially on the wild-type and position splits, current transfer learning doesn't consistently win. No single pLM architecture dominates. Scaling hasn't closed the gap yet.

That's not a criticism to any method. New benchmarks are precisely needed to see where we are at, and set some target of where we could go from here.