Unobserved Sequence Space Has Many Functional Proteins

In plain English

Protein models found useful candidates in unfamiliar sequence territory, but they did not reliably reproduce the detailed experimental fitness landscape.

What they found

• Each family contained functional new-to-nature sequences with low identity to known orthologs.
• Sequence models were useful but inconsistent and did not reliably capture local or global experimental fitness patterns.

The catch

• The record is a preprint and has not completed peer review.
• This summary is based on the abstract; methods and supplementary analyses were not independently rechecked.

Methods at a glance

Study type

Experimental protein-fitness benchmark

Biological system

Three protein families

Organism

Natural orthologs and synthesized new-to-nature proteins

Dataset

Functionally labeled natural and new-to-nature protein sequences

Sample size

Three protein families; sequence count not reported in the abstract

Model architecture

Potts models and protein language models

Training objective

Score protein function across distant sequence space

Code

Not reported in the abstract

Data

The authors describe an assembled functionally labeled dataset; release details were not verified.

Tasks and metrics

Tasks: Function prediction; Fitness-landscape reconstruction

Metrics: Experimental fitness; Prediction agreement

Baselines: Natural protein orthologs; Potts-model scores

Principal findings

• Many low-identity new-to-nature sequences retained measured function, while in-silico predictions remained inconsistent.

What remains unresolved

External validation: Experimental fitness measurements supplied the validation layer

• The record is a preprint and has not completed peer review.
• This summary is based on the abstract; methods and supplementary analyses were not independently rechecked.