Back

pLM representations unlock metagenomic space beyond homology

Le Breton, L.; Heurtel-Depeiges, D.; Millar, D. C.; Zetzsche, L. E.; Vernon, R. M.; Langmead, C. J.; Chandar, S.; Fournier, Q.

2026-07-30 bioinformatics
10.64898/2026.07.28.739874 bioRxiv
Show abstract

Metagenomic sequencing has uncovered billions of proteins from uncultured microorganisms, vastly expanding the known protein space. Yet most remain functionally inaccessible because existing annotation methods depend on close homologs or accurate structure predictions. Here, we show that protein language models (pLMs) can unlock this diversity only when their training data are appropriately curated. We introduce Residue Embedding Diversity (RED), a metric for protein quality assessment orders of magnitude cheaper than likelihood, and a calibration task that measures model alignment with natural evolutionary distributions. We discover a fundamental trade-off between evolutionary calibration and structural modeling, establishing training data composition as a primary determinant of pLM behavior. Finally, we successfully retrieve diverse enzyme candidates from billions of metagenomic sequences and validate their expression in vivo.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

1
Nature Methods
○ 385 papers in training set
Top 0.5%
17.1%
Similar papers in this journal
2
Science
● 477 papers in training set
Top 0.3%
15.1%
Similar papers in this journal
3
Cell Systems
○ 201 papers in training set
Top 0.1%
12.7%
Similar papers in this journal
4
Proceedings of the National Academy of Sciences
● 2444 papers in training set
Top 4%
7.9%
Similar papers in this journal
50% of probability mass above
5
Nature Communications
○ 5641 papers in training set
Top 20%
7.9%
Similar papers in this journal
6
Nature
○ 645 papers in training set
Top 4%
4.1%
7
Nature Biotechnology
○ 172 papers in training set
Top 1%
3.3%
8
Nucleic Acids Research
◐ 1281 papers in training set
Top 6%
2.8%
9
Science Advances
● 1243 papers in training set
Top 13%
2.7%
10
Nature Machine Intelligence
○ 70 papers in training set
Top 1%
2.7%
11
Bioinformatics
◐ 1204 papers in training set
Top 6%
2.5%
12
PLOS Computational Biology
● 1863 papers in training set
Top 13%
1.9%
13
Molecular Systems Biology
○ 162 papers in training set
Top 1%
1.7%
14
Cell
○ 431 papers in training set
Top 8%
1.1%
15
eLife
● 5828 papers in training set
Top 59%
1.1%
16
Cell Genomics
○ 172 papers in training set
Top 3%
1.1%
17
Cell Reports Methods
○ 165 papers in training set
Top 3%
1.1%
18
Briefings in Bioinformatics
◐ 354 papers in training set
Top 6%
1.0%
19
Genome Biology
○ 637 papers in training set
Top 8%
1.0%
20
Protein Science
○ 246 papers in training set
Top 4%
0.8%
21
Molecular Biology and Evolution
◐ 542 papers in training set
Top 5%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.