Risk Governance · Study

The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models

NUWA research note · independent review pending

First published: 2026-10-19Source checked: 2026-10-28

Original source ↗Cite this record ↓CBRN →

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Evidence update Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

Initial suppression and durable safeguards are different claims.

Research question

Is initial suppression from pretraining filtering also a durable safeguard?

Design and assumptions

The authors evaluate pretraining filtering across several sizes of the ESM-2 protein language model, distinguishing initial proxy-task performance from durability under subsequent adaptation. Each model uses one training seed. Operational recovery methods are excluded from this record.

Source findings

Initial suppression and durability under later adaptation are distinct claims. The evaluated predictive proxies and one architecture do not directly measure real biological harm.

NUWA interpretation

Safeguard evidence should address selectivity, initial effect and durability. Its connection with Deep Ignorance is complementary argument structure, not independent replication on the same model.

Limits and contrary evidence

The study uses one architecture and one seed per trained model. Proxy prediction outcomes do not establish real biological harm or transfer to all language models.

Research and decision relevance

NUWA recommendation: Compare safeguard scope and durability with unlearning and access-control evidence, retaining architecture and adaptation boundaries.

Precise source locations

The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models · Abstract; §4.4 limitations; conclusion; text lines 00271Original resource captured 2026-10-28T07:25:52.921030+00:00; sha256:83a3c96f121b7de5bcb6ba87f388d6ac39e33416fd00b70cbcf68d3235ac7670

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Evidence update · 2026-10-28

Model-assisted reading · independent evidence and bilingual editorial review pending. Values are source findings or inventory counts, separate from the trend result series.

Institutions and roles

Affiliation not indexed in this review

Authors (12)

Aaron Maiwald · Roderick McNeill · Jaeyoung Lee · Matthew A. Clarke · Lennart Justen · Alejandro Acelas · Ivo Andrews · Giles Harper-Donnelly · Adam Moore · Sophie Rose · Oliver M. Crook · Felix Jackson

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{protein-filtering,
  title = {The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models},
  url = {https://arxiv.org/html/2610.26651},
  author = {Aaron Maiwald and Roderick McNeill and Jaeyoung Lee and Matthew A. Clarke and Lennart Justen and Alejandro Acelas and Ivo Andrews and Giles Harper-Donnelly and Adam Moore and Sophie Rose and Oliver M. Crook and Felix Jackson},
  year = {2026},
  date = {2026-10-19},
  version = {arXiv:2610.26651v1; version header 19 Oct 2026; Original resource captured 2026-10-28T07:25:52.921030+00:00; sha256:83a3c96f121b7de5bcb6ba87f388d6ac39e33416fd00b70cbcf68d3235ac7670},
  note = {arXiv:2610.26651v1; version header 19 Oct 2026; Original resource captured 2026-10-28T07:25:52.921030+00:00; sha256:83a3c96f121b7de5bcb6ba87f388d6ac39e33416fd00b70cbcf68d3235ac7670}
}

NUWA record · BibTeX

@misc{protein-filtering-nuwa,
  title = {The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/protein-filtering/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--protein-filtering.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--protein-filtering.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Aaron Maiwald, Roderick McNeill, et al.. The Effectiveness and Durability of Pretraining Data Filtering in Protein Language Models. 2026-10-19 ; Abstract; §4.4 limitations; conclusion; text lines 00271

Search NUWA Index