Medium-sized protein language models perform well at transfer learning on realistic datasets
Open accessPublished version of a preprint
Abstract
Protein language models (pLMs) can offer deep insights into evolutionary and structural properties of proteins. While larger models, such as the 15 billion parameter model ESM-2, promise to capture more complex patterns in sequence space, they also present practical challenges due to their high dimensionality and high computational cost. We systematically evaluated the performance of various ESM-style models across multiple biological datasets to assess the impact of model size on transfer learning via feature extraction. Surprisingly, we found that larger models do not necessarily outperform smaller ones, in particular when data is limited. Medium-sized models, such as ESM-2 650M and ESM C 600M, demonstrated consistently good performance, falling only slightly behind their larger counterparts—ESM-2 15B and ESM C 6B—despite being many times smaller. Additionally, we compared various methods of compressing embeddings prior to transfer learning, and we found that mean embeddings consistently outperformed other compression methods. In summary, ESM C 600M with mean embeddings offers an optimal balance between performance and efficiency, making it a practical and scalable choice for transfer learning in realistic biological applications.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Highly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764 · see citing works
Evolutionary-scale prediction of atomic level protein structure with a language model
Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. (2022) · doi:10.1101/2022.07.20.500902 · Cited by 5,523 · see citing works
Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., et al. (2021) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2016239118 · Cited by 3,286 · see citing works
Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., et al. (2019) · doi:10.1101/622803 · Cited by 3,280 · cited 1× · see citing works
“BERT (Bidirectional Encoder Representations from Transformers) models, such as Evolutionary Scale Modeling 2 (ESM-2) 9 , ProtTrans family (ProtT5, ProtBERT, ProtAlbert) 10 , and ProtFlash 11 , are pre-trained masked protein language models (pLMs) that have proven effective at producing biologically…”
Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., et al. (2022) · doi:10.1101/2022.12.09.519842 · Cited by 2,369 · see citing works
ProtTrans: Toward Understanding the Language of Life Through Self-Supervised Learning
Elnaggar, A., Heinzinger, M., Dallago, C., Rehawi, G., Wang, Y., Jones, L., Gibbs, T., Feher, T., Angerer, C., Steinegger, M., et al. (2022) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2021.3095381 · Cited by 2,329 · see citing works
De novo design of protein structure and function with RFdiffusion
Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., et al. (2023) · Nature · doi:10.1038/s41586-023-06415-8 · Cited by 2,325 · see citing works
Scaling Laws for Neural Language Models
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020) · arXiv · Cited by 1,749 · cited 2× · see citing works
“ESM-2 has the widest range of model sizes and therefore is a logical starting point for an analysis on the impact of model size.”
UniRef: comprehensive and non-redundant UniProt reference clusters
Suzek, B. E., Huang, H., McGarvey, P., Mazumder, R., & Wu, C. H. (2007) · Bioinformatics · doi:10.1093/bioinformatics/btm098 · Cited by 1,613 · see citing works
Unified rational protein engineering with sequence-based deep representation learning
Alley, E. C., Khimulya, G., Biswas, S., AlQuraishi, M., & Church, G. M. (2019) · Nature Methods · doi:10.1038/s41592-019-0598-1 · Cited by 1,131 · see citing works
Large language models generate functional protein sequences across diverse families
Madani, A., Krause, B., Greene, E. R., Subramanian, S., Mohr, B. P., Holton, J. M., Olmos, J. L., Xiong, C., Sun, Z. Z., Socher, R., et al. (2023) · Nature Biotechnology · doi:10.1038/s41587-022-01618-2 · Cited by 1,078 · cited 2× · see citing works
“On the other hand, generative pre-trained transformer (GPT) models, such as ProGen 12 , are designed to generate novel protein sequences and have shown potential to create functional proteins.”
Modeling the language of life – Deep Learning Protein Sequences
Heinzinger, M., Elnaggar, A., Wang, Y., Dallago, C., Nechaev, D., Matthes, F., & Rost, B. (2019) · doi:10.1101/614313 · Cited by 606 · see citing works
Learning the protein language: Evolution, structure, and function
Bepler, T., & Berger, B. (2021) · Cell Systems · doi:10.1016/j.cels.2021.05.017 · Cited by 483 · cited 1× · see citing works
“The pLM models learn about protein properties during pre-training, where the models are trained to predict residues that have been masked in their input sequences 13,14 .”
ProGen2: Exploring the boundaries of protein language models
Nijkamp, E., Ruffolo, J. A., Weinstein, E. N., Naik, N., & Madani, A. (2023) · Cell Systems · doi:10.1016/j.cels.2023.10.002 · Cited by 480 · see citing works
Learning protein sequence embeddings using information from structure
Bepler, T., & Berger, B. (2019) · arXiv · doi:10.48550/arxiv.1902.08661 · Cited by 219 · see citing works
Feature Reuse and Scaling: Understanding Transfer Learning with Protein Language Models
Li, F.-Z., Amini, A. P., Yue, Y., Yang, K. K., & Lu, A. X. (2024) · doi:10.1101/2024.02.05.578959 · Cited by 51 · see citing works
Flagged first, then by citations
- Authors
- Vieira, L. C., Handojo, M. L., & Wilke, C. O.
- Affiliations
- The University of Texas at Austin (US)
- Corresponding author
- Wilke, Claus O.
- Venue
- Scientific Reports
- Published
- 2025-07-01
- Volume
- 15
- Issue
- 1
- Pages
- 21400-21400
- ISSN
- 2045-2322 (ISSN-L)
- PMC
- PMC12217344
- OpenAlex
- W4411865423
- Cited by
- 41
- References
- 16
- Citations by year
- 2026: 28 · 2025: 13
- FWCI
- 12.24
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- Scientific Reports (publishedVersion, cc-by)
- bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- PubMed Central (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- National Institutes of Health (R01 AI169462, R01 GM088344)
- Topics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
- Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › vaccines and immunoinformatics approaches
- Keywords
- protein language models, Earth system models, model size, transfer learning, feature extraction, embedding compression, protein sequence representation, protein structure prediction, protein evolution
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
