Show Notes
Longo LM et al., PNAS - This episode summarizes a PNAS study introducing CLSS, a contrastive two-tower protein language model that coembeds domain sequences, structures, and subsequences into a shared 32-dimensional latent space. Trained self-supervised on one million ECOD domains, CLSS aligns sequence and structure modalities, yields compact embeddings that recapitulate ECOD and CATH hierarchies, outperforms several state-of-the-art PLMs on ProteinShake classification tasks, and powers an interactive viewer for exploring protein space. Key terms: contrastive learning, protein sequence, protein structure, protein domains, protein embeddings.
Study Highlights:
The authors developed CLSS, a contrastive two-tower model that coembeds full domain sequences, structures, and sampled subsequences into a single latent space. CLSS embeddings recapitulate expert ECOD and CATH hierarchical labels despite never using those labels during training. A subsequence-trained variant (CLSS-sub) meaningfully embeds fragments, and CLSS outperforms comparison PLMs on downstream ProteinShake classification benchmarks. Visualizations of CLSS maps reveal a strong partitioning of domains by cofactor binding and other functional annotations.
Conclusion:
CLSS demonstrates that sequence and structure can be jointly organized into a compact, informative embedding space that captures domain hierarchy, subsequence reuse, and functional preferences; these embeddings enable efficient downstream classification, visualization, and potential applications in database search, alignment, protein design, and evolutionary analysis.
Music:
Enjoy the music based on this article at the end of the episode.
Article title:
Contrastive learning unites sequence and structure in a global representation of protein space
First author:
Longo LM
Journal:
PNAS
DOI:
10.1073/pnas.2532702123
Reference:
Longo LM, Yanai G, Axel G, Kolodny R, Ben-Tal N. Contrastive learning unites sequence and structure in a global representation of protein space. PNAS. 2026;123(32):e2532702123. doi:10.1073/pnas.2532702123.
License:
This episode is based on an open-access article published under the Creative Commons Attribution 4.0 International License (CC BY 4.0) – https://creativecommons.org/licenses/by/4.0/
Support:
Base by Base is independent and ad-free — no sponsors, no paywall. If an episode was worth your time, chip in and keep the papers audited and the original songs coming:
❤️ Support monthly: https://buy.stripe.com/cNifZhclVebvagk2JDgEg01
☕ One-time donation: https://donate.stripe.com/7sY4gz71B2sN3RWac5gEg00
More at basebybase.com
On PaperCast Base by Base you'll discover the latest in genomics, functional genomics, structural genomics, and proteomics.
Episode link: https://basebybase.com/episodes/clss-contrastive-sequence-structure-protein-space
QC:
This episode was checked against the original article PDF and publication metadata for the episode release published on 2026-08-11.
QC Scope:
- article metadata and core scientific claims from the narration
- excludes analogies, intro/outro, and music
- transcript coverage: Audited transcript portions describing CLSS concept and motivation, two-tower architecture, CLSS-sub subsequences, training on ECOD domains, evaluation on ProteinShake, TSNE visualizations, cofactor and zinc-binding patterns, and limitations.
- transcript topics: CLSS concept and motivation; Two-tower architecture: sequence tower (ESM2) and frozen structure encoder (ESM3); CLSS-sub subsequences (20-60 residues); Self-supervised training on ~1 million ECOD domains; Evaluation on ProteinShake and ECOD/CATH hierarchies; t-SNE visualizations showing coembedding of sequence and structure
QC Summary:
- factual score: 10/10
- metadata score: 10/10
- supported core claims: 8
- claims flagged for review: 0
- metadata checks passed: 4
- metadata issues found: 0
Metadata Audited:
- article_doi
- article_title
- article_journal
- license
Factual Items Audited:
- CLSS coembeds sequence and structure into a single 32-dimensional latent space
- Two-tower architecture: sequence tower initialized like ESM2 and a frozen ESM3 structure encoder
- CLSS-sub trains on random contiguous subsequences of length 20-60 residues
- Training data consists of ~1 million ECOD domains in a self-supervised setting
- CLSS-full overlaps sequence and structure embeddings on TSNE maps; other PLMs show separation
- CLSS outperforms state-of-the-art PLMs (ESM3, ProstT5, ProTrek) on ProteinShake SCOP-label tasks
QC result: Pass.
