PepHarmony: a multi-view contrastive learning framework for integrated sequence and structure-based peptide representation

01 / Research question

Peptide structures are flexible and complex. Sequence representations should benefit from structural information.

02 / Innovation

Align sequence and structure representations by multi-view contrastive learning; transfer structural information into the sequence encoder.

03 / Mechanism

ESM sequence and GearNet structure encoders align through contrastive training; downstream prediction keeps sequence-only input. The VAE variant additionally explores generative learning.

04 / Experimental setup

Manuscript v1 collects 17,300 PDB peptides shorter than 50 residues and AlphaFold structures; confidence-filtered AlphaFold subsets support pretraining. CPP contains 1,162 positive and 1,162 negative samples; PROSO-II supports solubility. ESM2, fine-tuned ESM and GearNet provide baselines. Exact downstream split proportions are unavailable in the inspected main text.

Author manuscript v1 · 2024 · Tables 1, 3 · Source ↗

05 / Results

Table 1 reports CPP AUC 0.874 versus ESM2 0.864; affinity Spearman 0.447 trails fine-tuned ESM at 0.473. Final 2026 journal results are not substituted for these v1 values.

ORIGINAL FIGURE / PAPER SOURCE

How the method works.

During training, an ESM sequence encoder aligns with a GearNet structure encoder. Downstream inference uses only the sequence encoder: structural information enters the representation through training.

REPORTED COMPARISON / SAME-PAPER CONDITIONS

Comparisons, including the tradeoffs.

Author-reported results from the public 2024 manuscript; values belong to that version rather than the final-journal tables. Check original tables ↗

Swipe horizontally or focus and use arrow keys for all columns →

CPP classification · Table 1

Model	Accuracy	F1	AUC ↑
ESM2	0.723	0.756	0.864
ESM-finetune	0.770	0.755	0.839
GearNet-finetune	0.764	0.753	0.842
PepHarmony-CL	0.790	0.766	0.874
CL + VAE	0.777	0.764	0.871

CPP contains 1,162 positive and 1,162 negative examples; structure confidence filtering is explored in the manuscript. Affinity SCC is 0.473 for ESM-finetune and 0.447 for PepHarmony, so superiority is not universal.

ABLATION / WHAT CHANGES THE RESULT?

Take the mechanism apart.

Swipe horizontally or focus and use arrow keys for all columns →

Condition	Task / metric	Reported value
Without confidence sorting	CPP AUC	0.843
With confidence sorting	CPP AUC	0.874
Diffusion initialization	CPP AUC	0.855
CL + VAE	CPP AUC	0.871

Source: CPP columns of Table 3, 2024 manuscript. Some early affinity and solubility entries conflict across tables or prose; disputed values are excluded.

A BOUNDED CASE / REPRESENTATIONS IN CONTEXT

Back to a peptide question.

Can structural information help a sequence representation?

CPP is a concrete test in the manuscript: add sequence–structure alignment during training, keep sequence-only downstream inputs. Same-paper CPP AUC supports a benefit on that task; the affinity counterexample shows that benefits remain task-dependent.

A public task case, not a therapeutic claim for an experimental candidate.

Manuscript Figures 4–5 & Tables 1, 3 ↗

06 / Ablation & limitations

Table 3 CPP AUC rises from 0.843 without confidence ordering to 0.874 with ordering; diffusion initialization gives 0.855. Predicted-structure confidence and short-peptide flexibility affect representation learning. Early affinity/solubility tables contain inconsistencies, so disputed entries are excluded. Contrastive alignment does not prove uniformly superior structural understanding.

The linked author manuscript is an earlier version; the final journal citation is listed above. Representation benchmarks do not demonstrate therapeutic efficacy or deployment.

07 / My contribution

First author.

08 / Paper & code