PepLand: a large-scale pre-trained peptide representation model for a comprehensive landscape of both canonical and non-canonical amino acids

01 / Research question

Peptide representations need to accommodate non-canonical amino acids beyond the usual sequence alphabet.

02 / Innovation

Learn at atom and fragment scales with a heterogeneous graph, masked-fragment self-supervision and canonical-to-non-canonical pretraining.

03 / Mechanism

AdaFrag combines amino-acid fragmentation with BRICS side-chain fragments. Atom, fragment and junction relations exchange information in a heterogeneous graph. Masked pretraining first learns canonical chemistry, then adapts to noncanonical peptides.

04 / Experimental setup

The 2023 manuscript uses 7,924,509 first-stage samples and 8,977 noncanonical second-stage peptides. Downstream tasks include CPP (1,162 positives/1,162 negatives), short-sequence PROSO-II solubility and peptide affinity. ESM and small-molecule encoders are compared. Exact downstream split ratios are not stated in the inspected main text.

Author manuscript · 2023 · Tables 2, 3, 5 · Source ↗

05 / Results

Noncanonical affinity PCC reaches 0.873 after two-stage training. Canonical CPP AUC is 0.838 versus ESM 0.885. These are 2023 manuscript values, not the 2025 journal tables.

ORIGINAL FIGURE / PAPER SOURCE

How the method works.

AdaFrag adds BRICS side-chain fragments to amino-acid fragmentation, placing canonical and noncanonical chemistry in one atom/fragment heterogeneous graph. Masked learning builds representations, then two-stage pretraining adapts them to noncanonical peptides.

REPORTED COMPARISON / SAME-PAPER CONDITIONS

Comparisons, including the tradeoffs.

Author-reported results from the public 2023 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 →

Canonical CPP classification · Table 2

Model	AUC ↑
ESM	0.885
PepLand	0.838
ESM + PepLand	0.885

PepLand alone trails ESM on this canonical CPP task; noncanonical tasks are a major focus. No cross-paper leaderboard is implied because splits and protocols are not aligned.

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
Amiibo	Noncanonical affinity SCC	0.708
AdaFrag	Noncanonical affinity SCC	0.749
Amiibo → AdaFrag	Solubility F1	0.696 → 0.693
One stage → two stages	Noncanonical affinity PCC	0.827 → 0.873
One stage → two stages	Canonical affinity PCC	0.496 → 0.450

Source: 2023 manuscript Tables 3 and 5. Side-chain fragmentation and two-stage training do not improve every metric.

A BOUNDED CASE / REPRESENTATIONS IN CONTEXT

Back to a peptide question.

1SMF

Public manuscript Figure 8: 6 samples, Spearman 0.829. A curated SKEMPI case excluded from training.

View public structure ↗

5XCO

Public manuscript Figure 8: 10 samples, Spearman 0.718. A small retrospective case, not prospective candidate discovery.

View public structure ↗

06 / Ablation & limitations

Tables 3–5 compare fragmentation, masking and training stages. Two stages increase noncanonical-affinity PCC from 0.827 to 0.873, but canonical-affinity PCC falls from 0.496 to 0.450. Representation quality remains task-dependent; small retrospective cases do not establish prospective candidate discovery.

Representation learning does not establish an experimentally confirmed therapeutic candidate. Generalization depends on the exact task, split and evaluation conditions.

07 / My contribution

First author.

08 / Paper & code