01 / The problem
Peptide AI / 2025
PepLand.
Multi-view heterogeneous graphs for peptide representations across natural and non-natural amino acids.
Briefings in Bioinformatics · 26(4), bbaf367
02 / The core idea
Learn at atom and fragment scales with a heterogeneous graph, masked-fragment self-supervision and canonical-to-non-canonical pretraining.
How the idea flows
FIGURE 01 / ORIGINAL CONCEPT SKETCH
- Peptide molecular graph
- Atom + fragment views
- Masked-fragment learning
- Canonical → non-canonical adaptation
- Representation → task prediction
03 / What the public evidence shows
Published evaluations cover cell penetrability, solubility and peptide–protein binding affinity. Ablations examine fragmentation, masking, training stages and view fusion.
Source: PepLand ↗04 / Where the claim stops
Representation learning does not establish an experimentally confirmed therapeutic candidate. Generalization depends on the exact task, split and evaluation conditions.
An editorial scope note, not an unverified quotation of author limitations.