The model had learned the mechanistic basis of peptide function.
For decades, the study of bioactive peptides—those small but potent molecules that defend, regulate, and heal within living systems—has been slowed not by a lack of data, but by a lack of integration. A research team at Zhejiang University has answered that fragmentation with PepAnno, a unified AI platform that predicts the biological functions of peptides by reading both their chemical sequence and their three-dimensional shape. In doing so, the platform does something rare in computational biology: it not only renders a verdict, but explains its reasoning, mapping the molecular features that drive each prediction back onto the peptide's physical structure.
- Drug discovery pipelines have long been stalled by the need to shuttle peptide data across disconnected tools—one for structure, one for function, one for properties—each operating in isolation.
- Most existing predictors function as black boxes, offering functional scores without revealing which parts of a peptide molecule are actually responsible for its biological activity.
- PepAnno fuses transformer-based sequence analysis with graph neural network structural processing, bridging the two through cross-attention to produce a single, unified molecular representation.
- Hierarchical transfer learning allows the model to overcome the scarcity of training data for rare peptide types by building outward from the well-studied antimicrobial category to six others.
- A case study of Human Neutrophil Peptide-1 demonstrated that the model independently recovered known mechanistic principles—positively charged residues for antimicrobial action, hydrophobic aromatic residues for anticancer activity—without being explicitly taught them.
- The platform is now publicly accessible via web interface, consolidating prediction, visualization, and a curated knowledge base into one encrypted, browser-compatible environment.
Peptides are among the body's most versatile molecular tools—short chains of amino acids capable of killing bacteria, suppressing tumors, and penetrating cell membranes. More than seven thousand naturally occurring peptides circulate in the human body, yet identifying which ones do what, and understanding why, has long been a bottleneck in therapeutic research. The problem was not only biological complexity but infrastructural fragmentation: researchers were forced to move between separate tools for sequence analysis, structure prediction, and functional annotation, while the predictors themselves rarely explained their conclusions.
A team at Zhejiang University has released PepAnno, a web-based platform designed to collapse that fragmented workflow into a single interface. Its core innovation lies in architecture: rather than analyzing peptide sequences alone, PepAnno first predicts each peptide's three-dimensional shape using ESMFold, then encodes that structure as a geometric graph in which amino acid residues become nodes connected by edges representing spatial proximity, backbone bonds, and local sequence context. A transformer network reads the sequence while a graph attention network reads the structure, and a cross-attention mechanism bridges the two—allowing each representation to query and inform the other. The result is a model that understands both what a peptide is made of and how it folds.
The platform predicts seven functional categories: antimicrobial, anticancer, anti-inflammatory, antiviral, antihypertensive, anti-angiogenic, and cell-penetrating. To address the imbalance in available training data—antimicrobial peptides are extensively catalogued while others remain scarce—the team used hierarchical transfer learning, pre-training on a large antimicrobial dataset before fine-tuning for each additional function using a head-reset strategy that prevents prior decision boundaries from contaminating new tasks. Benchmarking shows PepAnno matches or outperforms specialized single-function predictors across categories.
What distinguishes PepAnno most sharply from its predecessors is interpretability built into the architecture itself. The system extracts attention weights from its cross-attention layer and projects them onto the peptide's 3D structure, illuminating which residues the model weighted most heavily for each prediction. A study of Human Neutrophil Peptide-1 made this concrete: the model correctly predicted its known antimicrobial and anticancer properties, and when its attention patterns were visualized, positively charged residues dominated the antimicrobial signal—consistent with electrostatic interactions with bacterial membranes—while hydrophobic aromatic residues drove the anticancer prediction, matching experimental evidence about membrane insertion. The model had recovered mechanistic logic without being explicitly taught it.
For researchers, the practical implications are significant. Hypothesis-driven peptide discovery depends on understanding not just whether a molecule is active, but which structural features govern that activity. By making its reasoning visible and anchoring predictions in three-dimensional space, PepAnno transforms computational output from a score into a starting point for experimental design. The team plans to expand the platform's functional categories over time, but the foundational shift is already in place: peptide analysis now arrives with an explanation attached.
Peptides are small molecules with enormous therapeutic promise. Over seven thousand naturally occurring peptides circulate through the human body, many with antimicrobial, anticancer, anti-inflammatory, or antiviral properties. They bind to cell surface receptors, trigger immune responses, disrupt bacterial membranes, and kill cancer cells—all while remaining short enough to synthesize chemically. Yet identifying which peptides do what, and why, has remained a bottleneck. Researchers studying bioactive peptides have long faced a fragmented workflow: one tool to predict function, another to calculate physical properties, a third to visualize structure. Most existing predictors operate as black boxes, offering predictions without explaining which parts of a peptide molecule drive its biological activity. The field needed integration.
A team led by researchers at Zhejiang University has released PepAnno, a web-based platform that consolidates peptide analysis into a single interface. The system is built on a novel deep learning architecture that does something previous tools largely avoided: it fuses sequence information with three-dimensional structural data. Most peptide predictors work with sequences alone—strings of amino acids. PepAnno goes further. It predicts each peptide's 3D shape using a method called ESMFold, then represents that structure as a geometric graph where amino acid residues become nodes connected by edges encoding backbone bonds, local sequence context, and spatial proximity. A transformer network processes the sequence while a graph attention network processes the structure. A cross-attention mechanism then bridges the two, allowing semantic features to query spatial context and vice versa. The result is a unified representation that captures both what a peptide is made of and how it folds.
The platform predicts seven major bioactive functions: antimicrobial, anticancer, anti-inflammatory, antiviral, antihypertensive, anti-angiogenic, and cell-penetrating properties. To overcome the challenge of imbalanced training data—antimicrobial peptides are well-studied while others are rare—the team employed hierarchical transfer learning. They pre-trained the model on a large dataset of 8,387 antimicrobial peptides, then fine-tuned it for each target function using a "head reset" strategy that prevents the model from carrying over decision boundaries from one task to another. Comprehensive benchmarking shows PepAnno matches or exceeds specialized predictors designed for individual peptide types, all while operating within a single framework.
Beyond prediction scores, PepAnno provides what the field has lacked: mechanistic interpretability. The system extracts attention weights from its cross-attention layer and maps them onto the peptide's 3D structure, highlighting which residues the model considers most important for each predicted function. A case study of Human Neutrophil Peptide-1 (HNP-1) illustrates the power of this approach. HNP-1 is known to have antimicrobial and anticancer properties. PepAnno correctly predicted both, but also suggested anti-inflammatory, antiviral, and anti-angiogenic potential. More importantly, when researchers mapped the model's attention patterns to the 3D structure, they found that positively charged residues received high attention for antimicrobial and antiviral predictions—consistent with electrostatic interactions with negatively charged bacterial and viral membranes. Hydrophobic aromatic residues lit up for anticancer predictions, matching experimental evidence that these residues govern membrane insertion and target binding. The anti-inflammatory prediction highlighted residues known to mediate protein-protein interactions rather than broad membrane disruption. The model had learned, without explicit instruction, the mechanistic basis of peptide function.
The platform itself is designed for accessibility. Users submit a peptide sequence through a web interface and receive automated analysis: physicochemical properties, secondary and tertiary structure predictions, functional predictions across all seven categories, and interactive 3D visualizations. A curated repository of peptide databases and computational tools is integrated into the platform, providing a centralized knowledge base. All data is encrypted, temporary, and deleted after thirty days. The system runs on Django and is compatible with major browsers.
PepAnno represents a shift in how computational biology approaches prediction. Rather than treating models as opaque decision-makers, the team built interpretability into the architecture itself—not as an afterthought, but as a native feature. By anchoring predictions in 3D structure and making attention weights visible, the platform empowers researchers to formulate mechanistically meaningful hypotheses about why peptides behave as they do. For drug discovery, where understanding mechanism accelerates development, this transparency matters. The team plans to expand the platform with additional functional categories and target-specific predictions, but the core innovation is already clear: peptide analysis no longer requires jumping between tools, and predictions now come with explanations.
Notable Quotes
Most existing predictors operate as black boxes, offering limited mechanistic insight into how specific spatial motifs govern biological functions.— PepAnno research team, describing the problem the platform addresses
By enabling one-click peptide annotation, PepAnno offers an efficient and interpretable solution for large-scale peptide analysis and facilitates downstream experimental design and peptide-based drug discovery.— PepAnno abstract
The Hearth Conversation Another angle on the story
Why does structure matter so much for peptide prediction? Couldn't you just use the sequence?
Sequence tells you what a peptide is made of, but not how it folds or which parts of the molecule actually interact with targets. A positively charged residue buried inside the structure behaves very differently from one on the surface. The 3D shape determines function.
So you're saying the same sequence could fold differently and have different activity?
Exactly. And more subtly, the same peptide can have multiple functions—antimicrobial and anticancer, say—through different structural mechanisms. You need to see the geometry to understand which residues drive which activity.
How does the model learn to map attention to the right parts of the structure?
It doesn't learn that explicitly. We extract attention weights from the cross-attention layer—the mechanism that fuses sequence and structure—and project them onto the 3D coordinates. The model learned to attend to functionally relevant residues because those residues were predictive during training. We're just making that learning visible.
What happens when you test on peptides the model hasn't seen before?
We use strict cross-validation and length-stratified evaluation. The model performs consistently across peptides of different lengths, from very short to over 100 amino acids. But it's optimized for the typical bioactive peptide range—5 to 100 residues. Longer sequences would dilute the signal.
Why does the hierarchical transfer learning matter?
Because antimicrobial peptides are abundant in the literature, but anti-inflammatory peptides are rare. If you train on imbalanced data, the model overfits. By pre-training on the large antimicrobial dataset and then resetting the decision boundary for each new task, we prevent negative transfer—the model doesn't force anti-inflammatory peptides into antimicrobial patterns.
What's the practical impact for a researcher?
Instead of running five different tools and manually integrating results, they submit one sequence and get structure, properties, seven functional predictions, and visual explanations of why the model made each prediction. That's hours of work compressed into minutes.