Morphology-Aware Peptide Discovery via Masked Conditional Generative Modeling

The paper introduces PepMorph, a masked conditional generative modeling pipeline that successfully designs novel peptides with targeted fibrillar or spherical self-assembly morphologies, achieving an 83% validation success rate through coarse-grained molecular dynamics simulations.

The Gist

Imagine you are a master chef trying to invent a new recipe for a cake. But instead of just wanting any cake, you have a very specific vision: you want a cake that is perfectly round like a beach ball, or perhaps one that stretches out long and thin like a pretzel.

The problem is that there are billions of possible ingredients and mixing combinations. Trying to guess the right recipe by tasting random mixtures would take a lifetime. This is exactly the challenge scientists face when designing peptides (tiny building blocks of proteins) that stick together to form materials. They want to control whether these tiny blocks clump into spheres (like bubbles) or fibers (like spaghetti).

Enter PepMorph, a new "AI chef" developed by researchers at the Technical University of Munich. Here is how it works, broken down into simple concepts:

1. The Problem: The "Needle in a Haystack"

Peptides are made of a chain of 20 different amino acids (like letters in an alphabet). Even a short chain has billions of possible combinations.

• The Old Way: Scientists used to guess and check, or use rigid rules. It was slow, and often they found recipes that made the right shape but were toxic or didn't stick together well.
• The New Way: They needed a smart system that could look at the "ingredients" (amino acids) and predict the "shape" of the final cake before baking it.

2. The Solution: The "Morphology-Aware" AI

The researchers built PepMorph, which is like a super-smart recipe generator. But it has a special trick: Conditional Masking.

Think of this like a "Mad Libs" game where you can fill in some blanks but leave others empty.

• The User's Request: You tell the AI, "I want a peptide that is 7 letters long, has a neutral charge, and loves to stick together."
• The AI's Job: It fills in the rest of the recipe for you. It doesn't force you to specify every detail (which might be impossible to know). It just takes your specific clues and generates a brand-new, unique recipe that fits those clues.

3. The Secret Sauce: "Shape Proxies"

How does the AI know if a recipe will make a sphere or a fiber? It can't see the final 3D shape immediately. Instead, it uses proxies (clues).

• Imagine you want to build a tower. You know that if you use long, flat bricks, it will likely be tall and thin. If you use round, bouncy balls, it will likely be a pile.
• PepMorph looks at the "flatness" or "roundness" of the individual peptide pieces before they stick together. It uses these clues to steer the AI toward making either fibers (long chains) or spheres (compact balls).

4. The "Double-Check" Filter

The AI is creative, but it can sometimes get a little wild. So, the researchers added a filtering funnel:

1. The Aggregation Check: First, they ask, "Will this actually stick together?" If the AI generates a recipe that falls apart, it gets tossed.
2. The Shape Check: Next, they run a computer simulation (like a video game physics engine) to see if the peptide actually forms the shape you asked for.
3. The Result: Out of thousands of generated recipes, they picked the top 15 for spheres and 15 for fibers. When they ran the simulations, 83% of them actually formed the correct shape!

5. Why This Matters

This isn't just about making cool shapes.

• Medicine: If you want to deliver a drug to a specific spot in the body, you might need a spherical "bubble" to carry it.
• Materials: If you want to build a strong scaffold for growing new tissue, you might need long, fiber-like structures.

The Bottom Line:
Before PepMorph, finding the right peptide was like trying to find a specific needle in a haystack by blindfolded guessing. PepMorph is like giving the blindfolded person a metal detector and a map. It allows scientists to design materials from the bottom up, ensuring they have the exact shape and function needed for life-saving applications, all without wasting time on recipes that won't work.

It's a bridge between the abstract world of computer code and the physical world of new, life-changing materials.

Technical Summary

1. Problem Statement

Peptide self-assembly offers a powerful bottom-up strategy for creating biocompatible materials for biomedical and energy applications. However, designing peptides that self-assemble into specific aggregate morphologies (e.g., fibrils vs. spheres) is a significant challenge due to:

• Vast Sequence Space: The combinatorial diversity of amino acids makes exhaustive screening intractable.
• Data Scarcity: While datasets exist for predicting aggregation propensity (AP), datasets linking sequences to validated self-assembly morphologies are fragmented, inconsistent, and lack standardized labels.
• Limitations of Existing Generative Models: Previous approaches often rely on fixed conditioning (requiring all parameters to be specified) or metaheuristic searches (like genetic algorithms) that tend to converge on high-scoring motifs, reducing sequence diversity and failing to explore alternative assembly pathways.
• The "Morphology Gap": Even if a peptide is predicted to aggregate, predicting what shape it will form (fiber, sphere, sheet) based solely on sequence is difficult without direct morphology labels.

2. Methodology

The authors introduce PepMorph, an end-to-end pipeline that generates novel peptide sequences conditioned on morphology proxies (isolated peptide descriptors) rather than direct morphology labels.

A. Dataset Curation (PepMorph Dataset)

• Sources: Merged existing aggregation propensity datasets (Wang et al., Teijlingen & Tuttle) containing ~121k sequences with AP values, and added ~39k random peptides to reduce compositional bias.
• Total Size: 161,120 unique peptides (3–10 amino acids).
• Feature Engineering:
  • Aggregation Descriptors: Aggregation Propensity (AP) and Self-Assembly (SA/no-SA) binary labels derived from Coarse-Grained Molecular Dynamics (CG-MD) SASA ratios.
  • Morphology Proxies: Computed from isolated monomer structures predicted by PEP-FOLD. These include:
    • $\beta$-strand assignment (proxy for fibrils).
    • Hydrophobic moment (proxy for amphiphilicity/spheres).
    • Net charge.
• Handling Missing Data: Since 3D descriptors cannot be computed for very short peptides or if PEP-FOLD fails, the dataset contains missing values.

B. Generative Model: Masked Conditional VAE

The core of PepMorph is a Transformer-based Conditional Variational Autoencoder (CVAE) with a masking mechanism (inspired by VAEAC).

• Architecture:
  • Encoder: Maps peptide sequences to a latent space $z$.
  • Decoder: An autoregressive Transformer that generates sequences.
  • Conditioning: Instead of requiring a full vector of descriptors, the model accepts an arbitrary subset.
• Masking Mechanism:
  • A descriptor vector $c$ is paired with a binary mask $m$ (1 = observed, 0 = missing).
  • A Multi-Layer Perceptron (MLP) creates a condition summary $s = \phi([c \odot m, m])$.
  • This summary $s$ conditions both the latent prior $p(z|s)$ and the decoder $p(y|z, s)$.
  • Benefit: Users can specify only specific constraints (e.g., "I want a fibril, so I set $\beta$-strand=Yes") while leaving other properties (e.g., net charge) free, preventing the model from collapsing to trivial solutions or failing when data is incomplete.
• Training Objective: A variational lower bound with auxiliary reconstruction losses for the mask, binary descriptors, and continuous descriptors to ensure the model faithfully encodes the provided context without imputing missing values.

C. Validation Pipeline

1. Generation: Sample sequences conditioned on target morphology proxy ranges (e.g., high hydrophobic moment for spheres, $\beta$-strand presence for fibrils).
2. Dual-Stage Filtering:
   • Stage 1: Filter for aggregation propensity (AP $\ge$ 1.8) and SA probability ($\ge$ 75%) using trained regressors/classifiers.
   • Stage 2: Re-compute descriptors from PEP-FOLD structures; retain only sequences where predicted metrics match the conditioning targets within $\pm$10%.
3. CG-MD Simulation: Top candidates (15 per target) undergo 500 ns Coarse-Grained MD simulations (MARTINI 3 force field).
4. Morphology Assessment:
   • Quantitative: Ratio of Moments of Inertia (RMOI). $RMOI \le 0.35$ (fibril), $RMOI \ge 0.75$ (sphere).
   • Qualitative: Visual inspection of aggregate shapes.

3. Key Contributions

• Novel Framework: First generative model for peptides that explicitly steers self-assembly toward specific morphologies using partial conditioning on isolated monomer descriptors.
• Masked Conditioning: Introduces a flexible masking mechanism allowing users to constrain any subset of properties, addressing the "missing covariate" problem in peptide design.
• Comprehensive Dataset: Created a large-scale, unified dataset (161k peptides) linking sequences to aggregation metrics and computed 3D descriptors.
• End-to-End Validation: Successfully bridged generative design with physics-based simulation, demonstrating that morphology-aware conditioning significantly outperforms random or length-only generation.

4. Results

• Generative Performance:
  • Novelty: Generated sequences are highly novel (97.7% not in training set) with low similarity to training data (<10% identity).
  • Diversity: High pairwise diversity; the model does not collapse to a few templates even under strict conditioning.
  • Condition Matching: 55% of generated sequences met all target descriptors simultaneously. Matching rates decreased as the number of constraints increased (trade-off between flexibility and strictness).
• Morphology Steering (CG-MD Validation):
  • Success Rate: 83% of the validated candidates (30 peptides total) adopted the intended morphology (spherical or fibrillar) upon visual inspection.
  • Aggregation: 100% of PepMorph candidates formed aggregates (AP $\ge$ 1.8).
  • Baseline Comparison:
    • Random Sequences: 0 spheres, 4 fibrils found.
    • Length-Only Conditioning: 3 spheres, 7 fibrils found.
    • PepMorph (Full Conditioning): 15 spheres, 15 fibrils found.
    • Conclusion: Conditioning on morphology proxies is essential for efficient discovery.
• Control Cohorts: Candidates that passed aggregation filters but failed descriptor matching showed significantly lower morphology success rates, proving that the proxy descriptors effectively steer morphology beyond just ensuring aggregation.

5. Significance and Future Directions

• Paradigm Shift: Moves peptide design from "sequence $\to$ function" to "sequence $\to$ structure $\to$ function," enabling the design of materials with specific geometric properties.
• Scalability: The framework is architecture-agnostic and can easily incorporate new descriptors (e.g., antimicrobial activity) or target new morphologies (sheets, nets) without retraining the core model.
• Limitations & Future Work:
  • Proxy Reliability: The current proxies (isolated monomer properties) are imperfect; some sequences with high hydrophobicity formed fibrils instead of spheres in atomistic simulations.
  • Data Sparsity: Rare features (like $\beta$-strands in short peptides) are underrepresented, limiting the model's ability to satisfy those specific constraints.
  • Next Steps: Integrating machine-learning interatomic potentials for higher-fidelity training labels and potentially conditioning directly on simulation-derived morphology metrics (like RMOI) in a closed loop.

In summary, PepMorph demonstrates that leveraging isolated peptide descriptors as proxies within a masked conditional generative model is a viable and powerful strategy for the rational design of peptide self-assemblies with controlled morphology.