PepMorph: Conditional Peptide Design

• PepMorph is an end-to-end peptide discovery framework that uses a Transformer-based CVAE to generate sequences with targeted aggregate morphologies.
• It employs a binary masking mechanism on peptide descriptors to condition generation, achieving high success rates and statistical diversity.
• The framework integrates coarse-grained MD simulations for validation, enabling reliable design of peptide-based biomaterials in biomedical and energy applications.

PepMorph is an end-to-end peptide discovery framework designed to generate novel sequences that reliably self-assemble into specified aggregate morphologies (such as fibrillar or spherical forms). The system addresses longstanding barriers in peptide material design, where the mapping from sequence to aggregate shape is highly nontrivial due to complex dependencies on geometric and physicochemical properties. By leveraging a Transformer-based Conditional Variational Autoencoder (CVAE) equipped with a masking mechanism, PepMorph enables conditional generation of peptides under arbitrary design constraints and validates resulting sequences through coarse-grained molecular dynamics simulations. This approach achieves high success rates in realizing intended morphologies, with robust statistical diversity and fidelity, positioning PepMorph at the forefront of application-driven peptide discovery (Costa et al., 2 Sep 2025).

1. Framework Motivation and Significance

PepMorph was created to overcome two major limitations in peptide design: the difficulty of discovering sequences with targeted self-assembly behavior, and the intractability of sequence-based exploration with direct experimental validation. Peptides possess vast combinatorial sequence space due to the 20 canonical amino acids, and minor variations can induce fundamentally different supramolecular outcomes. Traditional aggregation screening approaches are limited by sparse coverage and noisy simulation predictions.

PepMorph integrates morphology as an explicit design objective by compiling a large dataset containing aggregation propensity values, self-assembly labels, and 3D structure-derived descriptors. This enables guided generation of aggregate-prone peptides with control over their eventual morphology, such as favoring spherical or fibrillar outcomes.

A plausible implication is that PepMorph's pipeline bridges the gap between unconstrained generative exploration and fully targeted material design—facilitating bottom-up synthesis of peptide-based biomaterials for biomedicine and energy applications.

2. Dataset Compilation and Descriptor Engineering

The PepMorph corpus is assembled from multiple aggregation propensity datasets (notably Wang et al. with ~62,000 entries; Teijlingen et al. with ~60,000), supplemented by ~39,000 randomly sampled peptides. After deduplication and enrichment, the combined set includes approximately 161,000 unique peptides, each annotated with:

• Aggregation propensity (AP) from MD ratio-of-solvent-exposure,
• Binary self-assembly (SA/no-SA) labels,
• Isolated descriptors from predicted 3D structures (PEP-FOLD): sequence length, β-sheet content (binary), hydrophobic moment, net charge, and additional geometric variables.

Descriptor coverage is partial (about 50% for 3D-derived metrics), creating statistical imbalances such as sparsity in β-sheet-positive peptides. These challenges are programmatically addressed via the model's masking mechanism during data preprocessing and downstream generation.

3. Conditional Generation Architecture

PepMorph employs a Transformer-based CVAE, wherein peptide descriptors function as conditioning variables. The framework's haLLMark is a binary masking strategy allowing partial or full specification of design constraints:

• Let $\mathbf{c} \in \mathbb{R}^d$ be the descriptor vector, %%%%1%%%% the mask, with $m_i=1$ for specified elements.
• The context summary is computed as $$s = \phi([\mathbf{c} \odot \mathbf{m}, \mathbf{m}])$$, using a multilayer perceptron $\phi(\cdot)$ and elementwise multiplication.
• The masked context $s$ parameterizes the conditional prior $p(\mathbf{z} | s)$ for latent variable $\mathbf{z}$.

The decoder is an autoregressive Transformer, conditioned jointly on a latent token (from $\mathbf{z}$) and a condition token (from $s$). Cross-attention ensures the generated peptide sequence honors the prescribed constraints while permitting unconstrained variation elsewhere.

Training minimizes a composite loss: $$L = L_{\mathrm{rec}} + \beta L_{\mathrm{KL}} + \lambda_{\mathrm{mask}} L_{\mathrm{mask}} + \lambda_{\mathrm{bin}} L_{\mathrm{bin}} + \lambda_{\mathrm{cont}} L_{\mathrm{cont}}$$ where $L_{\mathrm{rec}}$ is the cross-entropy reconstruction loss on amino acid tokens, and $L_{\mathrm{KL}}$ is the Kullback-Leibler divergence for the encoder posterior $q(\mathbf{z}|\mathbf{y})$ against the masked prior $p(\mathbf{z}|s)$, using a closed formula for diagonal Gaussians: $$L_{\mathrm{KL}} = \frac{1}{N} \sum_n \left[ \frac{1}{2} \sum_k \left( \log \frac{\sigma^2_{\text{prior}}}{\sigma^2_{\text{enc}}} + \frac{\sigma^2_{\text{enc}} + (\mu_{\text{enc}}-\mu_{\text{prior}})^2}{\sigma^2_{\text{prior}}} - 1 \right) \right]$$ Auxiliary terms ($$L_{\mathrm{mask}}, L_{\mathrm{bin}}, L_{\mathrm{cont}}$$) regulate the condition summary and descriptor encoding.

4. Statistical Performance and Simulation-Based Validation

PepMorph's generative outputs are validated across several axes:

• Novelty: ~13,000 generated sequences contain only ~300 exact matches with training data.
• Diversity: Mean pairwise edit distance is such that one third of sequence positions differ from closest training entries.
• Condition Matching: Effectiveness drops from 84.58% (single condition) to 24.55% (all six conditions) as constraints increase; essential features such as peptide length and aggregation propensity are reliably controlled.

Molecular validation applies a filter–recompute–simulation protocol:

1. Dual-stage filtering by predictive models (AP, SA) and recomputed descriptors,
2. 3D structure prediction using PEP-FOLD,
3. Coarse-grained MD simulations of top 15 candidates for each target morphology,
4. Classification by Ratio of Moments of Inertia (RMOI): $RMOI = L_x / L_z$, where $L_x$, $L_z$ are smallest/largest eigenvalues—spherical for $RMOI\ge0.75$, fibrillar for $RMOI\le0.35$.

Visual cross-validation of simulation outputs yielded an overall 83% success rate in matching intended aggregate morphology after filtering. The generated candidates displayed lower average sequence similarity (<10%) compared to competitor methods (>40%).

5. Applications in Biomedical and Energy Domains

PepMorph's ability to produce peptides that aggregate into designated morphologies enables several practical applications:

• Biomedical Materials: Custom assemblies for tissue scaffolds, drug delivery systems, biosensation platforms.
• Energy Materials: Rational design of peptide-based structures with tailored electronic/photonic properties (emergent from aggregate geometry).

The underlying conditional generative strategy, enabling descriptor-driven navigation of sequence space, could be extended to new material platforms and morphologies beyond fibrillar/spherical classes.

A plausible implication is that the integration of partial conditioning and robust validation cycles allows systematic exploration of designable biomaterial space.

6. Methodological Limitations and Prospects for Enhancement

Challenges include underrepresentation of certain descriptor combinations (notably β-sheet-positive structures), limitations in structural prediction (PEP-FOLD), and the expressiveness of the unimodal diagonal Gaussian prior in capturing highly multimodal condition–morphology mappings.

Future work will focus on:

• Expanding peptide length and increasing descriptor granularity,
• Addressing data sparsity via curated acquisition and expert-guided descriptor selection,
• Upgrading model priors to multimodal or mixture-of-Gaussian forms to better handle partial observability,
• Integrating higher-fidelity molecular models (e.g., machine-learning interatomic potentials) with the CG-MD validation step.

This suggests the pipeline will become more robust, accurate, and extensible for complex material architectures.

7. Summary and Outlook

PepMorph represents a technically rigorous framework for conditional peptide discovery, integrating masking-enabled generative modeling with simulation-based validation for morphology-specific sequence design. Its data-driven approach, unique handling of partial constraints, and demonstrated 83% morphology control fidelity present significant advances in application-driven peptide design. The model's capacity for novelty, diversity, and descriptor-based controllability positions it as a methodological reference for future exploration of sequence–structure–function relationships in supramolecular peptide chemistry and engineering (Costa et al., 2 Sep 2025).