Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3

Technical Summary: Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3

Problem Statement
Antimicrobial resistance (AMR) causes over a million deaths annually, yet the discovery of novel antibiotic classes has stagnated since 2000. While Antimicrobial Peptides (AMPs) offer a promising solution due to their multi-mechanistic action and slower resistance emergence, current generative AI models face two critical limitations:

1. Structural and Chemical Constraints: Existing models predominantly generate $\alpha$-helical sequences composed of canonical L-amino acids. These short helical peptides often suffer from conformational instability and rapid proteolytic degradation. Real-world therapeutic applications require non-canonical chemistries, such as D-amino acids and N/C-terminus modifications (e.g., amidation), to ensure stability, which current models cannot generate.
2. Training Instability: Prior conditional GAN approaches (e.g., AMPGAN v2) suffer from severe training instability, with only ~10% of random initializations producing usable models, often collapsing to repetitive single-token outputs.
3. Workflow Scalability: Translating in-silico designs into testable candidates requires repetitive filtering and database verification steps that scale poorly when performed manually.

Methodology
The authors introduce AMPGAN v3, a multi-objective conditional Generative Adversarial Network (GAN), and PepCraft, a multi-agent framework for end-to-end discovery.

• AMPGAN v3 Architecture:

  • Expanded Vocabulary: The model is trained on a curated dataset of 34,275 experimentally validated AMP sequences (from DBAASP v3), explicitly including D-amino acids and terminus modifications. Sequences are encoded with special tokens for N/C-terminus modifications and sequence boundaries.
  • Dual-Discriminator Design: To resolve training instability, the authors separate adversarial and activity-aware supervision into two specialized discriminators:
    • $D_{adv}$: An unconditional discriminator that judges whether a sequence is plausibly peptide-like (distributional judgment).
    • $D_{mic}$: A conditional discriminator trained exclusively on real sequences to regress Minimum Inhibitory Concentration (MIC) values based on target species (potency judgment).
  • Training Objective: The generator is optimized via a weighted combination of four losses: length loss ($L_{len}$), similarity loss ($L_{sim}$), adversarial loss ($L_{adv}$), and MIC regression loss ($L_{mic}$). The model uses Gumbel-Softmax to handle discrete token sampling.

• PepCraft Agentic Framework:

  • A Planning Agent orchestrates specialized executor agents to automate the discovery pipeline.
  • Generating Agent: Translates natural language objectives into structured calls for AMPGAN v3 and structure prediction tools (SimpleFold).
  • Filtering Agent: Applies sequential physicochemical and structural screens (charge, hydrophobicity, stability index).
  • Verifying Agent: Performs BLAST homology searches against SwissProt and DBAASP to assess novelty and recover known activities.
  • The system iterates, with the Planning Agent re-invoking executors with revised parameters if intermediate results deviate from user objectives.

Key Contributions

1. AMPGAN v3 Model: The first generative model for AMPs that supports D-amino acids and terminal modifications, significantly expanding the accessible chemical space beyond canonical L-amino acids.
2. Stability Improvement: By decoupling adversarial and activity supervision, the model achieves a success rate of ~70% (21/30 runs) compared to ~10% for AMPGAN v2, mitigating mode collapse.
3. Agentic Workflow: The introduction of PepCraft demonstrates how generative AI can be composed with agentic systems for scientific discovery, allowing for adaptive control flow and automated curation.
4. Experimental Validation: The paper provides a small-scale but real-world validation of the entire pipeline, from generation to wet-lab testing.

Results

• In Vitro Validation: Five candidates spanning three structural classes ($\alpha$-helix, $\beta$-hairpin, random coil) and incorporating non-canonical chemistry were synthesized and tested.
  • Two candidates showed activity against Gram-positive strains.
  • The best candidate, HY-P60322, achieved an MIC of 8 µg/mL against Bacillus subtilis and 16 µg/mL against Staphylococcus epidermidis.
  • Notably, the active candidates did not inhibit the species used as conditioning targets during generation (e.g., HY-P60322 was conditioned on E. coli but active against B. subtilis), indicating that species conditioning did not translate to species-selective activity in this specific instance.
• In Silico Performance:
  • AMPGAN v3 achieved the highest predicted AMP rates (92.7% by AMP Scanner v2, 95.1% by amPEP) among compared models (AMPGAN v2, HydrAMP, AMP-Designer).
  • Embedding space analysis (ESM2 + PCA) showed AMPGAN v3 significantly improved fidelity (reverse KL divergence) over its predecessor while maintaining coverage, though AMP-Designer (pretrained on 600K+ peptides) showed lower divergence due to its large protein prior.
• Agentic Efficiency:
  • The multi-agent PepCraft framework required fewer tool calls than single-agent baselines for task sizes up to 20 candidates.
  • By routing the Planning Agent through a large model and Executors through smaller, cheaper models, the multi-agent approach reduced total output token usage significantly compared to single-agent systems that must route all tokens through a large model.
  • Qualitative analysis showed the agent's prioritization recommendations aligned with the actual in vitro outcomes (flagging HY-P60322 for prioritization and others for deprioritization).

Significance and Claims
The paper claims that AMPGAN v3 and PepCraft represent a step toward "examining, on a small but real scale, how generative and agentic AI compose in therapeutic peptide discovery." The authors emphasize that their contributions allow for the generation of peptides with the chemical stability required for therapeutic viability (via non-canonical amino acids) and demonstrate a workflow where agentic systems can orchestrate generation, filtering, and verification with minimal human intervention. The work is modest in scope, presenting a preliminary framework where the alignment between agent recommendations and wet-lab results is encouraging but requires further prospective evaluation. The authors note that future work will focus on refining target-aware generation and expanding the chemical vocabulary to include post-translational modifications like cyclization.