Skip to content

Agentic Discovery of Non-Canonical Antimicrobial Peptides with AMPGAN v3

🕒 Published (v1): 2026-06-15 16:39 UTC · Source: Arxiv · Venue: ICML 2026 · link

Why this paper was selected

ICML 2026 top venue; agentic discovery harness for non-canonical peptide generation

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AMPGAN v3 is a multi-objective conditional GAN that extends AMP generation to non-canonical chemistry (D-amino acids, N/C-terminus modifications) via dual specialized discriminators. The authors pair it with PepCraft, a multi-agent orchestration framework in which a Planning Agent iteratively dispatches Generating, Filtering, and Verifying executor agents. Five in vitro–tested candidates validated the pipeline, with the best reaching MIC 8 µg/mL against B. subtilis.

Problem

Prior generative AMP models are restricted to canonical L-amino acid sequences, precluding the chemical modifications (D-amino acids, amidation) essential for proteolytic stability in real therapeutic contexts. Separately, post-generation curation (physicochemical filtering, database cross-referencing) is repetitive and scales poorly with output volume when done manually; no agentic system existed for end-to-end AMP discovery.

Method

AMPGAN v3 formulates generation as adversarial multi-objective learning. The generator \(G_\theta: \mathcal{Z} \times \mathcal{C} \rightarrow \mathbb{R}^{L \times V}\) takes a latent vector \(z \sim \mathcal{N}(0, I)\) and condition tuple \(c = (s, m, \ell)\) (species, MIC target, length). FiLM-based conditioning modulates the latent code: \(z' = \sigma(\alpha) \cdot z + \gamma(c) \odot z + \beta(c)\), followed by transposed convolutions and a Transformer encoder. Two discriminators are kept separate: \(D_\text{adv}\) is an unconditional real/fake classifier; \(D_\text{mic}\) is trained exclusively on real sequences to regress MIC conditioned on species, and is frozen during generator updates. Discrete samples are obtained via Gumbel-Softmax. Total generator loss is \(\mathcal{L}_G = \lambda_\ell \mathcal{L}_\text{len} + \lambda_s \mathcal{L}_\text{sim} + \lambda_a \mathcal{L}_\text{adv} + \lambda_m \mathcal{L}_\text{mic}\).

PepCraft is a multi-agent harness: a Planning Agent \(P\) receives a user objective \(\theta_0\) and prior report \(r_{t-1}\), selects an executor \(e\), issues a natural-language instruction \(i_e^t\), receives report \(r_t = e(i_e^t)\), and iterates until the objective is fulfilled. Specialized executors are: (1) Generating Agent (calls AMPGAN v3 + SimpleFold), (2) Filtering Agent (charge, hydrophobicity, length, then structure-dependent filters), (3) Verifying Agent (BLAST against SwissProt and DBAASP for novelty/homology).

Key Contributions

  • First generative AMP model supporting D-amino acids and terminal modifications (amidation), expanding chemical vocabulary beyond canonical L-amino acids.
  • Dual-discriminator architecture that separates distributional realism from activity-aware supervision, resolving GAN training instability: AMPGAN v3 succeeds on ~30/30 seeds vs. AMPGAN v2's 3/30.
  • PepCraft: a multi-agent Planning Agent framework that orchestrates end-to-end AMP discovery with adaptive re-invocation of specialized executors.
  • In vitro validation of 5 structurally diverse (α-helix, β-hairpin, random coil) candidates; PepCraft prioritization recommendations aligned with wet-lab outcomes.

Results

  • In vitro activity: 2/5 synthesized candidates active against Gram-positive strains; best candidate HY-P60322 achieved MIC/MBC = 8/8 µg/mL vs. B. subtilis and 16/32 µg/mL vs. S. epidermidis.
  • Training stability (N=30 seeds): AMPGAN v3 achieves near-perfect stability vs. AMPGAN v2's 3/30 success rate.
  • Predicted-AMP rate (N=2,000 unconditional generations, two external classifiers): AMPGAN v3 92.7%/95.1%, vs. AMPGAN v2 80.0%/83.4%, HydrAMP 81.0%/83.0%, AMP-Designer 92.4%/83.9%.
  • Generated sequences concentrate within AMP-active physicochemical ranges (charge \([+1,+9]\), GRAVY \([-1,+1]\)); only Arg and Trp are slightly overrepresented (~2.5 pp each).

Limitations

  • Target-species conditioning did not produce species-selective activity: both active candidates inhibited Gram-positive strains not targeted during conditioning, suggesting conditioning translates to potency tendencies but not species specificity.
  • Validation scale is small (5 peptides); no Gram-negative activity observed.
  • PepCraft evaluation is qualitative (alignment with in vitro outcomes) rather than a controlled ablation against non-agentic baselines.
  • GAN training still requires structured tokens to avoid the single-token collapse failure mode; full stability analysis of the revised architecture is not quantified beyond pass/fail counts.

Relevance to Harnesses / Meta-Harnesses

PepCraft is a textbook Planning-Agent meta-harness: a coordinator that decomposes an objective into natural-language instructions, dispatches typed executor agents, inspects reports, and adaptively re-invokes executors with revised parameters — the defining loop of iterative orchestration harnesses. The key harness design choice is strict executor specialization (generate / filter / verify) with the planner as the sole control-flow authority, separating domain logic from orchestration logic. The paper provides a small-scale but end-to-end example of integrating a generative model as a tool-bearing executor within such a harness, with the notable empirical result that the harness's prioritization aligned with wet-lab outcomes. For harness researchers, this is a concrete case study of agentic AI composing with a domain-specific generative model under real-world validation constraints.