How AI Uncovered 386 Novel Antimicrobial Peptides from Global Venoms

Database Construction

Four venom repositories—ConoServer, ArachnoServer, ISOB, and VenomZone—were merged to create a global venom database containing 16,123 unique venom proteins from six taxonomic groups (snakes, scorpions, spiders, cone snails, sea anemones, insects). A sliding‑window approach (8–50 amino acids) generated more than 40 million venom‑encrypted peptides (VEPs).

APEX Deep‑Learning Model

APEX is a sequence‑to‑function neural network trained on internal peptide data and public antimicrobial peptide (AMP) entries from DBAASP. It predicts the minimum inhibitory concentration (MIC) of each peptide against 34 bacterial strains, enabling strain‑specific activity estimation.

Screening Pipeline

Generate peptide fragments (8–50 aa) from the 16,123 venom proteins using a sliding window.

Predict MIC values for each fragment with the APEX model.

Remove peptides that show high sequence similarity to known AMPs, retaining structurally novel candidates.

Applying a median MIC ≤ 32 µM threshold yielded 7,379 VEPs; similarity filtering reduced the set to 386 low‑similarity candidates. Fifty‑eight of these were chemically synthesized for laboratory testing.

Experimental Validation

Antimicrobial assays against a panel of pathogenic bacteria showed that 53 of the 58 tested peptides (91.4 %) inhibited at least one strain. All candidates derived from the ArachnoServer subset were active, highlighting spider venom as a rich source. Peptides with low hydrophobicity and low net charge displayed weaker activity, indicating the importance of these physicochemical properties for membrane interaction.

Circular dichroism (CD) spectroscopy revealed that VEPs are largely unordered in aqueous solution but adopt α‑helical conformations in membrane‑mimicking environments (SDS micelles or TFE/H₂O mixtures), a characteristic shared by many known AMPs.

Fluorescence uptake assays identified 23 peptides that efficiently permeated the outer membrane; Arachnoserver‑18, ConoServer‑6, and ConoServer‑7 showed the strongest effects. Membrane depolarization experiments indicated that most VEPs disrupt the cytoplasmic membrane, supporting a primary mechanism of membrane destabilization.

Technical Implementation

Running APEX requires Python 3.9, a compatible PyTorch version, and dependencies such as numpy, scipy, and matplotlib. Input peptide sequences are stored in a plain‑text file; the model is executed via a command‑line script that outputs a CSV file with predicted MIC values for each bacterial strain.

Key Outcomes

Global venom database: 16,123 proteins, >40 M VEPs.

APEX predictions: strain‑specific MIC for 34 bacteria.

Final candidate set: 386 low‑similarity VEPs, 58 synthesized, 53 experimentally active.

Reference

Nature Communications, “Computational exploration of global venoms for antimicrobial discovery with Venomics artificial intelligence”. DOI: https://www.nature.com/articles/s41467-025-60051-6