Antimicrobial peptides (AMPs) are rich and structurally diverse molecules that can kill pathogens by either disrupting their membranes or interacting with their intracellular targets. The lack of bacterial resistance has encouraged their therapeutic avenues against antibiotic-resistant infections.^1-2 However, developing the next generation of AMP-based antibiotics includes tackling certain barriers (low metabolic stability, proteolytic degradation, poor oral bioavailability and high toxicity) without compromising their therapeutic promises – a multi-objective optimisation.

Machine learning algorithms intertwine predictive and generative models to design optimised AMP sequences rationally.⁴ Despite a vast structural diversity, most considered AMP candidates fold into alpha-helical structures. Simple observations or voluntary model bias, the limitations within these models lie in the diversity of peptide sequences and biological information. Here, we audited the structural landscape in GRAMPA, the largest and vastly uncharted repository of 5980 AMPs after benchmarking four protein structure prediction methods (Jpred4, PEP2D, PSIPRED, AlphaFold2).⁴  The dataset contained mostly loose helices (65.1%), random coils (17.8%), while β-stranded and mixed structures accounted for the rest. We assessed 13 existing machine learning predictors for antimicrobial activity, revealing their preferences for alpha-helical sequences.⁵