Cobra toxins yield modest antimicrobial candidate through AI-guided peptide mining

A team of researchers set out to test whether artificial intelligence could unlock hidden antimicrobial compounds buried inside cobra venom. They pulled cardiotoxin sequences from public databases, fed them through machine learning algorithms designed to spot bioactive peptide fragments, and synthesized 14 candidate molecules in the lab. Then came the hard part: seeing whether the computational predictions matched reality.

The work, published in NPJ Drug Discovery, represents a systematic attempt to bridge a gap in drug discovery. Snake venoms contain thousands of proteins and peptides evolved to disrupt cells and immune systems—properties that could theoretically be repurposed as antibiotics. But traditional venom research relies on painstaking biochemical isolation from actual snake venom, a slow and expensive process. Recent advances in DNA sequencing have flooded public databases with venom protein sequences, creating an opportunity: if you could computationally extract the active fragments and predict which ones work, you could skip the wet lab work and go straight to synthesis and testing.

The researchers used two strategies to identify candidates. The first extracted conserved regions from eight well-characterized cobra cardiotoxins and ran them through the Antimicrobial Peptide Activity algorithm, yielding seven native peptides. They then designed two modified versions of one promising fragment, CTX-p5, adding either extra positive charge or a fatty tail to improve membrane binding. The second strategy built a consensus sequence from 92 cobra cardiotoxins and designed five more peptides based on the most conserved amino acids. Machine learning tools flagged CTX-p6 and CTX-p14 as the most promising candidates based on physicochemical properties, but experimental results told a different story.

All 14 peptides were synthesized using standard solid-phase chemistry, purified to above 95 percent, and their structures predicted using AlphaFold2. Most were disordered in water but adopted partial helical structures in membrane-mimicking conditions—though the computational predictions of structure did not always match what the experiments showed. When tested against both Gram-negative and Gram-positive bacteria, the results were sobering. Most peptides showed no measurable antibacterial activity at concentrations up to 500 micromolar. Those that did work required minimum inhibitory concentrations of 125 to 500 micromolar, roughly 10 to 100 times higher than clinically used antimicrobial peptides.

CTX-p5 emerged as the clear winner, consistently showing the strongest activity across multiple bacterial strains. Microscopy and biochemical assays revealed its mechanism: it punches holes in bacterial membranes, increasing permeability and disrupting the electrical gradient that bacteria use to power their cells. The peptide caused relatively little damage to human red blood cells at low doses, though toxicity increased at higher concentrations. CTX-p7, which had been chemically modified with a fatty tail, showed higher toxicity to human cells and was dropped from further consideration.

The authors are careful about what this means. They have proof of principle that combining toxin sequence mining, machine learning, and experimental validation can identify membrane-active peptides from venom. But they also demonstrate the limitations of current prediction methods. Most candidates failed. The ones that worked require concentrations too high for practical use as drugs. CTX-p5 is not yet a therapeutic lead—it is an early-stage research compound that will need significant refinement before anyone would consider giving it to a patient. The work suggests that better AI models, improved structure prediction, and more sophisticated experimental validation might eventually unlock venom-derived antimicrobials against antibiotic-resistant bacteria. For now, it shows what is possible and what remains hard.