Proteins run the show of life. They move where the body needs them at exactly the right times and in the correct amounts. They give each cell its particular identity and unique abilities. But inside some proteins lies a hidden weapon.

Deep within their unique jumbles of amino acids, some proteins encode short sequences of amino acids that, when released, have powerful antibiotic effects.

These small molecules belong to the family of peptides called antimicrobial peptides or host defense peptides (HDPs). They may only be eight to 50 amino acids long, but HDPs provide an ancient evolutionary form of innate immunity. On top of fighting off microbial infections in organisms as diverse as bacteria, plants, frogs, and mammals (1), they can modulate the immune system, heal wounds, and even fight off cancer cells (2-4).

Scientists have known about HDPs since the 1930s. René Dubos, a microbiologist at the Hospital of The Rockefeller Institute for Medical Research, discovered in 1939 that the peptide gramicidin from Bacillus brevis bacteria protected mice from Streptococcus pneumoniae infections (5). In the 1990s, scientists realized that while the mammalian protein lactoferrin had some inherent antimicrobial activity, when proteins in the stomach cleaved it, it split into smaller peptides with an even stronger antimicrobial punch (6). With antibiotic resistant infections projected to become the leading cause of death worldwide, scientists are mining the biological world for new HDPs.

“Proteins are encoded in amino acid code. If we figure out ways to systematically explore that code to try to find interesting potential drugs, that’s a huge really unexplored area,” said Cesar de la Fuente, a computational biologist and microbiologist at the University of Pennsylvania. The human body is “basically like a huge treasure trove of new potential antibiotics.”

With the development of new computational tools combined with artificial intelligence, scientists are not only searching the human proteome for new HDPs, but they are also creating new ones. These newly discovered and designed HDPs present a new hope for combating antibiotic resistance.

Hidden human peptides

When searching for HDPs, scientists have a few clues they can follow. In addition to being short, these peptides contain a high proportion of hydrophobic and positively charged amino acids.

“The positive charge allows the molecule to interact electrostatically with the negatively charged bacterial membrane,” explained de la Fuente. “Then the hydrophobic domain allows it to translocate inside the membrane, creating a pore, which usually leads to bacterial cell death.”

In 2017, biochemist and computational biologist Eugenio Notomista and his team at the University of Naples Federico II used these characteristics to design an algorithm to identify hidden HDPs within proteins (7). De la Fuente joined up with Notomista’s team, and together they used the HDP finding algorithm to identify multiple cryptic peptides within the stomach protein pepsin A (8). The HDPs from pepsin A had no toxic effects on human cells, and they significantly reduced the number of Pseudomonas aeruginosa bacteria present in a mouse skin infection.

Inspired by this success, de la Fuente decided to search the entire human proteome for new HDPs. He and his team found more than 2,000 in their search (9). They then synthesized 55 of these HDPs and tested their antimicrobial activities against eight bacterial pathogens from the World Health Organization’s pathogen watch list. These include clinical isolates of Escherichia coli, Staphylococcus aureus, and Klebsiella pneumoniae among others. Of the 55 HDPs tested, 63.6% of them were active against these pathogens in vitro.  Two of those HDPs chosen for further testing reduced the level of bacteria in two different mouse skin infections. 

“It's quite amazing that you can have essentially a computational tool to explore biology,” said de la Fuente. “But then to actually validate that and combine that machine intelligence with human intelligence… that's really what excites me.”

De la Fuente and his team also discovered that while bacteria eventually developed resistance to conventional antibiotics, they never became resistant to the HDPs the researchers tested. HDPs are unique in this way because they target the bacterial membrane as a whole.

Proteins are encoded in amino acid code. If we figure out ways to systematically explore that code to try to find interesting potential drugs, that’s a huge really unexplored area. 
- Cesar de la Fuente, University of Pennsylvania

“For bacteria to develop resistance, they would need to mutate a lot. They would need to undergo many mutations to allow them to change the membrane in such a way that the peptides will not be active against them,” said de la Fuente. Mutating a conserved structure like the bacterial membrane is not a good use of energy, so the bacteria don’t do it.

While a lack of resistance makes these peptides exciting candidates in the race to discover new antibiotics, stability is a major roadblock when developing these compounds for therapeutic uses.

“If you incubate the peptides in human serum, they are completely degraded upon 24 hours,” said Angela Arciello, a biochemist and HDP researcher at the University of Naples Federico II. “Research literature is very rich in papers describing the potentialities of these peptides, but if you go to see clinical trials, there are very few peptides because there are these problems that we have to try to overcome.”

In a recent study, Arciello and de la Fuente identified new HDPs hidden inside the human protein apolipoprotein B (10). The naturally derived peptides killed a variety of drug resistant bacterial species both in vitro and in mice, and they also impaired the bacteria’s ability to form treatment resistant biofilms.

For Arciello and de la Fuente to get these peptides into the hands of patients, they needed to make them more stable. All amino acids exist in a particular orientation or handedness. In nature, most amino acids exist in a left handed or L-form, but the mirror image form D-amino acids are more resistant to cleavage by proteases.

To make their lead HDP candidate more likely to succeed in clinical trials, Arciello and de la Fuente replaced all of the peptide’s L-amino acids with the same amino acids in the D-form. They also reversed the order of the peptides so that the mirror image peptide maintained its same bacteria fighting activity.

When the researchers tested this optimized peptide in mice with a drug resistant skin infection, they found that it completely cleared the infection using half the dose of the natural peptide. De la Fuente plans to continue testing this engineered HDP in toxicity studies, and he hopes to eventually move it into clinical trials. In the meantime, Arciello and her team formulated the optimized HDP into a hydrogel as a way to one day help treat skin infections (11).

Coding compounds for the future

With the success of his team’s algorithm to identify hidden antimicrobial peptides, de la Fuente wondered if they could create new antibacterial HDPs using artificial intelligence. Their first challenge was how to encode a peptide in a way that a computer could understand.

“We had to translate the chemical complexity of, in this case, a peptide or peptides into the binary code of ones and zeros,” said de la Fuente. Then they had to train the computer to create new HDPs. “In the end, we decided to copy really the greatest engine that we have for innovation, which is evolution itself,” he added.

Instead of waiting millions of years for nature to select a great antibacterial peptide, de la Fuente’s system created new molecules on the order of days to weeks. The researchers synthesized some of the molecules that the computer created and discovered that one killed drug resistant bacteria both in a dish and in mice (12). This was the first antibiotic created by a computer that effectively treated infected mice.

Scientists are embracing the potential of machine learning and artificial intelligence to create and discover molecules with antibiotic potential. James Collins, a bioengineer at the Massachusetts Institute of Technology, and his team recently used an artificial intelligence based approach to screen more than 100 million molecules for potential antibiotic activity (13). They identified a compound called halicin, which will likely enter clinical trials soon.

“We're very excited about efforts in our lab and in other groups where now deep learning is being used as a design tool,” said Collins. By “enabling us to begin to put together molecular building blocks to create compounds with desired properties… we will much better explore a very large chemical space.”

The use of artificial intelligence and machine learning in antibiotic discovery is still in its infancy, but scientists like de la Fuente and Collins are excited to explore its potential.

“Machines are getting a little bit better at both designing and discovering things that actually could be beneficial and that actually work,” said de la Fuente. “To see that sort of hybridization between machines and humans, I think is quite exciting.”

References

1. Zasloff, M. Antimicrobial peptides of multicellular organisms. Nature  415, 389-395 (2002).
2. Mansour, S.C. et al. Host defense peptides: front-line immunomodulators. Trends Immunol  35, 443-450 (2014).
3. Pushpanathan, M. et al. Antimicrobial Peptides: Versatile Biological Properties. Int J Pept  2013, 675391, (2013). 
4. Riedl, S. et al. Membrane-active host defense peptides – Challenges and perspectives for the development of novel anticancer drugs. Chem Phys Lipids  164, 766-781 (2011).
5. Dubos, R.J. Studies on a Bactericidal Agent Extracted from a Soil Bacillus: II. Protective Effect of the Bactericidal Agent Against Experimental Pneumococcus Infections in Mice. J Exp Med 70, 11-17 (1939).
6. Sinha, M. et al. Antimicrobial lactoferrin peptides: the hidden players in the protective function of a multifunctional protein. Int J Pept  2013, 390230 (2013).
7. Pane, K. et al. Antimicrobial potency of cationic antimicrobial peptides can be predicted from their amino acid composition: Application to the detection of “cryptic” antimicrobial peptides. J Theor Biol  419, 254-265 (2017).
8. Pane, K. et al. Identification of Novel Cryptic Multifunctional Antimicrobial Peptides from the Human Stomach Enabled by a Computational–Experimental Platform. ACS Synth Biol  7, 2105-2115 (2018).
9. Torres, M.D.T. et al. Mining for encrypted peptide antibiotics in the human proteome. Nat Biomed Eng  6, 67-75 (2022).
10. Cesaro, A. et al. Synthetic Antibiotic Derived from Sequences Encrypted in a Protein from Human Plasma. ACS Nano  16, 1880-1895 (2022).
11. Cesaro, A. et al. Novel Retro-Inverso Peptide Antibiotic Efficiently Released by a Responsive Hydrogel-Based System. Biomedicines  10, 1301 (2022).
12. Cardoso, M.H. et al. A Computationally Designed Peptide Derived from Escherichia coli as a Potential Drug Template for Antibacterial and Antibiofilm Therapies. ACS Infect Dis  4, 1727-1736 (2018).
13. Stokes, J.M. et al. A Deep Learning Approach to Antibiotic Discovery. Cell  180, 688-702.E13 (2020).