AI-designed antimicrobial peptides: new promise to fight antibiotic resistance
AI-assisted techniques can find new antimicrobial peptides in just few months
The most alarming threat of the decades to come is not the global warming, rise of the machines or zombie-apocalypse. It is antibiotic resistance. Uncontrolled and excessive use of antibiotics has lead us to the point, when more and more multi-resistent strains of bacteria emerge. Eventually we risk to return into the 19-th century when none of “common” bacterial infections are actually treatable. Antibiotic resistance already claims about 700,000 lives annually. Global spread of “superbacteria”, resistant to all know antibiotics, would be the end of modern civilization as we know it thus the search for new antimicrobial agents is one of the hottest topics in the drug discovery.
Antimicrobial peptides (AMPs) are considered as one of possible answers to the antibiotics crisis. The AMPs are short peptides with 12–50 amino acids, which were first found in various animals to combat pathogenic bacteria. Natural AMPs are very diverse in terms of structure and chemical properties, but the most of them are cationic (positively charged) and amphiphilic (possessing both water-loving and water-hating regions).
The antimicrobial peptides are studied actively for decades. My own post-doctoral project 15 years ago was devoted to the natural cell-penetrating peptides. By that time researchers were severely limited by computational resources and algorithms, but the mechanisms of action of AMPs were established successfully. Such peptides are acting by disrupting bacterial lipid membranes: positively charged amino acids anchor them to the negatively charged anionic lipids, which are abundant in bacteria. After that the peptides penetrate the lipid bilayer and aggregate inside forming either water-filled pores or unstructured “piles of logs” which compromise the membrane integrity.
There is an astronomical number of possible peptides of given length and it is notoriously hard to find the rules for their rational design. However, such enormous chemical space could be the key point in fighting the drug resistant. If we find a way to design a dozen of effective antimicrobial peptides each months the bacteria would just not be able to adapt to them equally fast. The problem is that currently we are very far from such efficiency.
Recent paper in the Nature Biomedical Engineering journal proposes a pragmatic approach to the design of antimicrobial peptides. The method combines the power of AI and deep learning generative autoencoder model with the classical in silico molecular simulations.
Application of these scheme allowed to discover two novel AMPs with a broad spectrum of antimicrobial action in just 48 days — an unprecedented efficiency!
The authors utilized the idea of Variational autoencoder (VAE) in the form of Wasserstein Autoencoder (WAE). In this method the peptide generation problem is expressed as a density modeling mathematical problem: the model samples the peptide sequence space in such a way that the regions with high probability density are mostly involved. The density estimation algorithm is tuned to assign a high likelihood to known peptides and to “punish” random meaningless sequences.
The number of known AMP sequences is relatively small (about 9,000), which is not enough for efficient training of VAE/WAE encoder. Thus the authors also used 1.7 million peptide sequences from the UniProt database. This allowed sampling all known protein families and all known three-dimensional folds from many organisms, which improved the model quality dramatically.
Generation of peptide sequences was further controlled by a set of binary (yes/no) attributes of interest, such as antimicrobial function and toxicity. Conditional Latent Space Sampling (CLaSS) approach was used for such filtering.
The AI generator produced approximately 90,000 AMP sequences. The 163 candidates were classified to have antimicrobial function, broad-spectrum efficacy, presence of stable secondary structure and toxicity. This set of peptides was subject to the classical coarse-grained molecular dynamics simulations. The goal was to assess the peptide–membrane interactions and to select those peptides, which are likely to bind strongly to the membranes. The number of contacts between the positive residues and membrane lipids was used as predictor of antimicrobial activity.
After this in silico high-throughput screening only 20 peptide sequences remained in the set of promising candidates. They were transferred to the wet lab characterization of antimicrobial activity against Gram-positive Staphylococcus aureus and Gram-negative E. coli model bacteria strains. Two best sequences have shown very good antimicrobial activity and low toxicity.
The authors emphasize that even this first proof-of-concept study demonstrated quite impressive 10% success rate among generated peptide candidates. In combination with a rapid turnover of the whole pipeline (only 48 days) this looks very promising for industrial-grade search for antimicrobial peptides.
The Receptor.AI is not standing apart from the challenges of antibiotic resistance. We are currently developing a project on the rational AI-assisted design of antimicrobial peptides with changeable conformation. In addition our AI-based drug discovery platform could be used to generate highly credible lead small molecules for microbial target proteins. We also utilize combination of AI-based molecule generation and DTI assessment with subsequent in silico screening by means of molecular dynamics and molecular docking simulations. This approach generates small number of high-quality drug candidates, which have high success probability in pre-clinical and clinical trials and minimizes time and costs of the drug discovery projects.