AI-based methods can help in studying antibiotic-resistant bacteria

Image

– Bacteria have always had molecular mechanisms to fight other microorganisms, including the ability to produce antibiotics. The introduction of antibiotics for prevention and treatment of infections exacerbated the selection pressures in bacteria – the human use of antibiotica became a strong driver for their evolution. Antibiotic resistance genes, acquired through mutations in existing genes or horizontal gene transfer, are the primary cause of bacterial resistance. To my knowledge, AI methods based on language models have not been used earlier to identify antibiotic resistance genes, or to predict the susceptibility of bacteria to different antibiotics.

The thesis contains three papers. In the first, the researchers did a wide characterization of the diversity of the resistance genes in nature including both well-studied established resistance genes found in clinics and computationally predicted, so called latent genes, not currently found in existing repositories. Bioinformatics and statistical methods were used to handle and analyze the very big amount of raw data, 150 terabytes, with a wide range of environmental samples from different species. It was discovered that most of the resistance genes found in these environments were latent genes, and also that many latent genes were widely spread between environments and present in pathogens, thus posing a threat to humans. More about this can be read in an earlier article.

Better results with AI

The second paper is a continuation of the first but uses AI methods to identify the resistance genes. The idea is to use NLP (natural language processing) tools to analyze protein data in order to identify both known and novel resistance genes. The transformer-based model analyze fragments’ sequences in a context-dependent way, which so to speak is the “language” of the resistance genes. The results of this method in identifying fragments associated with antibiotic resistance genes seem to be better than for any other, also compared to the one used in the first paper. This paper has yet not been published, so there will be reasons to come back with further details on this.

Finally, a new AI-based method to predict antibiotic susceptibility profiles is presented, through a combination of transformers and neural networks. When you are treated for a bacterial infection and the case is not emergent, a sample of the bacteria is tested for different antibiotics to see which it is susceptible to, i. e. dies from. To prevent a long and inefficient treatment you need an antibiotic that is known to work, and these tests can take time. The ideal would be to be able to predict the results of many antibiotics from the result of only a few in a multi-output model. Also, different bacteria behave differently for different people, so patient data should also be used. The model in this paper complements missing resistance information and seems to work for most antibiotics with a high confidence level, though better for quinolones antibiotics than for cephalosporins. This paper also awaits its publication but is currently under review.

Wants to combine mathematics and medicine

Juan studied actuarial sciences at bachelor level in Mexico and then took his master’s degree in Uppsala. But he has also always been excited about biology and medicine – his father is a doctor – and so took a bachelor’s degree in life science and a master’s degree in biomathematics as well, to try to combine the subjects. His dream has been to help human health with mathematical models, and in this he thinks AI can give much support. In the future, Juan would like to continue to explore AI in combination with multiomics analysis, in academia or in industry, maybe as a postdoctor. But nothing is yet decided.

– The best part of my doctoral years has been the opportunity to travel to research schools and conferences, to present posters and give talks. There are grants you can apply for and other support for this and it has been amazing! The main difficulties have been my own negative thoughts – do I really belong here, what is the future like – the imposter syndrome. But all people at the mathematics department have been really kind, they care for the students and are always easy to reach when needed.