Forecast: Predicting Antibiotic Resistance from Susceptibility Data

Background

We are engaged in a continual arms race in combatting antibiotic resistance¹. It seems inevitable that the launch of any antibiotic is met with the emergence of resistance. However, the speed at which resistance is likely to emerge and disseminate is largely unknown. As such, a key challenge is being able to predict the temporal, geographic and genetic basis for resistance development for new and old antibiotics². The Atlas dataset (https://atlas-surveillance.com/login) has curated data for almost 1M samples over a 20 year period from diagnostic labs globally. Data mining efforts from the Atlas data^3,4 with cohesive outputs (eg understanding resistance emergence, transmission, geographic hotspots) from this longitudinal susceptibility data curation project have been limited. We wish to exploit this dataset to address “How can we predict the emergence of resistance based on existing global datasets on antibiotic susceptibility and genome information?”. In previous work (L. Jurcaga, MSc thesis, U Edinburgh 2025), we have already demonstrated that the Atlas data can be used to link trends in tigecycline (novel glycycline) resistance emergence, geographic hotspots of high rates of resistance and co-linkages to other existing or new antibiotics. As such, we wish to extend this analysis using the expanded Atlas dataset (2025), with a specific focus on pathogens posing a severe and urgent threat to human health (eg Klebsiella pneumoniae) against new antibiotics such as cefiderocol, ceftazidime-avibactam with the following aims.  

Aims

1. We will utilize the Atlas susceptibility data to mine for susceptibility trends linked to isolates, antibiotics, geographic location for WHO priority pathogens E. coli and Klebsiella spp. We will investigate the dataset for statistically significant trends in resistance development and shifts in susceptibility overall and in individual countries. 

2. Based on the trends and geographic locations identified in Aim 1 we will then focus on Klebsiella spp.. We will acquire temporally and geographically matched genome sequences from NCBI Bioproject to analyse genetic changes linked to alterations in susceptibility. Here, we will undertake both analyses of (1) targeted (known genes) (2) de novo (unknown) genes. The alterations in the unknown genes will be undertaken by sifting the strain pools into key epidemic lineages and undertaking a gene presence or absence or SNP-based GWAS analyses across the population. 

3. Based on genes identified in Aim 2, we will generate a subset of these known and de novo mutations to undertake targeted antibiotic resistance evolution experiments to demonstrate alterations in susceptibility to different antibiotics. 

4. We will train probabilistic machine learning models based on relations identified in Aim 1 to trial against a retrospective collection of susceptibility data of bacterial samples collected from multiple hospitals.

Ultimately, this work will allow us to move from retrospective analysis to probabilistic models that combine real-time susceptibility, genomic surveillance data and quantify the uncertainty in their predictions. For individual patients, such models could support patient stratification, helping clinicians tailor antibiotic treatment to the respective resistance profiles. At the population level, they could flag where resistance to new antibiotics is likely to emerge next, informing stewardship and helping preserve the effectiveness of these drugs for longer.

Training outcomes

The Atlas dataset is open source and there is no need for us to obtain ethical approval. The access of the anonymised susceptibility datasets from multiple hospitals.  

References

1: Ho et al, Antimicrobial Resistance- A concise update, Lancet Microbe (2025) 

2. Cesaro et al, Challenges and applications of artificial intelligence in Infectious Diseases and Antimicrobial Resistance, npj Antimicrobials and Resistance (2025) 

3. Valavarasu V et al, Prediction of Antibiotic Resistance from Surveillance Data using Machine Learning, Sci Reports, (2025) 

4. Catalan, P., et al, Seeking Patterns of Antimicrobial Resistance in Atlas, an open raw MIC database with patient metadata, Nature Communications (2022)