Antimicrobial peptides
From peptide sequences to reliable predictions: understanding what our models can tell us, and what still needs an experiment.
Small molecules, complex questions
Antimicrobial peptides are a diverse group of molecules studied for their ability to act against microorganisms. Their activity depends on sequence and molecular properties, but also on the organism and the conditions in which a peptide is tested. A single activity label can hide much of that context.
Computational tools can help prioritise sequences for further investigation. To be useful, however, their predictions need to hold up beyond the dataset used to build the model. That makes the choice of training examples, negative data and evaluation strategy central to the biological question.
Questions we explore
- What distinguishes a useful peptide predictor from a model that performs well only on a familiar dataset?
- How does the choice of negative examples influence a benchmark?
- Where do predictions agree with experimental observations, and where do they fail?
Our approach
We develop and examine computational methods for peptide research, with a particular focus on fair evaluation and reproducibility. Our benchmarking work asks how the construction of negative datasets influences apparent model performance.
We also connect predictions with experimental testing. Our study of selected short amyloids predicted to be antimicrobial illustrates why a model score alone does not establish biological activity. Together, these projects help explain both the possibilities and the limits of sequence-based prediction.
Tools, projects & people
Find AmpGram and AMPBenchmark in our software and databases catalogue, or explore the full publication list.
Publications & research stories
A closer look at AMP benchmarks
How negative data selection can bias the evaluation of antimicrobial peptide predictors.
Putting predictions to the test
Experimental testing of selected short amyloids predicted to have antimicrobial activity.
The peptide prediction landscape
A review of peptide activity prediction tools, availability and reproducibility.