Team:Evry/Modeling

Machine learning

We are currently working on a machine learning approach, in order to predict which antigen will be presented efficiently to the MHC-I and MHC-II.

We plan to create a pipeline allowing us to select the best candidate tumoral antigen to use for a vaccine. We already have a few ideas on criterias that could be used to choose such an antigen:

  • Antigens of oncolytic viruses (HPV for instance) can be tumor-specific.
  • The antigen can be a mutated version of a gene that is found in healthy tissue: for instance, a mutated p53 gene.
  • Testis-like antigens (normally found in gonad) can be expressed by a tumor because they are growth factors. However, using this type of antigens could have side effects on the gonad: evaluating if it is worth it to take such a risk would therefore be needed.

Once we are able to determine the most relevant targets, we won't express the whole corresponding proteins. We only need to express short fragments of them, corresponding to putative cleavage sites by the proteasome to link to MHC-I and MHC-II. Some tools are available to predict those sites, even if their accuracy is far from optimal. Those predictions could also be improved.

Scroll to top To top