The current in silico methods used for selection leave a lot to be desired https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1000048 https://www.cell.com/immunity/fulltext/S1074-7613(17)30042-0 They are also biased.
The ATLAS platform ($GNCA) should be able to overcome this. It enables them to identify relevant MHC class I (CD8+) and/or II (CD4+) restricted neoantigens. Once the mutations have been identified, they clone and express these into E. coli strains (+/- listeriolysin O to enable presentation via both MHC class I or class II). Then for each patient, these get co-cultured with their APCs (monocyte derived dendritic cells) and then T-cells are added and incubated overnight. After this, antigen-specific responses are detected by comparing cytokine production in response to each of these against a non-immunogenic control protein.
Dr. Rosenberg and his group at the NCI use a similar method ($IOVA have access to it). Tandem minigene constructs are used rather than E. coli strains. Also analysis of activation molecules (4-1BB and/or OX40) is undertaken. If a single construct shows reactivity then each point mutation in it gets converted back to wild type and the same process is run again. This time they are looking for no reactivity.