Iterative identification of papers for developing the classifier. The iterative process was initiated with papers selected with rule-based filters for expert review. We then leveraged the SPECTER embedding to select papers for review based on their distances in the embedding space. Once an initial training set was labeled, we fit an SVM classifier with the 316 abstracts and used it to select 86 additional papers. An SVM classifier fitted to the 401 abstract, predicted 91 highly likely papers to validate the classifier. This gave a final training set of 492 papers, 301 with immune signatures, 8 of which were ambiguous (where reviewers did not agree on the presence or absence of immune signatures)
