Immunodominant T cell epitopes preferentially targeted in multiple individuals are the critical element of successful vaccines and targeted immunotherapies. However, the underlying principles of this “convergence” of adaptive immunity among different individuals remain poorly understood. To quantitatively describe epitope immunogenicity, here we propose a supervised machine learning framework generating probabilistic estimates of immunogenicity, termed “immunogenicity scores,” based on the numerical features computed through sequence-based simulation approximating the molecular scanning process of peptides presented onto major histocompatibility complex (MHC) by the human T cell receptor (TCR) repertoire. Notably, overlapping sets of intermolecular interaction parameters were commonly utilized in MHC-I and MHC-II prediction. Moreover, a similar simulation of individual TCR-peptide interaction using the same set of interaction parameters yielded correlates of TCR affinity. Pathogen-derived epitopes and tumor-associated epitopes with positive T cell reactivity generally had higher immunogenicity scores than non-immunogenic counterparts, whereas thymically expressed self-epitopes were assigned relatively low scores regardless of their immunogenicity annotation. Immunogenicity score dynamics among single amino acid mutants delineated the landscape of position- and residue-specific mutational impacts. Simulation of position-specific immunogenicity score dynamics detected residues with high escape potential in multiple epitopes, consistent with known escape mutations in the literature. This study indicates that targeting of epitopes by human adaptive immunity is to some extent directed by defined thermodynamic principles. The proposed framework also has a practical implication in that it may enable to more efficiently prioritize epitope candidates highly prone to T cell recognition in multiple individuals, warranting prospective validation across different cohorts.