AI Pipeline Finds Hidden Cancer Targets That Trigger Real T Cell Attacks

Cancer immunotherapy depends on identifying neoantigens — peptide fragments derived from tumor-specific mutations that are displayed on the surface of cancer cells by MHC (major histocompatibility complex) molecules, where they can be recognized and attacked by T cells. Finding the right neoantigens is technically difficult because most computational predictions rely on genomic data alone and generate high false-positive rates. This study addresses that bottleneck by grounding neoantigen discovery in direct mass spectrometry evidence of what peptides tumors actually present.

The team assembled a pan-cancer immunopeptidomics atlas from 531 samples spanning 14 cancer types and 29 normal tissue types. Mass spectrometry identified 389,165 canonical peptides (derived from standard protein-coding sequences) and 70,270 noncanonical peptides (derived from non-coding regions, alternative reading frames, fusion genes, and other atypical sources). A critical finding was that noncanonical peptides were presented at comparable abundance levels to canonical peptides across cancer types — challenging the common assumption that noncanonical sources are negligible and suggesting they represent a largely untapped pool of immunotherapy targets.

Tumor-specific peptides showed statistically significant differences in biochemical features compared to those found in normal tissues, including differences in hydrophobicity, charge, and MHC binding motifs. These distinctions formed the biological rationale for the MaNeo (machine learning-based neoantigen) pipeline. MaNeo integrates multiple layers of evidence: somatic mutation data, MHC-peptide binding predictions, immunopeptidomics-derived presentation likelihood, and T cell recognition scores. Benchmark analyses demonstrated that MaNeo outperformed existing neoantigen prioritization tools in accurately identifying both shared (public) and tumor-specific (private) neoantigens from canonical and noncanonical sources, achieving superior area under the curve metrics compared to state-of-the-art alternatives.

For experimental validation, the authors applied MaNeo to cancer cell lines and identified three high-priority neo-peptides. These candidates were synthesized and tested in co-culture experiments with human T cells. All three neo-peptides induced significantly increased T cell proliferation and cytotoxic T lymphocyte (CTL) activity, with T cells effectively killing tumor cells expressing the cognate neoantigens. Crucially, the activated T cells did not damage normal cells displaying wild-type peptides, demonstrating tumor selectivity — the defining requirement for a viable immunotherapy target.

The pan-cancer peptide atlas and the MaNeo pipeline together represent a substantial methodological advance for personalized cancer vaccine development. By anchoring predictions in real immunopeptidomics data rather than purely computational binding affinity scores, MaNeo reduces false positives and surfaces noncanonical neoantigens that would otherwise be invisible. Limitations include the fact that validation was conducted in cell lines rather than animal models or patients, and the atlas's coverage, while broad, does not yet capture the full heterogeneity of tumor mutational landscapes across individuals and cancer subtypes.