MimicNeoAI

MimicNeoAI is a pipeline for discovering immunogenic tumor antigens, developed by Tao Chen and colleagues at the Hefei Institutes of Physical Science, Chinese Academy of Sciences with collaborators at Anhui University, Anhui Medical University, the University of Science and Technology of China, and Hong Kong Baptist University, first posted as a preprint in June 2025. Its premise is that tumor-resident microbes are an under-exploited source of therapeutic targets: peptides from bacteria colonizing a tumor can be presented on HLA molecules and sometimes resemble mutation-derived neoantigens closely enough to prime cross-reactive T cells.

The learned component at the center of the pipeline is a bidirectional LSTM ensemble that scores the immunogenicity of a peptide-HLA pair. It is trained once on curated epitope corpora and then applied without refitting to peptides called from new sequencing data, whether from microbial genomes, somatic mutations, or unannotated open reading frames. Around it sit conventional bioinformatics stages: host-read depletion and microbial profiling, somatic variant calling with pVACseq-based neoantigen discovery, shared HLA typing, and a sequence-similarity step pairing microbial epitopes with mutation-derived ones to flag mimicry.

Like DeepPROTECTNeo, MimicNeoAI runs end to end from raw reads to a ranked list, but where that framework centers on receptor recognition of tumor mutations, this one asks which peptides in the tumor microbiome are worth pursuing at all. The toolkit is released under Apache-2.0, with trained models and the microbial reference database archived on Zenodo under CC BY 4.0.

#Key Features

• Dual-source epitope discovery: Microbial and mutation-derived peptides are called through separate upstream arms and scored by one shared model, so candidates from both sources land on a single comparable scale.
• Molecular mimicry detection: Sequence-similarity comparison between microbial and mutation-derived epitopes identifies microbial peptides that mimic tumor neoantigens — the pairs most likely to support cross-reactive anti-tumor responses.
• Physicochemically grounded encoding: Residues are represented by 12-dimensional vectors derived from a PCA over 566 physicochemical properties in AAindex1, supplemented by eight peptide-level descriptors including polarity, volume, charge, hydrophobicity, and isoelectric point.
• HLA-aware scoring: Full-length allele sequences from IPD-IMGT/HLA are encoded alongside the peptide, and alleles without a full-length match receive a neutral score rather than a spurious prediction.
• Wet-lab-anchored validation: Predictions were checked against single-cell TCR sequencing and molecular dynamics simulations of HLA-epitope-TCR complexes.

#Technical Details

The immunogenicity model is a BiLSTM ensemble taking a tokenized peptide and HLA sequence over a 22-token amino acid vocabulary, with physicochemical features passed through a 256-node dense layer before fusion and softmax calibration. Training data combine a human set compiled from IEDB spanning 117 HLA-I alleles and 12,039 peptides, giving 15,482 balanced peptide-HLA complexes, with a microbial set from MicroEpitope contributing 28,246 balanced peptide-HLA-I and 1,754 peptide-HLA-II complexes. Optimization used Adam with cross-entropy loss over 200 epochs per fold in five-fold cross-validation, roughly 20 to 30 minutes per fold on a 48 GB NVIDIA L20 GPU.

Cross-validation yields a mean ROC-AUC of 0.909 (s.d. 0.006) and PR-AUC of 0.908 (s.d. 0.007). On a held-out set of 42 peptide-HLA complexes — 21 literature-verified immunogenic microbial complexes and 21 generated non-immunogenic controls, with training peptides excluded — the model recovered 19 of 21 true positives and 17 of 21 true negatives, for 0.857 accuracy, 0.826 precision, 0.905 recall, and an F1 of 0.864. Against IEDB's immunogenicity tool, DeepNeo, and DeepImmuno on the same evaluation, it reached a ROC-AUC of 0.94 versus 0.35 to 0.58, and a PR-AUC of 0.90 versus 0.37 to 0.45.

#Applications

The pipeline targets immunotherapy programs looking beyond mutation-derived neoantigens, which are scarce in tumors with low mutational burden. Applied to a colorectal cancer cohort, MimicNeoAI found that microbial peptides yielded 153 high- or moderate-immunogenicity candidates against 75 from somatic mutations, despite coming from a peptide pool nine times smaller, with no counterpart in matched normal tissue. Single-cell TCR sequencing of patient blood showed clonally expanded clonotypes preferentially recognizing 75% (9 of 12) of the top predicted epitopes, and molecular dynamics simulations found predicted immunogenicity tracking with complex stability.

#Impact

MimicNeoAI supplies a computational route to a hypothesis that has largely been argued case by case: that microbes inside tumors are a systematic, tumor-specific antigen reservoir, and that mimicry between microbial and tumor peptides can be enumerated rather than stumbled upon. One trained predictor spanning microbial, cryptic, and mutation-derived peptides is what makes the comparison between antigen sources meaningful, since a single scale applies throughout. The clinical evidence rests on a single-patient validation, with multi-center cohorts identified as the necessary next step, so the immunogenicity numbers are a strong internal benchmark rather than an established clinical result. The code carries an Apache-2.0 license and the model and reference bundles are openly archived, though the preprint is posted with all rights reserved and has not been peer reviewed.