bacNeo: a multi-omics computational toolkit for identifying bacteria-derived neoantigens in human cancers

Background

The intratumor microbiome has emerged as a defining hallmark of cancer biology. Tumor-resident bacteria modulate tumor growth, inflammation, genome instability, and the tumor immune microenvironment, and a subset of these microbes generates peptides that are loaded onto human leukocyte antigen (HLA) molecules and recognized by tumor-infiltrating T cells. Kalaora and colleagues first demonstrated that bacteria-derived HLA-bound peptides on melanoma cells can reactivate anti-tumor immunity, establishing the conceptual basis for bacterial neoantigens as a therapeutic axis distinct from classical tumor-mutation-derived neoantigens. Despite this potential, systematic computational discovery of bacterial neoantigens has remained restricted to a handful of cancer types. Established neoantigen pipelines such as pVAC-seq, nextNEOpi, PGNneo, and ProGeo-neo were engineered for endogenous mutations and offer no end-to-end support for exogenous bacterial inputs. Bacterial homology approaches add coverage but lack the tuned models needed to simulate prokaryotic translation, and tools that explore microbial mimicry, such as MimicNeoAI, require paired whole-transcriptome and whole-exome sequencing, which limits their use on routine clinical cohorts.

A unified toolkit for bacterial neoantigen discovery

To close this gap, the team developed bacNeo (bacterial neoantigen), a multi-omics computational toolkit released as a single Anaconda-managed package with one-step installation. The framework comprises three integrated modules that can be executed independently or as a coordinated pipeline:

BACC (Bacterial Components). BACC removes host reads using BWA-MEM or HISAT2, then classifies bacterial taxa via Kraken2, with utilities for raw counts, counts per million, relative abundance, single-cell expansion through cell-barcode mapping, and Newick/Tre-formatted phylogenetic outputs.

BACH (Bacteria-bound HLA). BACH extracts chromosome 6 reads from host sequencing and predicts sample-specific class I HLA alleles (HLA-A, HLA-B, HLA-C) via HLAscan, with optional class II prediction, supplying high-resolution genotyping that anchors downstream peptide-binding inference.

BACP (Bacterial Peptides). BACP simulates bacterial translation through CheckM2 (Prodigal-based ORF calling refined by artificial neural networks and gradient-boosted decision trees), filters peptides with CD-HIT, and slides 9-mer windows for HLA matching. When proteomic data are available, MaxQuant searches against a bacterial proteome reference. Candidate peptides are then ranked by SPARK, a composite score that combines TAP transport efficiency (TAP logIC50, modeled by the stabilized matrix method) and HLA-peptide binding affinity (BA rank from netMHCpan 4.1) through a percentile, multiplicative, and logarithmic transformation.

Validation across metagenomic, single-cell, and multi-omics cohorts

The authors benchmarked bacNeo on three independent datasets. On a 162-sample gastric metagenome (PRJNA1067082), BACC recovered the expected dominance of Helicobacter pylori alongside Serratia surfactantfaciens and Prevotella intermedia, confirming concordance with established microbiome profiling. On gastric single-cell RNA sequencing (PRJNA1183290), BACC mapped bacterial reads to cell barcodes and revealed differential bacterial loads across malignant epithelial and immune populations, including a strong signal in fibroblasts and stem-like cells. On a 98-patient Chinese esophagogastric junction adenocarcinoma cohort (PRJNA788008 and PXD030667), bacNeo simultaneously ran BACC, BACH, and BACP: it identified 19 HLA-A, 38 HLA-B, and 20 HLA-C alleles (HLA-A11:01 as the most prevalent allele, consistent with published Chinese-population genetics), generated TreeViewer-ready phylogenies, and detected 3,017 unique bacterial proteins spanning 136 species, of which 89 proteins appeared in more than 20% of tumor samples. Co-detection analysis showed that high-confidence bacterial neoantigens preferentially originate from species with concordant RNA-level abundance, lending biological coherence to the prioritization.

SPARK and experimental validation

To quantify predictive accuracy, the authors assembled experimentally detected 9-mer peptide-HLA pairs and processed them through BACP. Strong binders were defined as BA rank below 0.5 with normalized TAP logIC50 below 0.0. The SPARK index, which is the logarithmic transform of the multiplicative percentile of TAP and BA rank, separated experimentally validated binders from random peptides and produced an area under the ROC curve of 0.81. A three-dimensional distribution of SPARK against TAP logIC50 and BA rank illustrates how the composite score amplifies discrimination among high-performing candidates while compressing low-performing ones, providing an interpretable ranking scheme for downstream candidate selection.

Significance for cancer immunotherapy

The biological feasibility of bacteria-derived neoantigens depends on intratumor bacterial load, HLA expression, antigen-processing competence, and the route of presentation. Tumors with measurable bacterial burden, preserved HLA expression, and active antigen-presenting cells, such as colorectal cancer, pancreatic ductal adenocarcinoma, gastric cancer, and lung cancer, represent the most plausible settings for bacNeo-based stratification. By providing a reproducible computational reference for bacterial neoantigen calling, bacNeo can complement existing endogenous-neoantigen pipelines, support therapeutic cancer vaccines, inform CAR-T design against microbe-associated targets, and refine patient stratification for immune checkpoint blockade.

Availability

bacNeo is open-source under the MIT license. Source code, tutorials, and pre-processed test data are available on GitHub at https://github.com/ZhangLab-Fudan/bacNeo. Pre-constructed reference databases are deposited at Synapse under accessions syn66327848 and syn66514464.