The presence of bacteria within tumour tissue is increasingly recognised as a potential contributor to cancer biology. Intratumoural microorganisms may interact with epithelial and immune cells, influence inflammatory signalling and local metabolism, and contribute to changes in the tumour microenvironment. Understanding these host–microbe interactions first requires confidence that microbial signals detected by sequencing represent bacteria genuinely present in the tissue.

This is particularly challenging in tumour sequencing datasets because microbial reads usually represent only a very small fraction of the total sequencing data, while the vast majority originate from the human host. Residual human sequences can be incorrectly classified as microbial, creating false signals and potentially leading to incorrect biological interpretation.

To address this problem, we developed HumanFilt, an open-source computational workflow designed to improve removal of human-derived sequencing reads before microbial analysis. HumanFilt combines complementary filtering approaches and sequentially screens reads against multiple human reference resources, reducing the likelihood that residual host sequences are incorrectly assigned to microbial species.

The workflow was evaluated using simulated datasets in which the origin of human and microbial reads was known, allowing us to determine whether stringent host removal could be achieved without losing genuine bacterial sequences. HumanFilt substantially reduced residual human reads while preserving the microbial signal.

We subsequently applied the workflow to whole-genome sequencing data from ten mucinous rectal adenocarcinoma cases with matched normal tissue, focusing on Fusobacterium, a bacterial genus associated with colorectal cancer. More rigorous host filtering reduced apparent microbial background in normal tissue while retaining tumour-associated Fusobacterium signals.

Importantly, sequencing findings were assessed alongside immunofluorescence analysis of tumour tissue, providing independent evidence for the presence and localisation of Fusobacterium. Agreement between sequencing-based detection and tissue-based imaging increased confidence that the bacterial signal reflected microorganisms genuinely present within the tumour rather than artefacts generated by residual human reads.

This distinction is important for physiological studies of cancer. Before investigating whether tumour-associated bacteria influence immune activation, epithelial signalling, inflammatory pathways, cellular metabolism or other features of the local tissue environment, their presence must first be established reliably. HumanFilt provides a quality-control framework that strengthens this first step and supports more confident investigation of host–microbe interactions in cancer physiology.