Objective: Samples subjected to metagenomic sequencing can be accidentally contaminated during wet lab steps (DNA extraction, library preparation) by DNA from an external source (e.g.: lab reagents) or from other samples processed simultaneously (well-to-well contamination). These can lead to biased results and eventually to false conclusions if not detected. Although a critical issue, well-to-well contamination remains understudied. A few tools have been developed but suffer from several limitations, such as a lack of sensitivity. Methods: By inspecting species abundance profiles of published cohort samples, we identified specific patterns associated with well-to-well contamination. Here, we introduce CroCoDeEL, a tool that automatically searches for these patterns in order to accurately identify contaminated samples. CroCoDeEL relies on a supervised model pre-trained with semi-simulated sequencing data. Our approach does not require negative controls, works with related samples that may naturally share strains (e.g.: mother/child), discriminates contamination sources from contaminated samples and estimates contamination rates. Results: Benchmarks across three public cohorts comprehensively curated by human experts revealed that CroCoDeEL accurately identifies contaminated samples and their respective contamination sources even at low rates if sequencing depth allows. Applying CroCoDeEL to 13 cohorts focusing on human colorectal cancer (CRC) revealed that contamination is a widespread issue. Conclusion: We believe that our work underlines the urgency to acknowledge and systematically address the cross-sample contamination phenomenon to ensure the robustness of studies based on metagenomic data. CroCoDeEL is freely available at https://forgemia.inra.fr/metagenopolis/crocodeel.