PhenoCal: Natural-Control Label Calibration for Cross-Cohort Machine Learning - reproducibility archive

PhenoCal: Natural-Control Label Calibration for Cross-Cohort Machine Learning — reproducibility archive accompanying the paper by A. de Ruvo, P. Castelli, A. Di Pasquale and N. Radomski, accepted at the 12th International Artificial Intelligence Symposium (AIS 2026; to appear in Springer Lecture Notes in Computer Science). PhenoCal calibrates a cohort-confounded continuous label (label-shift-by-source) without shared reference measurements, by exploiting near-duplicate cross-cohort instances as natural controls: cohorts form a comparison graph, near-duplicate pairs give matched-pair label differences whose medians estimate per-cohort offsets by weighted least squares (identifiable up to one anchor), and the calibration uncertainty is propagated into instance-level probabilistic labels. This archive contains: phenocal/ — the dependency-free Rust reference implementation of the graph calibration method (also maintained at github.com/genpat-it/phenocal); scripts/ — 15 numbered Python analyses reproducing the synthetic experiments, the τ-sensitivity and residual diagnostics, and the downstream-ML evaluability and leakage controls; data/ — hashed cgMLST (1,748-locus) and wgMLST (6,785-locus) allele-profile matrices and the MIC label tables; dashboard/ — a self-contained interactive HTML report of the real-data run (comparison graph, per-edge twin pairs, offsets with propagated intervals, before/after distributions, per-isolate probabilities); one-shot runners (reproduce_synthetic.sh, reproduce_realdata.sh) and requirements.txt. Case study: benzalkonium-chloride minimum inhibitory concentrations of Listeria monocytogenes measured across five laboratory cohorts, with genomes as the similarity ruler. The underlying whole-genome sequences are not redistributed here: they are publicly available in ENA/SRA and can be re-typed to regenerate the profiles. Downstream modelling was run with GenomicBasedClassification / GenomicBasedRegression v1.3.0. Released under the MIT license.