Yeast Protein Localization

## Overview The Yeast dataset is a canonical multi-class classification benchmark from UCI, mirrored on OpenML as dataset 181. It contains 1,484 instances of Saccharomyces cerevisiae proteins, each described by 8 numeric features derived from amino acid sequence signals, with a target label indicating one of 10 cellular localization sites (CYT, NUC, MIT, ME3, ME2, ME1, EXC, VAC, POX, ERL). Format: ARFF. Time coverage: static, derived from sequences available circa 1996. ## Schema - mcg — numeric — McGeoch's method for signal sequence recognition - gvh — numeric — von Heijne's method for signal sequence recognition - alm — numeric — score from ALOM membrane-spanning region prediction - mit — numeric — score from discriminant analysis on amino acid content of N-terminal region of mitochondrial and non-mitochondrial proteins - erl — numeric — presence of HDEL substring (ER retention signal), binary-valued - pox — numeric — peroxisomal targeting signal in C-terminus - vac — numeric — score of discriminant analysis on amino acid content of vacuolar and extracellular proteins - nuc — numeric — score of discriminant analysis of nuclear localization signals of nuclear and non-nuclear proteins - class_protein_localization — nominal — target: one of 10 localization sites ## Sources - OpenML dataset 181 — https://www.openml.org/d/181 — Public Domain (CC0) - Originally from UCI Machine Learning Repository (Horton & Nakai, 1996) — https://archive.ics.uci.edu/ml/datasets/Yeast ## Methodology Features were computed by Paul Horton and Kenta Nakai using a combination of expert-designed signal sequence recognition methods (McGeoch's, von Heijne's), membrane-spanning region prediction (ALOM), and discriminant analysis over amino acid composition of relevant protein regions. The target localization labels were assigned based on then-current biological annotation. OpenML hosts the data unchanged in ARFF format. ## Known gaps & limitations The dataset is heavily class-imbalanced (CYT and NUC dominate; ERL has only 5 instances). Features reflect 1990s-era bioinformatics signal-prediction methods and are not state-of-the-art relative to modern sequence-embedding approaches. Labels reflect biological knowledge as of 1996 and may not match current SGD annotations. Source does not document additional gaps; buyers should validate empirically for any non-benchmark use. ## Intended use & out-of-scope - IS for: tabular ML benchmarking, multi-class classifier evaluation, imbalanced-classification studies, teaching examples, AutoML test suites. - NOT for: production protein-localization prediction (use modern sequence models), and not deduplicated against standard ML benchmark suites — leakage risk if used to train models evaluated on OpenML/UCI tabular benchmarks. Original supplier listing: Yeast Protein Localization (UCI/OpenML) Classic multi-class classification benchmark predicting cellular localization sites of yeast proteins from 8 sequence-derived numeric features. 1,484 instances, 10 classes.