TabFlux — microbiome classification, with TabPFN-3.5 on board

TabFlux uses TabPFN-3.5 to classify microbial community profiles and test whether predictions transfer to studies kept out of training. Code: https://github.com/iLivius/TabFlux Documentation and setup: https://ilivius.github.io/TabFlux/ Built on mlr3, this R/Quarto workflow starts from amplicon count tables or shotgun taxonomic profiles, plus taxonomy and sample metadata. When study identifiers are supplied, whole studies stay together in the evaluation and tuning splits. Feature selection and tuning use only each outer fold's training data. TabPFN-3.5 is the default learner. Its pretrained weights run locally; labelled training profiles provide context for predicting held-out samples. There is no dedicated model hyperparameter search. Random forest and XGBoost receive the same splits, selected features and class-frequency decision rule. The public example is food-category classification using curated Food Metagenomic Data (cFMD): 3,680 samples, 57 datasets, seven categories and five grouped outer folds. Mean balanced accuracy ± SD across folds: TabPFN-3.5 0.662 ± 0.121, XGBoost 0.634 ± 0.103, random forest 0.626 ± 0.157. These are fold averages, not pooled out-of-fold scores. TabPFN has the highest fold mean here; one run does not establish general superiority. The video explains the evaluation design and shows real results, including a dataset on which all three learners mostly fail to transfer. Outputs include an HTML report, fold/group/class metrics, out-of-fold predictions, saved final models and SHAP explanations. SHAP describes model behaviour, not biological causes. 0:00 Why microbial profiles are hard to learn from 0:32 What goes in 0:57 Can it work on an unseen study? 1:43 TabPFN-3.5 inside the same evaluation 2:30 A public example: food metagenomes (cFMD) 4:03 What comes out cFMD results: https://ilivius.github.io/TabFlux/cfmd/res... Learner comparison: https://ilivius.github.io/TabFlux/cfmd/com... Version DOI: https://doi.org/10.5281/zenodo.23041258 TabFlux code: Apache-2.0. TabPFN-3.5 weights have a separate Prior Labs licence. Grouped evaluation measures transfer without removing confounding. Shotgun-profile support is at demo stage in v1.6.0; external test sets at SGB level are not yet supported. Livio Antonielli and Lukas Pucher, AIT Austrian Institute of Technology. MICROSUPPRESS, Austrian Science Fund (FWF) P 36288. cFMD: Carlino et al. (2024), Segata Lab, CC BY 4.0. TabPFN-3.5: Prior Labs. Built on mlr3. Workflow map drawn with nf-metro. Figure notes: fold-score and decision-rule plots come from the TabFlux documentation. The per-category recall chart is redrawn from its published values. The per-class SHAP figure is redrawn from the published figure, with bar lengths measured from it. The t-SNE was computed for this video from the run's input table: all 3,680 samples and 7,847 SGBs, log10 relative abundance, pseudocount 1e-6, no feature selection. Rtsne 0.17; PCA to 50 dimensions, perplexity 30, theta 0.5, 1,000 iterations, seed 42. No new predictive-performance scores were computed for the video. Narration: synthetic voice, Kokoro-82M (Apache-2.0). Fonts: Nunito and Inter (SIL OFL). Made with Claude under Livio Antonielli's supervision; drawn and rendered with HTML/SVG, Chromium and ffmpeg.