Drug Transporter ML

I built Python machine-learning tools for a question posed by transporter researchers: which molecular properties help distinguish compounds handled by OAT1 and OAT3? The work contributed to the 2020 Journal of Biological Chemistry paper, "Unique metabolite preferences of the drug transporters OAT1 and OAT3 analyzed by machine learning." My contribution was the software and modeling implementation, developed alongside researchers who supplied the biological context.

The repository now serves a second purpose: it shows what can go wrong when a research implementation becomes an archive. A September 2026 audit found that successful classifier calls had been mistaken for verification of the complete feature-search program. The program could label a result with two feature names while evaluating every column. That changed this case study: the useful story includes the research design, the defect, and the evidence needed to repair it.

OAT1 and OAT3 are kidney transporters with overlapping roles, yet the paper investigates molecular differences among their substrates. Its reported finding links OAT3 preference with more structurally complex compounds, including more rings and chiral centers. The paper is the source for that scientific result. The maintenance checks described here concern software behavior and do not repeat the study.

For the implementation, each molecule becomes a row of numeric descriptors rather than an image or a sequence. A model predicts a transporter class from those descriptors. The engineering question is whether searching feature subsets or evolving transformations can produce a useful, inspectable classifier. A high score alone would not explain which inputs it used, whether selection overfit the evaluation data, or whether another researcher could run the same program.

Three datasets, each a CSV of molecules with 32 shared descriptor columns and a transporter-class label: OAT1OAT3Small.csv (51 metabolites, 3-class: 17 Both, 12 OAT1-only, 22 OAT3-only), OAT1OAT3Big.csv (73 metabolites, same 3-class scheme: 17 Both, 21 OAT1, 35 OAT3), and OAT1OAT3OATP.csv (137 metabolites combining kidney and liver transporters, 6-class: 73 kidney rows plus 64 liver rows split 29 Both OATP, 20 OATP1B1, 15 OATP1B3).

The shared loader produces a matrix, labels, and column names. Three modules consume that interface: a feature-subset search using Random Forest and Decision Tree classifiers, a NEAT classifier, and a NEAT feature transformer feeding a Random Forest. Utilities supply the classification-accuracy function, population setup, checkpointing, and interactive prompts. The intended architecture is compact enough to inspect without a service, database, or deployed API.

The legacy directory retains earlier implementations, including a five-classifier search, dataset-specific variants, parallel search, and an ID3 decision tree. Keeping those versions preserves the development history. It does not establish which exact source, environment, and outputs produced every published number. A consolidated implementation should not silently stand in for a pinned historical experiment.

The brute-force method is intended to enumerate every size-k subset, fit two classifiers on that subset, and save their cross-validation scores next to its column names. Exhaustive enumeration removes one search heuristic, but it does not remove statistical uncertainty. Choosing the best subset on the same folds used to score it can overfit the selection criterion. An independent performance estimate needs an outer evaluation procedure or untouched data.

The module called feature-selection GA evolves a neural network that consumes the raw descriptors and predicts a class. Its fitness is training-set accuracy, so it is searching a classifier rather than reporting held-out performance. The feature-engineering GA instead evolves six transformed features for a downstream Random Forest. Its fitness uses leave-one-out or ten-fold cross-validation, depending on the dataset. Evolution still repeatedly selects against that score. Neither a rising fitness curve nor a winning genome alone establishes generalization.

The utility named auc computes the fraction of correct class predictions. It is classification accuracy, not area under a receiver-operating-characteristic curve. Preserving that distinction matters more than the function's historical name: otherwise readers can mistake a training objective for a different statistical measurement.

The July audit exercised data loading and classifier helpers and reported scratch-run scores. It did not establish that the complete subset-search loop was correct. Reviewing the consolidated source at commit 4cd3f66 in September exposed the missing connection: feature_indices determined the CSV labels, but both eval_classifier calls received the full data matrix. A plausible accuracy and a plausible pair of feature names could therefore appear together without describing the same experiment.

A second defect affected completeness. Results were buffered in batches of ten, but the program never wrote the final partial batch. With 32 single-feature combinations it would omit two rows. With 496 two-feature combinations it would omit six. Checking that a CSV existed, or inspecting only its first rows, would miss that failure.

Startup also depended on the operator's working directory. The documented package command conflicted with absolute imports, datasets and configs used relative paths, and output directories were assumed to exist. The four NEAT configs omitted a key required by the previously tested neat-python 2.0.0 environment. These are separate failure modes: a config-only patch cannot verify the search or make the advertised workflow reproducible.

Repair a88149c sends the selected columns to the classifiers, flushes the CSV remainder, makes package imports and resource paths consistent, creates output directories, and updates the NEAT configuration contract. Its first three regression tests use a small matrix with distinctive column values and inspect the actual subset-search loop. They fail against the original implementation and pass after the repair: the recorded inputs match the names, and both a three-row partial batch and a twelve-row batch-plus-remainder are fully saved.

The full suite contains ten passing checks. It also starts the documented command, loads all three datasets, parses all four configs, and saves and restores a checkpoint in an initially absent output directory. Both real GA entrypoints complete one generation using two genomes and six artificial rows. That fixture uses seed 1729, two-tree Random Forests, and intercepted plot calls. It tests execution through the fitness functions without fitting models to the biomedical datasets or claiming research performance.

The checks pass on Windows with Python 3.12.14 and a recorded dependency lock, including a clean source snapshot using the same isolated environment. Those checks intercepted graphics. A later local repair, 6f1dfac, fixes optional network pruning and passes native SVG rendering checks on artificial graphs. Interactive desktop plotting remains unverified. The earlier repair is available on a review branch, not the default branch; the visualization repair has not been pushed. A full paper reproduction would additionally need the original experiment protocol, seeds, selection procedure, environment, and retained outputs. It remains a separate undertaking.

This page previously called the brute-force path correct end to end. That statement is withdrawn. Historical helper-level accuracy ranges cannot prove that the orchestrator supplied the right columns or saved every result. The same boundary applies to scratch-run charts: without a retained runner and raw measurements, they are insufficient evidence for a new quantitative comparison.

The architecture and differing evaluation objectives remain useful to explain. The published paper remains the source for its own findings. The maintenance audit makes the archive more inspectable by identifying the failing contracts and the checks needed to repair them, without relabeling a software fix as a scientific replication.

The durable lesson is to verify the relationship between inputs, labels, and saved outputs, not just whether a model returns a number. In a search pipeline, the feature names are part of the experiment identity. If they do not match the matrix passed to the estimator, a clean-looking table can describe experiments that never happened. Completeness is another contract: the number of retained rows should agree with the declared search space.

This remains an archived research project rather than a deployed predictor. No new full feature sweep, evolutionary study, or independent biomedical validation was performed for this portfolio audit. Small-sample uncertainty and repeated selection against cross-validation remain relevant even after implementation defects are fixed. Repairing the code makes the next experiment possible to inspect. It does not predetermine its result.