# Comparator predictions on the tan70 test fold

This directory carries per-compound prediction probabilities for the two
external comparators used in the reverse-leak audit and the head-to-head
comparison (manuscript Tables 3, 3b, S2, S3, and Figure 4).

| file | comparator | citation |
| --- | --- | --- |
| `ctoxpred2_tan70_predictions.csv`   | CToxPred2   | Arab et al. 2024 — `https://github.com/issararab/CToxPred2` |
| `cardiogenai_tan70_predictions.csv` | CardioGenAI | Kyro et al. 2025 — `https://github.com/gregory-kyro/CardioGenAI` |

Both files were generated by running each comparator's released code with
its default checkpoints on the tan70 test fold SMILES from this repository
(`data/compounds/compounds.csv` joined to `data/splits/tan70.csv` on
`row_idx`). The reverse-leak audit (manuscript line 199 and following) is
computed against the InChI-keys in these files vs. each comparator's
training-set InChI-keys.

## Schema (both files)

| column | dtype | notes |
| --- | --- | --- |
| `row_idx`    | int   | canonical row index — joins to `data/compounds/compounds.csv` and `data/labels/labels_v1.csv` |
| `inchikey`   | str   | 27-character standard InChI-key (duplicated here so this file is self-contained) |
| `valid`      | bool  | did the comparator successfully score this compound? (compounds that fail comparator-specific SMILES parsing get `valid=false` and the prob columns are filled with whatever the comparator emitted — usually NaN or 0.0) |
| `herg_prob`  | float | hERG blocker probability (post-sigmoid output of the comparator) |
| `nav15_prob` | float | Nav1.5 blocker probability |
| `cav12_prob` | float | Cav1.2 blocker probability |

All probabilities are matched against the manuscript's deployed CardioSafe
output thresholds (default = 0.5) for the head-to-head MCC computation.
The hERG probability head from both comparators is trained at the 10 µM
threshold, so it is compared against our `herg_blocker_10um` label; the
1 µM CardioSafe head has no direct comparator equivalent (see Table S2
hERG 1 µM rows: the comparator probability is used for both 10 µM and 1 µM
matched against our respective ground truths, by paper convention).

## Coverage

| comparator   | n total | n valid |
| ---          | ---:    | ---:    |
| CToxPred2    | 46,326  | 46,326  |
| CardioGenAI  | 46,326  | 46,230  |

(CardioGenAI fails to score 96 of the 46,326 tan70 test compounds — these
are filtered out before any aggregate metric computation.)

## Reproducing the head-to-head numbers

Tables S2 and S3 (the comparator panels pre- and post-de-leak) are built
on top of these CSVs. The full paired-bootstrap output is shipped at
`data/supplementary/tables_s2_s3_paired_bootstrap.json` so you can verify
the aggregated numbers without re-running the bootstrap.