Blame for README.md

---
license: cc-by-nc-4.0
task_categories:
- tabular-regression
tags:
- protein-protein-interaction
- binding-affinity
- interactome
---

# OccupancyFlow data

Data for the OccupancyFlow pipeline: sequence-based binding-affinity prediction and the
competition-adjusted weighting of the human interactome.
Code: https://github.com/GainGod-Xu/OccupancyFlow

## Contents

| path | rows | what it is |
|---|---|---|
| `affinity/train_6655_no_overlap_s79_s90.csv` | 6,655 | chain pairs with a measured Kd, converted to ΔG (kcal/mol). Curated from PPB-Affinity: chain-pair decomposition, interface filtering, deduplication, then removal of every PDB entry shared with S79/S90 |
| `affinity/test_s79.csv` | 111 | S79 benchmark (110 labelled), sequences re-extracted per chain from RCSB |
| `affinity/test_s90.csv` | 89 | S90 benchmark |
| `interactome/edges_RSW_20231018_filtered.csv` | 328,433 | human interaction network (Entrez gene IDs + symbols) after removing proteins without a reference sequence |
| `interactome/proteins_RSW_20231018_clean.csv` | 17,219 | protein table: sequence, PaxDb abundance (ppm), identifiers |
| `interactome/predictions_esmc_300M_nM.csv` | 328,433 | predicted ΔG and Kd per interaction (`esmc_300M`, current), plus abundances |
| `weights/interactome_weights_esmc_300M.csv` | 328,422 | competition-adjusted weights `w_ij`, complex and free concentrations - current selected encoder (`esmc_300M`); the weight file the manuscript and disease-module analyses use |
| `negatome_check/predictions_negatome_random.csv` | 6,352 | predictions for Negatome human non-interactions, Negatome pairs that are also reported interactions, and random pairs of the same proteins |
| `negatome_check/per_protein_comparison.csv` | 283 | per-protein median ΔG for reported partners vs Negatome partners |
| `fixed_test_sets/test_s79_fixed.csv` | 111 | S79 with the 1E96 label restored — all 111 chain pairs labelled |
| `fixed_test_sets/test_s90_fixed.csv` | 90 | S90 with 3H6G added (chains A–B) — all 90 complexes |
| `official_supplementary/` | — | the upstream sources as published: PPI-Affinity SI files 3 and 4 (labels), the ProAffinity chain list, the PPAP pdbbind90 chain list |

## Defects in the benchmark sources, and what was fixed

Audited against the published supplementary files; the full write-up is in
[docs/data_provenance.md](https://github.com/GainGod-Xu/OccupancyFlow/blob/main/docs/data_provenance.md).

- **S79 loses one complex in the official file.** PDB code `1E96` is stored as `1.00E+96` — Excel
  read it as scientific notation before publication. Any pipeline joining on the PDB name drops it,
  so published S79 numbers are most likely computed on 78 of 79 complexes. The label survives
  (7.4 kcal/mol) and is confirmed by the ProAffinity pKd for the same entry (5.5686 → 7.60).
  `fixed_test_sets/test_s79_fixed.csv` restores it.
- **S90 loses one complex to the chain list.** The labels cover 90 complexes; the PPAP chain list
  covers 89. The missing entry, 3H6G, has an unambiguous interface (chains A–B, 355 heavy-atom
  contacts within 4.5 Å, minimum distance 2.75 Å) and is a homodimer, which may be why it was
  excluded. `fixed_test_sets/test_s90_fixed.csv` adds it back.
- **The two official files use opposite sign conventions**: SI-File-3 stores |ΔG| (positive,
  4.3–18.6), SI-File-4 stores ΔG (negative, −18.1 to −3.3). Files here are all ΔG (negative).
- **Label quality is good**: cross-checking S79 against the independent ProAffinity pKd values
  gives a median absolute difference of 0.032 kcal/mol over the 78 shared complexes.

Keep the original `affinity/test_s79.csv` and `test_s90.csv` when comparing against published
numbers, which were computed on the uncorrected sets.

## Things to know before using these files

- **ΔG sign convention**: negative means binding. Kd in nM.
- **Affinity values in the interactome files are predictions**, not measurements, and they are
  compressed: SD 0.64 kcal/mol against 2.68 for the measured training labels. They order
  interactions better than they calibrate them.
- **The predictions do not distinguish interactions from non-interactions** (Negatome AUC 0.481);
  the model was trained only on binders. See `negatome_check/`.
- **Abundance conversion** uses k = 1.881 nM/ppm (hepatocyte GAPDH: 3.39e7 copies/cell, 3,000 fL,
  divided by the PaxDb whole-organism abundance 9,978 ppm).
- **44 titin (TTN) interactions are missing** from the weighted network: the 34,350-residue
  sequence broke an intermediate spreadsheet export. They are present in the `interactome/edges`
  file and absent downstream.
- **Leakage control in the affinity split is at PDB-entry level**, not sequence identity: 2 of 111
  S79 chain pairs have both sequences present in the training set (S90: none).

## Sources

PPB-Affinity (Liu et al., Sci Data 2024) · PPI-Affinity S79/S90 (Romero-Molina et al., JPR 2022) ·
PaxDb (Wang et al. 2015) · Negatome 2.0 (Blohm et al., NAR 2014) via `Synthyra/NEGATOME`.