Cross-Assay Normalization of Protein–Protein Binding Affinities

Binding affinities for the same protein–protein mutation, measured by different experimental assays, are not directly comparable. This repository calibrates affinity measurements from 20 assay method codes onto a single common scale — surface plasmon resonance (SPR) — so that measurements pooled from heterogeneous sources can be used together, and quantifies how much that calibration improves between-assay agreement.

The calibration is built on SKEMPI v2, which covers protein–protein complexes generally rather than one interaction class.

Because a cross-assay comparison needs the same mutation measured more than once, the pipeline keeps only mutations carrying two or more distinct Method values. That is the binding constraint on scale: 339 mutations, drawn from 813 individual measurements, out of the 7,036 non-n.b. rows in SKEMPI v2. Every number and figure in this repository rests on those 339.

No complex-type filter is applied, and the 339 are mixed. By SKEMPI's Hold_out_type label, 18.0% are antibody–antigen, 8.3% protease/inhibitor and 4.1% TCR/peptide–MHC; the remaining 69.6% carry no label and are dominated by hormone–receptor and enzyme–inhibitor systems — growth hormone / hGH binding protein, interferon alpha-2 / its receptor, TEM-1 beta-lactamase / BLIP and Ras / RBD together account for well over a third of all measurements used. The results below therefore describe protein–protein binding broadly, and should not be read as antibody-specific.

Per-assay fold-change distributions before (left) and after (right) calibration onto SPR. PDF

Agreement between assay pairs — colour is −log10(q), size is the number of shared mutations. PDF

Method

For each mutation the raw quantity is a log10 affinity fold change

y = log10( Kd_mut / Kd_wt )

computed per assay method, giving a matrix of mutations (rows) × assay methods (columns). Non-binding mutants (Affinity_mut (M) == "n.b.") and non-positive ratios are excluded.

SPR is taken as the reference assay. For every other method m, an ordinary least-squares fit is performed over the mutations measured by both m and SPR:

y_SPR  =  alpha + beta * y_m

That fit is then applied to the whole of column m, mapping it onto the SPR scale. Methods overlapping SPR on fewer than 3 mutations are dropped to NaN rather than extrapolated.

Agreement before and after calibration is reported as Pearson and Spearman correlation matrices, with significance assessed by a permutation test (pearson_permutation_test, 10,000 permutations) and Benjamini–Hochberg FDR correction.

Assay codes follow the Method field of SKEMPI v2: BI, CSPRIA, ELFA, ELISA, EMSA, ESMA, FL, IAFL, IAGE, IARA, IASP, ITC, KinExA, RA, SE, SFFL, SFPF, SP, SPR and the composite code SPR;SFFL. See the SKEMPI v2 documentation for their definitions.

Layout

data/
  raw/           SKEMPI v2 tables — supplied, not produced here (see data/raw/README.md)
  processed/     every matrix the pipeline computes
src/
  calibrate.py   the calibration: raw -> processed + most figures
  figures.py     correlation heatmaps from the processed matrices
results/
  calibration/                    per-assay before/after scatter vs SPR
  correlation_affinity/           pairwise assay correlation, affinity
  correlation_ddg/                pairwise assay correlation, ddG
  correlation_ddg_no_nonbinding/  same, excluding non-binders
  summary/                        heatmaps, FDR bubble, before/after boxplot
docs/
  literature.txt            background reading

The flow is one direction: data/raw → src/calibrate.py → data/processed → src/figures.py → results/.

This repository contains the normalization study only. The originating project also held ~270 GB of third-party database staging directories (IEDB, TDC, CATH, HSPVdb, BioLiP and others) and AlphaFold prediction output; none of that is included here. See Data below.

Reproducing

pip install -r requirements.txt
python src/calibrate.py     # data/raw -> data/processed + results/
python src/figures.py       # data/processed -> results/summary/

Both scripts anchor themselves to the repository root, so they can be run from any working directory.

src/calibrate.py writes:

Output	Contents
data/processed/affinity_calibrated.csv	calibrated matrix, mutations × assays
data/processed/calibrated_corr.csv	assay × assay correlation after calibration
data/processed/affinity_corrarray.csv	correlation before calibration
data/processed/affinity_q_matrix.csv	BH-FDR q-values per assay pair
data/processed/energy_difference_correlation.csv	Spearman correlation of ddG
data/processed/delta_delta_G_*_corrarray.csv	ddG correlations, with/without non-binders
results/calibration/	per-assay before/after scatter vs SPR
results/correlation_*/	pairwise assay correlation figures
results/summary/	FDR bubble, value plot, before/after boxplot

src/figures.py reads data/processed/affinity_corrarray.csv and writes heatmap.pdf, clusteredheatmap.pdf, bubble.pdf and correlation_scatter.pdf into results/summary/.

Data

The SKEMPI v2 tables needed for the calibration are included. The bulk inputs — PDB structure sets, AlphaFold output and the per-database dumps — are not redistributed here. Obtain them from their upstream sources (SKEMPI v2, IEDB, SAbDab, TDC, BindingDB, BioLiP, IMGT) under those projects' own terms, which differ from the license of this code.

Status and caveats

• The calibration fits a single global linear map per assay. It does not model per-complex or per-interface effects, and assays overlapping SPR on few mutations give correspondingly unstable fits — ELFA correlates with SPR at r ≈ 0.39 against ELISA at r ≈ 0.93.

License

Code in this repository is released under the MIT License (see LICENSE). The third-party datasets it consumes (SKEMPI v2, IEDB, SAbDab, TDC, BindingDB, BioLiP, IMGT) are not covered by that license and remain subject to their own terms.