AlphaFold's confidence scores used to predict which disordered protein segments bind LC8

The work

Technical · from the record

AlphaFold-Multimer, run locally through ColabFold, was used to co-predict structures of the LC8 dimer together with client peptide sequences, and AlphaFold's own confidence metrics (confidence score, interface pLDDT, and three PAE measures, in 1-client and 2-client contexts) were used to decide whether a sequence binds LC8. A genetic algorithm fitted a composite threshold over 10 of these score parameters; the paper reports an AUROC of 0.9274 for the composite threshold against 0.9104 for the best single parameter, an exclusive threshold with a 5.5% false-positive and 28.1% false-negative rate, and an inclusive threshold with a 27.9% false-positive and 7.5% false-negative rate. Scores were then compared with 62 experimentally measured binding-site affinities using linear regression and adjusted mutual information, and the authors use qualitative Bayesian reasoning to argue that AlphaFold has learned an energy function that matches the natural one near minima but is inaccurate elsewhere. Applying the thresholds to 24 proteins known to bind LC8, parsed into 16-residue segments, gave 247 predicted sites under the exclusive threshold and 776 under the inclusive threshold, reduced to 133 and 213 after filtering by predicted affinity; the predicted sites are untested.

How AI was used

Binder and nonbinder client sequences were collected from the literature and from AAA anchor mutants, and a Python script parsed protein sequences into subsequences and wrote FASTA inputs pairing each client with one or two LC8 copies. ColabFold, installed on a local server, ran AlphaFold-Multimer on these inputs in both 1-client and 2-client configurations, returning mostly 25 structures per prediction (a few early runs returned five, some later runs 100); approximately 7,500 predictions were run, including 641 nonbinding and 333 binding sequences with clients 10 to 71 residues long, plus more than 6,000 predictions scanning 24 proteins in 16-residue segments with 8-residue overlap. Python scripts extracted five score types per run — confidence score, average interface pLDDT, LtoP PAE, PtoL PAE and LC8 dimer interface PAE — assigned binding status, and linked 2-client to 1-client results. A genetic algorithm optimised the AUROC of a composite threshold over all 10 resulting parameters, with learning curves generated by training on different fractions of the data and 10 iterations per fraction; the reported thresholds came from the highest-AUROC run trained on 90% of the data, and were reported as an exclusive and an inclusive set. Scores for predictions matching 62 sequences with known affinities were regressed against affinity in Excel using LINEST, and adjusted mutual information was computed with scikit-learn under three affinity binning schemes and three score binning schemes. The strongest linear relations for TQT and non-TQT anchors were written as equations to estimate affinities for newly predicted sites, which were filtered at 40 µM and screened against structured-domain assignments from the AlphaFold Database, then compared with LC8Pred predictions. ChatGPT was used in writing the analysis scripts.