Answers are grounded in this paper and cite their sources.
AlphaFold 3 interface confidence scores are increasingly used not to judge a single prediction but to rank candidate nucleic acid sequences against a protein. Whether they support that use was tested on HMGA1a, an intrinsically disordered splicing regulator for which three natural RNA sites, two published loss-of-function mutants and two residue-level experimental restraints are available. Eight RNAs carrying an identical 11-nucleotide binding site, differing only in six nucleotides of downstream tail at fixed GC content, gave ipTM job medians from 0.150 to 0.430, a span of 0.280 that is reproducible across seeds and not attributable to noise (Kruskal-Wallis on 40 models, p below 5e-05; between-sequence SD 0.096 against a pooled within-job SD of 0.030). Over the same protein, replacing the real presenilin-2 site with the published decoy mutant that abolishes exon 5 skipping changed the score by 0.030 (exact Mann-Whitney U, five seeds each, p = 0.0238). Sensitivity to biologically irrelevant flanking sequence therefore exceeds sensitivity to a function-abolishing substitution by a factor of 9.3. The real site was not ranked above six composition-matched random sequences drawn from the enumerated permutation space (rank 4 of 7, p = 0.571), and that null panel spans 0.140, itself 4.7 times the functional difference. A consensus derived from the three natural sites and broken at a shared position by both published mutants did not move the score when written into twelve backbones and tested paired. In a thirteen-chain U1 snRNP assembly an overall ipTM of 0.620 was carried by the components present in deposited structures, while HMGA1a and the target RNA reached chain_ptm values of 0.08 and 0.01. Applying the identical perturbation to a folded cold shock domain whose complex is in the training data gave a span of 0.080 against 0.280 (variance ratio 15.6, exact p = 0.00016), with the flank-to-core sensitivity ratio inverted, 0.44 against 2.00. That domain's own cognate RNA nevertheless ranked third of seven against composition-matched scrambles. The failure has an identifiable cause. Submitting one twelve-base-pair sequence as DNA and as RNA gives an ipTM of 0.770 for the DNA, the highest value in the study, and the model reproduces the deposited complex of this protein with this duplex to 2.40 Å, against an internal seed-to-seed scatter of 1.22 Å. The contact distribution and the groove widths match the experimental structure, and the Arg-Gly-Arg core of the AT-hook inserts 3.5 Å inside the phosphate shell, against 3.4 Å in the experiment. On the same sequence as RNA the minor groove widens from 7.3 to 10.1 Å and its floor moves from 3.1 to 5.2 Å from the helical axis; the core then penetrates 0.1 Å, and no core atom of any model lies inside the C1' shell. The canonical binding mode is geometrically unavailable on an A-form duplex, and no complex of an AT-hook with RNA has been deposited, so the predictor has no alternative mode to apply. The mechanism itself was then tested. Eight splice sites of differing strength were each predicted inside the complete U1 snRNP with and without HMGA1a, with the number of canonical U1 base pairs to the splice site as the pre-registered endpoint. The difference is exactly zero in all eight pairs (exact sign-flip, p = 1.0000, floor 0.0078), and the dependence on splice-site strength is identical in the two arms (rho = 0.932, p = 0.0143 in both). Within a job, superimposed on the Sm ring, U1 snRNA moves 1.9 Å between models and HMGA1a 40.1 Å; its chain_ptm is 0.080 for all eight RNAs and in half the models it touches the target nowhere. The predictor does not place this protein, so the null result reports its own limits rather than the behaviour of the protein. Comparing nucleic acid sequences by interface score requires a drawn, composition-matched null distribution; a difference between two chosen sequences, however reproducible across seeds, is not evidence of recognition.
No takes yet. Share an insight, caveat, or question.
Kenji Ohe (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: