Joseph F. Thole

Recent work suggests that AlphaFold (AF)-a deep learning-based model that can accurately infer protein structure from sequence-may discern important features of folded protein energy landscapes, defined by the diversity and frequency of different conformations in the folded state. Here, we test the limits of its predictive power on fold-switching proteins, which assume two structures with regions of distinct secondary and/or tertiary structure. We find that (1) AF is a weak predictor of fold switching and (2) some of its successes result from memorization of training-set structures rather than learned protein energetics. Combining >280,000 models from several implementations of AF2 and AF3, a 35% success rate was achieved for fold switchers likely in AF's training sets. AF2's confidence metrics selected against models consistent with experimentally determined fold-switching structures and failed to discriminate between low and high energy conformations. Further, AF captured only one out of seven experimentally confirmed fold switchers outside of its training sets despite extensive sampling of an additional ~280,000 models. Several observations indicate that AF2 has memorized structural information during training, and AF3 misassigns coevolutionary restraints. These limitations constrain the scope of successful predictions, highlighting the need for physically based methods that readily predict multiple protein conformations.

Abstract Trisomy for human chromosome 21 (Hsa21) results in Down syndrome (DS), one of the most genetically complex conditions compatible with human survival. Assessment of the physiological consequences of dosage-driven overexpression of individual Hsa21 genes during early embryogenesis and the resulting contributions to DS pathology in mammals are not tractable in a systematic way. A recent study looked at loss-of-function of a subset of Caenorhabditis elegans orthologs of Hsa21 genes and identified ten candidates with behavioral phenotypes, but the equivalent over-expression experiment has not been done. We turned to zebrafish as a developmental model and, using a number of surrogate phenotypes, we screened Hsa21 genes for effects on early embyrogenesis. We prepared a library of 164 cDNAs of conserved protein coding genes, injected mRNA into early embryos and evaluated up to 5 days post-fertilization (dpf). Twenty-four genes produced a gross morphological phenotype, 11 of which could be reproduced reliably. Seven of these gave a phenotype consistent with down regulation of the sonic hedgehog (Shh) pathway; two showed defects indicative of defective neural crest migration; one resulted consistently in pericardial edema; and one was embryonic lethal. Combinatorial injections of multiple Hsa21 genes revealed both additive and compensatory effects, supporting the notion that complex genetic relationships underlie end phenotypes of trisomy that produce DS. Together, our data suggest that this system is useful in the genetic dissection of dosage-sensitive gene effects on early development and can inform the contribution of both individual loci and their combinatorial effects to phenotypes relevant to the etiopathology of DS.

Understanding how the crowded and complex cellular milieu affects protein stability and dynamics has only recently become possible by using techniques such as in-cell nuclear magnetic resonance. However, the combination of stabilizing and destabilizing interactions makes simple predictions difficult. Here we show the potential of Danio rerio oocytes as an in-cell nuclear magnetic resonance model that can be widely used to measure protein stability and dynamics. We demonstrate that in eukaryotic oocytes, which are 3–6-fold less crowded than other cell types, attractive chemical interactions still dominate effects on protein stability and slow tumbling times, compared to the effects of dilute buffer.