Loading…
Per-residue prediction of where a receptor domain can be inserted into a protein without breaking it, for building allosteric and inducible switches.
No providers recorded yet. Browse all providers
Dropping a photoreceptor or a ligand-binding domain into the middle of an enzyme is one of the most direct ways to make that enzyme switchable: a trigger changes the inserted domain's conformation, and the change propagates to the host. The catch is finding the insertion site. Most positions destroy the host protein's fold or activity, the tolerant ones are not reliably predicted by surface exposure or loop annotation, and the usual remedy is a library screened across thousands of variants — feasible for one protein, not for a design campaign.
ProDomino, from Benedict Wolf, Jan Mathony, Dominik Niopek and colleagues at Heidelberg University with the Technical University of Darmstadt, predicts insertion tolerance per residue from sequence. Its key move is a training set that sidesteps the scarcity of experimental data: rather than learning from insertion mutagenesis screens, it learns from evolution, mining natural proteins in which one structural domain has been acquired inside another. Those intradomain insertions mark positions that tolerated a domain and remained functional over evolutionary time, and there are enough of them to train on.
The model itself is a small head over frozen ESM-2 representations. The biotechnological payoff is what distinguishes it: model-nominated sites yielded working optogenetic and chemogenetic switches on the first attempt in every protein tried, including CRISPR nucleases.
The dataset was built from InterPro CATH-Gene3D annotations, filtered for cases where one domain annotation interrupts another, then intersected with UniRef50 at 50 % maximum pairwise identity and restricted to proteins under 2,048 residues: 174,872 sequences spanning the tree of life, 202 insert domain superfamilies and 168 parent domain types. Insert domains were deleted; the junction and its two flanking residues were labelled positive and all other positions unknown. Sequences were embedded with the 3-billion-parameter ESM-2 model into 2,560-dimensional per-residue representations, and a two-layer perceptron with ReLU activations was trained on top under cross-entropy loss with sequence-specific masking — each iteration updates on the known positive plus one randomly drawn negative — to cope with the absence of true negatives. The released model uses the strictest of three splits, one representative per parent-insert domain combination, leaving 174 training and 46 validation sequences, checkpointed at 1,500 steps.
Against a published AraC domain insertion screen the model reaches an AUROC of 0.84; against Cas9 transposon insertion data it reaches 0.71, with several high-scoring regions absent from the published hits. Testing those directly, all four dCas9-LOV2 hybrids retained activity and three became photosensitive. Across predictions in puromycin and chloramphenicol acetyltransferase, experiments confirmed the model in 78 % of cases. Engineered switches included CAT-LOV2 variants with a 20-fold optical density difference between light and dark, a Cas12a-LOV2 hybrid at N1153 with 3-fold light-dependent reduction in editing, and Cas12a-GR2 hybrids reaching around 70 % of wild-type editing under cortisol while dropping to near the detection limit without it.
The intended users are protein engineers building biosensors and inducible effectors. Given a target sequence, the model returns a probability trace whose peaks become the handful of constructs worth cloning, replacing a library screen. The authors' own use — inserting the Avena sativa LOV2 photoreceptor or a circularly permuted glucocorticoid receptor ligand binding domain at nominated sites — produced light- and drug-inducible antibiotic resistance enzymes and genome editors in roughly six months. Running it requires a local ESM-2 3B and a GPU; the repository ships a 39 MB checkpoint, a conda environment and an example notebook under CC BY-NC 4.0.
ProDomino's transferable idea is the training signal: where experimental data for a design problem is too scarce to learn from, natural sequence archives may already contain a semi-synthetic proxy for it. The transfer test is genuine — validation proteins come from unrelated families, not held-out folds of the training distribution — and single-chain inducible variants of a Cas12a ortholog had not previously been made. The caveats are stated plainly: low scores are less reliable than high ones, with half of low-scoring Cas12a sites still tolerating insertion, and the work remains a preprint under a non-commercial license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.