Loading…
Structure-based protein-RNA screening that reads position-wise nucleotide preferences off a single complex structure and ranks libraries in seconds.
No providers recorded yet. Browse all providers
An RNA-binding protein almost never recognizes a single sequence. It recognizes a family of them, with a graded preference at every position, and that preference is what a position weight matrix records. Measuring one experimentally — by RNA Bind-n-Seq, RNAcompete-S or HTR-SELEX — means synthesizing and assaying a library for each protein. Structure-based scoring functions such as FoldX and the Rosetta energy functions can in principle read the preference off one complex structure instead, but they have to thread each candidate sequence onto the binding site and re-minimize the complex, so a library of 10^5 hairpins is days of compute rather than minutes.
PRIS, developed by Tingjun Hou's group at Zhejiang University and released as a preprint in September 2026, replaces that per-sequence energy calculation with a single forward pass. It is a framework of two heads over one shared feature extractor: PRISeq predicts a probability distribution over nucleotide types at each RNA position of a bound complex, and PRIScore predicts residue–nucleotide distance distributions, which is used to pick native-like poses when no crystal structure is available. Because PRISeq needs only one input structure to produce the whole position-wise profile, scoring a candidate sequence afterwards costs little more than a table lookup.
The two heads compose into a workflow. Where no complex has been solved, AlphaFold 3 generates candidate poses, PRIScore re-ranks them, and the selected structure goes to PRISeq. AlphaFold 3 is an upstream dependency here, not part of the model — PRIS predicts binding preference, not structure.
The pocket is a residue-level graph of everything within 10 Å of the bound RNA, with per-residue hidden states from ESM3 among the node features; the RNA is a nucleotide-level graph over backbone atoms only. Both pass through a feature extractor that places graph attention inside the stable, non-dissipative ODE framework of an Anti-Symmetric Deep Graph Network — the authors call the combination A-GAT — alongside sparse k-Maximum Inner Product attention, whose top-k selection approximates full attention at linear memory cost. PRISeq closes with a softmax over 12 nucleotide types; PRIScore uses a mixture density network over the continuous residue–nucleotide distance. Training used 9,110 protein–nucleic acid complexes from the PDB, then fine-tuning on PRA328, 328 complexes carrying affinity labels from PDBbind and PRA310, under a loss adding a Pearson-correlation term on predicted versus measured affinity. Chains sharing more than 40% identity with any test protein were clustered out with MMseqs2.
On a PWM benchmark of 15 crystal structures covering 108 nucleotide positions, PRISeq reached 65.6% accuracy and an MAE of 0.75, against 42.9% and 1.04 for NA-MPNN and 32.7% and 1.17 for FoldX. In virtual screening it achieved an EF₀.₅% of 14.40 against MS2 and 9.88 averaged over NELF-E and GFP aptamer libraries.
The practical use is triaging RNA libraries before synthesis: ranking aptamer candidates against a protein target, narrowing a SELEX search space, or scoring designed hairpins for a packaging or reporter system. Separately, PRISeq lets structural biologists read a binding motif off a complex — its predicted PWM for splicing factor SRSF1 recovered the purine-rich preference and 3'-terminal C of the experimental motif — which matters for the many RNA-binding proteins with a solved structure but no selection data.
PRIS is a preprint awaiting peer review, and its evaluation is retrospective: every library it screens had already been measured, and no designed sequence was tested at the bench. It also scores one target at a time and does not model specificity, so a sequence it ranks highly is not thereby shown to avoid off-target proteins; the authors name this as the main direction for future work, along with evaluation beyond the three targets used here. What it does establish is that a preference model trained once on general protein–nucleic acid structures transfers to unseen targets without refitting, and that doing so collapses a screening cost measured in GPU-days to one measured in seconds.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.