In a preprint, the Pythia neural network separated the contributions of RNA sequence and structure to protein binding: cell experiments showed that at all 11 sites where structure contributed more and the original RNA bound the protein, mutations that preserved structure kept binding closer to its original level than mutations that disrupted it
In a preprint, the Pythia neural network separated the contributions of RNA sequence and structure to protein binding: cell experiments showed that at all 11 sites where structure contributed more and the original RNA bound the protein, mutations that preserved structure kept binding closer to its original level than mutations that disrupted it
On October 13, Hani Goodarzi, a researcher at Arc Institute and the University of California, San Francisco, and colleagues posted a preprint, a manuscript that has not yet undergone peer review, on bioRxiv. They mapped the sites where 67 proteins bind RNA in human cells, trained the Pythia neural network on these data, and tested its predictions using synthetic RNAs with mutations.
RNA-binding proteins influence how a cell processes an RNA molecule, how long it survives, and how much protein the cell produces from it. Binding sites are identified using CLIP: ultraviolet light crosslinks proteins to RNA, an antibody captures the protein, and sequencing reads the RNA fragments attached to it. A protein can recognize the RNA sequence or the structure of the strand: A pairs with U, and G pairs with C or sometimes U, allowing a region of the strand to fold into a hairpin.
The earlier ENCODE map covered 150 proteins in two cancer cell lines, with a separate antibody required for each protein. In the authors’ method, ecCLIP, every protein carries the same tag: a fragment of the fluorescent protein mNeonGreen is inserted into each protein’s gene. A single antibody against the tag can therefore capture all the proteins, making their binding maps comparable. The researchers identified about 9,1 million binding sites for 67 proteins in human embryonic kidney cells; 40 of these proteins are absent from ENCODE. Adding these data to the training of the Borzuki neural network improved its predictions of RNA lifetime from DNA sequence.
Goodarzi and colleagues studied RNA structure in 2012, using a grammar, a set of rules describing paired regions of an RNA strand, to identify eight structural elements associated with messenger RNA stability. Pythia incorporates these rules into a neural network. One branch reads the sequence directly. The other receives the same sequence through a predefined layer that marks where each allowed base pair could form along the strand; subsequent layers select the pairs that matter. A separate Pythia model was trained for each protein to predict whether it would bind an RNA fragment. At each site, the two branches indicate whether sequence or structure contributes more to the prediction.
The researchers tested this separation using an MPRC experiment. If a protein recognizes RNA structure, replacing both bases in a pair with another allowed pair, such as replacing A–U with G–C, should preserve binding, whereas making the same number of substitutions that disrupt pairing should cause the protein to detach. They selected 30 binding sites for five proteins whose binding depended on structure to different degrees. For each site, they made synthetic DNA segments that cells transcribe into RNA: the original sequence, a version with substitutions that preserved base pairing, and a version with the same number of substitutions that disrupted it. They introduced these segments into cells and measured how much of each RNA remained bound to the protein. The original RNA bound the protein at 19 sites. At 11 of these, structure contributed more than sequence to Pythia’s prediction. At all 11, binding by RNAs with preserved base pairing remained closer to the original level than binding by RNAs with disrupted pairing. At the remaining 8 sites, where structure did not contribute more than sequence, there was no significant difference.
The researchers used the same method to test 88 combinations of a protein and a naturally occurring single-base substitution at its binding site. Genomic studies associate these substitutions with human traits, such as cholesterol levels. In 24 cases, the substitution significantly changed binding. The correlation between Pythia’s predictions and the measurements was 0,33 (maximum 1). According to the authors, the measured change provides a testable hypothesis for how the substitution affects the trait.