RyboDyn Finds a Second Human Proteome Hiding in Cancer's Dark Transcriptome
A San Diego company has mapped millions of unannotated RNAs to tens of thousands of previously unknown proteins, then targeted the tumor cells with a tailored antibody
Of roughly 13,600 drug-target pairs in the global preclinical and clinical pipeline, a quarter rely on just 38 protein targets. The rate at which new targets enter the development pipeline has fallen from around one hundred a year a decade ago to roughly thirty in 2024. While design capabilities keeps growing, the list of things worth targeting has not.
RyboDyn, a San Diego company working out of Lilly Gateway Labs, is making the case that a large part of that list has been sitting in plain sight. In a preprint posting to bioRxiv this week, the company describes what it calls a cryptic human proteome, tens of thousands of peptides that reference annotations never captured.
The scale is the headline. Applied to lung and colorectal tumors, matched healthy tissue, and cancer cell lines, RyboDyn's sequencing platform RyboCypher resolved roughly 6.6 million dark RNA loci, more than 97% of which are absent from existing non-coding RNA databases. Those transcripts yielded about 16 million candidate open reading frames. Searching those predictions against roughly half a billion mass spectra drawn from 2,229 patient samples returned about 80,000 cryptic peptides at a false discovery rate below 1%, of which roughly 10,000 are cancer-associated or cancer-upregulated. The company has assembled the result into a database it calls CypherAtlas. Compared against the microprotein and peptidein catalogs published in Nature earlier this year, the overlap is close to nil.
The cryptic loci also land where the industry already spends its money. They cluster into the protein classes drug hunters have chased for decades, deubiquitinases, E3 ligases, and transcription factors in cellular proteomics, and cell adhesion molecules, transporters, and receptors in membrane preparations.
"Oncology is overdue for another checkpoint-inhibitor moment," said Corey Dambacher, president and co-founder of RyboDyn and senior author of the study. "That doesn't come from finding a different way to go after the same three dozen targets. We can really only accomplish that by discovering new biology, and RyboCypher is built to find it."
The obvious objection is that a search space of 16 million candidates will produce noise. RyboDyn uses the ESM-2 protein language model to measure how surprising a sequence looks to a model trained on known proteins. The score ran higher for cryptic proteins than for the canonical reference set, and that elevation was concentrated almost entirely in the novel regions absent from any annotation. The sequences still sat well below shuffled and random controls, and structure prediction largely returned confidently folded models.
"These proteins were, in the most literal sense, perplexing to the best protein language models available," said Imad Ajjawi, chief executive and co-founder. "AI could not have guessed this biology existed, because it was never in the training data. Yet the sequences are nowhere near random. As proteins they're statistically novel, but structurally sound."
With its own proprietary data, RyboDyn is building a multimodal model, DarkCypher. The sell is that no public model can learn a layer of biology that was never in anyone's training corpus.
The preprint's proof of concept is cYBX1, a cryptic protein expressed from the YBX1 locus. Canonical YBX1 is a well-known oncogenic regulator and has proven stubbornly undruggable. The cryptic version gives rise to a tumor-restricted peptide-MHC complex on the cell surface. TCR-mimic antibodies raised against it bound at 626 picomolar and showed no measurable binding to a nearest-neighbor peptide differing by three amino acids. One antibody recognized the target across three HLA-A*02 isoforms, which the company estimates raises the addressable population above forty-five percent in Western European and North American populations. Formatted as an antibody-drug conjugate, the lead killed tumor cells in vitro.
The prevalence data are where investors should look hardest. RyboDyn's top-ranked cell-surface candidate is upregulated roughly 7.5-fold in ~80% of the lung squamous cell carcinoma patients assessed (a cohort of over 100), with significant tumor upregulation also seen in clear cell renal cell carcinoma (ccRCC) and pancreatic ductal adenocarcinoma (PDAC). By comparison, a well-known blockbuster target like HER2 is actionable in only 15-20% of breast cancers.
"This work establishes the principles and validates the concepts RyboDyn is pursuing," said Gordon B. Mills, director of the SMMART trials at the Knight Cancer Institute at Oregon Health and Science University.
The usual qualifiers apply. The work is a preprint and has not been peer reviewed, the sequencing chemistry behind RyboCypher is licensed from Oregon Health and Science University and remains undisclosed, and the tumor killing is in vitro with no animal data reported. What RyboDyn has done is move the hard part. Finding new targets used to be the scarce step, and the company now claims one to five high-value druggable candidates per indication. Converting any one of them into an asset is the older problem, and it is the one the field will judge RyboDyn on.
