How We Design mRNA for Aging Biology
Three integrated modules — comparative genomics, AI sequence design, LNP formulation prediction — running in series. Evolutionary signal in, ranked delivery-ready candidate out.
Scanning 40+ long-lived species for conserved variants
The starting point is a curated genome dataset assembled from species with documented longevity phenotypes: exceptional lifespan, cancer suppression, or resistance to age-related disease. We run multispecies variant alignment across all 40+ species simultaneously — not pairwise against a human reference, but a full cross-species comparison.
Each variant is assigned a conservation score weighted by phylogenetic distance — variants shared across distantly-related lineages are upweighted. We then apply a functional annotation filter to retain variants in proteins with known or predicted roles in aging-relevant pathways: proteostasis, DNA repair, mitochondrial dynamics, inflammation suppression, and growth-longevity tradeoff regulation.
The output is a ranked list of candidate longevity proteins — ordered by conservation evidence, pathway relevance, and estimated upregulation tractability via mRNA.
Predicting safe upregulation from sequence
Given a target protein from the genomics module, the sequence design system generates candidate mRNA sequences optimized across multiple axes: codon optimization for translation efficiency in human neurons, 5′ and 3′ UTR engineering for stability and half-life, and minimization of innate immune activation motifs.
The ML model is trained on curated protein upregulation literature to predict which sequence configurations are likely to achieve target expression levels without triggering off-target effects. Off-target prediction is integrated at the design stage — not as a post-hoc filter — meaning sequences that score poorly on specificity are deprioritized before they reach in vitro testing.
We do not claim accuracy benchmarks without experimental validation. The model guides prioritization; wet-lab data will validate or falsify.
Modeling BBB-crossing formulations in silico
Delivering mRNA to the central nervous system remains one of the hardest problems in the field. The blood-brain barrier excludes most systemically-administered nanoparticles through a combination of tight junctions, efflux pumps, and restricted transcytosis pathways. Lipid nanoparticle formulation parameters — ionizable lipid pKa, lipid molar ratio, PEG density, particle size — strongly influence whether a formulation reaches the brain.
Our computational approach treats LNP formulation as a parameter search problem. We use predictive models trained on published LNP-CNS data to rank formulation configurations by predicted BBB penetration, encapsulation efficiency, and release kinetics. This narrows the wet-lab search space from thousands of candidates to a focused set for experimental validation.
caVos does not claim proprietary validated BBB-crossing formulations. The models guide which candidates to test first — wet-lab data is the source of truth.
From Variant to Validated Candidate
The pipeline runs left to right: genomic variant data in, in vitro testing targets out. Each stage narrows the candidate space before the next stage begins. In vitro validation — the rightmost node — is where computational predictions meet biological reality.