Cipher: Inverse Drug Target Discovery from Human Genetic Variation

How do you know the cure for cancer isn’t you? Somewhere in a biobank there is a person whose genome already solved a disease that kills everyone else. One woman with extremely low LDL cholesterol and a loss-of-function variant in PCSK9 led to a drug class generating over a billion dollars a year. The bottleneck was never chemistry. It was knowing which gene to target, and the field has been waiting on luck to say. Problem: Drug development costs roughly $2.6B and twelve years per approved drug, and only about 14% of candidates entering clinical trials reach approval, much of it traceable to targets chosen from cell models rather than human outcomes. Human genetic evidence roughly doubles the odds of clinical success, yet no automated system takes a disease name and returns a ranked, mechanistically explained target list grounded in population genetics. Approach: Cipher inverts the usual direction of inference: rather than simulating forward from genome to predicted phenotype, it starts from an outcome that already occurred in a living person and works backward to mechanism. Four layers. A population layer identifies individuals carrying variants in disease-relevant genes who do not develop the disease, using biobank-scale linked genotype and phenotype data. A variant layer applies a DNA language model to predict functional consequence for any observed variant, including those absent from clinical databases. A structural layer places the mutated protein into its AlphaFold-predicted conformation to assess altered binding interfaces. A network layer propagates that perturbation across a protein interaction graph via graph convolution, with the goal of learning which pathway between variant and protective phenotype carries the signal, how strongly each edge contributes, and in which direction. Results: A prototype implements the structural and network layers, substituting gnomAD constraint metrics for the biobank layer and ClinVar records for the variant layer. Low pLI paired with high observed loss-of-function carrier count identifies genes humans already tolerate losing, the prerequisite for safe inhibition. Evaluated retrospectively on hypercholesterolemia, type 2 diabetes, and ADHD, the prototype surfaced 23 targets with approved or clinical-stage drugs and recovered the correct mechanism and modality for 16, each match adjudicated by an LLM against the approved drug's documented mechanism. It ranks PCSK9 first (obs_lof 57, pLI 0.0000, STRING experimental 0.982 with LDLR), recommending a monoclonal antibody against the LDLR-binding domain, matching evolocumab and alirocumab. Four cases failed traceably: for GCK, AKT2, and HNF4A the pipeline inferred the wrong therapeutic direction, marking the boundary of what constraint metrics resolve without expression or signed-interaction data. Limitations: The graph convolution currently runs untrained, aggregating features over topology rather than fitting outcomes; training it on known target and non-target pairs is the next step. Candidate discovery draws on curated databases that themselves aggregate clinical evidence, biasing the pool toward drugged genes. Removing that confounder requires the biobank layer, which is why it sits first. Context: Prior work applies sequence models to variant effects and graph networks to biomarker discovery separately. The contribution is the combined pipeline and the evaluation protocol: benchmarking against approved drugs rather than GWAS hits, asking directly whether the system would have found the drugs we know work.