Verge Genomics (now Verge Labs) published a retrospective on how its first-generation discovery engine, CONVERGE 1.0, took a novel target from human tissue data to a Phase 1b trial in amyotrophic lateral sclerosis (ALS), and what the trial showed the platform had missed.
Target discovery. CONVERGE ran a multi-omic analysis of nearly 1,000 tissue samples from ALS patients and healthy controls. It flagged PIKfyve, an enzyme in the endolysosomal system that cells use to recycle and dispose of internal cargo, as a top target: in ALS brain tissue the gene network around PIKfyve was profoundly dysregulated. PIKfyve inhibition cleared protein aggregates and improved survival in motor neurons grown from ALS patient stem cells, and reduced neurofilament in a mouse model. Verge then developed VRG50635, a brain-penetrant PIKfyve inhibitor.
Biomarker from the network. In the PIKfyve gene network derived from human tissue, GPNMB was one of the genes most tightly coupled to PIKfyve. Verge found GPNMB released by cells and elevated in ALS patient blood and spinal fluid, and the drug raised GPNMB in cells, patient-derived motor neurons, rodents and healthy volunteers in proportion to exposure. This gave a pharmacodynamic readout of target engagement before any ALS patient was dosed, and was used to set the dose for the Phase 1b.
Trial result. The Phase 1b enrolled 54 ALS patients, with efficacy defined as a 30% drop in plasma neurofilament light chain (NfL). The drug reached the brain at roughly twice the concentration needed to fully inhibit PIKfyve, but plasma NfL rose instead of falling, which Verge attributes to the drug's mechanism pushing cell contents out, and the drug missed its primary efficacy endpoint. A third of patients could not tolerate the lowest dose; of 48 baseline measures, only plasma NfL and GFAP separated them from the others, both elevated, which tracked more advanced disease.
What the platform missed. Verge names three gaps: CONVERGE 1.0 was trained mostly on bulk post-mortem brain tissue to recover the average ALS-versus-control signal, so it could not resolve patient-level heterogeneity; it could not connect the blood samples collected during the trial back to brain biology; and it could not predict how biomarkers would move once a drug was given.
CONVERGE 2.0. Verge is training generative AI models on its dataset of more than 12,000 human tissue samples across 6,000 patients and 15 million single-cell profiles, with clinical and genetic annotations, plus the trial data. The aim is a "virtual biopsy": an inferred molecular picture of a patient's brain from a routine blood draw, used to match patients to therapies and predict disease course and biomarker response.
Details