vBx-1.0

vBx-1.0, nicknamed "Virtual Biopsy," is a multimodal foundation model for precision neurology developed by Verge Labs and released on June 16, 2026. The central problem it addresses is that the molecular state of a living patient's brain is largely inaccessible: brain tissue cannot be routinely sampled, so the disease-relevant signals that would guide therapy selection are hidden. vBx-1.0 reconstructs that molecular state from measurements obtainable in a living person, including blood RNA-seq, genetics, and clinical data, and uses it to predict disease progression, therapy response, and on-target toxicity.

What makes the model novel is that the blood-to-brain mapping was not learned through task-specific supervision but emerged from a self-supervised contrastive objective over matched patient modalities. Rather than treating blood and brain as separate prediction targets, vBx-1.0 fuses any available combination of a patient's blood RNA-seq, brain RNA-seq, and genetics into a single 512-dimensional embedding that captures shared and modality-specific biology.

Positioned alongside virtual-cell models such as scVI and scGPT, vBx-1.0 differs by operating at the level of whole patients across multiple diseases and tissues, with a deliberate focus on translational neurology endpoints rather than generic cell-state representation.

#Key Features

• Blood-to-brain reconstruction: Recovers disease-relevant brain module activity from blood alone, a capability that emerged from the training objective rather than explicit supervision.
• Unified patient embedding: Fuses any combination of blood RNA-seq, brain RNA-seq, and genetics into one 512-dimensional representation usable for downstream prediction.
• Modality-masking contrastive objective: Unlike CLIP, it preserves both modality-specific and shared signal by masking modalities and pulling the masked view toward the full-modality view in latent space.
• Lightweight and fast: At roughly 90M parameters it runs on a single GPU and embeds 1,000 patients in 16 seconds on a laptop.
• Translational endpoints: Directly targets patient stratification, treatment-response prediction, and zero-shot on-target toxicity assessment.

#Technical Details

vBx-1.0 is a multimodal transformer with approximately 90 million parameters. It was trained on 3,272 human brains from 1,586 patients across five diseases, spanning brain and blood expression, genetics, and clinical trajectory, with every patient contributing multiple modalities. This corpus is a subset of Verge's larger atlas of more than 12,000 brain samples across 6,500 patients. The model obeys data-scaling laws, with gains driven by more patients and modalities rather than more parameters. On reconstructing brain state from blood, it sets state of the art, recovering Parkinson's substantia-nigra modules at an average Pearson of 0.40 (versus 0.27 for baselines and 0.08 for scVI), Alzheimer's cortex at 0.41, and DLPFC at 0.49, with single-gene reconstruction at 0.25 versus 0.20 for a linear model. For Parkinson's L-DOPA response in the PPMI cohort, it identifies responders with 69% accuracy versus 52% all-comers, enabling roughly a 43% reduction in trial size at fixed power. In a zero-shot toxicity test on Verge's Phase 1b trial of VRG50635 (a PIKfyve inhibitor for ALS, NCT06215755), baseline blood from 33 patients predicted a PIKfyve brain module whose low activity flagged later adverse events, an enrollment screen that would have cut dropout from 39.4% to 26.1%.

#Applications

vBx-1.0 is aimed at CNS drug developers and clinical trialists who need to enrich for likely responders, anticipate on-target toxicity, and reduce trial size and dropout. By turning routine blood and genetics into a proxy for inaccessible brain state, it supports patient stratification, treatment-response prediction, and biomarker discovery, including interpretable signals such as a substantia-nigra vascular module linked to levodopa non-response. The model is available in preview only through Verge's CONVERGE platform to partner drug developers.

#Impact

vBx-1.0 reframes precision neurology around a reconstructable molecular state of the living brain, demonstrating that a compact, self-supervised multimodal model can outperform virtual-cell baselines on translational endpoints while running on commodity hardware. Its retrospective gains on real trials, a 33% relative enrichment in responder identification and a 34% reduction in projected dropout, point to concrete value for trial design. Important limitations remain: the model is closed, with no peer-reviewed paper, preprint, public code, or weights, and v1.0 covers only genomics, transcriptomics, and clinical data, leaving proteomics, imaging, and longitudinal modeling on the roadmap.