Nexus

A single blood draw yields hundreds of millions of transcript-level observations of cell-free RNA — fragments shed into circulation by tissues throughout the body, reflecting what those tissues are actively expressing rather than what is written in the genome. The signal separating early disease from ordinary variation is spread thinly across that whole transcriptome rather than concentrated in a handful of markers, and labeled patient cohorts run to thousands of samples, not millions. A supervised model trained directly on such a cohort spends most of its scarce labels learning what circulating RNA looks like in the first place. Nexus pays that cost once: a self-supervised transformer pretrained on cfRNA expression profiles without disease labels, whose representation is intended as the substrate that downstream supervised models start from.

Nexus is being developed by Network Bio, a Palo Alto biotechnology company that launched on 19 August 2026 with $50 million in financing from Section 32, Thiel Bio, Founders Fund and other life science and AI investors. The model was announced the same day alongside a collaboration with NVIDIA whose stated purpose is to scale Nexus from research-cohort training to population scale, using NVIDIA's accelerated computing, the BioNeMo Recipes training collection and Parabricks for sequence processing. The model is in active development rather than finished: Network Bio reports early engineering milestones only, and no checkpoint, code, evaluation or preprint has been released.

The approach has a direct lineage. Exai-1, in Nature Machine Intelligence, and Orion, in Nature Communications in 2024, both applied transformer and generative models to circulating RNA; Orion reported 94% sensitivity and 87% specificity for early-stage lung cancer detection. Both came out of Exai Bio, the liquid-biopsy company co-founded by Network Bio technical co-founder Hani Goodarzi, and Orion's first author Mehran Karimzadeh is now Network Bio's founding AI engineer. Nexus generalizes that per-indication work into one pretrained model. The name is generic — an unrelated multi-scale simulator of gene-regulatory dynamics also goes by Nexus; this entry covers Network Bio's cfRNA model.

#Key Features

• Label-free pretraining objective: Trained self-supervised on cfRNA expression profiles, so its representation is shaped by the structure of circulating transcriptomes rather than by any one disease's labels.
• Pretrain once, specialize many times: Network Bio describes Nexus as the substrate for downstream supervised models across oncology and non-oncology indications, not as a diagnostic in its own right.
• Harmonized multi-site training data: Samples come from a research network of academic biobanks — Mass General Brigham, the University of Pennsylvania and the University of Colorado Anschutz — under common selection criteria, quality standards and cross-site harmonization, which is what makes multi-center cfRNA usable as a single corpus.
• Tissue paired with blood: The underlying datasets link patient-derived tissue to matched blood samples and longitudinal clinical outcomes, anchoring circulating signal to the tissue it came from.
• GPU-accelerated data preparation: Secondary analysis, alignment, quantification and quality control gate how fast biobank access becomes training data; NVIDIA Parabricks cuts per-sample processing time and compute cost ahead of training.

#Technical Details

Nexus is a transformer trained with a self-supervised objective over cfRNA expression profiles. Training runs on NVIDIA BioNeMo Recipes, an open-source collection of Transformer Engine-accelerated model implementations and training recipes that scale transformer pretraining across GPUs with fully sharded data parallelism. Network Bio reports improved training throughput and reduced time to convergence — the only quantitative result the company has published about Nexus. Parameter count, layer count, context length, tokenization scheme and pretraining corpus size have not been disclosed, and no biological benchmark for Nexus has been reported; the 94%/87% lung cancer figures in the announcement belong to the earlier Orion model, not to Nexus.

#Applications

The intended use is as an upstream representation for clinical prediction. Network Bio's biobank network spans oncology, immunology, metabolic, cardiovascular and autoimmune disease, and the company positions Nexus as the starting point for supervised models across those areas: early disease detection from a blood draw, biomarker discovery, patient stratification, and target and therapeutic development for pharmaceutical partners. Network Bio has signed a $30 million-plus co-development and licensing agreement with a Fortune-100 healthcare company to apply its AI architecture to novel disease signatures.

#Impact

Cell-free RNA has been modeled at cohort scale before, but pretraining a general-purpose foundation model on it has been limited less by architecture than by access to harmonized multi-center samples — the gap Network Bio's biobank network is assembled to fill. Whether Nexus delivers on that premise is not yet answerable from public evidence: it was announced without a preprint, without a released checkpoint, without code, and without any evaluation of the pretrained model on a biological task. The training corpus is patient-derived clinical material governed by institutional agreements, so it is unlikely to be redistributable even if the model itself is later opened.