SandboxAQ and NVIDIA are piloting a drug discovery stack that grounds every answer in a biomedical knowledge graph and returns it fast enough for a scientist to use in conversation. The system pairs the AQ Knowledge Graph, an auditable store of biomedical evidence, with NVIDIA BioNeMo orchestration and GPU-accelerated inference. A scientist asks a mechanistic question in natural language and receives a fully traceable evidence pack in seconds.
Because the stack runs through BioNeMo Inference Runtime, it deploys where the data already lives: on a partner's own GPUs behind their firewall, or inside their cloud account. That pairing gives target and biomarker teams the provenance they need to defend a decision and the speed they need to explore one.
Target and biomarker discovery stalls on a mundane problem: the evidence that justifies a decision sits scattered across genomics, proteomics, transcriptomics, longitudinal clinical cohorts, and a literature base too large for any team to read. A language model will answer confidently over that landscape. Defending the answer to a reviewer takes more, namely the mechanistic pathway, the source database, the PMIDs, and the dataset version behind every claim. Programs commit real budgets on these answers, so the evidence has to hold up.
We built the AQ Knowledge Graph as an auditable substrate for discovery. It is a heterogeneous biomedical knowledge graph whose base layer spans roughly 188,000 nodes and 6.7 million edges, harmonized from 15 external public databases. Each program then adds its own proprietary omics and clinical layers on top.
Our architecture can and has also been used at scale. For example, for one customer our knowledge graph holds about 6.7 million relationships across 222,000 biological entities and integrates more than 100 TB of longitudinal clinical and multi-omic data spanning 10 years for over 12,000 participants.
Network-based prioritization, using random walk with restart via multiXrank across 15,703 genes, recovered 49 clinically validated PD targets and ranked UCHL1 first, SNCA fourth, and LRRK2 seventh. Our graph neural network work goes further. Heterogeneous graph transformers, with nodes grounded in pretrained domain representations (ESM-2 for genes and proteins, ChemBERTa and PubMedBERT for compounds and ontology terms), prioritize roughly 60% of known PD targets within the top 10% of genes screened. We benchmarked these models against degree-based and graph-diffusion baselines to rule out structural shortcuts and confirm that the rankings reflect biology.
The pattern extends beyond neuroscience. In a separate immunoscience program, we assembled a custom graph of more than 6 million edges, confirmed its fidelity by recovering 97% of a known signaling pathway (115 of 119 nodes), and then narrowed 21,000 candidates to 14 high-potential biomarkers.
Our user interface for the AQ Knowledge Graph sits on top of the graph and lets a scientist query it directly. This agentic Graph-RAG interface takes a natural-language question, plans an approach, decomposes the question into Cypher queries against the graph, resolves the relevant entities, retrieves the supporting subgraphs, and synthesizes an evidence pack in which every claim traces back to a node, an edge, a source database, and its PMIDs.
Grounding raises answer quality measurably. Against a fixed benchmark of 99 gold answers drawn from the graph, the interface reached 0.70 recall at cited evidence, compared with 0.28 for the same model working without the graph. It scored hits@10 of 0.60, where the ungrounded model scored 0, and earned an LLM-judged correctness of 4.76 out of 5, compared with 3.39.
Grounding earns its trust at a cost in latency, and in an agentic system that cost compounds at every step.
A single collaborator-grade question shows the problem. Asking whether there is evidence that alpha-synuclein is directly involved in mitochondrial vesicle biology took three to four minutes end-to-end in live demos, because the interface plans, issues multiple graph queries, expands neighborhoods, and only then synthesizes an answer. Synthesis alone measured about 42 seconds per query, against roughly 16 seconds for an ungrounded model. Faithfulness is what that extra time buys, and the bill lands squarely in the interaction loop.
GNN scoring adds a second cost center. Ranking an entire proteome, 15,703 genes with heterogeneous message passing, together with explainability subgraphs of roughly 1,000 nodes each generated by masking nodes to explain a prediction, currently runs as an overnight batch well outside a live conversation.
Annotation at scale multiplies all of it. Our roadmap calls for systematic biological and clinical annotation across entire patient-subtyping datasets, which turns a single three-minute query into thousands of them.
Latency shapes how scientists actually work. A three-minute round trip forces them to batch their questions and walk away. A sub-30-second round trip lets them interrogate a hypothesis, follow the graph where it leads, and disprove their own idea in a single sitting. Interactivity is what turns a system that generates evidence packs into one that changes what a scientist believes.
The pilot we are proposing with NVIDIA composes four layers, each assigned to what it does best.
NVIDIA BioNeMo agents orchestrate the workflow. They handle planning and tool routing and call model endpoints as containerized inference microservices, with human approval gates at the points that matter.
The knowledge-graph interface handles retrieval. It resolves entities, retrieves subgraphs, and constructs the evidence pack, keeping every downstream claim traceable.
GPU-accelerated graph transformers do the ranking. PyTorch Geometric heterogeneous graph transformers score and rank targets and biomarkers, moving proteome-scale ranking and explainability out of the overnight batch and into the interactive loop.
The AQ Knowledge Graph stays the source of truth. Every assertion resolves to a node, an edge, a source, and a version.
Served through BioNeMo Inference Runtime, the same stack runs where the data already lives: on a partner's own GPUs behind their firewall as NIM microservices, or inside their cloud account. For pharma partners whose proprietary cohorts have to stay in the building, that deployment property is what makes the project possible in the first place.
OpenFold3, co-developed with the OpenFold Consortium and fully open-sourced under Apache 2.0, ships as a GPU-optimized inference microservice through BioNeMo. AQAffinity, our structure-free affinity model built on OpenFold3, followed it. Our SAIR dataset of more than one million co-folded protein-ligand structures with IC50 labels went into BioNeMo's DiffDock NIM and measurably improved docking accuracy. And our CUDA-accelerated DMRG collaboration delivered speedups of more than 80x over CPU in quantum chemistry simulation.
Our internal evaluation of the NVIDIA-optimized inference path for OpenFold3 is the reason we expect these gains to carry over. Three techniques did most of the work:
The largest gains landed exactly where our agentic workload lives, in many small, latency-sensitive calls. Adaptation took real effort: the optimized implementation stands alone and so lags a fast-moving upstream repository, template preprocessing needed attention, and the driver requirements were strict.
We will report full end-to-end numbers that reflect what a scientist experiences at the keyboard, measured under stated test conditions on representative A100-class and H100-class hardware:
In the near term, we are consolidating the knowledge-graph interface into a single library that carries its production features: reflection loops, streaming synthesis, provenance-rich Cypher recipes, and structured output enforcement. From there, we scale annotation across patient-subtyping cohorts and extend the pattern from Parkinson's disease to additional programs.
The larger goal is a fully agentic loop. A scientist states an objective, the system plans and executes graph retrieval and GPU-scale ranking, and it hands back a ranked, mechanistically explained, and fully traceable shortlist of candidates in about the time it takes to read the question back.
Grounding makes agentic discovery trustworthy. Inference performance makes it usable. This pilot pursues both at once.
Explore BioNeMo Inference Runtime
Explore SandboxAQ Drug Discovery
