Acute myeloid leukemia (AML) shows the problem well: patients who share a diagnosis, and sometimes the same mutations, can still differ in the biology that drives their disease and shapes how they respond to treatment.
Acute myeloid leukemia (AML) shows the problem well: patients who share a diagnosis, and sometimes the same mutations, can still differ in the biology that drives their disease and shapes how they respond to treatment.
OCEAN investigation view: evidence across sources, competing hypotheses ranked by weight of evidence, and molecule exploration with NVIDIA BioNeMo.
Biolevate built OCEAN to run research like an investigation. It gathers evidence from many layers of biology and many sources, lets competing explanations confront each other, and brings forward the areas worth pursuing with the reasoning and the evidence attached.
Biolevate built OCEAN to run research like an investigation. It gathers evidence from many layers of biology and many sources, lets competing explanations confront each other, and brings forward the areas worth pursuing with the reasoning and the evidence attached.
Research run as an investigation
Research run as an investigation
Built on Biolevate's core technology, which coordinates more than 100,000 AI agents across complex workflows, OCEAN runs agentic workflows working toward the same research objective. Connected by one shared memory, these agents draw on proprietary data alongside over 20 curated sources spanning transcriptomics, protein networks, human genetics, pharmacology, safety, and biomedical literature.
Built on Biolevate's core technology, which coordinates more than 100,000 AI agents across complex workflows, OCEAN runs agentic workflows working toward the same research objective. Connected by one shared memory, these agents draw on proprietary data alongside over 20 curated sources spanning transcriptomics, protein networks, human genetics, pharmacology, safety, and biomedical literature.
An investigation moves through stages a researcher can follow on screen: evidence intake, signal discovery, mechanism comparison, hypothesis ranking, molecule exploration and expert review. At each stage, agents propose explanations and test them against each other. When the evidence contradicts an explanation, OCEAN drops it and records why. Every inference stays linked to the data, the source and the reasoning step behind it.
An investigation moves through stages a researcher can follow on screen: evidence intake, signal discovery, mechanism comparison, hypothesis ranking, molecule exploration and expert review. At each stage, agents propose explanations and test them against each other. When the evidence contradicts an explanation, OCEAN drops it and records why. Every inference stays linked to the data, the source and the reasoning step behind it.
What the investigation found in AML
What the investigation found in AML
In AML, this approach surfaced a transcriptomic phenotype marked by high AGR2 expression and associated with endoplasmic-reticulum stress biology. The signal adds a functional layer to mutation-based classification: it may help explain why patients with similar genetic profiles behave differently, and it gives researchers a concrete direction for treatment-response studies.
In AML, this approach surfaced a transcriptomic phenotype marked by high AGR2 expression and associated with endoplasmic-reticulum stress biology. The signal adds a functional layer to mutation-based classification: it may help explain why patients with similar genetic profiles behave differently, and it gives researchers a concrete direction for treatment-response studies.
Researchers reviewed the evidence chain at each stage: which data supported the signal, which alternative explanations were set aside, and whether the result justified a treatment-response hypothesis. The agents organize and connect the evidence. NVIDIA NIM, part of NVIDIA AI Enterprise, is a set of easy-to-use microservices designed for secure, reliable deployment of high-performance AI model inferencing across clouds, data centers, and workstations. The researchers own the biological interpretation and decide what is fit for publication.
This discovery resulted in a patent application covering Biolevate's fresh approach and serves as the foundation for multiple other discoveries currently undergoing clinical validation, laying the groundwork for what comes next in precision oncology.
Accelerated knowledge retrieval and molecular validation
OCEAN relies on intensive computation and optimization, with NVIDIA NIM microservices taking on three jobs in the workflow.
The first is finding evidence across thousands of documents. On the biomedical retrieval benchmarks Biolevate uses to qualify search, moving to NVIDIA Nemotron Embed 1B improved ranking quality by about 8%, with one model in place of the previous dense-and-sparse pair.
The second is reading at scale: agents extract facts from articles, patents and reports by the thousands. On private evaluation sets whose answers have never been published, NVIDIA Nemotron 3.5 Super VL matched the frontier model Biolevate uses in production on routine metadata extraction. NVIDIA Nemotron 3.5 Super VL is a 120B open MoE model activating just 12B parameters to deliver maximum compute efficiency and accuracy for complex multi-agent applications. The plan is to run it alongside frontier models. Biolevate's LLM gateway keeps the hardest reasoning on frontier models and can send high-volume, well-scoped extraction tasks to Nemotron, including for customers who need models running in their own environment. Biolevate also plans to explore with NVIDIA how tuning and fine-tuning can shorten the model's reasoning.
The third is computationally testing a hypothesis at the molecule level. Biolevate is integrating NVIDIA NIM microservices as specialized sub-agents called by OCEAN’s agents: Boltz-2 for structure and binding prediction, GenMol for molecule generation, and NVIDIA BioNeMo open models nvDock for pose generation and DualBind for binding-affinity evidence. OCEAN records each call with its model version, inputs, outputs and raw scores, so molecule-level results join the same evidence chain as the biology behind them.
Biolevate OCEAN architecture with NVIDIA components, from the research question to the decision dossier. Components marked "evaluating" or "EA" are under evaluation.
From evidence to a portfolio of R&D decisions
From evidence to a portfolio of R&D decisions
An OCEAN investigation ends with a portfolio of R&D decisions: hypotheses ranked by weight of evidence, the evidence for and against each one, the explanations that were dropped, and the experiment worth running next. With NVIDIA models in the workflow, that chain of evidence extends from disease understanding to the first steps of molecule design. Connecting these investigations with experimental feedback offers a path for each research cycle to inform the next.
An OCEAN investigation ends with a portfolio of R&D decisions: hypotheses ranked by weight of evidence, the evidence for and against each one, the explanations that were dropped, and the experiment worth running next. With NVIDIA models in the workflow, that chain of evidence extends from disease understanding to the first steps of molecule design. Connecting these investigations with experimental feedback offers a path for each research cycle to inform the next.
Biolevate builds evidence-grade AI for regulated life sciences operations. The platform combines scientific evidence, regulatory knowledge, company context, governed workflows, and AI agents to help teams automate complex work while preserving traceability, auditability, reproducibility, and human control. Biolevate enables life sciences organisations to move beyond isolated AI tools and build knowledge-driven automations that can safely support regulated work at scale.
Biolevate builds evidence-grade AI for regulated life sciences operations. The platform combines scientific evidence, regulatory knowledge, company context, governed workflows, and AI agents to help teams automate complex work while preserving traceability, auditability, reproducibility, and human control. Biolevate enables life sciences organisations to move beyond isolated AI tools and build knowledge-driven automations that can safely support regulated work at scale.
Biolevate builds evidence-grade AI for regulated life sciences operations. The platform combines scientific evidence, regulatory knowledge, company context, governed workflows, and AI agents to help teams automate complex work while preserving traceability, auditability, reproducibility, and human control. Biolevate enables life sciences organisations to move beyond isolated AI tools and build knowledge-driven automations that can safely support regulated work at scale.