NYB.AI Explores NVIDIA Nemotron for Scientific Literature Annotation in Vecura

A closer look at a natural-product literature pilot in which NYB.AI evaluated how NVIDIA Nemotron 3.5 Super VL could help verify compound records against their source publications.

A molecular research project may begin with findings in a published paper, continue through the investigation of a compound's biological activity, and eventually involve molecular modelling or computational screening. Although these steps contribute to the same investigation, the relevant information and tools often sit in separate systems, leaving researchers to reconcile evidence and prepare data as they move between them.

Although AI has strengthened individual analytical tasks, researchers still need to establish whether the information passed from one stage of an investigation to the next reflects the original scientific evidence.

NYB.AI is exploring that challenge through Vecura, including a pilot with NVIDIA Nemotron 3.5 Super VL that examines how AI-assisted literature annotation could contribute to a more consistent foundation for computational research.

Bringing Scientific Knowledge and Computational Tools Together

Research findings appear across publications, molecular illustrations, biological databases, and computational outputs; their usefulness depends not only on the information they contain, but also on whether it can be interpreted and carried into subsequent analysis.

Researchers frequently need to move between these sources, confirm that records refer to the same biological or chemical context, and prepare data in formats that different tools can use.

Vecura is an agentic AI platform that brings together scientific models, biological data, and specialized computational tools for molecular discovery and life science research, helping researchers work across those activities within a shared environment.

Its capabilities cover applications such as compound screening, molecular docking, bioactivity prediction, and toxicity assessment, with agentic workflows designed to coordinate the tools involved in each investigation.

Vecura is designed to reduce the technical effort involved in moving between analytical steps while preserving the context researchers need to evaluate results; the pilot discussed below examined compound annotation and did not measure time savings or improvements in research outcomes.

Exploring NVIDIA Nemotron for Scientific Literature Analysis

NVIDIA Nemotron 3.5 Super VL can process both language and visual information, which makes it relevant to scientific papers containing written findings, figures and molecular illustrations. The results reported in this article, however, come exclusively from a text-based annotation pilot and do not assess the model's visual capabilities.

A paper may describe an experiment in its methods section and present related findings in figures or molecular diagrams, so future multimodal research workflows could benefit from interpreting those materials together. That possibility remains separate from the performance measured in NYB.AI's current evaluation.

For now, the more immediate question is whether AI can help researchers interpret and organize the information already captured in published literature, without losing the context or source evidence on which an annotation depends.

NYB.AI therefore evaluated Nemotron in a controlled compound-annotation workflow that relies on scientific text extracted from published papers, rather than treating untested visual capabilities as a demonstrated outcome.

The work forms part of Vecura's broader exploration of how scientific knowledge might be prepared for use alongside specialized computational tools, with clear provenance and further validation remaining central to that process.

To test this approach in a measurable setting, the team examined whether Nemotron could verify existing compound records against what the original publications actually reported.

Putting Nemotron to the Test: Understanding Scientific Literature

Building a useful natural-compound library requires more than recording which compounds appear in a paper. A compound may have been isolated from an organism, detected without isolation, purchased as a reference material or synthesized for an experiment, and those distinctions affect how its provenance should be recorded.

If an annotation fails to distinguish these cases, later users may draw the wrong conclusion about where a compound came from or what the original experiment established.

NYB.AI evaluated Nemotron on the task of checking these details against scientific publications, using an annotation pipeline that combined model-generated judgments with subsequent consistency and evidence checks.

The pilot covered 50 natural-product research papers across 434 pages, with 516 compound records identified at an earlier stage of the data-processing workflow.

TABLE 1:  SCIENTIFIC ANNOTATION PIPELINE

Each eligible paper went through three annotation rounds, generally involving four model calls, with additional calls for larger compound sets or incomplete responses. Automated checks then assessed the consistency of the annotations and whether the supporting evidence could be traced to the source text.

The evaluation used text extracted through MinerU rather than direct image input. Nemotron assessed the previously identified compounds, including whether they were studied, how they were obtained and what biological sources were reported, while also identifying passages that supported its annotations.

This distinction between model interpretation and downstream validation is important: a generated annotation is useful only when its scientific meaning is consistent and the supporting information can be traced to the source publication.

What the Initial Results Showed

Following document screening, 466 compound records were eligible for scoring, and the Nemotron–MinerU text workflow returned results for 461 of them (99%). The results were compared with reference annotations reviewed during NYB.AI's internal pilot, with each percentage measuring agreement on a specific scientific curation task.

TABLE 2:  NYB.AI INTERNAL PILOT — ANNOTATION AGREEMENT USING NVIDIA NEMOTRON 3.5 SUPER VL

In this pilot, the workflow agreed with the reference annotations more often when identifying whether a compound had been studied than when classifying its exact route of acquisition. This difference is scientifically meaningful: detecting a compound in an extract does not establish that it was isolated in purified form, and a curated library needs to preserve that distinction.

These findings remain preliminary: the benchmark represents a single internal run, and its reference annotations were reviewed from the outputs of an earlier pipeline rather than established through an independent blind assessment. Additional evaluation would be needed before generalizing performance to other literature collections.

How Literature Annotation Could Support Vecura Research Workflows

The potential value for Vecura lies in making scientific records more interpretable before researchers use them in further analysis. Details about a compound's origin, experimental context and supporting literature could help researchers assess the relevance of existing evidence when choosing subsequent computational investigations, although that downstream benefit was not measured in this pilot.

By improving how compound information is checked and organized, the approach is intended to reduce manual preparation and make it easier to carry reliable scientific context into later research steps; confirming any time savings or changes in research quality would require separate evaluation.

The pilot provides an initial assessment of annotation agreement against the reviewed reference set, rather than evidence that downstream screening, modelling or discovery outcomes have improved.

Nemotron 3.5 Super VL also offers vision-language capabilities that could be investigated for figures, molecular imagery and other visual evidence in future work, but the benchmark described here used MinerU-extracted text rather than image input.

These avenues of development align with Vecura's aim to connect information from scientific literature with specialized computational tools, while maintaining a clear distinction between capabilities under exploration and outcomes that have been tested.

Further evaluations will be needed to determine whether these approaches make research workflows more efficient or support better-informed decisions; for now, the pilot establishes a defined starting point for assessing how Nemotron handles scientific compound annotation.