AI Agents Are Only As Good As The RNA Knowledge Behind Them

By William Soliman, Ph.D., BCMAS, Healthcare AI Expert, Scientist, Entrepreneur

The recent launch of Jolunara Intelligence, an AI company focused on governed, human-aligned systems for life sciences, has sharpened a question that is becoming urgent across RNA therapeutics: What does an AI system actually need to know before we trust it to influence a scientific or clinical decision?

The RNA field is producing extraordinary volumes of information. Sequence libraries, structural data sets, preclinical findings, clinical outcomes, manufacturing records, and scientific publications are expanding rapidly. In theory, this should create ideal conditions for artificial intelligence. Yet giving an AI system access to more information does not necessarily make the system more scientifically intelligent.

The distinction is important. Data can tell a model what has been observed. Knowledge must explain what the observation means, under which conditions it applies, how it relates to other findings, and where its predictive value ends. Now that the world of medicine has had its first major mRNA FDA approval, we need to answer this important question.

More Data Does Not Automatically Mean More Knowledge

Advancing RNA recently explored why AI-designed mRNA depends on biologically relevant training data. That is an essential starting point. Models trained on short reporter constructs, immortalized cell lines, unmodified RNA, or artificial delivery conditions may perform well in development and still fail to predict how a therapeutic-length, chemically modified RNA will behave in a patient. But even relevant data is not sufficient by itself.

Consider a study reporting that a particular sequence improves protein expression. An AI system might identify and reproduce that association. A scientifically useful system must go further. It should know whether the sequence was evaluated in vitro or in vivo, which cell type was used, how the RNA was modified, what delivery system was employed, how expression was measured, and whether the finding has been replicated. In RNA therapeutics, context is not supplementary information but rather it is part of the biological result.

The same sequence may behave differently depending on secondary structure, untranslated regions, nucleotide modifications, poly(A) tail characteristics, cell type, innate immune activation, route of administration, and delivery vehicle. Manufacturing conditions and impurity profiles may further influence stability, potency, and tolerability. An AI system that separates a result from those conditions may retrieve the correct fact while supporting the wrong conclusion.

Scientific Retrieval Is Not Scientific Understanding

Large language models are exceptionally good at locating, organizing, and summarizing information. Those capabilities can save scientists enormous amounts of time. But retrieval should not be confused with understanding. In this context, “understanding” should be treated as a testable operational standard, not as an anthropomorphic description of the technology. A system demonstrates useful scientific understanding when it can connect a claim to its source, identify the biological and experimental conditions surrounding it, distinguish observation from inference, recognize conflicting evidence, and communicate when the available information does not justify a conclusion. That requires a knowledge architecture, not simply a larger repository. This is why companies like Jolunara are interesting because by relying on data that closely resembles the human beings that would be carrying out this job function, we have a better representation of a “human” effect for that particular AI agent’s task.

A strong scientific knowledge architecture connects entities such as RNA sequences, targets, pathways, cell types, delivery systems, assays, manufacturing attributes, and clinical outcomes through explicitly defined relationships. Ontologies and knowledge graphs are one way to create this structure. For example, researchers are developing RNA-KG integrated information from more than 60 public databases into a semantically consistent network of RNA-related entities and relationships. The significance of this work is not merely the consolidation of more data; it is the preservation of meaning across previously disconnected sources.

Provenance And Uncertainty Must Travel With The Answer

A scientific claim should never become detached from its provenance. An AI-generated answer that cannot show where its evidence originated, how recently the source was updated, or what experimental conditions produced the finding has limited value in a regulated scientific environment. Every important assertion should carry a traceable evidence chain. Was the source peer reviewed? Was the finding derived from an animal model, a primary human cell, or a clinical population? Does the evidence show correlation or causation? Is the conclusion broadly accepted, supported by a single study, or actively disputed?

Negative and contradictory findings are equally important. Scientific knowledge rarely develops as a clean collection of settled facts. It evolves through competing hypotheses, methodological limitations, failed experiments, and revisions. Systems trained primarily on published positive results may create an artificially confident picture of the evidence. Trustworthy AI must therefore represent uncertainty, not hide it. Sometimes the most scientifically responsible response is not a prediction but an acknowledgment that the available evidence is incomplete or does not transfer to the proposed use case.

AI Governance Is Part Of Scientific Validity

In life sciences, AI governance is often discussed as a matter of data privacy, cybersecurity, or regulatory compliance. Those issues matter, but governance must also extend to scientific validity.

Organizations need defined standards for which evidence enters an AI system, how sources are ranked, how conflicting findings are represented, who may modify the underlying knowledge base, and when scientific experts must review an output. They also need version control so that conclusions can be reconstructed as the evidence changes.

This is especially important as AI moves from summarizing literature to recommending experiments, selecting candidates, supporting manufacturing investigations, or informing clinical strategy. The closer the system comes to influencing a consequential decision, the stronger the requirements for traceability, validation, and human oversight should become.

Human expertise is not a temporary bridge until models improve. Domain experts determine whether the system is asking the right question, applying evidence appropriately, and operating within acceptable scientific boundaries.

The Next Competitive Advantage

RNA developers should not begin their AI strategy by asking which model to deploy. They should begin by defining the decision the system is expected to support and the biological knowledge required to support it responsibly.

That means identifying the relevant entities and relationships, preserving experimental context, establishing evidence hierarchies, capturing uncertainty, and validating performance prospectively against real scientific decisions. Only then should organizations determine which model or interface sits on top of that foundation.

The next competitive advantage in RNA therapeutics may not belong to the organization with the largest model or the most data. It may belong to the organization that can transform fragmented scientific information into governed, traceable, and biologically meaningful knowledge.

AI can help scientists navigate complexity at a scale no human team could manage alone. But it should not be expected to manufacture understanding from disconnected information. If we want AI to become a trustworthy scientific partner, we must first give it a disciplined representation of what is known, why it is believed, where it applies, and what remains uncertain. Ultimately, AI will only be as reliable as the biological knowledge architecture behind it.

About The Author

William Soliman, Ph.D., is the founder & CEO of the Accreditation Council for Medical Affairs (ACMA) and the founder & CEO of White Manna Capital Partners, a biotech/pharma focused hedge fund. The ACMA is the leading life sciences accreditation, certification, and training company in the world and established the first ever certification standards for prior authorization, reimbursement, pharma sales, medical science liaisons, and medical affairs professionals. Soliman is considered a pharmaceutical industry futurist. In March 2021, he testified before the United States Congress’ Energy and Commerce Health Subcommittee about the pharmaceutical industry and the importance of professional standards for those who directly engage healthcare providers, like sales representatives. Soliman is a former pharmaceutical executive who held leadership roles at several Big Pharma companies, including Merck, Johnson & Johnson, AbbVie, and Gilead. He is routinely featured on media outlets such as NewsNation, Fox News, ABC News, Forbes, Al Jazeera, Yahoo! Finance, Yahoo! Business TV, and more. Soliman received his Ph.D. from Columbia University and his bachelor’s degree from New York University.