Venky Soundararajan, PhD: What the Semaglutide-COVID Data Can't Yet Show

Venky Soundararajan, PhD, is cofounder and chief scientific officer of nference, the real-world evidence company behind a retrospective analysis linking prior semaglutide use to lower 30-day mortality, hospitalization, and acute kidney injury after COVID-19 and influenza infection. Soundararajan has authored more than 50 scientific papers and patents and has led multiyear data-science collaborations examining real-world epidemiological outcomes, including the impact of mass vaccination during the COVID-19 pandemic.

He spoke with Contagion about what the semaglutide findings can and cannot establish, how the study addressed healthy-user bias, and what evidence would be needed before the results could influence prescribing.

Transcript edited for clarity.

Contagion: For a clinician seeing these findings, what's the most important thing the data actually shows versus what it can't yet tell us about the mechanism driving those outcomes in COVID-19 and influenza?

Soundararajan: The most important finding is the consistency across 2 biologically distinct respiratory infections. Prior semaglutide exposure was associated with lower 30-day mortality, hospitalization, and composite severity after both COVID-19 and influenza, with acute kidney injury also significantly lower in both cohorts. That cross-virus consistency is compatible with a host-directed effect rather than something SARS-CoV-2-specific, but it does not establish immune modulation as the mechanism. We measured clinical outcomes, not cytokines, immune-cell phenotypes, viral kinetics, or other mechanistic biomarkers. As an observational study by design, we can only surface striking correlates and cannot draw any causal inference, which will require prospective targeted validation.

Contagion: The data suggest semaglutide's benefit could extend beyond glycemic control into immune modulation or systemic inflammation. What would a prospective study need to look like to actually test that hypothesis?

Soundararajan: I would want a prospective randomized study with serial sampling before and after infection, measuring inflammatory cytokines, immune-cell states, endothelial and renal injury markers, viral burden, glycemic control, and body weight alongside prespecified clinical outcomes. A low-dose semaglutide arm would be particularly interesting because, in our exploratory analyses, neither greater pre-infection weight loss nor higher semaglutide dose tracked with lower event rates. That observation is hypothesis-generating, not proof of weight-independent biology.

Contagion: GLP-1 users, as a population, have been suggested to be more likely to be health-seeking, better-monitored, and more likely to be on other cardioprotective therapies. How did you approach the healthy user bias problem in this study design?

Soundararajan: We tried to address healthy-user bias in our study. Specifically, we used metformin as an active comparator and matched 1:1 on age, sex, race, BMI, HbA1c, vaccination, prior infection, antiviral use, major comorbidities, health care utilization, and, for COVID-19, variant era; all matched covariates achieved standardized mean differences below 0.10. We also saw no association with the negative-control outcome of urinary tract infection. Still, observational matching cannot eliminate unmeasured differences in socioeconomic status, frailty, adherence, health care engagement, or therapies we did not explicitly match.

Contagion: From a data science standpoint, how do you think about the risk of over-interpreting real-world associations when a drug class is this broadly promising, and what's the bar of evidence you'd want to see before findings like these shape prescribing decisions?

Soundararajan: With a drug class generating this much enthusiasm, I think the evidentiary bar should become higher, not lower. Multimodal real-world evidence is most useful here for detecting reproducible signals and defining hypotheses, and prospective biological targeted experiments will be required to convert any compelling routine care association into causal mechanism testing. Before these findings should change prescribing specifically for respiratory-infection protection, I would want replication with other active comparators and independent international health care datasets beyond the US, where our study focused on routine care signal inference, followed by prospective randomized evidence with mechanistic biomarkers and prespecified clinical end points. Our current study supports that next step toward assessing the suitability of new indications, such as influenza or COVID-19 infection-associated outcome improvement, for prescribing semaglutide.