Artificial Intelligence (AI) promises to revolutionize biomedical research with unprecedented opportunities to reduce drug development costs, accelerate timelines, and improve clinical trial success rates. On the other hand, AI learning algorithms depend on high-quality training data as the Garbage In, Garbage Out (GIGO) concept is considerably amplified through AI (key word: AI Hallucinations).

I raised questions about the quality of the biomedical training data for AI models in a recent article (AI, GIGO, and the Biopharmaceutical Venture). In my opinion, low success rates of drug development programs and high irreproducibility rates of peer-reviewed research studies put data integrity center stage of any serious attempt at building AI models for the life sciences.

While data integrity, in the view of the computer scientists I spoke with, largely pertained to data management challenges (e.g., data format and metadata inconsistencies, incomplete data, lack of data provenance and data lineage), as a biomedical scientist, I content the life sciences face even more profound data integrity issues. These include extreme data heterogeneity and complexity, incomplete data of unknown extent, pervasive data silos, and, above all, ethical considerations.

In the biomedical sciences we have a big problem with an extremely poor signal-to-noise ratio that fundamentally undermines the reliability of foundational data. Much of this "noise" stems from the inherent complexity and variability of biological systems, experimental limitations, but more crucially, human and systemic biases. Scientists often tend to validate rather than falsify hypotheses, which inadvertently creates an echo chamber effect: flawed initial data points and wrong hypotheses are reinforced by peer scientists rather than challenged.

The deeply rooted issues of groupthink and confirmation bias necessitate a paradigm shift in how we approach data integrity in the life sciences to build more reliable foundational models. To address this, AI offers a transformative potential: to act as a "red-teaming" tool for research claims. Such an AI red-teaming tool would not just perform consistency checks but would be designed to identify subtle and systemic anomalies that human bias or the sheer volume of information might obscure. And it could be incorporated into the peer-review process.

An AI-enabled Red Team would add substantial scrutiny to the peer-review process of research publications. The Red Team could act as the third “peer reviewer” focusing on contradictions, missing controls and logical inconsistencies. AI can analyze a body of literature much larger than any expert ever could to flag experiments that need more cross-examination. Such a Red Team could also identify contradictions not only with previous studies but also across domains, for example when genetic traits and protein biochemistry results do not match up.

A Red Team could add additional scrutiny to the peer-review process, for example, through temporal trend analysis by evaluating how findings evolved over time in a research field. A Red Team could also review materials & methods sections for completeness and clarity, raising reproducibility and standardization across studies. One could even imagine an AI-enabled Red Team that simulates experiments based on previously published methods and data, to then compare the simulated outcomes to the reported outcomes.

While machine learning and pattern recognition algorithms are already assisting in detecting plagiarism and managing the publishing process, I believe it is possible to build an advanced AI that (or should I say who?) can identify patterns in the vast amount of human knowledge that contradict the specific hypothesis underlying a novel research study without falling to the human fallacy of confirmation bias and groupthink.

I invite interested biomedical and computational scientists to share their reactions and ideas on this very critical topic, as we all seek solutions for a more reproducible future.

#artificialintelligence #science #scicomm