How Elucidata uncovered what was missing in Alzheimer’s research

AI systems trained on publicly available summaries alone are working with an incomplete picture of the scientific record. To measure exactly how much gets missed, Wiley worked on a proof of concept with Elucidata, a data-centric AI company, to analyze 21,926 Alzheimer's research articles using Knowledge Graphs.

For AI systems designed to support drug discovery or clinical research, those missing connections aren't footnotes. They can be the difference between a research direction worth pursuing and one that quietly leads nowhere.

Introduction

Elucidata builds AI-ready data infrastructure for target discovery teams. Its platform, Polly, helps drug discovery teams find, harmonize, and integrate biological data from public and proprietary sources, turning it into program-specific Knowledge Graphs that enable AI-powered target validation.

Elucidata’s outputs depend directly on the quality of the content flowing through its pipelines. That reliance is what made our trusted knowledge so valuable, and what made this comparison worth doing.

Alzheimer's disease was chosen as the research focus deliberately. It's one of the fastest-growing areas of biomedical research, a field where we hold an extensive content position, and one where the disease mechanisms remain complex and far from fully understood. If access to full-text articles was going to make a measurable difference anywhere, this was an area it would show it.

The problem

Most AI systems operating in scientific research work from what's publicly accessible: abstracts, summaries, metadata, and the slices of knowledge that appear in open repositories. This isn't unusual. It's simply the practical reality of working with research literature where a significant portion sits behind paywalls.

The assumption built into this approach is that the abstract captures the essentials. That what's behind the paywall is detail, not substance. That the summary gives you enough to reason from.

But here's the problem with "almost right." It doesn't announce itself. An AI system drawing on incomplete evidence doesn't produce obviously flawed outputs. It produces plausible ones. Confident ones. Outputs that reflect the structure of good reasoning while resting on a map with gaps in it.

In Alzheimer's research, where disease mechanisms involve many interacting biological processes, where the relationship between amyloid proteins, neuroinflammation, synaptic function, and cardiovascular conditions is still being actively mapped, those gaps can point a research program in the wrong direction. Not toward a clearly wrong hypothesis, but toward a plausible one that happens to be incomplete.

That's the specific challenge we set out to measure with Elucidata.

The solution

We worked with Elucidata to design a controlled experiment to test how much a system's map of scientific knowledge changes when it can read the research and not just the summaries. The experiment worked like this.

A dataset of 21,926 Alzheimer's research articles was divided into two comparable reading lists. The first simulated free-only access: open-access articles in full, but with Wiley's paywalled articles represented by summaries only. The second simulated full-text access: the same open-access articles, but with our paywalled articles included in full.

Elucidata then built a separate Knowledge Graph for each reading list. Knowledge Graphs are a method of representing scientific understanding not just as a collection of facts, but as a structured map of relationships between concepts. They capture how ideas connect, which makes them well suited to measuring not only what a dataset contains, but what it understands.

By comparing the two graphs side by side, the experiment could quantify precisely what changed when the complete articles were added.

Results

The findings were unambiguous.

The full text surfaced 11.4% more connections between scientific concepts: 16,300 additional relationships within the Alzheimer's literature. And critically, 1 in 10 of those connections were entirely absent from the free literature, relationships with no presence in the abstracts at all. Full text also added 7.2% more concepts: approximately 1,600 terms found nowhere in the summaries.

Two examples illustrate what "completely new" actually means in practice.

The first involves protein fibrillation, the process by which proteins misfold and clump together into insoluble fibers. The full-text content surfaced a meaningful link between atrial fibrillation and Alzheimer's disease, a connection that abstracts never mentioned.

This relationship, absent from the free literature, is the kind of finding that could redirect cardiovascular research toward a new area of investigation, or open a line of inquiry around shared mechanisms that no abstract-only analysis would have surfaced.

The second involves synaptic plasticity, the brain's capacity to form and reorganize neural connections, and its relationship to amyloid proteins, which are central to Alzheimer's disease.

Richer detail around this relationship only emerged in the full-text dataset. For teams working on therapeutic strategies targeting amyloid or synaptic function, that detail isn't background context. It's the kind of mechanistic understanding on which development decisions get made.

These aren't isolated findings specific to Alzheimer's. The same dynamic applies wherever disease mechanisms are complex and the literature is dense. In any domain where researchers are building on a partial map and don't know it, the cost of those missing connections is likely to compound quietly over time.

Conclusion

This proof of concept demonstrates something that's easy to assert and harder to prove: the gap between what an AI system believes and what the full scientific record actually contains is measurable, meaningful, and consequential.

When AI systems work from abstracts alone, they don't produce wrong answers, they produce incomplete ones. They’re only “almost right”. An argument can hold together logically while resting on evidence that's missing a significant portion of its foundation. And in environments where decisions carry real weight, whether a clinical trial design, a drug discovery target, or a regulatory submission, an AI system working from an incomplete map is a liability that's difficult to detect precisely because the outputs it generates appear so plausible.

What this partnership with Elucidata makes visible is the mechanism behind that problem. AI reasoning is only as good as the evidence it can reach. Improving AI-assisted research isn't simply a matter of adding more data. It's a matter of improving the quality of relationships and evidence that shape scientific understanding at the source.

Closing that gap is what our full-text content is for.

The questions this raises extend far beyond Alzheimer's, into every domain where incomplete evidence, dressed up as sufficient, can steer work in the wrong direction. If an AI system can't recognize what it can't see, what does responsible AI-assisted research require?