This is one of the most common questions we receive when we talk about building knowledge graphs directly from PubMed.
Every researcher knows that not all papers agree: one study might report a positive correlation between two entities, while another finds a negative one under different conditions. In biomedical research, such conflicts are everywhere.
So how do we reconcile them?
We developed a method called Probabilistic Semantic Reasoning (PSR) to address exactly this challenge.
When we extract relations from millions of articles, each relation (e.g., drug–gene or gene–disease) may appear in many papers. PSR treats each appearance as a piece of evidence, not a definitive truth. It assigns a probability to that evidence based on the extracted context and relation type, and then combines all of them to infer how likely the relation truly exists.
In other words, instead of deciding “yes or no,” PSR asks “how likely?”
But the real power of PSR comes from what happens next. It uses these probabilities to reason through indirect connections. For instance:
* Drug A is related to Gene B.
* Gene B is related to Disease C.
* PSR calculates how likely Drug A affects Disease C through Gene B, and does this for all possible intermediate paths.
This reveals potential causal links even when no paper connects them directly.
Regarding conflicts, some relation types clearly contradict each other — like “positive correlation” and “negative correlation” under the same condition. PSR quantitatively evaluates all the relation types, weighs the evidence, and determines which one is most probable for each entity pair.
By doing this, PSR provides a scientifically interpretable way to handle conflicting information. Instead of picking a single source, PSR integrates all evidence into a model that captures uncertainty and diversity of real-world research.
This is one of the key reasons why our PSR framework can outperform AI co-scientists built purely on large language models (LLMs).
LLM-based agents are impressive at generating hypotheses or summarizing findings, but they usually do not infer probabilities between entity pairs, nor do they explicitly use all available evidence when reasoning about relationships.
In contrast, PSR systematically aggregates every piece of published evidence and transforms it into a unified, interpretable probabilistic framework, ensuring that every inference is grounded in data, not just language patterns.
That’s how our system reasons more like a scientist than a chatbot:
It reads everything, weighs evidence quantitatively, and forms explainable conclusions with traceable confidence levels.
This probabilistic reasoning forms the backbone of our AI-powered knowledge graph and enables reliable, reproducible discovery even when the literature disagrees.
#AI #KnowledgeGraph #BiomedicalAI #DrugDiscovery #Pharmacovigilance #InformationExtraction #LLM