Reading Between the Lines: Challenges in Biomedical Relationship Extraction and the promise of LLMs

‍

As discussed in a recent blog post by Dr Ben Sidders, (CSO at Biorelate), for Artificial Intelligence (AI) to fully realise its promise in drug discovery, we must leverage the prior knowledge from decades of scientific and medical research. To generate useful structured data from scientific literature, it is important to be able to extract meaningful relationships between biomedical entities such as interactions between proteins, gene-disease associations, or drug side effects. The previous blog in this series focused on the challenges of biomedical Named Entity Recognition (NER) and how Large Language Models (LLMs) are advancing that issue. This article focuses on how accurate biomedical Relationship Extraction (RE) requires not only identifying relevant entities, but also understanding the complex and evolving ways in which scientific text describes the relationships between them.
‍

Sentence complexity

Scientific writing is inherently dense, with authors often using long, syntactically complex sentences with multiple clauses, making it hard to parse relationships correctly via traditional Natural Language Processing (NLP) techniques. Interpreting complex sentences requires careful logical unpacking. For instance, the sentence “Bicuculline, a GABA_A receptor antagonist, decreased the inhibition of pyramidal neurons, resulting in an increase in their firing rate” does not simply mean that Bicuculline is an inhibitor or has a direct negative effect. Rather, GABA_A receptors normally suppress the firing of pyramidal neurons, and Bicuculline blocks this process. This means that in the presence of Bicuculline, pyramidal neuronal activity actually increases. This kind of nested logic – where decreasing an inhibition leads to an increase in activity – can easily confuse traditional NLP models tasked with relationship extraction, as they must correctly interpret both the direct and indirect actions described. Misinterpretation of such constructs can flip the extracted prediction or functional relationship, leading to significant errors in downstream applications. 
‍

Beyond logical nesting, biomedical sentences often describe several overlapping relationships at once, further complicating automatic extraction. For example, in the sentence “U0126 inhibits MEK1/2, blocking the activation of ERK1/2 and subsequent phosphorylation of transcription factors like ELK-1, which normally induces the expression of genes promoting cell proliferation”, there are at least four interconnected relationships: inhibition, activation, phosphorylation, and induction of gene expression – and more if you count the indirect relationships. Such sentences require careful parsing to determine which entities are acting upon others and in what sequence. Traditional NLP methods, which often depend on keyword matching or rigid syntactic patterns, can easily misinterpret or overlook these intertwined relationships, resulting in incomplete or inaccurate relationship data. 
‍

Variety of relationship types

Relationships found in biomedical literature also cover a wide variety of types, each with their own linguistic patterns and biological intricacies. For instance, the word “reduces” in “Amlodipine reduces blood pressure by relaxing blood vessels” means that Amlodipine lowers blood pressure, whereas in a biochemical context, e.g. “NADH reduces pyruvate into lactate”, the word “reduces” refers to a redox reaction where electrons are transferred. Understanding relationships often requires specialised biomedical knowledge to resolve such ambiguities and interpret findings correctly. This is a challenge for traditional NLP systems that rely on keyword matching, fixed patterns, or shallow syntactic parsing and can result in the extraction of misleading or incorrect relationships. 
‍

Negation, speculation, and uncertainty

Another nuance of scientific literature is the presence of negation, speculation, and uncertainty. Text often includes statements that are tentative, hypothetical, or negative. For example, an author might state, “Very-low-fat diets may be associated with increased risk of metabolic syndrome”. Traditional NLP systems struggle with these scenarios, as they may extract a relationship simply based on keyword proximity and overlook crucial cues indicating negation or uncertainty. This can result in erroneous or overstated findings in biomedical databases, ultimately impacting downstream research and decision-making.
‍

Implicit, and cross-sentence relationships

Many relationships described in biomedical literature are implicit – i.e. they are not explicitly described and must be inferred from the surrounding context, or by piecing together information from multiple sentences. For example, one interacting entity might be named at the start of a section, but anaphora can occur, where the entity that it interacts with is described a paragraph or two later. The first entity might be referenced in the later paragraph using terms like "it" or "the disease", complicating relationship extraction. Additionally, crucial contextual details like the mouse model or cell line used to validate a relationship may appear in entirely separate sections of the paper (e.g., methods section) from where the relationship is discussed (results and discussion section). As a result, NLP systems focused on sentence-level analysis often struggle to accurately extract these relationships, since key information may be distributed across multiple sentences or sections.
‍

LLMs: A new era for biomedical relationship extraction

Collectively, these examples highlight just a few of the challenges of performing RE in the biomedical domain. When performed accurately, RE can empower researchers with deeper, causal insights into drugs, their targets, and disease biology. Traditional NLP methods often fall short, but this is where LLMs offer significant promise.
‍

LLMs, fine-tuned on carefully curated biomedical literature, address many of the longstanding challenges in relationship extraction thanks to their advanced contextual understanding and ability to process complex language. Unlike traditional NLP, these fine-tuned LLMs can accurately parse dense, multi-clause sentences and untangle overlapping or nested relationships, making sense of intricate relationships. Their exposure to biomedical literature enables them to recognise subtle differences in scientific terminology, thereby minimising ambiguity, improving precision, and increasing recall. Biorelate find that task-specific fine-tuning is crucial for achieving a high level of performance in as demanding a task as biomedical relationship extraction.
‍

These models are also adept at identifying cues of negation, uncertainty, and speculation, which reduces the risk of extracting misleading or incorrect relationships. With their capacity to analyse longer context windows, domain-specific LLMs have the potential to link information spread across multiple sentences or sections of a paper, resolve references such as pronouns or general terms, and incorporate critical experimental details found outside a single paragraph. These capabilities are illustrated in Figure 1, which visualises how Biorelate’s fine-tuned LLM can identify relationships across a larger section of text, compared to the output of a more traditional, sentence-bound NLP pipeline.
‍

Together, these capabilities allow biomedical LLMs to extract richer and more reliable relationships from scientific literature at scale, ultimately supporting more robust knowledge discovery and faster, better-informed decision making. Read more about the promise of LLMs in this blog post. ‍

References

Perretta, Fernando, Antongiovanni, Norberto, Jaurretche, Sebastián, Early Renal Involvement in a Girl with Classic Fabry Disease, Case Reports in Nephrology, 2017, 9543079. https://doi.org/10.1155/2017/9543079

‍

‍