PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Models as Biomedical Hypothesis Generators: A Comprehensive Evaluation

Biqing Qi, Kaiyan Zhang, Kai Tian, Haoxiang Li, Zhang-Ren Chen, Sihang Zeng, Ermo Hua, Hu Jinfang, Bowen Zhou

arXiv:2407.08940v2cs.CL

TL;DR

Growing biomedical literature makes it difficult to keep pace with discoveries and generate hypotheses, while untraceable LLM training data complicates zero-shot evaluation. This paper evaluates LLMs using publication-date-partitioned biomedical data and reports that they can generate validated hypotheses on unseen literature, with uncertainty-oriented interactions improving zero-shot generation.

  • Problem

    Growing biomedical literature makes it harder for researchers to keep pace with discoveries and generate hypotheses, and untraceable training data prevents assurance that test data do not overlap training data.

    Source: §1 · p. 1

  • Method

    The paper builds publication-date-partitioned biomedical background-hypothesis data and evaluates instructed models across prompting, fine-tuning, tool-use, and multi-agent settings.

    Source: §1 · p. 2, §2.2 · p. 3

  • Results

    LLMs generated novel and validated hypotheses on unseen literature, while multi-agent interactions and tool use facilitated diverse candidates and improved zero-shot generation; few-shot learning and tool use did not always improve performance.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings support LLMs as potential aids in biomedical hypothesis generation and offer insights for further research in the field.

    Source: §5 · p. 10

    The authors identify uncertainty-strategy optimization, including reinforcement learning, as future work to address limitations in LLM knowledge discovery.

    Source: §6 · p. 10

Abstract

from arXiv · show

The rapid growth of biomedical knowledge has outpaced our ability to efficiently extract insights and generate novel hypotheses. Large language models (LLMs) have emerged as a promising tool to revolutionize knowledge interaction and potentially accelerate biomedical discovery. In this paper, we present a comprehensive evaluation of LLMs as biomedical hypothesis generators. We construct a dataset of background-hypothesis pairs from biomedical literature, carefully partitioned into training, seen, and unseen test sets based on publication date to mitigate data contamination. Using this dataset, we assess the hypothesis generation capabilities of top-tier instructed models in zero-shot, few-shot, and fine-tuning settings. To enhance the exploration of uncertainty, a crucial aspect of scientific discovery, we incorporate tool use and multi-agent interactions in our evaluation framework. Furthermore, we propose four novel metrics grounded in extensive literature review to evaluate the quality of generated hypotheses, considering both LLM-based and human assessments. Our experiments yield two key findings: 1) LLMs can generate novel and validated hypotheses, even when tested on literature unseen during training, and 2) Increasing uncertainty through multi-agent interactions and tool use can facilitate diverse candidate generation and improve zero-shot hypothesis generation performance. However, we also observe that the integration of additional knowledge through few-shot learning and tool use may not always lead to performance gains, highlighting the need for careful consideration of the type and scope of external knowledge incorporated. These findings underscore the potential of LLMs as powerful aids in biomedical hypothesis generation and provide valuable insights to guide further research in this area.

1 Introduction

The introduction frames biomedical hypothesis generation as increasingly difficult amid growing literature and untraceable LLM training data. It presents a temporal evaluation framework and reports that LLMs can generate validated hypotheses on unseen literature, while uncertainty-oriented methods can improve zero-shot generation.

1 Introduction

  • The evaluation uses biomedical background-hypothesis pairs split by publication date and tests instructed models in zero-shot, few-shot, and fine-tuning settings, including tool use and multi-agent interactions.The temporal split is intended to mitigate data contamination.Source: §1 · p. 2
  • LLMs generated novel and validated hypotheses on literature unseen during training, while multi-agent interactions and tool use improved zero-shot generation and candidate diversity.Few-shot learning and tool use did not always improve performance, so external knowledge type and scope matter.Source: §1 · p. 2
  • The work presents temporal biomedical instruction data and experiments as a rigorous evaluation of zero-shot and few-shot hypothesis generation.It also characterizes the resulting hypotheses as evidence of higher-order reasoning and potential knowledge discovery.Source: §1 · p. 2
  • The authors report that LLM and human assessments using multidimensional hypothesis-quality metrics are correlated, suggesting a role for LLMs in hypothesis evaluation.The paper describes the metrics as developed for evaluation with GPT-4 and human assessors.Source: §1 · p. 2
  • The paper proposes a multi-agent framework in which agents with distinct roles collaborate using tools to analyze factors influencing hypothesis proposers.The framework is intended to support collaborative analysis in hypothesis generation.Source: §1 · p. 2

2 Preliminary

The section formalizes hypothesis generation as producing a hypothesis from background knowledge and an instruction, then describes a date-based dataset construction pipeline. The resulting temporal splits support evaluation while leaving extraction accuracy and possible bias as concerns for further assessment.

2.1 Problem Definition

  • Given instruction I and background X_i, the task assesses model M by generating the corresponding hypothesis Y_i.The paper formalizes this as text completion over background-hypothesis pairs extracted from medical papers.Source: §2.1 · p. 3

2.2 Dataset Construction

  • The dataset pipeline splits literature before and after January 2023 to control test-data visibility, with training literature before that cutoff and test literature after it.The process composes a paper set, summarizes literature, generates background-hypothesis pairs, filters low-quality data, and partitions by publication time.Source: §2.2 · p. 3
  • The authors note that ChatGPT extraction accuracy and possible extraction biases warrant further evaluation, with manual annotation of a subset suggested as a check.This qualifies the reliability of the automated dataset construction process.Source: §2.2 · p. 4
  • The seen dataset contains 2,700 pre-January 2023 pairs, while the unseen test set contains 200 pairs from papers published in August 2023.The seen pairs are partitioned into 2,500 training and 200 validation examples, as well as a seen test set.Source: §2.2 · p. 4

3 Can LLMs Truly Generate Zero-Shot Hypotheses?

The evaluation compares LLM hypothesis generation across prompting and fine-tuning settings, including seen and unseen literature, and finds that external knowledge does not consistently improve performance. Zero-shot generations show higher novelty, while few-shot settings tend toward higher verifiability; uncertainty and human evaluation provide further evidence about these trade-offs.

3.1 Experiment Setup

  • The study evaluates premier LLMs across experimental settings using word-overlap metrics and four quality dimensions, with GPT-4 and expert assessments.The dimensions are novelty, relevance, significance, and verifiability; prompting includes zero-shot and sampled or similarity-retrieved 5-shot examples, alongside fine-tuning.Source: §3 · p. 4, §3.1 · pp. 4–5

3.2.1 Results of Zero-shot Setting

  • Zero-shot prompting yields higher novelty, while few-shot literature examples improve verifiability, revealing a trade-off between these qualities.Nearly all models, particularly WizardLM models and Openchat-v3.2-super, show enhanced novelty in zero-shot settings; few-shot prompts also improve word overlap for most models, especially gpt-3.5-turbo and WizardLM models.Source: §3.2.1 · p. 5, §3.2.2 · p. 6, Figure 4 · p. 6
  • LLMs tend to achieve better BLEU performance on the unseen test set than the seen test set, contrary to the expected seen-set advantage.The authors hypothesize that the intricacies of hypothesis generation may impede effective use of parameterized knowledge.Source: Figure 3 · p. 5, §3.2.1 · p. 6

3.2.2 Results of External Knowledge

  • External knowledge from few-shot examples, domain adaptation, and instruction tuning does not consistently improve performance across all metrics.Instruction tuning improves BLEU, ROUGE, and verifiability but reduces novelty and significance; domain adaptation improves word overlap while GPT-4 evaluations suggest it helps related few-shot use but diminishes zero-shot ability.Source: §3.2.2 · pp. 6–7
  • Similarity-retrieved few-shot examples are more likely to bolster verifiability but may reduce novelty, whereas random examples may heighten uncertainty.Random examples can diverge from the background topic or domain and potentially confuse models.Source: §3.2.2 · p. 7

3.3 Quantitative Analysis on Uncertainty

  • Zero-shot hypotheses have higher uncertainty and novelty scores, while few-shot hypotheses have lower uncertainty and higher verifiability.The paper uses SelfBLEU as an uncertainty indicator, with lower scores signifying greater uncertainty and diversity.Source: Figure 5 · p. 6, §3.3 · p. 8

3.4 Human Evaluation and Case Study

  • Human and GPT-4 evaluations correlate strongly, with both Pearson and Spearman coefficients above 0.7.Three biomedical experts assessed 100 unseen-test examples using four metrics scored from 0 to 3; the evaluation focused on top-performing models because of cost constraints.Source: Table 1 · p. 7, §3.4 · p. 8

4 Can agent collaboration enhance LLMs’ zero-shot generalization?

The paper tests whether tool use and multi-agent collaboration can help LLMs generate hypotheses in zero-shot settings. Multi-agent interaction can improve hypothesis proposal by introducing uncertainty, while tool use alone has minimal impact.

4 Can agent collaboration enhance LLMs’ zero-shot generalization?

  • The study aims to improve hypothesis generation by leveraging uncertainty through tool use and multi-agent collaboration, then reports preliminary validated results on the constructed dataset.The framework is inspired by scientific discovery and allows optional human involvement.Source: §4 · pp. 8–9, §4.1 · p. 9

4.1 Multi-agent Framework

  • The framework uses an Analyst, Engineer, Scientist, and Critic in an iterative process that extracts background elements, retrieves information, formulates hypotheses, and scrutinizes them.Critic feedback can loop back for refinement or prompt further Analyst work.Source: §4.1 · p. 9
  • The Engineer uses search engines, mainly PubMed, while post-January 2023 literature is excluded to control visibility of the unseen test set.The experiments use ReAct or OpenAI function calling, and models may decide whether to invoke tools after considering the background.Source: §4.1 · p. 9

4.2 Experiment Results

  • Division of labor and interaction among agents can significantly enhance hypothesis proposal by introducing uncertainty, and tool use can further enhance the multi-agent framework.The reported tool-use results show minimal improvement for LLMs in general, with ReAct performing slightly worse than OpenAI function calling.Source: §4.2 · pp. 9–10

5 Conclusion

The study evaluates LLMs for biomedical hypothesis generation, emphasizing zero-shot generalization and generation of novel, validated hypotheses on unseen literature. It introduces a multi-agent framework and multidimensional evaluation metrics, while identifying future directions for improving performance and adoption.

5 Conclusion

  • The corpus shows that LLMs can generate novel and validated hypotheses even when evaluated on unseen literature.The study focuses on zero-shot generalization in biomedical hypothesis generation.Source: §5 · p. 10
  • A multi-agent framework is introduced to support collaborative analysis during hypothesis generation.The study also develops multidimensional metrics to evaluate hypothesis quality.Source: §5 · p. 10
  • Future work could integrate domain-specific knowledge and diverse data sources, and develop user-friendly interfaces.The authors frame these as directions to enhance performance and adoption.Source: §5 · p. 10

6 Limitations and future works

The authors identify factual hallucinations and the need for broader evaluation as limitations requiring further study. They also describe the study's focus on hypothesis generation and outline future additions to its tools and uncertainty strategies.

6 Limitations and future works

  • The impact of factual hallucinations on hypothesis formulation requires further investigation, as does evaluation of knowledge discovery from multiple perspectives.The authors identify both as limitations and areas for further exploration.Source: §6 · p. 10
  • The current work focuses on hypothesis generation, using Web search and PubMed rather than covering the full knowledge discovery cycle.The authors propose integrating bioinformatics analysis tools in future work.Source: §6 · p. 10
  • The authors plan to incorporate knowledge graphs and explore more dynamic tools, including multi-tool collaborations.These are proposed future additions intended to improve hypothesis quality and model performance.Source: §6 · p. 10
  • The authors propose optimizing uncertainty-introduction strategies through reinforcement learning to improve hypothesis generation.They state this approach is intended to address limitations in the knowledge discovery process.Source: §6 · p. 10

A Process of Scientific Discovery

  • Scientific discovery includes data analysis, hypothesis formulation, experiment design and execution, and observation and reflection.The paper describes hypothesis formulation as guiding research-question selection, experiment design, and data analysis.Source: A Process of Scientific Discovery · p. 15

B.1 Data-Driven Scientific Discovery

The related-work discussion identifies limitations in prior scientific-discovery datasets and methods, especially guarantees about test-data visibility. It also notes limited exploration of LLM agents for scientific discovery and describes table-based evaluations across models, prompting formats, and hypothesis-quality measures.

B.1 Data-Driven Scientific Discovery

  • Earlier data-driven discovery work used dated or limited corpora, required human knowledge or explicit context, or did not propose new hypotheses; the reviewed literature lacked strict test-data visibility guarantees.The paper contrasts those approaches with datasets partitioned by publication date to ensure test data independence from LLMs.Source: §B.1 · p. 15

B.2 LLM-driven Autonomous Agents

  • The literature review finds little exploration of agents, particularly multi-agent systems, for scientific discovery.The paper positions its own effort as a preliminary use of tools and multiple agents to enhance hypothesis proposing and analyze influencing factors.Source: §B.2 · p. 16
  • Tables 2 and 3 assess instructed models with zero-shot and few-shot prompts across novelty, relevance, significance, verifiability, and their average.Asterisks denote few-shot prompts retrieved from training samples similar to the input background; the passages provide no table values.Source: Table 2 · p. 16, Table 3 · p. 17

C Implementation Details

  • The implementation-details section covers the constructed dataset and open-source models.The passage introduces these details but does not provide further specifics.Source: C Implementation Details · p. 17

C.1 Details of Dataset

The dataset discussion compares publication and text distributions across training, seen, and unseen data, while the model discussion describes selection criteria. The automatic-evaluation discussion points to detailed results tables and ROUGE-score figures without reporting their values in these passages.

C.1 Details of Dataset

  • The seen and unseen test sets show minimal differences in embedded background and hypothesis texts despite coming from distinct time periods.The texts were embedded and reduced with t-SNE; the accompanying figure describes slight distribution differences attributable to publication dates and minimal variance between ChatGPT- and GPT-4-constructed datasets.Source: §C.1 · p. 17, Figure 8 · p. 18

C.2 Details of Models

  • Open-source models were selected based on Alpaca-Eval leaderboard performance rankings before September 2023.The authors gathered pre-training, supervised-learning-corpus, and release-date information to ensure unseen test data was not visible to the models.Source: §C.2 · p. 17

C.1 Details of Dataset

  • The training and seen-test publication distributions rise year by year through January 2023, while the unseen set was sampled from August 2023.The unseen-set sample emphasizes the latter part of August; Figure 7 also presents topic distributions of abstracts in training and test datasets.Source: §C.1 · p. 17, Figure 7 · p. 18

D.1 Automatic Evaluations

  • The evaluation appendix points to detailed model and prompting results in Tables 2 and 3, multi-agent and tool-use results in Table 6, and seen/unseen ROUGE scores in Figure 9.These passages identify the reported materials but do not provide their outcome values.Source: §D.1 · p. 18

D.2 Human Evaluations

Human evaluations show ChatGPT assessments correlate better with human scores than word-overlap metrics, while the section also describes the literature corpus overview and generated-hypothesis examples.

D.2 Human Evaluations

  • ChatGPT evaluations correlate more highly with human scores than BLEU and ROUGE-L, which correlate weakly with comprehensive assessments of generations.Table 5 reports Pearson and Spearman correlations for automatic metrics against average human scores; Figure 1 is not part of this request’s evidence.Source: §D.2 · p. 18
  • Table 4 overviews related literature in LLM training corpora with publication dates to assess test-data non-visibility.The table caption notes that data marked with an asterisk was generated by people talking to ChatGPT, with date marking consistent with ChatGPT.Source: Table 4 · p. 19
  • Figure 9 displays ROUGE scores on both seen and unseen datasets.Source: Figure 9 · p. 19

E Case Study

  • The case-study section presents generated hypotheses from various models and walks through ChatGPT evaluation step by step.Source: E Case Study · p. 19

E.1 Generated Hypothesis

The section compares generated hypotheses, including a myocardial-infarction biomarker example, and reports unseen-test results for individual and multi-agent systems with and without tools.

E.1 Generated Hypothesis

  • For an August 2023 myocardial-infarction biomarker paper, GPT-3.5-turbo in zero-shot proposed the power-law-index hypothesis from background alone.GPT-3.5, Llama-2-70b-chat, and WizardLM-70B-V1.0 offered distinct hypotheses, while PMC-LLaMA-13B and WizardLM-13B-V1.2 gave straightforward interpretations without new content.Source: §E.1 · pp. 19, 21
  • Table 5 reports Pearson and Spearman correlations between automatic metrics and human evaluations.Source: Table 5 · p. 20
  • Table 6 reports that multi-agent systems and tool use enhance hypothesis generation on the unseen test dataset.The two tool-use variants differ in implementation: ReAct and OpenAI function calling.Source: Table 6 · p. 20
  • The multi-agent example uses Analyst, Scientist, and Critic roles to generate hypotheses across six rounds.Source: §E.1 · p. 21

INSTRUCTED LLMS CASE

This section presents instructed-LLM and multi-agent hypothesis examples alongside ChatGPT metric annotations, and documents the prompts used for collaboration and evaluation, including feedback and approval steps and mappings between evaluation metrics and prompt formats.

INSTRUCTED LLMS CASE

  • The generated MI hypotheses connect stage-specific mechanical markers with infarction, while Llama-2 adds fibrosis- and extracellular-matrix-based explanations not present in the background.The background reports AUC > 0.7 for αshort and αlong in early MI and AUC > 0.6 for E2, E3, and Etotal in mid- and late-stage MI; Llama-2 proposes mechanisms involving fibrosis, collagen, and proteoglycans.Source: INSTRUCTED LLMS CASE · p. 21
  • The multi-agent case refines a broad hypothesis about telemonitoring and wearable devices into claims about specific outcomes, early rhythm detection, timely intervention, and personalized treatment.The final hypothesis names reduced hospitalizations, improved quality of life, and decreased mortality, and states that studies support telemonitoring’s effectiveness in reducing detection time and improving outcomes.Source: INSTRUCTED LLMS CASE · pp. 24–25

CASE OF CHATGPT AS ANNOTATOR

  • ChatGPT’s example scores the generated hypothesis 2 for novelty, 3 for relevance, and 3 for significance.The explanations characterize it as extending the background with mechanisms and testable implications while remaining strongly related to the provided context.Source: CASE OF CHATGPT AS ANNOTATOR · p. 27

F Prompts Design and Finetuning

  • The appendix introduces prompts for LLM hypothesis generation and ChatGPT evaluation.The prompt sets are described in the following zero-/few-shot and multi-agent subsections.Source: F Prompts Design and Finetuning · p. 28

F.1 Zero- and Few-shot Prompts

  • Zero-shot and few-shot generation prompts are provided, with few-shot examples obtained through random sampling or similarity retrieval.The prompts are presented in Tables 10 and 11; the context examples are manually constructed from literature published before January 2023.Source: §F.1 · p. 28, Table 11 · p. 29

F.2 Prompts for Multi-agent Collaboration

  • The multi-agent prompt design assigns Analyst, Engineer, Scientist, and Critic roles for background analysis, literature search, hypothesis formulation, and evaluation.The Engineer is instructed to iteratively search and organize findings into clues; the Scientist crafts a grounded hypothesis, and the Critic evaluates novelty, coherence, and validity.Source: §F.2 · pp. 28, 30
  • The collaboration prompts instruct the Analyst to give specific feedback and either revisit the research background or suggest insights for the Engineer’s next search iteration.When a hypothesis meets expectations and desired standards, it is presented and approved in a structured format.Source: §F.2 · p. 30
  • Role-design and environment-setting prompts for multi-agent hypothesis proposing are presented in Tables 12 and 13.Source: Table 12 · p. 30, Table 13 · p. 31

F.3 Prompts for ChatGPT Evaluation

  • ChatGPT evaluation prompt formats for novelty, relevance, significance, and verifiability are presented in Tables 14–17, respectively.Source: §F.3 · p. 31, Table 14 · p. 31

F.4 Finetuning Implementation

The fine-tuning procedure uses three epochs and selects checkpoints based on performance on the seen test dataset; the supplied tables identify ChatGPT evaluation prompts for three metrics.

F.4 Finetuning Implementation

  • Fine-tuning runs for three epochs with batch size 8, maximum sequence length 2048, and learning rate 3e-5.Early stopping is used, and the best checkpoint is selected by seen-test performance.Source: §F.4 · p. 31
  • Table 15 contains prompts for ChatGPT evaluation on relevance.Source: Table 15 · p. 32
  • Tables 16 and 17 contain ChatGPT evaluation prompts for significance and verifiability, respectively.Source: Table 16 · p. 32, Table 17 · p. 32

Related papers

Loading 2407.08940v2…