Benchmarking and Enhancing Large Language Models for Biological Pathway Reasoning

Using RAG

Thank you for your insightful feedback! Our current graph database already leverages textual content for retrieving nodes and edges, which aligns with the concept of "retrieval from a pathway database." A potential Retrieval-Augmented Generation (RAG) approach could indeed incorporate textual modalities to provide additional contextual information in the retrieved content.

In practice, existing literature or pathway databases are more accessible and widely used, making this a highly practical solution. Our method, PathSeeker, is adaptable to any database as long as graph navigation—particularly local subgraph searching—can be developed based on the database. We believe this capability is critical for efficiently navigating graph-structured databases.

On the other hand, using pure graph data provides an idealized framework for formalizing the reasoning process, specifically for testing the capacity of LLMs to perform deductive reasoning on pathway graphs. Since this work focuses on Pathway Reasoning, we believe a graph-form database offers the most suitable foundation for evaluating the biological pathway reasoning capabilities of LLMs.

Open-ended answer evaluation method

Thank you for the insightful question! For the evaluation of open-ended tasks, we utilized LLaMA-3.1 405B, which we found to provide evaluation quality comparable to GPT-4. Additionally, we experimented with other models, including GPT-3.5 and LLaMA-3.1 70B, and obtained the following results verified by human assessment:

The observed inconsistencies between model and human evaluations often arise in cases where the generated answer is close to the ground truth but expressed in a general manner or lacks specific details. This highlights a trade-off between evaluation precision and computational cost. For instance, GPT-3.5 is more cost-efficient, while GPT-4 offers higher accuracy at a greater expense.

Regarding evaluation methods like rule-based approaches, similarity metrics, or ROUGE scores, they are not well-suited for open-ended generation tasks. The key challenges include:

1. Matching Biological Entities: For instance, evaluating answers that equate terms like coatamer protein II complex and COPII requires advanced engineering efforts that rule-based methods struggle to handle.

2. Different Expressions of the Same Fact: For example, the standard answer might state, "NleF slows the intracellular trafficking of tsVSVG from the endoplasmic reticulum to the Golgi," while the model-generated response suggests, " NleF causes a delay or blockage in the anterograde transport trafficking of tsVSVG, leading to changes in its intracellular localization." While semantically equivalent, such variations are difficult to assess using rule-based, similarity, or ROUGE metrics.

Given the additional costs and challenges of evaluating open-ended tasks, we developed the True/False task in BioMaze to address these problems, which offers both greater convenience and accuracy. The True/False questions are designed as probing forms of biological pathway queries, which, while still challenging as our experiments demonstrate, simplify evaluation.

Pathway graph limitations (potential sources of error when using the pathway graph data)

Thank you for your suggestion! To enhance the clarity of the error categories and provide more detailed insights, we have included representative examples for each error category in Appendix A.2 of the revised draft (due to space constraints in the main paper). In particular, we present error cases related to the pathway graph-augmented method, as illustrated in Figures 10 and 12.

One source of errors in the pathway-augmented method arises from the inherent complexity of graph data, especially in pathways with self-circulatory or multi-branch structures. For example, in Figure 10, the question asks: "What is the effect of heparin deficiency on the formation and degradation of Ang II in these peritoneal cell cultures?" Here, the model's reasoning process considered the pathway involving the degradation of Ang II but overlooked the more critical pathway concerning the conversion of Angiotensin I to Angiotensin II. This omission led to an incorrect conclusion.

The challenge arises from the textual representation of pathway graphs. Although we developed a DFS-based graph sequentialization method to better capture graph features, sequential LLMs still face difficulties in understanding and reasoning about complex graph structures. This limitation is especially pronounced when they need to perform deductive reasoning across multiple branches or navigate self-circulatory pathways simultaneously.

Handling multi-step reasoning decline during cot

Thank you for this inspiring idea! To address this, we conducted additional experiments by designing a hierarchical reasoning method. This approach requires the LLM to first outline the pathway potentially involved in the question as a reasoning plan, and then conduct reasoning based on this self-proposed pathway. This method, which we denote as CoT-Self-Pathway, is similar to the pathway-augmented reasoning method, except that the pathway is self-generated by the LLM rather than provided by an external database.

Below are the results comparing CoT-Self-Pathway to standard CoT and our PathSeeker on the True/False task:

As the results indicate, CoT-Self-Pathway generally achieves performance comparable to standard CoT. We observed that the self-generated pathways tend to be more abstract and less detailed or comprehensive compared to pathways retrieved from graph databases. In some cases, these self-proposed pathways are more contextually aligned with the question, which can be an advantage.

The performance of this method is primarily constrained by the quality of the pathways generated by the LLM itself. Since the LLM-generated pathways lack the details and accuracy of those from a dedicated pathway graph database, CoT-Self-Pathway typically does not perform as well as pathway-augmented methods such as PathSeeker.

Pathway graph searching improvement

Thank you for your insightful observation! The limitation of pathway graph augmentation, particularly omissions in reasoning, can indeed be mitigated through database expansion and enhanced graph search algorithms.

Our method, PathSeeker, specifically focuses on improving graph search efficiency by employing LLM-guided subgraph navigation, which has resulted in superior performance compared to other graph augmentation methods. For example, we conduct an error analysis including LLM-pruning graph BFS method ToG. Below are the error type classification results from the open-ended task:

Percentage of Errors Across All Data in Open-Ended Tasks

These results indicate that as an LLM-pruning graph BFS method, ToG is more prone to errors from omissions in reasoning, due to its less efficient graph navigation strategy. As a result, PathSeeker illustrate higher accuracy in the BioMaze evaluation. We believe that further advancements in graph search efficiency could enhance pathway graph recall, ultimately boosting the model's reasoning capacity.

In this work, we primarily utilized the KEGG pathway graph database. Moving forward, we plan to incorporate additional pathway databases, such as Reactome, into the project. This expansion will allow us to cover a broader range of scenarios and improve the overall robustness of the method.

We sincerely hope that our response addresses your questions. We remain available to address any further questions you may have.

Data correctness validation

To ensure question quality, we employ a two-step process. First, we filter questions using an advanced language model ( LLaMa 3.1-405B) to assess their relevance and clarity. Subsequently, each question undergoes a final quality check by human reviewers.

To validate the answer quality, we require the LLM (LLaMa 3.1-405B) to answer the questions based on the original paper content rather than by themselves. The model is explicitly instructed to respond with Undetermined if it cannot confidently generate an answer. Each question is tested five times, and only questions that are consistently answered correctly (i.e., aligned with the intended label) and not marked as Undetermined in any of the trials are retained. This process helps eliminate questions with incorrect labels, ambiguous phrasing, or poor structure.

In the final stage, human experts perform an additional quality check to refine the questions further. Approximately 5% of the data is filtered out at this stage, primarily due to issues such as hint leakage in the question, overly complex phrasing (e.g., asking for multiple facts), or poorly defined structure. During this stage, human reviewers also verify label correctness, ensuring the dataset's overall reliability and usability.

Through this comprehensive validation pipeline—particularly the human review step—we strive to ensure high data quality, with a focus on minimizing LLM errors and enhancing the accuracy of ground truth answers.

Error categorization reason analysis

Thank you for your insightful suggestions! To improve the clarity of the error categories and provide more detailed insights, we have added representative examples of each error category in Appendix A.2 of the revised draft (due to space constraints in the main paper). While it is challenging to succinctly summarize the reasons behind each type of error, we believe these examples will help improve understanding.

Below, we briefly illustrate a few examples, but we strongly recommend referring to Appendix A.2 for a more comprehensive analysis.

1. Omission in Reasoning
   This error occurs when critical steps in the reasoning process are omitted, leading to an incorrect final answer.
   Example:
   Question: Does BAMBI enhance or inhibit Wnt-promoted cell cycle progression?
   Label: BAMBI increases Wnt-promoted cell cycle progression.
   Error Description:The model's reasoning only identified BAMBI as a target of beta-catenin but failed to account for its interactions with key components of the Wnt signaling pathway, such as LRP6, FZD10, and DVL1. This omission led to an incorrect conclusion.

2. Faulty Reasoning
   This error occurs when the reasoning path aligns with the question context but contains significant errors in deducing the events or relationships within that pathway.
   Example:
   Question: What is the effect of GogB-deficient Salmonella on tissue damage and colonization in the gut during long-term chronic infections?
   Label: GogB-deficient Salmonella cause extensive tissue damage and enhanced colonization in the gut during long-term chronic infections.
   Error Description:In this case, the model considered the NF-kappaB signaling pathway but failed to deduce that GogB deficiency leads to increased inflammation. This critical oversight resulted in an incorrect conclusion.

3. Unresolved Conclusion
   This error occurs when the model cannot provide a definitive answer, indicating uncertainty or suggesting that the answer is unknown.
   Example:
   Question: What is the effect of the E2A-Pbx1 fusion protein on the expression of the novel WNT gene, WNT-16, in pre-B ALL cells?
   Label: The E2A-Pbx1 fusion protein activates the expression of WNT-16 in pre-B ALL cells.
   Answer: Further investigation would be needed to determine the exact effect of the E2A-Pbx1 fusion protein on WNT-16 expression in pre-B ALL cells.
   Error Description:In this case, the model is unable to determine the specific effect of the fusion protein on WNT-16 expression.

4. Incomplete Answer
   This error occurs when the response omits essential details, such as specific effects or key components of the labeled answer.
   Example:
   Question: How is continual efferocytosis enhanced in macrophages?
   Label: Continual efferocytosis is enhanced in macrophages by the metabolism of apoptotic cell-derived arginine and ornithine to putrescine by macrophage arginase 1 (Arg1) and ornithine decarboxylase (ODC).
   Answer: When macrophages engulf apoptotic cells (ACs), they not only clear the ACs but also metabolize the engulfed cargo. This metabolic reprogramming involves the activation of specific metabolic pathways, such as glycolysis and fatty acid oxidation, to generate energy and metabolites necessary for the efficient clearance of ACs.
   Error Description:In this case, the model highlights metabolic reprogramming but fails to mention the critical role of arginine and ornithine metabolism in continual efferocytosis, resulting in an incomplete answer.

We sincerely hope that our response addresses your questions. We remain available to address any further questions you may have.

Motivation for using subgraph methods rather than the whole graph

Thank you for your insightful question! While it is true that an individual pathway file, such as MAPK, can fit within the context window of an LLM, typical and realistic biological research scenarios often do not have pre-given information of which specific pathway is relevant to the question. Furthermore, the reasoning process may involve interactions or activations spanning multiple human-splitted pathway maps.

To address this challenge, we combine all KEGG pathways into a single, comprehensive graph that serves as the augmentation database for all the questions in the BioMaze, as we described in "Pathway Graph Database" of Subsection 3.3. This combined graph is too large to fit within the LLM’s context window in its entirety. As a result, PathSeeker, along with the baselines we selected, focuses on methods capable of dynamically identifying and extracting relevant subgraphs from large graph databases. This ensures our approach is both scalable and well-suited to the complexities of real-world biological pathway reasoning.

We have modified the description of the graph database in the revised draft to clarify the motivation and methodology behind our approach.

Evaluation of cutting-edge models as backbone

Thanks for the suggestion of experiment! We conduct experiment with backbone LLaMa 3.1-405B, and here is the result on True/False tasks:

Here are some key observations:

1. Cutting-edge model achieve higher performance: Overall, LLaMa 3.1-405B achieves an 8% performance improvement compared to the 8B version.

2. Persistent gap in intervention scenarios: The results demonstrate that cutting-edge models exhibit varied performance under different settings. A noticeable performance gap remains between natural and perturbed cases. This indicates that even state-of-the-art models struggle with reasoning about interventions in biological pathways compared to their better understanding of natural pathway states.

3. Effectiveness of PathSeeker: Our proposed method, PathSeeker, still shows improvement comparing to CoT method, particularly in scenarios involving interventions.

We will continue updating this dataset by incorporating data from more recent publications that were not available during the model’s pretraining phase, aiming to enhance the evaluation of LLMs in realistic biological research scenarios involving pathway reasoning.

Further related work

Thank you for highlighting this relevant area of research! The graph-augmented baselines we explore—such as ToG, CoK, and G-Retriever—are indeed closely tied to graph-based retrieval methods, as discussed in Section 2 of the paper. We have incorporated additional related works, such as GraphRAG and GraphReader, in the revised draft.

Answer validation during data creating and filtering

To ensure question quality, we employ a two-step process. First, we filter questions using an advanced language model ( LLaMa 3.1-405B) to assess their relevance and clarity. Subsequently, each question undergoes a final quality check by human reviewers.

To validate the answer quality, we require the LLM (LLaMa 3.1-405B) to answer the questions based on the original paper's content. The model is explicitly instructed to respond with Undetermined if it cannot confidently generate an answer. Each question is tested five times, and only questions that are consistently answered correctly (i.e., aligned with the intended label) and not marked as Undetermined in any of the trials are retained. This process helps eliminate questions with incorrect labels, ambiguous phrasing, or poor structure.

In the final stage, human experts perform an additional quality check to refine the questions further. Approximately 5% of the data is filtered out at this stage, primarily due to issues such as hint leakage in the question, overly complex phrasing (e.g., asking for multiple facts), or poorly defined structure. During this stage, human reviewers also verify label correctness, ensuring the dataset's overall reliability and usability.

We sincerely hope that our response addresses your questions. We remain available to address any further questions you may have.

Additional result of evaluating cutting-edge models as backbone

We further evaluated LLaMA 3.1-405B on BioMaze open-ended tasks. We apologize for the delay, as to minimize potential model bias during evaluation, we employed GPT-4 as the evaluation LLM for LLaMA 3.1-405B. Our analysis of the evaluator demonstrated that GPT-4 exhibits evaluation accuracy comparable to LLaMA 3.1-405B, with both achieving 96% consistency with human. The performance results are as follows:

The key observations are similar to the result on True/False tasks:

1. Cutting-edge model achieve higher performance: Overall, LLaMA 3.1-405B achieves a 5% performance improvement compared to the 8B version.

2. Persistent gap in intervention scenarios: A noticeable performance gap remains between natural and perturbed/intervened cases. This is one of the key conclusions of our benchmarking. Despite the backbone model's stronger knowledge and reasoning abilities, interventions in biological systems still pose significant challenges for the LLM's reasoning.

3. Effectiveness of PathSeeker: Our proposed method, PathSeeker, demonstrates improved performance compared to the CoT method, particularly in scenarios involving interventions.

Data (metrics and units) presentation

Thank you for the helpful feedback. We apologize for any confusion caused by the description of our metric. We introduced the metric in Subsection 5.1. For True/False tasks, we compute accuracy averaged across the True and False labels to account for label imbalance in the dataset (50% for random guessing baseline). For open-ended tasks, the LLM is used to evaluate the accuracy of generated answers by comparing them to the ground truth and determining whether they are correct or incorrect. In this study, we use the LLaMa 3.1-405B model as the evaluator, with five in-context examples. The performance of the evaluator is further analyzed in Appendix A.8. We improved the metric description subsection.

We add the metric description to Table 2, 3, 6 and Figure 5 in the revised draft. Thank you for the feedback on this matter.

Meaning of lowest results in Tables 2 and 3

In Tables 2 and 3, higher metric values indicate better performance. In these results, we underline the lowest values in each dimension to highlight the more challenging setting. We further explain the meaning of the underline in the revised draft.

Additional baselines for analysis experiment

Thank you for your thoughtful suggestion. We chose CoT and PathSeeker as baselines to represent two distinct reasoning approaches: independent reasoning by LLMs and reasoning augmented by graph structures. As shown in the main experiment in Subsection 5.2, PathSeeker effectively utilizes pathway graphs, making it the most representative method for illustrating how graph-augmented reasoning works.

To further explore other graph augmentation methods, we also analyzed the ToG method. Below are the error type classification results from the open-ended task:

Percentage of Errors Across All Data in Open-Ended Tasks

These results indicate that as an LLM-pruning graph BFS method, ToG is more prone to errors from omissions in reasoning, likely due to its less efficient graph navigation strategy. Interestingly, the phenomenon that ToG with GPT 3.5 as the backbone performed worse than with LLaMA3 8B could be attributed to GPT 3.5's shorter context length (4096 tokens) compared to LLaMA3’s (8192 tokens), which limits the extent of graph navigation and may exacerbate pathway omissions.

Evidence for PathSeeker reducing gap between natural and intervened/perturbed groups

Thank you for this valuable suggestion! To better demonstrate how PathSeeker enhances intervention reasoning, we present a comparison of the performance gap between natural and intervened/perturbed groups below:

True/False Task: Natural - Intervened/Perturbed (Lower is Better)

Open-ended Task: Natural - Intervened/Perturbed (Lower is Better)

The results indicate that PathSeeker achieves smaller performance gaps between natural and intervened/perturbed groups compared to CoT in most scenarios. This suggests that leveraging pathway graphs improves reasoning for intervention cases.

Figure 4's fitting method and standard deviation

We fitted the lines using a third-order polynomial curve with NumPy's polyfit method.

Thank you for suggesting the inclusion of standard deviation! In the revised draft, we have incorporated the standard error, calculated from five independent test runs, into Figure 4. The results show that the performance gap is significant compared to the standard deviation, particularly for questions requiring a larger number of reasoning steps.

The observed phenomenon where "the gap between CoT and PathSeeker is very small" predominantly occurs for questions involving fewer reasoning steps. For questions requiring more reasoning steps, however, the gap becomes more pronounced.

We sincerely hope that our response addresses your questions. We remain available to address any further questions you may have.