| 摘要: | "隨著生醫文獻數量的快速增長，藉由準確辨識生物醫學領域中蛋白質交互作用(PPIs)以提供研究者快速捕捉文獻中的關鍵資訊成為了一項重要且艱鉅的任務。在過去，神經網路(Neural Network, NN)的突破推動了其在文字探勘任務中的廣泛應用，然而，直接將通用領域的方法應用於生物醫學方面仍存在限制；而近年來，隨著大型語言模型(Large Language Model, LLM)的發展與其基於大量文獻進行預訓練的優勢，使得LLM模型在各領域中能夠更有效地理解專業術語和上下文資訊，這樣的發展能夠實現對疾病的深入理解和治療方法的改進以及對新藥物開發研究的助益。
本研究旨在探討使用不同prompt指令於模型GPT-3.5及GPT-4來預測蛋白質之間的互動關係，並從中提出一個最適用於GPT模型的prompt提問方法。此外，方法中我們也針對較複雜的實體型態進行改善，例如：巢狀蛋白質結構及複合詞蛋白質的例外處理。我們在五個常用於效能比較且公開的PPI資料集(LLL、IEPA、HPRD50、AIMed及BioInfer)進行評估，實驗結果表明，本研究所提出的方法在效能上具有相當的準確度，尤其在LLL資料集中F_1-score為87.3%，僅次於多核方法中的DSTK模型；再者，相較於其他深度學習模型，GPT基於具有高度彈性的prompt提問功能及多項參數可供調整，我們相信這將為生物醫學研究者提供更多的便利性。"
"With the rapid growth of the number of biomedical literature, it has become an important and arduous task to accurately identify protein-protein interactions (PPIs) in the biomedical field to provide researchers with the ability to quickly capture key information in the literature. In the past, breakthroughs in neural networks (NN) have promoted its widespread application in text mining tasks. However, there are still limitations in directly applying general-field methods to biomedicine. In recent years, with the development of Large Language Model (LLM) and its advantages of pre-training based on a large amount of literature, the LLM model can more effectively understand professional terminology and contextual information in various fields. Such developments could lead to a deeper understanding of disease and improved treatments, as well as aiding research into the development of new drugs.
This study aims to explore the use of different prompt instructions in the models GPT-3.5 and GPT-4 to predict the interaction between proteins, and propose a prompt that is most suitable for the GPT model. In addition, we have also improved the method for more complex entity types, such as nested protein structures and exception processing for compound proteins. We conducted evaluations on five publicly available PPI data sets (LLL, IEPA, HPRD50, AIMed and BioInfer) that are commonly used for performance comparison. The experimental results show that the method proposed in this study has considerable accuracy in performance, especially in LLL. The F_1-score in the data set is 87.3%, second only to the DSTK model in the multiple kernels method; furthermore, compared with other deep learning models, GPT is based on a highly flexible prompt questioning function and multiple parameters that can be adjusted. We believe that this It will provide more convenience for biomedical researchers." |
