發展可解釋性人工神經網路以提取生化系統之資訊

Abstract

人工神經網路的技術近年來已經被應用在各個領域當中，包含影像辨識、輔助醫療診斷以及自然語言模型之基礎，然而目前較少能將結果與系統的物理化學知識進行連結並給予解釋。本論文探討可解釋式機器學習框架於生物分子感測與微生物培養系統中的分析與優化應用。內容涵蓋兩項主要研究，分別結合深度學習與主動式學習策略，以解決光譜分類與培養基配方設計中的核心挑戰。 論文的第一部分，我們設計了一套多尺度一維卷積神經網路（1D-CNN），用以分類霍亂毒素次單元B（Cholera Toxin B Subunit, CTB）與細胞膜結合後所產生的表面增強拉曼散射（SERS）光譜。透過調整卷積核大小等超參數以優化結構，我們成功提取多尺度光譜特徵，並在準確度、靈敏度、特異性與精密度方面優於傳統機器學習方法以及先前研究所提出之多尺度一維神經卷積網路。此外，我們應用了資料增強以緩解因生物樣本數據稀少而產生的過擬合問題。我們進一步提出基於卷積網路中活化層的解釋方法—單位元梯度加權類激活映射（Grad-AM），以可視化模型在分類決策中所依據的關鍵光譜區段，證明卷積網路其在低濃度樣本中也能有效捕捉具生物意義的特徵。 論文的第二部分，我們提出一套結合人工神經網路（ANN）與主動式學習（active learning）的框架，用於優化植物乳桿菌（Lactobacillus plantarum）的培養基配方。該模型整合經濟成本考量與設計特徵，成功預測出同時具備高活菌濃度與資源使用效率的配方。透過多輪迭代訓練與實驗驗證，人工神經網路模型展現出收斂性，並提出可使活菌濃度顯著提升之配方。綜合而言，這兩項研究展現了結合光譜學知識或物理性質的可解釋機器學習方法，在生化研究中的潛力，提供了在生物感測與生物製程工程中進行光譜分類、模型可視化與配方優化的實用框架。

Artificial neural network (ANN) has been widely applied in fields such as image recognition, medical diagnostics, and natural language modeling. However, their results are rarely linked to the physicochemical knowledge of systems, limiting the interpretability. This dissertation explores interpretable machine learning frameworks for analysis and optimization in biomolecular sensing and microbial cultivation systems. It encompasses two main studies integrating deep learning and active learning to address core challenges in spectral classification and culture medium formulation. The first part develops a multiscale one-dimensional convolutional neural network (1D-CNN) to classify surface-enhanced Raman scattering (SERS) spectra generated by cholera toxin B subunit (CTB) binding to cell membranes. By optimizing hyperparameters, such as convolutional kernel size, we extracted multiscale spectral features, achieving superior accuracy, sensitivity, specificity, and precision compared to traditional machine learning algorithms and prior multiscale 1D-CNNs. Data augmentation was applied to mitigate overfitting due to scarce biological data. We further proposed Grad-AM, a pixel-based gradient-weighted class activation mapping method, to visualize critical spectral regions driving classification decisions, demonstrating the model’s ability to capture biologically meaningful features even in low-concentration samples. The second part proposes an ANN and active learning framework to optimize Lactobacillus plantarum culture medium formulations. Integrating economic cost and feature engineering, the model predicted formulations with high viable cell density and resource efficiency. Through iterative training and experimental validation, the ANN demonstrated convergence and proposed formulations with substantially enhanced viable cell density. Collectively, these studies highlight the potential of interpretable machine learning, combining spectroscopic knowledge and physical properties, offering practical frameworks for spectral classification, model visualization, and formulation optimization in biosensing and bioprocess engineering.