一個結合條件生成de novo 適體結構設計、ADMET篩選AI模型與SPR 驗證的研發框架平台

Abstract

本研究提出一套以 aptamer 設計為核心，整合生成式人工智慧、結構生物資訊、計算篩選與表面電漿子共振（surface plasmon resonance, SPR）驗證的研究平台。研究動機源自傳統 SELEX 在巨大序列空間、PCR 偏差與實驗週期上的限制，使 aptamer 發現流程雖具生物可行性，卻仍面臨效率不足的問題。因此，本文以「蛋白特徵抽取－條件式序列生成－計算篩選－PK/PD 與 ADMET 評估－SPR 驗證」作為方法主軸，建立可串接計算預測與實驗端點的候選發現流程。 方法上，本文首先根據目標蛋白結構建立表面電荷、溶劑可及性、結合口袋與界面幾何等描述子，再將其作為條件輸入至生成式模型，以產生具結構導向意義的 aptamer 候選序列。資料集方面，本研究自 RCSB PDB 整理約 9,395 個 DNA–protein 複合體，轉換為約 173,577 筆 protein site–DNA pairing 預訓練樣本；微調階段則整合 aptamer 資料庫、文獻紀錄、已解析 aptamer–protein 複合體與 AlphaFold 輔助結構，形成約 958 筆 protein site–aptamer pairing 微調樣本。接續以序列層級過濾、分子對接、分子動力學進行 in-silico 篩選，並納入 PK/PD 與 ADMET 指標評估候選於藥物開發場景中的潛在可用性。最後，以 SPR 量測結合動力學參數，驗證模型候選是否在實驗端點上具備可解釋且可重現的表現。 本研究貢獻主要有三項：第一，將傳統文獻回顧、生成模型、結構篩選與生醫轉譯評估整合為一致的 aptamer 發現流程；第二，建立從蛋白特徵條件化生成、候選序列排序到 SPR 實驗驗證的核心方法；第三，明確呈現生成品質、模型可解釋性、結構一致性、PK/PD 與 ADMET 評估，以及 SPR 驗證結果之間的對應關係。整體而言，本研究旨在將 aptamer 候選發現由經驗導向流程推進為可工程化、可比較且可逐步驗證的設計平台。

This thesis presents an integrated aptamer design platform that combines generative artificial intelligence, structure-based bioinformatics, in-silico screening, pharmacokinetic/pharmacodynamic (PK/PD) evaluation, and surface plasmon resonance (SPR) validation. The study is motivated by the practical limitations of conventional SELEX workflows, including the vastness of sequence space, amplification bias introduced by PCR, and the time-consuming nature of iterative wet-lab selection. To address these issues, the proposed framework organizes the discovery process into a coherent pipeline consisting of target protein feature extraction, conditional sequence generation, computational screening, multi-objective ranking, and experimental binding analysis. The core methodology begins with extracting protein structural descriptors such as electrostatic patterns, solvent accessibility, putative binding pockets, and interface geometry. These descriptors are used as conditions for a generative model that produces aptamer candidates with target-aware sequence constraints. The generated candidates are then evaluated through sequence-level filtering, off-target screening, docking and molecular dynamics analysis, and PK/PD- and ADMET-oriented prioritization. Finally, selected candidates are validated by SPR to measure binding affinity and kinetic parameters, allowing computational predictions to be mapped to experimentally observable endpoints. The main contribution of this thesis is not only a candidate generation model, but also an integrated research architecture that connects molecular design, engineering reproducibility, and translational validation. The proposed structure clarifies where classical aptamer discovery methods end, where generative AI adds value, and how experimental evidence can close the loop for iterative optimization. This framework is intended to support future implementation, benchmarking, and wet-lab refinement in aptamer-based therapeutic and diagnostic development.