AI-based neoadjuvant immunotherapy response prediction across pan-cancer: a comprehensive review

Abstract

Neoadjuvant immunotherapy (NIT) has emerged as a transformative treatment strategy across various cancer types. However, due to the significant heterogeneity of tumors, patients exhibit highly variable responses to NIT, making the accurate preoperative identification of those who would benefit a pressing clinical challenge. In recent years, artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), has opened new pathways for predicting treatment response. AI-driven approaches have the ability to extract latent features from high-dimensional, multimodal oncological data, facilitating the construction of efficient predictive models that can optimize individualized treatment strategies. In this review, we systematically summarize existing AI-driven computational approaches for NIT response prediction, categorizing them into indirect and direct predictive paradigms. The indirect paradigm predicts clinically validated surrogate biomarkers to infer therapeutic response to NIT. In contrast, the direct paradigm leverages AI to analyze high-throughput data and establish data-driven biomarkers that directly predict clinical endpoints of NIT. Additionally, we categorize existing AI predictive models based on data modalities, spanning radiomics, pathomics, genomics, and multi-omics approaches, each providing distinct insights into tumor characteristics and treatment response. Despite notable progress, current predictive models still face significant challenges, which we broadly classify into biomarker-based and AI-based limitations. We further discuss potential strategies to address these challenges. This review systematically summarizes recent AI-based predictive models for NIT response across cancer types. By offering a structured analysis of current methodologies and challenges, we aim to guide future research and accelerate the integration of AI into precision immunotherapy.

Keywords: Artificial intelligence; Biomarkers; Multi-omics; Neoadjuvant immunotherapy; Response prediction.

Fig. 1

Historical evolution of artificial intelligence and representative algorithm architectures. (A) Timeline illustrating the major developmental phases of AI, from symbolic logic and expert systems (1950s), through machine learning (1980s) and deep learning (2000s), to the current foundation model (2020s). (B) Schematic representations of classical ML algorithms, including LR, KNN, SVM, RF, and GBDT. (C) Architectures of widely used DL models, specifically MLP, CNN, RNN, Transformer, and GNN

Fig. 2

AI-based single- and multi-omics computational frameworks for pan-cancer NIT response prediction through surrogate biomarker (outer) and direct modeling (inner)