《生命科学》 2025, 37(12): 1493-1504

基于生物学机制驱动的人工智能与基础模型

杨　帆^*

山东大学公共卫生学院，济南 250012

摘　要：

本综述审视了系统生物学中机理建模与先进机器学习的融合，我们将其定义为“知识注入式学习” (knowledge-infused learning) 范式。我们剖析了四种将生物学知识整合到计算模型中的主要模式：(1) 使用神经普通微分方程(neural ordinary differential equations) 和物理信息神经网络(physics-informed neural networks) 编码连续动态过程；(2) 利用基于生物通路的图神经网络(graph neural networks) 表示结构关系；(3)通过因果发现算法从观测数据中推断有向相互作用；以及(4) 通过大规模、自监督的基础模型(foundation models) 学习“生物学语言”。对于每种模式，我们都分析了其基础方法论，重点介绍了里程碑式的应用，并对其基本假设和实践局限性进行了严格的批判。我们认为，尽管每种方法都为特定挑战( 如处理不规则时间序列数据、整合多组学数据集或生成新颖假设) 提供了强大的解决方案，但它们的真正潜力并非孤立存在，而是作为统一的神经- 符号框架的组成部分时才能得以释放。本综述最后综合了这些主题，并为构建整合动态、结构、因果和大规模学习表征的混合模型指明了方向，旨在超越单纯的预测，实现真正的机理洞察。

通讯作者：杨　帆 , Email:fanyang@sdu.edu.cn

Biology-driven artificial intelligence and foundation models

YANG Fan^*

School of Public Health, Shandong University, Jinan 250012, China

Abstract:

This review examines the integration of mechanistic modeling and advanced machine learning in systems biology, which we define as the "knowledge-infused learning" paradigm. We dissect four primary modes of integrating biological knowledge into computational models: (1) encoding continuous dynamic processes using neural ordinary differential equations and physics-informed neural networks; (2) representing structural relationships with graph neural networks based on biological pathways; (3) inferring directed interactions from observational data via causal discovery algorithms; and (4) learning the "biological language" through large-scale, self-supervised foundation models. For each mode, we analyze its underlying methodology, highlight landmark applications, and provide a critical assessment of its fundamental assumptions and practical limitations. We argue that while each approach offers powerful solutions to specific challenges (such as handling irregular time-series data, integrating multi-omics datasets, or generating novel hypotheses), their true potential lies not in isolation but as components of a unified neuro-symbolic framework. This review concludes by synthesizing these themes and charting a course for building hybrid models that integrate dynamic, structural, causal, and large-scale learning representations, aiming to move beyond mere prediction and achieve genuine mechanistic insight.

Communication Author：YANG Fan , Email:fanyang@sdu.edu.cn