The convergence of artificial intelligence (AI) and life sciences has evolved from unidirectional tool-assisted research into a bidirectional, co-evolutionary paradigm characterized by mutual knowledge transfer and synergistic advancement. This review aims to systematically examine the latest progress in this interdisciplinary fusion, delineate the underlying logic of their interactive empowerment, and discuss the transformative potential and remaining challenges of this symbiotic relationship for future scientific research. The foundational logic of AI-life science integration rests on three pillars: the analogy between biological and computational information-processing paradigms, a shared mathematical language for modeling systemic complexity, and a spiraling cycle of bidirectional knowledge transfer from “AI understanding life” to “life inspiring AI”. Building on this foundation, the review first surveys the innovative development of Generative Biology across multiple scales. At the molecular level, breakthroughs such as AlphaFold-series models and diffusion-based protein design frameworks (e.g., RFdiffusion) have fundamentally shifted the paradigm from observation to de novo creation, enabling atomic-precision design of proteins, enzymes, and drug candidates. At the cellular level, single-cell foundation models and virtual cell frameworks leverage deep generative architectures to learn transferable cell-state representations, supporting perturbation response prediction and causal mechanistic reasoning. At the organ level, deep learning empowers medical imaging analysis, organ-on-chip integration, and digital twin construction for individualized intervention simulation. At the individual and ecological levels, AI-driven models integrate electronic health records, wearable sensor streams, wastewater-based epidemiological surveillance, and biodiversity monitoring to bridge data gaps across personal health, public health, and environmental stewardship. The review then examines Bio-inspired AI, encompassing brain-inspired learning rules and neuromorphic hardware that replicate neural computation with event-driven, energy-efficient architectures, biomimetic optimization algorithms including genetic algorithms, artificial immune systems, and swarm intelligence, as well as molecular computing and biological hybrid intelligence platforms that employ DNA circuits and living organoid networks as computational substrates. Despite remarkable progress, critical challenges persist, including the “black-box” accountability dilemma in clinical and regulatory contexts, the corrosive risks of AI hallucinations in molecular and clinical design, algorithmic adaptability to the inherent fuzziness and nonlinearity of biological systems, and multi-scale data heterogeneity. Addressing these challenges demands the establishment of robust “dry-wet closed-loop” validation systems, transparent and explainable model architectures, and globally coordinated governance frameworks that balance innovation incentives with biosecurity, privacy protection, and algorithmic fairness, thereby ensuring a safe and sustainable trajectory for this transformative convergence.
ZENG et al. (Wed,) studied this question.