Issue 01AI 바이오 논문
AC POST
AI 바이오 논문 목록
arxiv2026년 9월 20일 09:00

StressNET: 조직 내 기계적 응력 추론을 위한 적응형 딥러닝 모델

세포 간의 기계적 상호작용은 발달 및 재생 과정 중 조직 형태 형성의 핵심 요소이다. 세포 모양의 현미경 이미지로부터 세포 간 응력을 추론하는 계산 방법은 실험적 섭동 기술을 대체할 수 있는 비침습적 대안을 제공하지만, 기존의 모든 접근 방식은 명시적인 물리 모델에 의존한다. 본 연구에서는 기저 물리 모델을 가정하지 않고 조직의 기하학적 구조로부터 세포 간 기계적 응력을 직접 추론하는 Graph Neural Network (GNN)인 StressNET을 제시한다. StressNET을 학습시키고 벤치마킹하기 위해 합성 데이터셋을 생성하였으며, zebrafish neuromasts 및 Xenopus embryos의 실험적 응력 대리 지표와 비교했을 때 StressNET의 예측이 최첨단(state-of-the-art) 상관관계를 달성함을 입증하였다. 네트워크의 잠재 공간(latent space) 분석 결과, StressNET은 국소적인 세포 간 상호작용을 넘어 기계적 응력 분포의 전역적 조직 원리를 학습함을 보여준다. StressNET은 오픈 소스로 제공되며, in vivo 데이터에 맞춰 미세 조정(fine-tuning)할 수 있는 사전 학습된 모델을 제공하여 다양한 생물학적 시스템의 조직 역학 연구에 폭넓게 적용 가능한 도구이다.

Mechanical interactions between cells are fundamental to tissue morphogenesis during development and regeneration. Computational methods that infer intercellular stresses from microscopy images of cell shapes offer a non-invasive alternative to experimental perturbation techniques, yet all existing approaches rely on explicit physical models. Here we present StressNET, a Graph Neural Network (GNN) that infers intercellular mechanical stresses directly from tissue geometry, without assuming any underlying physical model. We generated synthetic datasets to train and benchmark StressNET, and demonstrate that its predictions achieve state-of-the-art correlation with experimental stress proxies in zebrafish neuromasts and Xenopus embryos. Analysis of the network's latent space reveals that StressNET learns global organizational principles of mechanical stress distribution, beyond local cell-cell interactions. StressNET is open-source and provides pre-trained models that can be fine-tuned on in vivo data, making it a broadly adaptable tool for studying tissue mechanics across biological systems.