[论文解读] BioDynaMo: an agent-based simulation platform for scalable computational biology research

Lukas Breitwieser, Ahmad Hesam+11 more|arXiv (Cornell University)|Jun 8, 2020

Scientific Computing and Data Management参考文献 38被引用 5

一句话总结

BioDynaMo 是一个高性能、开源的基于代理的仿真平台，通过利用并行和异构硬件，加速计算生物学研究，可在单台服务器上模拟高达12.4亿个代理，在笔记本电脑上模拟1200万个代理。与最先进的串行基线相比，其性能最高提升124倍，使用72个CPU核心时并行加速达76倍，支持C++和基于SBML的细胞内动力学。

ABSTRACT

Abstract Computer simulation is an indispensable tool for studying complex biological systems. In particular, agent-based modeling is an attractive method to describe biophysical dynamics. However, two barriers limit faster progress. First, simulators do not always take full advantage of parallel and heterogeneous hardware. Second, many agent-based simulators are written with a specific research problem in mind and lack a flexible software design. Consequently, researchers have to spend an unnecessarily long time implementing their simulation and have to compromise either on model resolution or system size. We present a novel simulation platform called BioDynaMo that alleviates both of these problems researchers face in computer simulation of complex biological systems. BioDynaMo features a general-purpose and high-performance simulation engine. The engine simulates cellular elements, their interactions within a 3D physical environment, and their cell-internal genetic dynamics. Cell-internal dynamics can be described in C++ code or using system biology markup language (SBML). We demonstrate BioDynaMo’s wide range of application with three example use cases: soma clustering, neural development, and tumor spheroid growth. We validate our results with experimental data, and evaluate the performance of the simulation engine. We compare BioDynaMo’s performance with a state-of-the-art baseline, and analyze its scalability. We observe a speedup of 20–124× over the state-of-the-art baseline using one CPU core and a parallel speedup between 67× and 76× using 72 physical CPU cores with hyperthreading enabled. Combining these two results, we conclude that, on our test system, BioDynaMo is at least three orders of magnitude faster than the state-of-the-art serial baseline. These improvements make it feasible to simulate neural development with 1.24 billion agents on a single server with 1TB memory, and 12 million agents on a laptop with 16GB memory. BioDynaMo is an open-source project under the Apache 2.0 license and is available at www.biodynamo.org. Author summary Computer simulations of biological systems are crucial to gain insights into complex processes of living organisms. However, the development of increasingly large and complex simulations is a difficult task, partly because a strong background in biology as well as computer science is required. In this paper, we introduce BioDynaMo, an agent-based simulation platform with which life scientists can create simulations that are three orders of magnitude faster than the state-of-the-art baseline. By taking advantage of the latest developments in computing hardware, we build a platform that is highly optimized. This enables the simulation of 1.24 billion agents on a single server and 12 million agents on a laptop. BioDynaMo places a lot of focus on hiding computational complexity and providing an easy-to-use interface, such that the life scientist can concentrate on biological aspects, rather than computational. BioDynaMo helps scientists to translate an idea quickly into a simulation by providing common building blocks, and a modular and extensible software design. We analyze the performance of the platform and demonstrate the capabilities with three example use cases: soma clustering, neural development, and tumor spheroid growth. The results support the view that BioDynaMo will help open up new research opportunities for large-scale biological simulations.

研究动机与目标

• 解决现有基于代理的仿真器未能充分利用现代硬件且缺乏可扩展性的问题。
• 减少在大规模生物系统仿真中所需的研发时间和计算负担。
• 提供一个灵活、模块化且用户友好的平台，使生命科学家能够专注于生物学本身，而非底层实现细节。
• 实现对复杂生物过程（如神经发育和肿瘤球体生长）的高分辨率、大规模仿真。

提出的方法

• 该平台使用通用仿真引擎来建模细胞代理、其三维物理相互作用以及细胞内遗传动力学。
• 细胞内动力学通过原生C++代码或SBML实现，支持与系统生物学工具的互操作性。
• 引擎针对多核CPU上的并行执行进行了优化，并支持超线程以提升可扩展性。
• 采用模块化、可扩展的软件架构，便于生物模块的复用与集成。
• 平台抽象了底层计算复杂性，为生物学家提供高层接口。
• 通过真实世界生物用例评估性能，并与最先进的基线进行对比。

实验结果

研究问题

• RQ1 通用仿真平台是否能在大规模生物仿真中显著超越现有基于代理的仿真器的性能？
• RQ2 BioDynaMo在多核CPU上扩展时，能否在保持细胞和遗传动力学高保真度的同时实现良好扩展性？
• RQ3 在串行和并行执行方面，BioDynaMo的性能与最先进的基线相比如何？
• RQ4 BioDynaMo能否在前所未有的规模下模拟神经发育和肿瘤球体生长等生物相关过程？
• RQ5 该平台在不需研究人员具备高性能计算专业知识的前提下，能否显著降低开发开销？

主要发现

• 在单个CPU核心上运行时，BioDynaMo相较于最先进的串行基线，性能提升达20至124倍。
• 使用72个物理CPU核心并启用超线程后，BioDynaMo实现67至76倍的并行加速。
• 结合串行与并行优化，BioDynaMo在测试系统上的性能至少比基线快三个数量级。
• 该平台成功在单台配备1TB内存的服务器上模拟了12.4亿个代理，证明了其在大规模神经发育仿真中的可行性。
• 同时，它也能在配备16GB内存的笔记本电脑上模拟1200万个代理，凸显其可移植性和高效性。
• 与实验数据的对比验证表明，所有三个用例（胞体聚集、神经发育和肿瘤球体生长）的仿真结果均具有生物相关性和准确性。