Agent-based Models to address the Crisis of Reproducibility and Precision Medicine

摘要 本提案旨在解决与开发和开发有关的根本方法挑战 使用多尺度模型(MSM)，进而可以解决当前的认知危机 作为一个整体影响着生物医学研究。我们提出了一种方法，通过这种方法，使用 MSM，特别是基于代理的模型(ABM)，提供了一种解释并最终 解决可重复性的危机，并在这样做的过程中提供一条通向“真正”精确度的容易处理的道路 医学(即正确的药物、正确的患者、正确的时机，以及如何设计这样的策略)。我们断言，这场危机 重复性的产生在很大程度上是因为相对于所有空间而言，真实世界的数据是稀疏的 可能的生物/病理表型(根据系统状态，特别是系统轨迹)；这 导致可以通过实验采集的样品与生物的真正丰富性之间的不一致 异质性。我们进一步提出，解决这种差异可以通过近似 以ABM的大尺度参数/轨迹空间探索为代表的系统行为景观 对于现实世界的系统来说。这种观点是新颖的，因为它强调了 由许多轨迹生成的多维空间/流形，而不是单个或高度- 由经典参数拟合/校准产生的选定轨迹子集。因此，验证目标 远离高保真/精度拟合(例如，对单个数据集的平均值进行拟合)，这有助于 稀疏性问题；相反，验证涉及概括覆盖的广度和分布 多个数据集的结果，这包含了异构性。鉴于系统动力学的重要性 以及轨迹对于特定状态的非唯一性，这一观点导致了我们断言 精准医学只有在彻底描述了行为的多样性之后才能实现，而且， 在没有现有的数学形式主义的情况下，为开发控制策略建立方向可以 通过对模拟数据进行进化计算和强化学习，可以达到最好的效果。

Summary This proposal seeks to address fundamental methodological challenges associated with the development and use of multi-scale models (MSMs), and by extension, can potentially address a current epistemic crisis affecting biomedical research as a whole. We propose an approach by which a novel perspective of using MSMs, and specifically agent-based models (ABMs), provides a means of explaining and eventually addressing the Crisis of Reproducibility, and, in so doing, providing a tractable path towards “real” Precision Medicine (i.e. right drug, right patient, right time, and how to design such a strategy). We assert that the Crisis of Reproducibility arises in great part because of the sparseness of “real world” data relative to the space of all possible biological/pathological phenotypes (in terms of system state and especially system trajectories); this leads to a discordance between what can be sampled experimentally and the true richness of biological heterogeneity. We further propose that addressing this discrepancy can be accomplished by approximating the behavioral landscape of a system using large-scale parameter/trajectory space exploration of ABMs as proxies for the real world system. This perspective is novel because it emphasizes the distribution and variability of multi-dimensional spaces/manifolds generated by many trajectories, as opposed to the individual or highly- selected subset of trajectories that result from classical parameter fitting/calibration. Thus, the validation target shifts away from high-fidelity/precision fitting (e.g. fitting mean values of a single dataset), which contributes to the sparseness problem; instead, validation involves recapitulating the breadth of coverage and distribution of outcomes across many datasets, which embraces heterogeneity. Given the importance of system dynamics and the non-uniqueness of trajectories to a particular state, this perspective leads to our assertions that true Precision Medicine can only be achieved after behavioral manifolds are thoroughly characterized, and that, without an existing mathematical formalism, establishing the direction for developing control strategies can best be achieved using evolutionary computing and reinforcement learning on simulation data.