Table 2. Limitations of alternative models and methods and strategies to overcome these limitations.

Human-based models/tools	Limitations	Strategies to address limitations
Epidemiological studies, randomized clinical trials	Inability to determine causality due to potential multiple interacting and confounding factors Difficult to compare studies designed according to different inclusion/exclusion criteria	Comprehensive assessment of multiple behaviors and risk factors and complex multivariate analyses to address conjoint confounding and effect modification. Application of machine learning and other techniques capable of non-linear and high-dimensional pattern recognition in large data sets.
		Possibility to create multi-center collaborations, taking advantage of common platforms
		Multiple intervention studies to test treatment effects in different types of populations
Patient-derived samples: CSF, blood/plasma, fibroblasts, and postmortem AD and control brain tissues	Storage and analytic methods are often not standardized, preventing inter-lab comparisons Poorly preserved brain tissues and long postmortem delays Samples are often not readily available for test for reproducibility and validation	Creation of multi-center collaborations to standardize methods & optimize distribution: e.g., 2-3 nationwide brain banks centers of excellence, with 24/7 autopsy services, short postmortem delays (2-3 hours maximum) and with standardized neuropathological protocols. Digitize neuropathology finds using standardized methods and creating an open-access database for additional analysis.
Neuroimaging techniques (e.g., MRI, PET, MRI tractography)	High costs; sometimes weak correlations between measures and clinical manifestations; sometimes difficult to quantify	Consider large-scale studies to improve correlations between imaging measurements and clinical manifestations
Synchrotron x ray fluorescence imaging	Requires ex vivo or post mortem brain tissue	Integrate this technology with other neuroimaging tools
Microfluidics/organ-on-chip	Some limitations with regard to transport and diffusion of nutrients and oxygen; individual organs, kept in isolation	Increase investment in research and development. Complement these technologies with neuroimaging data and/or other omics data sets
3D models (e.g., iPSCs, NPCs)	Not applicable for all purposes	Integrate 3D models with 2D models depending on applications and research goals
AD patient-iPSCs and their differentiated functional derivatives	Generating high-quality iPSCs is expensive and time consuming; a limited number of AD iPSC lines have been generated and thoroughly characterized so far	Cost is dropping over time; several entities (e.g., CIRM, NYSCF, etc.) are funding the development of hundreds of iPSC lines from AD patients
	They might be not fully representative of the complex physiology of the brain and/or of AD pathophysiology	Possibility to create co-culture systems with human microglial cells. Genome-editing technologies can be applied to create mutations related to the AD genetics, measure their impact on patient iPSC-derived neurons and design patient tailored treatments.
	Different reprogramming and QCs have been used, so comparisons between labs are difficult to make at this time	Several entities (e.g., CIRM, NYSCF, etc.) could standardize reprogramming methods allowing inter-lab comparisons
	Challenges with regard to penetrance, cell purity, degree and type of differentiated cells generated from iPSCs	Need to harmonize QC standards, which would be more feasible with the participations of dedicated entities
	Traditional reprogramming methods (e.g., integrating lentiviruses) and xeno-contamination might have affected the phenotype of the lines	Develop and adopt xeno-free techniques with non-integrating reprogramming vectors
	Epigenetic signatures of the somatic cell of origin might be retained in the reprogrammed iPSCs (NB: evidence that epigenetic traits get lost upon long term culture)	Possibility to directly reprogram fibroblasts into neurons
		Possibility to reprogram post-mitotic neurons and frozen brain tissue samples into iPSCs (to retain the neuronal epigenetic and pathologic background)
	iPSCs metabolic profile has not been investigated enough (which has special relevance in AD research)	Define QC metrics to establish metabolic features of iPSCs
	Still not clear how long iPSC-derived neurons should be kept in culture in order to mimic late-onset AD neurons and tissue pathophysiology; possible issues with the loss of aging-related transcriptional signatures and features.	Use AD brain tissues as benchmark models to define QC metrics suitable to assess neuronal and pathological features of differentiated iPSCs. Overexpression of aging-related genes (e.g., progerin) might help model AD in a dish. Direct conversion of aging donors' fibroblasts into neurons (iNs) can help retain aging-related transcriptional signatures.
	2D and 3D iPSC cultures might be characterized by different biological/cellular/molecular features and generate different responses	Define QC metrics to establish features of 2D vs. 3D iPSC cultures
	Not clear if AD-derived fibroblasts might be proven as suitable as their reprogrammed counterparts (i.e. iPSCs) to define molecular/cellular features of AD (e.g., metabolic profiles)	Define QC metrics to establish features of AD-derived fibroblasts vs AD-derived reprogrammed iPSCs
Non-mammalian/invertebrate models of AD	More phylogenetically distant from humans than mammalian species; might lead to intermediate validation steps in mammalian (non-human) species	Consider their suitability for basic research effort; less time consuming and less expensive than traditional animal models
		Investigate directly in human ex vivo tissues/cultures (rather than animals) to assess preclinical data (applying microdosing analysis)
PBPK and PD studies, IVIVE	PBPK, PD and IVIVE are currently applied mainly in toxicology.	Possibility to establish dedicated consortia with a multi-disciplinary approach (e.g., combining medical research and toxicology expertise).
Connectomics, computational analysis and modeling	Connectomics still in early development. Resolution too low. Very large data sets. Computational models are often restricted to simply mimicking observed phenomena and have no predictive value.	Develop techniques to study both individual and large cohorts necessary to recognize significant patterns. Increases in resolution, computational power and large-scale analysis algorithms are all rapidly improving. Encourage move to foundational computational simulations that explore the basic effects of cellular, network and system factors in aging and dementia. Use to elucidate and predict previously unrecognized changes in anatomy, physiology and cognition.
Various other omics: transcriptomics, proteomics, lipidomics metabolomics, exposomics, nutrigenomics, nutrigenetics, genomics, epigenomics	High costs	Costs of analysis are reducing. Possibility to establish dedicated consortia with a multi-disciplinary approach (e.g., combining molecular biology and biostatics expertise)

Abbreviations: CSF, cerebrospinal fluid; iPSCs, induced pluripotent stem cells; QC, quality control; NSCs, neural stem cells; MRI, magnetic resonance imaging; PET, positron emission tomography; PBPK, physiologically based pharmacokinetic; IVIVE, in vitro-in vivo extrapolation; PD, pharmacodynamics