Generative artificial intelligence-based algorithm to increase the predictivity of preclinical studies while keeping sample sizes small

将基础科学转化为新的临床有效物质似乎常常令人不满意。这是由于临床前研究的数据质量标准不足和样本量小。提出了一种基于人工智能（AI）的方法，在不增加临床前实验样本量的情况下，人工生成有效的附加数据。所提出的方法是基于生成模型（GM），它能够从一个非平凡的，可能是高维的分布中生成有效的数据，这些分布最初是未知的，通常只能在有限的程度上进行分析描述。因此，生成有效的新数据需要一种算法，该算法可以识别数据中的高维结构，然后使用该结构生成与可用数据一致的新数据。本项目旨在比较基于密度的聚类和数据生成中最有前途的方法，看看它们是否适合从太少的数据中生成更多但合适的数据。本文将基于大量的经验和初步的实验结果，提出一种基于紧急自组织映射（ESOM）的人工神经网络结构检测算法。它将与人工数据生成和基于距离的聚类的替代方法进行比较，包括但不限于生成对抗网络（GAN）、高斯混合模型（GMM）、隐马尔可夫模型（HMM）、naïve贝叶斯模型、潜在狄利克雷分配和（受限）玻尔兹曼机和DataBoost-IM，以及分层DBSCAN、自适应密度峰值聚类和自适应亲和传播聚类。最合适的方法应该发展成为临床前研究的可行方法。因此，工作方案将包括：(i)进一步发展和完善生成式ESOM方法的基础研究，包括对一种可能的替代方法进行比较评价；（ii）将其与用于数据生成和基于密度的聚类的其他当代生成式机器学习方法进行比较；（三）将最合适的方法应用于各种不同的生物医学数据集，包括为从临床前研究数据中生成有效新病例制定工作解决方案，同时并随后（四）实施基于esom的数据补全生成方法，作为免费提供的软件解决方案。因此，计划项目的主要目标是开发一种方法，该方法使用机器学习算法来有效和概率地生成实验数据，这些数据与原始数据一起，提供了从临床前实验中得出有效结论所需的高数据密度。

The translation of basic science into new clinically effective substances often seems unsatisfactory. This has been attributed to inadequate data quality standards and small sample sizes in preclinical studies. The development of an artificial intelligence (AI)-based method is proposed to artificially generate valid additional data without increasing the sample size in preclinical experiments. The proposed method is based on generative models (GM), which are capable of generating valid data from a non-trivial, possibly high-dimensional distribution, which are initially unknown and often can only be described analytically to a limited extent. The generation of valid new data therefore requires an algorithm that can recognize high-dimensional structures in the data and then use this structure to generate new data that is consistent with the available data.This project aims to compare the most promising approaches in density-based clustering and data generation for their suitability to generate more but appropriate data from too few data. Based on extensive previous experience and preliminary experimental results, a method based on emergent self-organizing maps (ESOM), as structure- detecting algorithm based on artificial neural networks, will be developed. It will be compared with alternative methods of artificial data generation and distance-based clustering including but not exclusively limited to Generative Adversarial Networks (GAN), Gaussian mixture models (GMM), hidden Markov models (HMM), naïve Bayesian models, latent dirichlet allocation and (restricted) Boltzman machines and DataBoost-IM, as well as hierarchical DBSCAN, adaptive density peak clustering, and adaptive affinity propagation clustering. The most suitable approach should be developed into a feasible method for preclinical studies.The work programme will thus include (i) basic research to further develop and refine the generative ESOM method including the comparative evaluation of a possible alternative method, (ii) its comparison with alternative contemporary methods of generative machine-learning for of data generation and density based clustering, (iii) the application of the most suitable method to various different biomedical data sets including the development of a working solution for the generation of valid new cases in data from preclinical research, accompanied and followed by (iv) the implementation of the ESOM-based generative method of data completion as a freely available software solution.Thus, the primary objective of the planned project is the development of a method that uses machine-learning algorithms for valid and probabilistic generation of experimental data that, together with the original data, provides the high data density necessary to draw valid conclusions from preclinical experiments.