Novel generative active learning algorithms for exploring the space of antimicrobial peptides to respond to antibiotics resistance

我们建议开发和评估一种新的机器学习(ML)方法，以积极探索潜在的抗菌肽(AMP)序列的巨大空间，目的是为设计一种快速、经济有效和通用的发现新的抗生素活性的管道做出贡献。AMP形成了一类分子，在自然界中被广泛用于对抗病原体，现在被用来解决抗菌素耐药性(AMR)，这是病原体发生突变，对所有已知抗生素产生抗药性的进化过程。AMR是一个主要的且日益严重的公共卫生问题，预计到2050年，在一切照旧的情况下，每年将有1000万人死亡。家畜使用的抗生素占世界使用的所有抗生素的66%，是抗生素耐药性的主要驱动因素。我们正在开发一种新的主动学习、强化学习和生成性建模方法，用于药物设计和分子遗传学方法，以实验性地评估ML产生的候选药物，目的是设计新的AMP来对抗牲畜中的主要细菌病原体。我们将整合生成候选AMP的ML系统和生物检测之间的整个交互序列，以筛选预测的候选对象。这些生成和测试循环被重申，以帮助基于生成流网络的新的ML方法通过适当地探索生物活性多肽空间的不同的有前景的区域来发现数据中的结构。由于多肽序列的多样性是巨大的(例如，有20^50个可能的50聚体多肽)，ML工具有望有效地搜索这个空间，但需要新的算法创新才能成功，特别是为了满足生成不同批次候选的需求。所提出的构建和快速筛选大型多肽文库的合成生物学方法对于ML驱动的AMP的发现至关重要。抑制细菌生长的活性AMP将通过DNA测序来识别，以通知下一轮ML。我们将为环化的AMP建立多肽库，这些多肽库通常具有更好的稳定性和更多的药物性质。然后，我们将应用ML方法来设计更有效的AMP组合。这一基于ML的抗生素发现管道将提供一种强大的新方法，以对抗抗菌素耐药性的全球威胁。

We propose to develop and evaluate a new machine learning (ML) methodology for active exploration of the vast space of potential antimicrobial peptide (AMP) sequences for the purpose of contributing to the design of a rapid, cost-effective and versatile pipeline for discovering new antibiotic activities. AMPs form a class of molecules that are widely used in nature to combat pathogens and are now used to address antimicrobial resistance (AMR), the evolutionary process by which pathogens have been mutating to become resistant to all known antibiotics. AMR is a major and growing public health concern, with 10 million deaths per year expected by 2050 with business as usual. Antibiotic administration to livestock animals accounts for 66% of all antibiotics used in the world, and is a major driver of antibiotic resistance. It is thus imperative that alternative antibiotic strategies are developed for livestock.We are developing a new active learning, reinforcement learning and generative modeling approaches for drug design and molecular genetic methodologies to experimentally evaluate the ML-generated candidates with the aim of designing new AMPs against major bacterial pathogens in livestock animals. We will integrate the whole sequence of interactions between the ML system generating candidate AMPs and the biological assays to screen the predicted candidates. These generate-and-test cycles are reiterated to help novel ML methods based on generative flow networks to discover structure in the data by appropriately exploring a diverse set of promising regions of bioactive peptide space. Because peptide sequence diversity is enormous (e.g., there are 20^50 possible 50-mer peptides), ML tools hold the promise of efficiently searching this space but require new algorithmic innovations to succeed, especially to address the need for generating diverse batches of candidates. The proposed synthetic biology methods for the construction and rapid screening of large peptide libraries is crucial for ML-driven AMP discovery. Active AMPs that inhibit bacterial growth will be identified by DNA sequencing to inform the next round of ML. We will generate peptide libraries for cyclized AMPs, which often have improved stability and more drug-like properties. We will then apply ML methods to design more potent combinations of AMPs. This ML-based antibiotic discovery pipeline will provide a powerful new approach to combat the global threat of antimicrobial resistance.