AI Forge: Trustworthy AI for Drug and Therapeutics

人工智能（AI）已经被包括医疗保健在内的一系列行业广泛使用。然而，缺乏对这些系统的隐藏内部如何操作的理解（定义内/定义外），引起了人们对我们可以或应该对每个特定工具的信任程度的担忧。无论系统是为最终用户提供服务（例如癌症诊断）还是作为产品开发工具（例如药物开发），可信度都是关键。2018年底，英国DeepMind开发的工具AlphaFold在很大程度上解决了一个困扰最优秀生物化学家数十年的问题-蛋白质折叠。在随后的几年里，人工智能已经证明，它不仅可以预测数千种蛋白质的三维结构，这些蛋白质允许生命发挥功能（即“折叠”它们），而且还可以创造新的独特的蛋白质（“反向折叠”它们）。这将使新的功能在医疗保健和超越有益的。蛋白质锻造是一个位于牛津的初创公司，专门从事蛋白质工程，并设计新的蛋白质为基础的疫苗和治疗。免费提供的新人工智能工具为在该领域工作的中小企业、学术界和制药公司提供了巨大的机会。设计新的蛋白质来治疗疾病通常是一个漫长而昂贵的过程。治疗开发使用少数天然存在的蛋白质之一，通常是抗体，与细胞或病毒的某些部分结合。疫苗使用死亡的生物体或生物体自身的蛋白质。人工智能现在提供了使用全新蛋白质作为治疗病原体、肿瘤和其他疾病的平台的机会。不幸的是，人工智能目前对药物开发管道的启动有很大帮助。用这些新方法在几秒钟或几分钟内就能设计出数百种可能的药物，要测试它们既不容易，也负担不起。因此，我们开发可信的工具来评估AI建议的许多药物中哪些可能会失败，就像大多数药物一样，在被公众批准用于临床之前通过严格的测试，这一点至关重要。该提议是一项可行性研究，将整个药物开发过程中的关键利益相关者与领先的AI专家聚集在一起。它将建立建立可信度的方法，并组成一个联合体，在其中可以评估数据、技能和开发这些工具的必要性。

Artificial Intelligence (AI) is already used extensively by a range of industries, including healthcare. However, a lack of understanding of how the hidden internals of these systems operate (in/out of definition), raises concerns about the level of trust we can or should place in each specific tool. Trustworthiness is key, regardless of whether the system is performing a service for an end user (e.g. cancer diagnostics) or as a tool in product creation (e.g. drug development).One of AI's headline successes came in late 2018 when AlphaFold, a tool created by UK's DeepMind, largely solved a problem that has taxed the best biochemists for decades - that of protein folding. In subsequent years, AI has demonstrated that it can not just predict the three-dimensional structure of the thousands of proteins that allow life to function (i.e. "fold" them) but also create new and unique proteins ("inverse fold" them). This should allow new functions beneficial in healthcare and beyond.The Protein Forge is an Oxford-based start-up specialising in protein engineering, and designing novel protein-based vaccines and therapeutics. The opportunities afforded to SMEs, academics and pharmaceutical companies working in this sector by the new AI tools that have been made freely available are enormous. Designing novel proteins to tackle a disease was typically a lengthy and expensive process. Therapeutic development uses one of a small number of naturally occurring proteins, usually an antibody, to bind to some part of a cell or virus. Vaccines use dead organisms or one of the organism's own proteins. AI now offers the opportunity to use entirely new proteins as platforms for treatments against pathogens, tumours and other conditions.Unfortunately, AI currently assists very much towards the start of the drug development pipeline. It is neither tractable nor affordable to test the many hundreds of possible drugs that can now be designed in seconds or minutes using these new methods. It is, therefore, key that we develop trusted tools to assess which of the many suggested by the AI are likely to fail, like most drugs, to pass the rigorous testing before they are cleared for clinical use by the public.This proposal is for a feasibility study bringing together key stakeholders from throughout the drug development process with leading AI experts. It will create methodologies to establish trustworthiness, and form a consortium in which data, skills and the need to develop these tools can be evaluated.