Inside the nascent industry of AI-designed drugs
来源：Nature | 发布时间：2023-06-09
Artificial intelligence tools are beginning to upend the drug discovery pipeline, with several new compounds entering clinical trials.
Drug discovery is expensive, inefficient, and fraught with failure. An estimated 86% of drug candidates developed between 2000 and 2015 did not meet their stated endpoints.
Despite this challenge, the use of artificial intelligence (AI) and machine learning to understand drug targets better and synthesize chemical compounds to interact with them has not been easy to sell. Alex Zhavoronkov would know. When the CEO and founder of Insilico Medicine, with offices in Hong Kong and New York, first started trying to raise funding nearly a decade ago, he struggled to find others who shared his vision.
“It was such a grand goal, but every time I went to a venture capitalist, they never gave me money,” says Zhavoronkov.
Even as recently as 5 years ago, his presentations had to explain to pharma collaborators why AI was so promising. Not anymore. Now he is at the forefront of drug discovery’s AI nascent revolution.
“We’ve managed to get here in three years, and we didn’t fail. And we did it multiple times,” Zhavoronkov says.
The persistence of Zhavoronkov and a small cadre of other startup founders, including Exscientia’s Andrew Hopkins and BenevolentAI’s Bryn Williams-Jones, means that not only are some of the biggest players in pharma already convinced of the utility of AI in drug development, but also some of these drugs are beginning their ultimate test in clinical trials (Table 1).Table 1 Selected AI-designed drugs in or entering clinical trials
“In the last couple of years, AI has gone from being hypothetically interesting to real programs moving towards the clinic,” says Williams-Jones. “There’s no shortcuts to drug discovery. We can have better informed ideas, but you still have to go through the rest of the [development] process.”
These trials are still in their early days, says Hopkins, so it is not yet clear which compound will cross the finish line first. But he is confident that the use of AI is leaving an indelible mark on drug development and promises to make the process better, faster, and cheaper, as well as enabling the development of more first-in-class compounds.
“We expect this year to see some major advances in the number of molecules and approved drugs produced by generative AI methods that are moving forward,” Hopkins says.
Entering trials
As AI-designed drugs enter clinical trials, pharma companies can see how their new compounds are paying off. The preliminary read-outs look promising. In June 2022, Exscientia announced preliminary results from a phase 1 trial of EXS-21546, a highly selective A2A receptor antagonist developed with Germany’s Hamburg-based Evotec. The small molecule has subsequently entered phase 1b/2 trials for patients with solid tumors carrying high adenosine signatures.
Exscientia’s next AI-developed candidate, a small molecule called EXS4318, is not far behind. A selective protein kinase C-theta (PKC-θ) inhibitor, designed for inflammatory and autoimmune conditions, EXS4318 has been licensed to Bristol Myers Squibb in a partnership worth up to US$1.2 billion, according to a company press release. The company has 16 other AI-designed drugs in its pipeline, including drugs for COVID-19, tuberculosis, malaria, and hypophosphatasia — a rare, inherited disorder that affects bones and teeth.
“It’s not just about using generative AI to help us to precision design an exact molecule,” Hopkins says, “but also actually helping us precision design which patients are responders and non-responders.”
What this would look like in practice, Hopkins says, is performing deep, multi-omics (single-cell proteomics, transcriptomics, and genomics) analyses of participants before the trial starts to identify multi-gene signature biomarkers. This will help the researchers to determine which participants are most likely to respond — and why. At the end of the trial, Exscientia will be able to go to regulators with a drug that consistently works well in a very defined patient population.
“This is where AI is going to lead as well. It’s not just about using AI to make drug discovery better, but about how we can create better drugs overall,” Hopkins says.
In January 2023, Insilico Medicine announced an encouraging topline readout of its phase 1 safety and pharmacokinetics trial of INS018_055, designed by AI for idiopathic pulmonary fibrosis, a progressive disease that causes scarring of the lungs. Their proprietary AI platforms identified a new target (which Zhavoronkov would identify only as ‘target X’) and a small molecule inhibitor, which was granted breakthrough status by the Food and Drug Administration (FDA) in February.
“It’s the first time anyone in our industry has developed a novel target of a molecule, and completed phase one trials, all the way with AI,” Zhavoronkov says. He expects phase two readouts in the first half of 2023. It is part of Insilico’s growing pipeline targeting diseases associated with aging. What makes Insilico’s work more impressive, according to Zhavoronkov, is that the company only began development on INS018_055 in February 2021.
“We have 31 therapeutic programs. In 2020, we had zero,” says Zhavoronkov.
AI for analysis
Recursion, a biopharma startup based in Salt Lake City, Utah, uses AI not to design molecules but to analyze data from millions of experiments and billions of microscopy images that their lab is gathering with the help of robots.
“Just like Google has all these cars driving around taking pictures that they turn into really useful maps for all of us, we’ve done the same thing with biology,” says Chris Gibson, Recursion’s co-founder and CEO.
Recursion is also working to develop a therapeutic agent for ovarian cancer that targets a gene that their AI systems indicated was part of the same pathway as CDK12, an existing target that has proved challenging to inhibit directly. In preclinical studies that target the CDK12-associated protein, 40% of mice showed a complete response. When the compound was paired with a PARP inhibitor, tumors were eliminated in four out of five mice. The company also has three other compounds in clinical trials for oncology and rare diseases: familial adenomatous polyposis, cerebral cavernous malformation, and neurofibromatosis type 2.
“Biology and chemistry are so broad and complex. Your goal isn’t to find everything. Your goal is to find something really good and advance it,” Gibson says.
Relay Therapeutics has developed an oral, small molecule inhibitor of FGFR2, a receptor tyrosine kinase that is overactive in certain cancers, such as intrahepatic cholangiocarcinoma. Existing FGFR inhibitors are not very selective, but the company is testing RLY-4008, which is only active against FGFR2. At the end of 2022, BenevolentAI completed a phase 2a trial for BEN2293, a topical ointment for the treatment of atopic dermatitis (eczema). The treatment was found to be safe but did not meet its secondary endpoint of reducing itch and inflammation, according to a company press release in April 2023.
BenevolentAI has also filed a clinical trial application with the UK Medicines and Healthcare Products Regulatory Agency (MHRA) for BEN-8744, a small molecule phosphodiesterase 10 (PDE10) inhibitor designed to treat ulcerative colitis. If approved, Williams-Jones says BenevolentAI plans on beginning a phase 1 trial in the first half of 2023. But for BenevolentAI, as for everyone else, he points out this is still early days.
“Biology is hard, and we don’t know very much in real terms,” says Williams-Jones. Every time scientists think that they have made a big step forward in simplifying the drug development process, he says, they stumble across two or three other issues that they did not expect.
The protein folding problem
Much of AI-driven drug discovery builds on protein folding. By the latter half of the twentieth century, biochemists had decoded some of the basics of protein structure tenets that now fill biology textbooks. A string of amino acids, proteins fold into complex, three-dimensional structures based on the atomic interactions between the backbone and amino acid side chains. This structure determines the protein’s function. As crystallography and electron microscopy began to crack open the atomic-level structures of proteins, biochemists began to wonder whether it might be possible to predict the final structure of a protein complex using only its amino acid sequence. The discovery of α-helices and β-sheets in the 1960s made the promise seem almost tractable.
Then reality began to sink in. Twenty simple amino acid building blocks could give rise to a dizzying array of proteins — greater than the number of stars in the universe, Baker says. Methods such as multiple sequence alignment (MSA) enabled structural bioinformatics experts to compare the amino acid sequences of numerous protein homologues to determine domains, disordered regions, and other elements of local secondary structure. But even the most advanced MSA methods could not reveal allosteric interactions, or how different α-helix regions were arranged next to each other.
AI and machine learning took a completely different approach. “Machine learning is based on the results you attain rather than a statistical model that describes the population,” Deane says. “It’s about finding predictive patterns in the data.”
Instead of applying the laws of physics to every single atom or bond, what if scientists began to look for similarities between proteins? If they could assemble a reasonably broad base of protein structures (gathered the old-fashioned way, through painstaking crystallography, X-ray diffraction, and electron microscopy techniques), then perhaps scientists could try to figure out the similarities between proteins and use that to predict a protein’s structure.
“With deep learning, you don’t really try and simulate the actual folding process. You’re not trying to find the lowest energy state. It’s more about pattern recognition,” Baker says.
The intellectual leap to this way of thinking was profoundly important, says Alan Lipkus, senior data analyst at Chemical Abstracts Service in Columbus, Ohio.
Weird molecules
By the early 2010s, computer scientists and computational chemists had developed the prototypes of groundbreaking AI systems such as RoseTTAFold and DeepMind’s AlphaFold. Most modern machine learning algorithms devoted to predicting protein structure contain four different modules: an input module that contains the amino acid sequence and structures from homologous proteins; a sophisticated neural network that uses pattern recognition algorithms to transform the amino acid sequence into spatial information of the protein; an output module that converts the spatial information into a preliminary three-dimensional structure; and a refinement process that enables fine-tuning. Using these algorithms, AlphaFold2 can predict single protein domain structures down to 2.1 Å, essentially solving the protein structure problem. It is a staggering accomplishment, Baker says, but he wants to move beyond it.
“By just predicting protein structure, you’re stuck with whatever exists in nature. You can’t make anything new. But now we can make all these brand-new proteins for cancer therapeutics and clinical trials. You can make all kinds of different things with protein design,” Baker says.
Beyond the basic science accomplishment, these advances have also given a huge leg up to pharma. Determining a protein’s structure was a major hurdle in designing the right molecule to alter its function. Determining the structure of a small molecule was simple compared to a protein. Even biologics designed by AI were a possibility, antibodies just being one specific type of protein. This progress did not remove the need for experiments and tinkering — no computer algorithm is yet that good — but it narrowed down the number of possibilities to help scientists prioritize molecules that were far more likely to have the desired effect without causing undue toxicity.
The molecules these AI systems helped to design, however, looked very different from compounds designed by medicinal chemists. When InSilico’s Zhavoronkov began pitching his AI therapeutic design service to pharma companies, he included examples of several molecules his system had built. Their novelty immediately grabbed the attention of potential pharma partners, some of whom helped provide series A and B funding rounds.
“They said to me: Alex, these molecules look weird. Tell us how you did it,” Zhavaoronkov says. “We did something in chemistry that humans could not do.”
And it is this weirdness that just might be AI’s biggest strength in pharmacology. Although the total number of possible chemicals in the universe — what some scientists refer to as chemical space — is vast, humans have only explored tiny slivers of this space. Synthetic chemists develop expertise working with certain types of compound or performing specific reactions, says Lipkus, leading to a few small areas of chemical space that are well mapped out. Most of chemical space remains terra incognita.
Many clinical trials test tweaks of existing drugs, which may give a slightly improved safety or efficacy. However, a much bigger prize is a first-in-class drug against an entirely new target, which AI-designed drugs are well-positioned for.
Lipkus and his colleague Todd Wills (now a senior vice president at Cass Information Systems) analyzed the novelty and creativity of pharmaceutical molecules using the chemical abstract service database of thousands of molecules, which “is probably the best representation of they known chemical universe”, Lipkus says. They compared the uniqueness of a molecule’s scaffold and shape, which they defined as the atom-to-atom connectivity that prunes back all but the most basic information about a compound’s structure. ‘Me too’ drugs, they pointed out, tend to consist of small alterations to a drug’s chemical side chains rather than large-scale shifts in molecular structure. A growing number of pharmaceutical compounds, they pointed out in a 2019 paper in the Journal of Organic Chemistry, are showing signs of creativity, with more unique structures and scaffolds. AI, Lipkus says, will only accelerate this trend.
“It’s one more piece of evidence that there’s value in looking for novel structures,” Lipkus says. “Talking to people in the drug industry, they want to break away from these scaffolds that have been used so heavily.”
AI tools also enable drug developers to explore the chemical world much more quickly.
“It allows us to explore a much broader slot or chemical space than we’d be able to using experimental methods on their own,” says Don Bergstrom, president of research and development at Relay Therapeutics.
Neglected diseases
AI-designed drugs are not just being developed for potential blockbuster status. In Geneva, Switzerland, the Drugs for Neglected Diseases Institute (DNDi) is using machine learning to create better drugs for conditions that predominantly affect the world’s poor, such as Chagas disease and dengue fever. Charles Mowbray, discovery director at DNDi, says the institute is also turning to AI strategies to guide its drug repurposing pipeline as part of its global efforts to develop therapies for neglected diseases. For such diseases, speed is critical; AI can help scientists generate hypotheses and test them more quickly.
“These tools don’t replace a scientist, they complement them,” Mowbray says. “[AI] enables them to have all the information at their fingertips, to ask the good questions, to refine their queries, and to iterate until they can figure out what they’re really after.” This synergy is true for machine learning across drug development, he adds.
Even as the impacts of AI in drug design are beginning to emerge in clinical trials, these strategies are joining other AI tools in clinical trial design, manufacturing, and more. There is no doubt that machine learning is profoundly reshaping the pharmaceutical industry, Lipkus says. As for how the effects of AI-developed drugs will play out, he is more circumspect, saying that is still up in the air.
“Nothing guarantees anything. Drug discovery is really difficult. I don’t know if people expect AI to just pop out the design of a molecule that’s your next blockbuster, says Lipkus. “It’s all kind of a crapshoot.”
AI制药正在变革药物研发
药物研发过程充满挫折和挑战。利用人工智能（AI）和机器学习（ML）能更好地发现药物靶点并合成化合物。
但AI制药的发展过程同样曲折，英矽智能（Insilico Medicine）首席执行官兼创始人Alex Zhavoronkov表示，“大约10年前，每次我去找风险投资家融资，他们从不给我钱。”
但现在大家对AI制药的接受度越来越高，这一技术赛道站在了医药研发革命的最前沿。
Zhavoronkov创办的英矽智能是全球最早成立的一批AI制药公司，目前已有一款药物处于临床II期试验。除英矽智能外，Andrew Hopkins创办的Exscientia、Bryn Williams-Jones创办的BenevolentAI等公司也都有产品处于临床阶段（下表1）。
来源：Nature Medicine
多款药物进入临床
EXS-21546是一种高选择性的A2a受体拮抗剂，由Exscientia与Evotec合作开发。2022年6月，Exscientia宣布了EXS-21546的I期试验的初步结果，随后该分子进入Ib/II期试验，适用于高腺苷特征的实体瘤患者。
Exscientia的另一款AI候选药物EXS4318是一种选择性PKC-θ抑制剂，用于治疗炎症和自身免疫性疾病，目前以12亿美元的价格授权给了Bristol Myers Squibb（BMS）。
2023年1月，英矽智能宣布了ISM001-055的I期安全性和药代动力学试验的数据，这是一项由AI设计的、用于治疗特发性肺纤维化的临床试验。Zhavoronkov说:“这是AI制药行业中，从新靶点发现、到临床I期试验都使用AI的首个案例。
Recursion总部位于美国犹他州，成立于2013年。不同于传统的、基于结构或基于配体的药物发现公司，Recursion以细胞图像为最主要的数据集。
Recursion的药物发现过程如下：首先，在实验室里通过各种方式使人体细胞生病，拍摄这些生病的细胞；然后，让机器学习程序来学习这些生病细胞与健康细胞的区别；最后，将各种药物作用于患病细胞，通过机器学习程序来判断细胞是否回归健康状态，从而判断药物的作用效果。
Relay Therapeutics开发了一种口服小分子FGFR2抑制剂。与现有的FGFR抑制剂相比，RLY-4008的优势是，只对FGFR2有选择性。
2022年底，BenevolentAI完成了BEN-2293的IIa期试验，这是一款治疗特应性皮炎的外用软膏。根据该公司在2023年4月发布的新闻稿，这种治疗方法被发现是安全的，但没有达到减少瘙痒和炎症的次要终点。
BEN-8744是一种小分子磷酸二酯酶10（PDE10）抑制剂，用于治疗溃疡性结肠炎。2022年12月，BenevolentAI向英国MHRA提交了BEN-8744的临床试验申请（CTA）。Williams-Jones表示，如果获得批准，BenevolentAI计划在2023年上半年开始I期临床试验。
蛋白质折叠问题
人工智能驱动的药物发现大部分建立在蛋白质折叠的基础上。计算生物学家David Baker表示，蛋白质的一/二/三/四级结构极其复杂，20个简单的氨基酸可以组合出极大量蛋白质，甚至比宇宙中的恒星数量还要多。
21世纪以来，计算机科学家和计算化学家已经开发出了突破性人工智能系统的原型，如David Baker的RoseTTAFold和DeepMind的AlphaFold。除了基础科学的成果之外，这些进步也极大地推动了制药行业的发展，因为确定蛋白质的结构是设计正确药物分子的主要障碍。
有时候人工智能设计的分子看起来与药物化学家设计的分子截然不同，但这种“怪异”可能是人工智能在药理学方面的最大优势。因为尽管宇宙中可能的化学物质总数（化学空间）是巨大的，但人类只探索了这个空间的一小部分，目前仍有大部分化学空间是未知的领域。
探索更多可能性
人工智能设计的药物不仅仅是为了潜在的重磅炸弹而开发的。
在瑞士日内瓦，被忽视疾病药物研究所（the Drugs for Neglected Diseases Institute，DNDi）正在利用机器学习为一些主要影响世界穷人的疾病（如恰加斯病、登革热）开发更好的药物。该研究所还在转向人工智能战略，以指导其药物再利用管线。对于这类疾病，研发速度至关重要，人工智能可以帮助科学家更快地提出假设并进行测试。
DNDi的研发主管Charles Mowbray表示：“这些AI工具不会取代科学家，而是对科学家的补充，人工智能使科学家掌握所有信息，提出好问题，并不断迭代，直到他们能找到真正想要的东西。