Turbocharging Phenotypic Screening with AI to Target mRNA Biology

While the pharmaceutical world is investigating biologics or mRNA-based therapies, Yochi Slonim, Israeli technology and biotech entrepreneur and public speaker on the subject of startup building and positioning, is sticking to the classics -- small molecules -- but his strategy comes with a twist.

Yochi is not a classically trained pharmaceutical scientist. His background is in software, where he has created and sold start-ups that were acquired by large corporations like HP, UGS, and BMC. Yochi took his expertise in software/artificial intelligence (AI) and co-founded Anima Biotech in 2014 to help create a novel way to identify mRNA-targeting drugs and their mechanisms of action.

Anima’s platform – called mRNA Lightning – has been the subject of recent partnerships with Takeda and Eli Lilly, amounting to more than $1.12 billion. mRNA Lightning overcomes the common barriers in traditional drug design, having the potential to elucidate new therapies for previously “undruggable” diseases. 

Small-molecule drug design works by fitting a novel molecule to a receptor’s binding pocket, matching and complementing distinguishing features of the targeted receptor.  For some rare and difficult-to-treat diseases, the binding pocket is ambiguous with relatively non-descript characteristics. Other diseases like Idiopathic Pulmonary Fibrosis involve an overproduction of a ubiquitous and essential protein, collagen, in only one organ. Giving a small molecule that shuts down or interferes with collagen production would result in off-target effects on collagen throughout the body, a seemingly worse alternative than the disease itself.

Anima Biotech has gone one step above targeting the protein; their small molecules target the mRNA transcript used to make it. Through leveraging artificial intelligence and phenotypic screening in live biology, the company can identify multiple small molecules that can turn up or turn down protein production with the selectivity and specificity of working on one organ system. Not only that but the drugs’ mechanisms of action are elucidated. Anima's technology produces “multiple shots on goal” in tandem with the knowledge of how these molecules work in live cells.

Below is our interview with Yochi, where he is telling a story of his journey into biotech, explaining the scientific rationale behind Anima’s focus on mRNA biology and describing how their AI-driven platform is capable of cracking biology secrets in a high-throughput way. 

Andrii: You developed your career as a successful serial entrepreneur in the area of information technologies with several multimillion-dollar exits, and then… biology? Biotech is undoubtedly among the most demanding industries for entrepreneurship, with much more significant overall business risk due to the complexity and poor predictability of biology systems (on top of all other “traditional” challenges startups face in any industry). Obviously, this wasn’t a spontaneous decision -- what was the catalyst, a life event, or a revelation that made you switch industries like that? 

Yochi: My journey towards the biotech industry was not the standard one, which connects to the very essence of what we do at Anima, a company at the intersection of biology and software. 

I was in the software industry for more than 20 years before co-founding Anima, and during that time, I created three companies that went through substantial exits. The first company, Mercury Interactive, became the world leader in automated software testing. We took the company public after four years and reached revenues of more than $1B annually. It was then acquired by HP for $4.5B and became their software division. After Mercury, I led the Products and Marketing division of Tecnomatix, a publicly traded NASDAQ company. I managed an organization of 500 people comprising four divisions and generated revenues of $100M until the company was acquired by UGS for $230M.

Mercury was about automatically finding bugs in software applications, but the product could not provide a description of the root cause of the error.  When I was 17, I developed an initial product that was like a “black box flight recorder” for software that records what is happening in the application, so when a software failure occurs, you could identify the problem. It was the year 2000, and the internet, online banking, and e-commerce industries were just becoming popular. The software running on these websites was not stable. Users were experiencing many errors, but with thousands of them sending transactions simultaneously, it was tough for the developers to understand why a particular fault had happened. I thought that, just like airplanes, online systems needed a black box. With this, I founded Identify Software to build the “Black Box of the Internet.” The company grew very fast, and in just five years, we reached more than 1,000 enterprise customers with $75M in revenues until we were acquired by BMC Software for $150M in cash.

My introduction to biotechnology came just before this exit, in late 2005, when I became co-founder and the first investor in Anima Biotech. I had a feeling that software had great potential for drug discovery. AI was not commonly used then, but neural networks were being used in image processing. The premise of Anima’s platform was based on phenotypic screening, a drug discovery process that generates millions of images showing the effect of molecules on the biology of mRNA. The combination of automated analysis of phenotypic screening data using AI and computational biology-based analysis tools was right within my expertise; however, I did not know much about mRNA biology. I started to study biology, and over the years, I became a dedicated student, taking more than 30 online courses. Anima developed the core of its science in the biochemistry labs of the University of Pennsylvania for over ten years. In 2015 we launched the platform into a drug discovery company after we achieved strong validation of the technology, including collaborations with 17 universities.

Andrii: Anima Biotech occupies an incredible niche -- targeting RNAs with small molecules. As far as I know, for a long time, RNA used to be a notoriously hard (if not impossible) object to target by small molecules -- in contrast to proteins which typically possess suitable “pockets” and have overall rigidity to their 3d structure. Most drug discovery is centered around targeting proteins. How has this changed lately? What is the scientific rationale behind Anima’s innovation? 

Yochi: Targeting mRNA is an indirect way to control protein expression. The idea is not new and, in fact, was invented over 30 years ago, but it took a very long time for such drugs to come to market. The first drugs were RNAi based, which are not small molecules and require injections. Recently, the COVID-19 vaccines dramatically raised the awareness of mRNA. This has become the #1 word in biology that almost everyone knows. There is a strong validation that mRNA biology is a “drug mine,” like a goldmine, for new drugs. The idea that we could do the same with small molecules is very attractive.

We can do this in two different ways. The first, as you mentioned, is to try to bind directly to the RNA molecule. Quite a few companies are trying to do that, but there are many challenges related to the chemistry of RNA and the selectivity of such drugs. The other approach is to discover molecules that bind to the various proteins that regulate mRNA across its life cycle. This approach targets proteins in the conventional binding manner into pockets; however, no one knows which proteins to target as mRNA biology is a very new field. Anima’s platform enables us to identify molecules that selectively modulate the mechanisms around mRNA biology. This approach has another significant advantage. Directly targeting the mRNA molecule is essentially a “one-way street.” By binding to the mRNA with a small molecule, the best outcome you could hope for is to interfere with its translation, but this approach cannot increase the expression of proteins. This approach is irrelevant for diseases caused by underexpression of a protein.

In contrast, cells have innate mechanisms that adjust mRNA translation either up or down in real time. By targeting these mechanisms of mRNA biology, you have a “two-way street” and can discover treatments for diseases caused by underexpression or overexpression of the target protein. This is of great interest to big Pharma as the conventional methods of directly inhibiting proteins cannot address such diseases.

The molecular targets of our discovered active compounds are proteins that control mRNA translation, splicing, re-localization, and essentially anything that happens to the mRNA post transcription. It is a novel target space with great potential. 

Andrii: Considering that you are coming from a computer science background, it is not surprising that the whole R&D model of Anima is centered around a technological platform employing artificial intelligence. What is this platform, and how does it operate? 

Yochi: Anima Biotech’s mRNA Lightning platform discovers small molecule mRNA drugs and their mechanisms of action. Our platform addresses these two elements with a novel approach to both. The first is our proprietary, high-scale automated phenotypic screening system that Anima has developed from the ground up around our unique expertise in mRNA biology. The phenotypic screening is unbiased and works in diseased cells expressing the protein of interest. This enables us to discover molecules that work through any of the previously mentioned mechanisms of action. In turn, we can find molecules of various chemistries that achieve control differently. This is a tremendous advantage as it provides projects with “multiple shots on goal.” The biggest challenge with phenotypic screening systems has always been that it is tough to identify the mechanism of action of active molecules. The screening yields millions of images but no understanding of what they really mean. This is where the second component of our platform comes in, and for the first time, we are using AI to tackle this enormous challenge. With our platform, we can rapidly elucidate the mechanisms of action of discovered molecules, understanding how they work and their molecular targets in this novel space. The AI technology is sophisticated, but we generally utilize an AI-driven image analysis that identifies possible MOAs for the active compounds. As the screening runs, hundreds of thousands of compounds are evaluated in millions of automated biology experiments. Each experiment tests whether the compound is active in controlling mRNA biology for the protein of interest. The results are visible in high-resolution images where we can see the effect of the compound on mRNA translation, mRNA relocalization, RNA binding protein activity, and more. Essentially, we can see anything that happens to the mRNA. This is mRNA biology imaging happening in real-time.

As an image is taken, it is uploaded to our cloud-based AI software, and the AI image analysis takes over. The system looks at the image and determines whether it has seen a similar image before. By now, we have a database of billions of images that came from dozens of projects on many different targets. The idea is like “facial recognition” technology applied to MOA elucidation. Indeed, we call this process “MOA facial recognition.” The idea is that if a compound works through a given MOA, the image will be similar to other compounds’ images with the same MOA. It is essential to understand that this is not a final determination of the MOA but provides a good idea of where to look next. This is done for every image and results in a starting point for all active compounds.

The next step utilizes our machine learning-guided system that takes the recommended MOAs and references our built-in information on the biology around them to suggest experiments and further investigate the potential MOA. By applying dozens of proprietary MOA assays that Anima has developed, the system acts like a “Compass” as it analyzes the results of experiments. A knowledge graph built around mRNA biology points us to the next experiment to run. As experiments are executed, results are integrated into the system, and the algorithm iteratively determines the path of experiments, navigating until the MOA is revealed. Finally, confirmatory experiments are performed on the proposed molecular target. With this proprietary technology, we leverage AI image analysis and machine learning-guided mRNA MOA assays to elucidate the mechanism of action rapidly. We have successfully employed our platform to discover the MOA of six different molecules in less than a year. This process has previously taken years for a single molecule, and many industry projects fail as the success rate of determining a compound’s MOA is relatively low. It is a novel use of AI in drug discovery, focused on elucidating the MOA of drugs discovered through phenotypic screening. We call this technology “MOAi.”

Andrii: Where do you get the high-quality biology data required to train models? Phenotypic screening involves a lot of complex experimentation. How does your company manage a “vet lab” experimentation -- in-house or via collaborations and outsourcers? 

Yochi: The great thing about phenotypic screening is that each run generates millions of images showing the activity of hundreds of thousands of compounds. We’ve run dozens of screens against many targets and have generated a database with billions of mRNA biology images. This is probably the single most significant source of visual information on mRNA biology that anyone has ever built. We use this to train the system, and in the training sets, we’ve included existing drugs and molecules from publications with known mechanisms of action. This makes the AI image analysis a self-learning and self-improving system. Our Compass machine learning guided system utilizes an integrated crowd-sourced knowledge graph that describes how specific proteins interact with others along pathways. We added our knowledge of mRNA biology to this graph, and this acts as a great source of insight for the machine learning algorithms to navigate the paths and connect the dots.

All of this technology was built in-house, and all biology experiments are fully guided and managed by our proprietary mRNA Lightning software at every step of the way. This ties back to the strong software roots of Anima, which complement our expertise in mRNA biology. 

Andrii: What is the growth strategy? Anima has a long list of drug candidates and also two high-profile industry collaborations -- with Takeda and Eli Lilly. Are you seeking growth through partnerships and asset licensing, or maybe the primary bet is on advancing internal pipelines to clinics? 

Yochi: There are so many potential targets, and our technology can be applied to almost any protein. We have been leveraging collaborations as a strategic way to develop our company while pushing our programs forward.  Collaborations give you money and credibility, and by now, Anima has become a recognized front runner in the fast-developing space of small molecule mRNA drugs. We have a very differentiated approach being the only company utilizing phenotypic screening, which enables our partners to de-risk projects and come up with novel drug candidates. Our ability to elucidate the MOA of drugs with MOAi technology is seen as a powerful advantage by partners. Our approach provides multiple strategies against a disease with an early understanding of the MOA, which helps move a project forward. We can optimize drugs coming out of phenotypic screens because we understand how they work early on in the development process. Unlike existing approaches that try to determine the MOA as the last step of a discovery process, our MOA elucidation starts as early as the first image that comes out of the screen and continues with every step. We think that this approach has a strong affinity to partnering, and we will continue to focus on this as a strategy going forward. For our programs, we chose targets where we have developed a great deal of expertise and where we have a strategic competitive advantage. For example, our idiopathic lung fibrosis (IPF) program is centered around the discovery of selective collagen I translation inhibitors. The drugs are at advanced preclinical stages, and they work through novel MOAs around tissue selective regulation of collagen translation. This means they work in the lungs but not in the liver, kidney, tendons, or any other tissue where collagen I is expressed. It is a great demonstration of the selectivity that can be achieved when targeting the regulatory mechanisms around mRNA biology. In contrast, if we were to target the collagen RNAs, the most abundant protein in the human body, a small molecule would hit collagen systemically. As another example, our three oncology programs are pursuing novel targets that were discovered using our MOAi technology.

Andrii: Right now, drug discovery is still a primarily serendipitous “trial and error” venture in many cases. A classical example is the target-based high throughput screening (HTS), a “brute force” approach, I would say. Do you believe that such technologies as AI can substantially improve the predictability of interventions in biology? Can we, in any foreseeable future, reach what some experts call “industrialization” of drug discovery when it would look more like a well-defined engineering/computation problem with good predictability rather than a “lucky breakthrough” game? 

Yochi: Anima’s approach can be seen as precisely that “industrialization” of drug discovery through phenotypic screening that runs in a fully automated biology lab. Hundreds of thousands of compounds are being run through millions of biology experiments. Each experiment tests a specific molecule at single cell level execution. The whole process is fully automatic and managed by our software utilizing advanced robotics and high-end microscopy. Notably, despite having so much software involved, there is no “IT footprint” in the lab. Instead, the system sends the results directly to the cloud, where Anima’s cloud servers handle the storage and analysis of billions of images safely and securely. In this environment, the MOAi technology runs automatically on every image, using MOA facial recognition to propose an MOA and determine the sequence of validating biology experiments to execute. Our tight loop between automated biology and AI software iterates until it lands on the MOA. When you look at the overall process, you can say that it is “industrializing” drug discovery by making it a software-managed workflow that connects biology, chemistry, and AI under one software-managed environment to deliver drugs at a much faster pace and higher success rate. Having said that, we must remember that biology remains unpredictable. While you can automate discovery, there will always be a need for clinical trials to assess how these molecules work in the human body. There are recent attempts to apply AI to the problem of predicting the outcomes of clinical trials, but this is the very beginning stage. I am not sure if that goal is set too high or if it can be achieved. Some companies discover drugs with AI, and while there have been reports of drug candidates being “invented” by way of machine learning, it is still unclear how these drugs work. At Anima, we decided to use the phenotypic screening approach for the discovery of active molecules and developed our proprietary AI software to solve the decades-long problem of elucidating the MOA of the active compounds. We think that this approach really has reimagined phenotypic screening as a key strategy for drug discovery.