Interviewed by William Leonard Pickard
Alex Zhavoronkov, PhD, is Founder and CEO of Insilico Medicine, a leading clinical stage biotechnology company developing next-generation artificial intelligence and robotics platforms for drug discovery, with headquarters in Cambridge, MA and facilities around the world. He has invented critical technologies for the creation of novel molecular structures, and pioneered the prediction of human biological aging.
Q: Tell us about Insilico’s work
AZ: Insilico has two main focus areas: developing software and discovering and developing novel drugs for aging-related diseases (thus enhancement in neurodegenerative states and in healthy adults). When it comes to software, most of our platforms are commercially available for startups and academics with free trial access, and for some specific projects we can provide software for free in the form of a collaboration. When it comes to drug discovery, academics and startups can contact us to test our drugs in a variety of experimental models.
We are also very collaborative in AI for aging research and sustainability. While I do not support irrational environmental activism where people drop out of school, stop traffic or impose draconian sanctions on fossil fuels, climate change is real and we need to develop new technologies to be able to clean up the environment and make it more sustainable. That is why we created a consortium to build and benchmark generative AI models in a variety of applications starting from CO2 capture and hydrogen storage but also branching into new fuels and sustainable agriculture. Scientists can apply to join this consortium to co-publish. We will be releasing the first models soon.
I also believe that it is very important to develop AI that understands aging. Biology does not stand still – it happens in time. Even if we do not get specific diseases, all humans will gradually lose function and die. Most of today’s disease models do not understand the changes in biological systems in time and instead treat individual examples like disconnected snapshots.
In 2015, we realized that age is a universal feature that every organism has, and started building deep learning models trained to predict age and other features. Now, we have several models, the most capable being Precious3GPT, a multimodal, multiomics, multispecies, and multitissue transformer-based Life Model. We just open sourced this tool last month and are beginning to build a community of active users. Once we get enough use cases, we will submit the paper to a high-profile journal. This effort is also expected to produce a few drugs that may have anti-aging properties. We encourage academic and industry scientists to use it, collaborate, and contribute to the advancement of aging research (and, through these means, to enhancement of cognition).
Q. Why is aging research so important for drug discovery in cognitive enhancement, and how can it help us accelerate drug discovery of novel drugs targeting neurodegenerative diseases?
AZ: The greatest problems with neurodegenerative diseases and drugs that enhance cognitive functions are the lack of relevant data points and the difficulty of validation. Getting high-quality data from tissue biopsies from patients and healthy patients is a challenge for every disease. Even in cancer, where biopsies are common and data is more abundant, we still face difficulties training large enough models.
With anything neuro, you cannot easily biopsy a healthy brain or even the brain of a patient with severe cognitive decline. It is too risky. We rely on animals and on post-mortem biopsies. Animals are very different from humans, so different that we still do not have a good model for most neurodegenerative diseases, and post-mortem biopsies come from a brain that is dead and was dying for some time. It is very difficult to formulate a good hypothesis and find a good target from this data. We need to have a good understanding of biology in time on many levels, in many tissues, and in many organisms. Without it, we will just continue shooting very expensive bullets in the dark.
In drug discovery for neurological diseases and cognitive enhancement, I am very bullish on two technologies: Life Models, which overcome the lack of relevant data points through computer modeling and Brain-Computer-Interfaces (BCIs) which translate brain activity into commands to an external device or interface. Precious3GPT is our first Life Model that can generalize across different data types in several species and perform some basic drug discovery tasks. But I think we are still too early with this model to go after a large neurological drug discovery program. To ensure that we become profitable enough as a biotech to eventually invest more resources in aging research, it will take us a few years to start a full-blown neuro drug discovery project with novel targets or mechanisms. For a case study of drug discovery in fibrosis, where the protein target was evaluated using the hallmarks of aging assessments, I recommend our recent paper in Nature Biotechnology.
Q. You have been in aging research for 20 years – did you discover promising drugs for Alzheimer’s or, more broadly, cognitive enhancement for the aging brain?
AZ: Often people ask me why after ten years I haven’t found an anti-aging drug that would work in the brain, and it is a difficult question to answer. The entire world could not find why and how Alzheimer’s or even ALS happens. My estimate is that the NIH alone spent over $60 billion, or over $100 billion in today’s dollars on neurodegenerative diseases across all institutes, between 1990 and 2023. The industry has probably spent significantly more than that and we still don’t have a good enough model of Alzheimer’s. The answer to this phenomenon is clearly not a problem that can be easily solved with just a little bit of time and a little bit more money. I started noticing that some executives in the biotech industry do not have real hope of seeing a “GLP-1 moment” for Alzheimer’s within their lifetimes, and now that I know how drug discovery works, I cannot blame them.
To answer your question – no, we do not have an effective neuro or aging drug yet. Some promising drugs are now nearing preclinical candidate stage and may work in neurodegenerative diseases – we are testing them now. Many of the drugs in Insilco’s pipeline score very highly in aging clocks, meaning the genes we are targeting are important for age prediction. Some of these targets are also implicated in several hallmarks of aging, so I do hope that these drugs will work in aging as well. Our primary objective is to get these drugs approved for their intended indications first.
I think that regardless of how ambitious you are, you need to play by the FDA rules and ensure complete compliance with the established clinical pathways. These rules are there for good reasons and, while we are often using aging research and AI as a platform for drug discovery, we need to focus on specific diseases and follow the standard development process. It is the same for every biotechnology company regardless of how they discovered their drugs.
Q. You’ve said, “We need to recognize medicinal chemists, even when some of those are not human.” How may your generative AI drug discovery platform be applied to small molecule development for cognitive enhancement, like improving attention, learning and memory, and intelligence?
AZ: At Insilico we have several tools to go after this problem. One, PandaOmics is a popular industry software tool for disease modeling and target discovery. It predominantly works on comparisons between norm and disease. The easiest way to apply it for cognitive enhancement is to study dementia to find promising new targets and to identify targets for repurposing. A case study in ALS is now in a fully-enrolled principal investigator initiated Phase II study. Another tool called Chemistry42: Generative Chemistry allows us to design molecules with desired properties like those that can optimize blood-brain barrier (BBB) penetration, a very important parameter for any kind of targeted drug for cognitive enhancement.
Q. Are you also working on cognitive enhancement through expression of genes by genomic editing or only through small molecule development?
AZ: We can do both. Biology42, our generative biologics platform relates to genomics. Chemistry42 our generative chemistry platform relates to small molecules. Both tools are very versatile. Chemistry42 is the most experimentally validated tool, with many papers published and 9 human clinical trials from molecules generated using this tool. In addition, Precious3GPT, our first Life Model, is trained on gene expression, methylation, protein expression, text and other data types. It can solve many tasks in drug discovery and aging research, and, arguably, enhancement of cognition. We open-sourced it to build a community, and to continue to get broader validation.
Q. At what point are your platforms self-training?
AZ: Many of our platforms rely on reinforcement learning (RL), where the already capable and pre-trained model generates the result and then another model or many other models evaluate the result to check if it satisfies the generation conditions. If the results are valid, the model that produced it gets rewarded. If the results are not perfect, the model gets corrected. This process can be done algorithmically or with experimental validation, where our robotics lab tests the output of the generative models. Additionally, expert feedback can contribute, where a community of experts test and train the models. Often, it is a combination of all three approaches.
Q. How do Brain-Computer Interfaces influence drug discovery of small molecules for cognitive enhancement in dementia?
AZ: BCI technologies are advancing at exponential rates since Elon Musk entered the field. He made it clear that going invasive at scale is not impossible. Once implanting BCIs becomes routine, even if only in patients, trying out new drugs will become much easier.
I am betting on a near-term bright future where our Life Models advance and get substantial experimental validation; and where most Alzheimer’s patients and patients with dementia get a BCI.
At Insilico we already have the most efficient drug discovery engine that can generate molecules with desired properties, and we will be ready for the future when we can formulate a hypothesis using a PreciousGPT Life Model, generate a drug, and then test it in a clinical trial with an invasive BCI biomarker readout.
Q. Last question, an important one: How do you and your lab handle ethical issues that might arise?
To ensure that we address any possible ethical challenges we adopt the policy of maximal transparency. That is why we regularly publish in academic journals (at the rate of approximately one paper every 10 days) and present at scientific and industry conferences. We also make our software commercially available to the pharmaceutical companies so that they can quickly acquire our capabilities.
