Founder, president, and CEO of Compugen. I am Danya Ben-Hail, an analyst at Jones Research. Compugen was doing computational target discovery long before AI and biotech became a buzzword. Today, we'll explore what it really takes to turn predictive biology into first-in-class immuno-oncology drugs. Eran, thank you for joining us.
Thank you, Danya for inviting me.
Let's start with the foundation. Compugen repeatedly uncovers first-in-class targets where our traditional discovery methods stall. Every AI company today claims a data advantage. To get things started, can you give a brief overview on Compugen's platform and what makes it unique?
First, as you mentioned, we are not just a recent AI story that's saying that we're doing AI just to have the buzzword in the name of the company. We're doing it long time ago. We have built the system. We have this engine called Unigen. This is one. I think the second part is the stage of the research and development that we are focused in, which is relatively unique. You gather the really nice group of companies in this set of Fireside chat with AI companies. Most of them are doing, for example, AI to identify the drug targets, to identify easier, maybe better antibodies, small molecules. We are focusing on the very first stage of the research and development, on the drug target itself.
Which target, which molecule, which protein in the human body we should target to really modify it to have eventually effect in cancer patients. There are not many companies that are focusing on this relatively challenging first part of bringing novel targets, there are definitely not many companies that have shown again and again that they can bring novel targets, they are generating clinical data, further validation by pharma companies collaborating on the targets. I think this is another differentiation that the platform is validated. Finally, it's not only the AI, the machine learning, the algorithms, the database that we have, that we build along years. It's really the end-to-end capabilities.
In one company, we have this knowhow, again, developed along years of using the right algorithm, right data sets, asking the right questions to identify the drug targets, also how to validate, how to eliminate target that should not proceed, how to choose the most promising one to take into the clinic, then in the clinic, where exactly to take them. Think also, this end-to-end approach is relatively unique for Compugen.
Yeah. No, that completely makes sense. Your tumor microenvironment, the proprietary algorithms, your decades of accumulated wet lab, you have the specific way to map the immune evasion mechanisms. All of that together gives you this advantage? Is there anything very specific and unique that you'd like to say, "This is the point of uniqueness." Or,
It's absolutely that. It's the combination of how these pieces reinforce each other. We have great algorithms. We're mapping the tumor microenvironment for years, from every different angle you can imagine, transcriptomics, proteomics, to try to learn about the human tumor microenvironment. Also, the convergence of disciplines, having data scientists, biologists, and clinicians sitting together in the same table, asking themselves, "What clinical problem we can solve?" Asking a specific biological question, how we can use the algorithms and data science to help us identify the targets to solve that problem. This is one. Second is that this is a flexible approach. The database continue and feeds itself. We continue to generate data all the time, from the preclinical data, from clinical studies we are doing. We're sequencing patients.
Whatever sample we can put our hands on, we're sequencing. This feeds back into the database to enable us to choose even better targets for the next round.
Yes. Given the high failure rate of novel biology, why is finding new targets still the most valuable application of your platform? How do you protect against the existential risk of novel target failing in phase II?
First, I don't think this is only a problem of a novel target which fails in the clinic.
Right
The biggest challenge is how to really predict clinical success in a preclinical package. For example, if you want to generate a better molecule for a known target, you need sometimes to be better than the existing one. You need to be differentiated. Even if your molecule is working because the target is validated clinically, you're not better than the existing drug. There are many challenges also for developing better drugs for known targets, especially, I think, these days with China competition rising up. I think that's really bringing the target itself, not just trying to develop better drugs or known targets, is something which is still a challenge. Something that we have a knowhow and advantage in how to do it.
Again, given the recent competition for China, I think this is still something which is unique and still have less competition and less competitive pressure from others.
Yeah. You're right. The industry has seen high-profile failures. I will point that especially in targets that looked great in silico but failed in humans. How does Compugen ensure its models are capturing real drug-responsive human biology rather than just finding statistical noise in massive multi-omics datasets?
This is a very important question. It has multiple layers of the way we ensure it. First of all, we are not looking into statistical noise in data sets. We are starting with a unique clinical question, we always, always make our analysis in the most relevant human samples. We never use proxies. We make sure that we have the right data set of human tumors from patients to answer the clinical questions that we have.
The targets that we are working on are never remains in silico. Comes, and discussed before, the end-to-end capabilities, how to build a biological package to really convince us, first of all, that this target is a valid target that can succeed in clinical settings. We put huge efforts into doing all the right experiments. Maybe I can take an example.
The COM503 that we licensed to Gilead, it's haven't yet proven clinical, or at least we didn't disclose any clinical data, at least the package was convincing enough for a few pharma companies to chase it, eventually we got this $850 million deal, 60 plus 30, all of that. The COM503, we started with a very unique clinical question about resistant mechanism to PD-1, we looked into patient samples in Unigen. For some of the samples, for that specific question, we collected more samples. We have a very good collaboration with hospitals in Israel, we get, every day, samples from hospitals, from surgeries. We made the discovery itself. A disease relevant, that source for the data in which we found the discovery. Also the validation stage.
We try to rely as much as we can on human samples directly ex vivo from patients, showing the activity of the drug in the most relevant system. Yes, sometimes mice could be relevant, but definitely the focus should be on the human systems, on a very rigorous validation and high bar for targets to move forward. Then, at the end of the day, also in the clinical settings, to really use the computational tools to identify the patient that could benefit, giving all the data that we have. I think this whole package, going from discovery in human samples all the way to the clinical settings, with a very strong biological package around it, experimental package, is key to what we do.
This is, in a way, taking this initial in silico prediction into a target with full biology around it that is sufficient for us to move forward with.
You're one of the first companies to bring computationally discovered targets into human trials. What have regulators taught you about how they evaluate preclinical packages that were rooted in predictive algorithms?
From our experience, we didn't have any issues because eventually, we're not bringing them the in silico package. We're bringing them the experimental package, the preclinical package. Lot of focus on safety. The packages that we generated for our computationally discovered targets had never any issues with regulators. We always had a nice path into phase I, INDs acceptance and all of that. If you generate the right package, the fact that the target initially was discovered in silico doesn't really matter much.
Right. Okay. With AI being implemented across the entire R&D path, what is the hardest biological constraint in drug discovery, whether it's tumor microenvironment complexity or immune pathway redundancy, that no algorithm can simply optimize away?
It's exactly that. I think that eventually, the ability to model the complexity of the human, in our case, immune system, but in general, the immune system, is extremely challenging. What we are doing to address that, as mentioned before, we are really modeling and mapping the tumor microenvironment from every angle you can imagine. We definitely use the recent AI tools, which allows us to do things that we couldn't have done before. Absolutely. We could navigate now in this complex, huge complexity of data in a way that we couldn't have navigated before, definitely. Still, we're not in the place we can say, "Okay, we can push on a button, and we get a target." This is a very complex system, very complex trial and error approach.
Definitely, the tools are improving every day, and we definitely can do today things we couldn't have done in the past.
Right. There is the biological redundancy, and that's something that AI probably can't fully predict today, and hopefully will get better at. Does that explain your shift into looking at combination therapies, like pairing your PVRIG inhibitors with TIGIT or PD-1 inhibitors? How does the platform account for these complex interactions?
First of all, I think that, in general, oncology is going into combinations. We know that a single drug can achieve, in some cases, significant effects, but in many cases, you need to combine different mechanisms. This is, again, where Compugen fits in. We are bringing new mechanisms that could work as monotherapy but could definitely also work in combination. For example, the biology of PVRIG, which we think is very relevant for ovarian cancer. The focus now in the MAIA study, and we will discuss it in a second, is on the monotherapy activity of COM701, the blocker of PVRIG that we identified computationally. The next steps could definitely also be combinations. This really depends on the patient population, what they can tolerate, the ability to combine.
Again, PVRIG, for example, COM701, the block of PVRIG, has a very good safety profile, so it's relatively easy to combine. I think combination is definitely something that is relevant. It's not necessarily for AI-discovered or non-AI-discovered targets.
Yeah. For PVRIG, what gave you the confidence to take it into preclinical and then clinical studies? What in silico signals convinced you to continue with that?
First of all, yes, PVRIG was identified computationally. There were no publications around it. The academic community didn't know PVRIG when they identified it, which is a good start, but definitely not the important part. The important part that we started to explore it, we identified it has a very, very different biology, different from TIGIT completely, from PD-1. We don't only have a new checkpoint. It's a checkpoint that could have different consequences when you block it compared to other checkpoints. This is what we saw preclinically, and this is exactly what we saw clinically. In patients, we have seen this unique biology translating into activity. As mentioned before, we're sequencing the patients.
We looked in patients treated with COM701 before. We take biopsy also, and samples also after treatment. We saw how COM701 can modulate the tumor microenvironment in a way that we think is unique to its biology. Also the clinical signals. We have, for example, a patient with PD-L1 negative ovarian cancer that failed all the standard of care treatment it could receive.
She received COM701 in monotherapy, and that computational prediction prolonged the life of that patient. She remained on the study for two years. The goal of the current MAIA-ovarian study is to take the signals we have seen. The drug is active, and PVRIG is not in silico prediction now. We know it's an active drug that could modulate and affect the progress of ovarian cancer tumors. We also have seen signals in other tumors.
The MAIA-ovarian study is a study we are doing that after seeing signals in the last line, platinum-resistant ovarian cancer, patient who failed everything, some of them had 10 prior lines. We are now taking the signals we have seen and their very good safety profile into an earlier line of platinum-sensitive patients, ovarian cancer, second, third line. The goal of the study is really to see, can COM701 in monotherapy, mediate significant monotherapy activity and prolong the progression-free survival of these patients who have no standard of care? These patients have a huge unmet need. In the second, third line, they receive the platinum. There is no treatment approved to maintain them from becoming platinum-resistant, and this is the goal of the MAIA-ovarian study.
Yeah. We expect interim data readout in first quarter 2027, correct?
Absolutely. We expect to have the interim analysis, which will include the meaningful data progression-free survival and any other clinical signals, by Q1 2027, yes.
Yeah. If the data is strongly positive, does Compugen take the leap into building its own late-stage clinical infrastructure, or does that milestone simply trigger another out-licensing event?
Yeah, it's a good question. It will depend on the actual data, magnitude of effect, whether we will think that, remember, again, there's a huge unmet need. Definitely thinking about this exact patient population and how we can go to registration fast will be the first priority. Either we can do it alone or maybe we think that the best thing would be to partner to move fast and aggressive into phase III. In addition, we're talking about an adaptive trial design, we can also add additional arms, to continue and explore the possibilities. For example, we can combine COM701 after showing the monotherapy with bevacizumab. We can combine it with ADCs and then open also more opportunities going earlier, later in ovarian cancer, and also into other indications.
Yes, especially taking into account it's a very competitive space right now.
It's a very competitive space, there are very few non-toxic agents which are fighting for that space. Patients after six cycles of platinum, some of them may want to have a bit of a easier drug than an ADC, which is eventually toxic. The pharma companies, which are fighting between themselves with the different ADCs, also want a way to differentiate. Having a combination partner which is combinable and ideally, again, the durability we have seen in the last line will translate here as well, could also be a matter of priority to differentiate in the competition for pharma companies.
Yeah. Makes complete sense. You've secured impressive validation with your AstraZeneca and Gilead partnerships, bringing in substantial milestone potential. Is the monetization strategy to continue, offload these assets early? Do you intend to keep full commercialization right, for future discovery assets beyond what you have currently in the pipeline?
I think it's really going to be program dependent. First of all, I think with the financial stability we have now after the monetization of small portion of the AstraZeneca royalty last year, we now have cash in the $29, we have financial stability, and this gives us the freedom to make our choices, because we don't have to rush and out-license necessarily to maintain the company alive. If we think for a specific program, if we generate additional data in a phase II or maybe even a phase III study, if it's a small and focused one, we will bring most of the value for ourself and for the shareholders. We'll do so.
If we think that we need now a very aggressive clinical program, and we need a pharma partner to move aggressively in multiple fronts, and this will be the path that will bring most of the value to the company and our shareholders, this will be the path. Eventually, it's going to be program dependent.
Yeah. To summarize this part, where does your platform create the sharpest shift in return on investment? Is there financial and operational leverage found in accelerating early target discovery, or is it at de-risking downstream clinical decisions? Where is it at?
I think, first of all, the platform is focusing on bringing new biology, bringing new targets. As mentioned before, pharma companies tend to herd, also biotechs, to herd around the same target, same biology, and we're bringing new biology that could really enable more and new options for patients.
Also, as discussed, the rigorous process that we are doing, and this is yet to be proven, but I think this is ongoing, should enable, bring us not only targets with new biology, but also targets which have more probability of success. We discussed before the COM701 data, and we have the readout by Q1 2027. We didn't discuss, for example, TIGIT. TIGIT is an example of target we identified long time ago. Definitely multiple failures for TIGIT. I think this is a case of a target that is definitely clinically active. It's clear that TIGIT is active.
We've seen it in multiple trials. Some of the initial assumptions that some of the drug makers had on TIGIT maybe overestimated the activity, maybe didn't choose, in this case, the right format. In this case, for example, I think that with the right format, with the bispecific antibody of AstraZeneca, they will leverage also the TIGIT biology that we initially identified. We have 11 phase III trials ongoing. It was interesting to hear AstraZeneca's management in the recent ASCO event, talking about the way they see the evolving data for rilvegostomig. They talked about stabilization of the responses. They showed in ASCO durable responses in early trials. They talk about combinability.
I think that also TIGIT, which is a target that we identified and had definitely a roller coaster of ups and downs, I think that AstraZeneca, with their way of developing it, the clinical strategy, the bispecific, will also eventually make the most potential out of this TIGIT computational discovery.
Can you just highlight the differences of your TIGIT asset compared to the prior failures, just to clarify to people listening in?
Absolutely. It's very important. Again, this is an example in which doing what we think is the wrong drug format could make a difference. Most of the initial developers who developed TIGIT had an Fc active. I would not go through all that biology, but that was, we think, was not the right format. It caused safety issues, it had challenges to combine, and eventually, it was difficult to keep patients for a long time on the study because of the safety.
Even maybe more importantly is the fact that AstraZeneca are the only ones who are developing TIGIT as a bispecific for PD-1 and TIGIT. This has some mechanistic advantages that we see for other bispecifics as well that could be more active, and then some evidence is for that, could be more active than PD-1/TIGIT combinations that others have done.
This also allows a different clinical strategy, much easier to combine, less burden of showing contribution of components of the bispecific. I think it's the format of the antibody, the clinical design strategy, and the combination around all these 11 phase III trials, most of them, or many of them, with ADCs, that will put AstraZeneca's TIGIT antibody, which is partnered from us, obviously, in a different position than the other TIGIT will.
Yes. Thank you for that. As we look into the second half 2026 and beyond, what are the most important clinical and strategic milestones investors should be watching?
First, obviously, is our own asset, COM701, the MAIA study. By Q1 2027, we're going to have the meaningful data to show if COM701 can really drive monotherapy activity in this population of ovarian cancer. rilvegostomig, the expectation from AstraZeneca for phase III readout, which is going to be the meaningful one, is after 2027, but we see all the time accumulation of data, like in the recent ASCO, showing again the durability, the safety, the potential of different combinations.
I think with the accumulation of data, it will show eventually that this molecule of TIGIT is doing something else. Then, obviously, the COM503, called now GS-0321, that we licensed to Gilead, we're already in the clinic for more than a year now, and the progress for phase I, we don't have yet disclosure for exactly when, typically when working with a pharma company.
Definitely keep an eye for the COMP903 readouts and the early pipeline. The computational discovery platform, we have financial stability. We have a validated engine that we are continue every day to leverage and bring more assets. Along the coming year, again, we don't have specific guidelines, but definitely, we will report on the early pipeline when time will come.
Great. Looking forward. To close up, we'll do a rapid-fire questions section. Let's get going. Most overhyped claim in AI drug discovery today?
I think that people that say that AI is going to dramatically increase clinical success rate and cure any disease are still a bit early. Things are moving fast, difficult to make predictions, but for now, human system is too complex to being solved with the push of a button.
Most underappreciated bottleneck?
The clinical development. The community must do something, especially in the U.S., but also in general, about making trials cheaper, faster. We have great targets. We need to test them. We need to see in patients if it works or not. For that, we need a better system for faster and more cost-effective clinical development process.
Hopefully AI can help with that.
Absolutely.
What is the one metric investors should focus on that actually captures platform value?
Eventually, it's looking at the totality of the assets that we have and that we will have and we will disclose, our ability to open new target space, new mechanism, to bring additional BD activities, to bring more clinical validation for our internal and partnered assets.
One proof point investors should demand over the next two to three years?
Eventually, the proof is in the pudding. I can sit here and tell about the processes we are doing and the rigorous validation. Eventually, an asset should show success in clinical translation. This is what we work for. This is what the investors should wait for.
Great. Well, thank you for the thoughtful discussion. What you've built at Compugen shows that computational biology isn't about shortcuts. It's about uncovering biology that was invisible to traditional methods. We're looking forward to the future updates. Thank you, everyone, for joining us today. Enjoy the rest of the upcoming sessions.
Thank you, Danya.