Good morning or good evening, everyone, and welcome to InSilico 2025 full year financial results conference call. I'm Leah Liu, the Head of Capital Markets at InSilico Medicine. Presenting to you today are Dr. Alex Zhavoronkov, our Founder and CBO, Dr. Ren Feng, our CEO and CSO, and Dr. Alex Aliper, the President. Before we get started, please be advised that today's conference call is being recorded. I like to remind you that the conference call contains forward-looking statements related to the likely future developments in the business of the company and its subsidiaries, such as expected future events, business prospects, or financial performance. The words expected, anticipate, continue, estimate, objective, ongoing, may, will, project, should, believe, plans, intent, vision, schedule, and similar expressions are intended to identify such forward-looking statements.
These statements are based on assumptions, analysis made by the company at the time of this conference call in light of its experience and its perception of historical trends, current conditions, expected future developments. However, actual results, performance, and financial conditions may differ materially from the company's expectations. The inclusion of such forward-looking statements should not be regarded as representations by the board or company that the plans and objectives will be achieved. Investors should not place undue reliance on such statements. Okay, now I will turn over the call to Alex to provide you with more details on our business updates and the 2025 full-year results. Alex, please.
Thank you, Leah. Hi, everybody. Welcome to our first earnings call after being listed. Next slide. Today, I wanted to ensure that I explain a little bit the company strategy, because in the AI drug discovery world, I haven't seen anyone properly present the objectives for an AI drug discovery company. So far, there are no clear performance metrics on how to evaluate an AI-powered drug discovery company, AI drug discovery algorithm, or any platform or system there is. I would like to explain our philosophy, how we approach drug discovery, and also AI, and how we are disrupting the industry. Also, why we believe that we can remain very productive and continue delivering on our promise of AI-powered drug discovery. First of all, on this slide, you can see the different stages of drug discovery and development.
That's coming from a very classical paper by Steve Paul in Nature Reviews Drug Discovery in 2010, "How to Improve R&D Productivity." We only added the target discovery and target disease hypothesis section here. On the top, you see probabilities of success, number of years, and millions of dollars. Since then, if you actually account for failures, the costs have increased and some of the timelines have not shrunk. We will use the 2010 estimates, even though nowadays, due to Eroom's law, the kind of opposite of Moore's law, we see some of those numbers increasing. Where I believe AI gives you the most advantage is zero to developmental candidate. Nowadays, there is a lot of talk about when are we going to see approvals from AI-powered drug discovery. I'm pretty sure there will be soon.
As a matter of fact, some of the drugs that have been approved in the past used machine learning and kind of primitive forms of AI already. It's kind of a sliding window always. Where we need to measure productivity is zero to PCC, one step before the regulated area. Next slide. Here I'd like to kind of represent it in the form of a diagram where we have speed racing. Imagine that Tesla is competing with a Ferrari and Lamborghini. Where is it going to compete? Well, if we are talking about the highway, you can actually talk about the acceleration from zero to 100, for example. That's where you can significantly accelerate, you can improve, you can add additional tools to your car.
Afterward, after you reach the highway, you have to move with the speed of traffic because there are traffic signs and also traffic cops. You don't want to break the limits. As a matter of fact, if you break, you can completely derail yourself and the company. AI can help you only very little after you reach IND. Actually very little after you reach PCC, because IND-enabling studies is the same 28-day GLP toxicity studies in two species that we have in our preclinical candidate package. It's just we do it in non-GLP manner. After that, it's actually better to hand over to minivans and trucks that have more resources and actually have the agility to transport more on a highway. That's our philosophy. We compete in this area between zero and PCC. Next slide.
I tried to make an attempt to actually create a formula for how to evaluate AI R&D productivity. You need to look at the number of preclinical candidates. Here you see the formula for measuring AI productivity, where we look at the number of developmental candidates that the company can develop, for example, every year in a time frame, divided by the cost of those developmental candidates and time. Preferably you need to include novelty on both target and molecule side, probability of success, basically looking at how you transition from one stage to another, or if you out-license, the value of the out-licensing deals. I hope that this formula is going to evolve, and please feel free to extend it or use it for other companies. In the early stages, investors and pharma partners alike may ask, "Okay, what's your compute power?
What's your raw data volume, teams and talents, algorithms and patents?" Once your algorithm starts churning out drugs, all that matters is going from zero to PCC in as fast as possible, as cheaply as possible, with the highest probability of success as possible. As you iterate, you increase the target novelty and small molecule novelty, and then you try to scale. AI is all about scaling. AI drug discovery is also about scaling and learning. Our InSilico Medicine stats are now 12 to 18 months typical for zero to PCC, with reasonably high novelty. Some of the molecules that we've designed are targeting targets that have never been in a clinic before, absolutely novel, like TNIK, QPCTL, and PHD1/2/4, the indication with a gut-restricted molecule. That's a truly amazing feat of molecular engineering and biology.
So far, over the past five years since we started drug discovery ourself, thanks to Dr. Ren. He is the true hero behind the in silico revolution, so to speak. We've managed to develop 28 developmental candidates, launched 12 of them into clinical stage, made 10 BD deals, and one phase IIa completed. Next slide. Every time you launch a PCC, you iterate and you learn. At InSilico, we try to learn from one company, which is SpaceX. When you ask SpaceX, "What's your secret sauce?" It is the ability to consistently launch to orbit with high probability of success at low cost. That's what we are trying to do with PCCs. That's why it's very important to measure the number of PCCs a year and constantly iterate. Every time you iterate, your AI improves.
Our challenge is that AI changes every six months. In the case that we don't maintain technology leadership, we're going to lose. We realized that the most important data in AI-powered drug discovery is actually the data coming from zero to PCC. Every time you launch a PCC, you generate around 1,000 benchmarks, and those benchmarks could be used to improve your AI and prospectively test your AI. That's why we're consistently improving on the AI platform and consistently launching new software tools. Our software business is actually not about revenue. Of course, we generate healthy revenue from that business, but it's more about validating many different AI models at scale. Instead of the need to synthesize and test ourself, we put the model out for the entire community to use. They use it, we get more confidence.
We push for a PCC, learn, iterate, make it better for the entire community. Every community player contributes to the success of the platform as well. That's why it's important to scale. Next slide. Here is another very important finding for all of you on the call, is that nowadays, what I found is that we are technically competing not against If you were to name a single competitor in the West that has very clear zero-to-PCC benchmarks across many PCCs, I wouldn't be able to name any. Many of the traditional Chinese pharmaceutical companies managed to reinvent themselves as innovative drug hunters, and some of them have hundreds of innovative drug programs with the ability to scale. Nowadays, we're essentially competing in China. We call it the China gym.
We compete on speed, cost, the molecular novelty, but also on the novelty of the target. AI gives us the ability to actually bet on more novel stuff, where you've got higher risk. We believe that we have higher confidence in both the molecule and the target. Our kind of differentiation is actually on the novelty side. We try to innovate on biology and chemistry at the same time. Next slide. Our company's strategy, our real secret sauce, is that we specialize in highly novel biology and chemistry, improve our AI, and now we actually decided that we cannot compete on scaling in foundation models with the top players in this field like the OpenAIs, the Googles, and the Anthropic of this world. Of course, in the Eastern world, that would be the Tencents and Alibabas.
Instead of trying to battle in this gladiatorial multi-billion dollar industry where you burn cash so quickly, and the models become obsolete, we decided to develop smaller AI models that can act as teachers to teach large foundation models AI for drug discovery. Actually now we are shifting our customer base to frontier AI model developers that really want to give you the comfort to access frontier AI-powered drug discovery capabilities from the comfort of your prompt window. That actually gives me the ability to deploy those tools at scale within my organization because the adoption gap between frontier AI and the bench work is actually pretty substantial. It takes people time to learn new AI skills if you deploy them as specialized software.
If you deploy them in a prompt window, they actually use it very well, very easily. We now are constantly improving and trying to build models that understand fundamental human and animal biology in time. We started working on life models together with foundation model developers, and you've seen some of the papers probably that just came out. We've been doing it for the entire 2025, and now we're trying to identify the next kind of GLP-1. Our strategy on the drug side also has evolved, and we are looking for new pipeline in the product targets. Next slide. The pipeline in the product targets is something that is supported by aging biology. This is kind of one of the unspoken secrets about InSilico.
We do a lot of work in longitudinal data analysis in humans and animals trying to identify common biological patterns that unite all of us. We really love the story of GLP-1s where the drug was developed for an indication, expanded into broader disease biology, and then became a lifestyle drug. I don't deny it. I'm on tirzepatide, so I'm on Mounjaro. I think it's one of the best drugs ever invented by man. By the way, don't buy, sell, or take any drugs based on what I say. This drug makes me extremely happy because I don't want to eat and still feel happy.
The idea for us is to find targets that work in similar fashion that affect everybody, and could in theory, be repurposed from a disease to a basic fundamental biology, preferably biology of aging, that will help all of us live longer, more productive lives, and of course, tackle more indications, making them much more profitable. Next slide. If you look at our current pipeline and look at where we purpose some of those targets. In oncology, we've identified this target and we're the first ones to bring it to clinic and decided to purpose it to fibrosis.
If you read the seminal papers that we've published around this target in Nature Biotechnology and other journals, you'll see that we actually identified it using a combination of AI and aging research, with a fibrotic angle, identified the indication where we could potentially outperform the standard of care, but the potential of the target is vast in multiple disease areas. As we progress through the clinical trial, we uncover additional areas of fundamental human biology that we can pursue with this target. We also have the second and the third generation of the molecule for this target as well. We have a pipeline of drugs that are progressing in time, helping unlock the full potential of this target. In other areas like QPCTL, same story.
We are going after IO, but in reality, this target, with the molecule at low dose, has the potential to work in cardiovascular diseases. IRAK3, we've already partnered our CNS penetrant molecule with a subsidiary of Fosun Pharma in 2025. Beautiful drug, beautiful high potential pipeline and product target. Same with TEAD. The idea is to create more of that, currently in our early stages, we have many more. Next slide. I'll walk briefly over all of the business highlights because mid-year last year, I became the Chief Business Officer of the company. As you know, we've done reasonably well on BD recently, you should expect greatness in 2026. On the software side, our revenue has increased, but again, as I've mentioned, we do not care about revenue on the software side as much as we care about the usage of the platform.
I would have loved to open source most of what we have. Just to get more community usage of many of our tools. We actually open source quite a bit. What we've noticed is that if you open source, people use it less, because the specialist users really need to understand the value of the product, evaluate it, go through the budget reviews, then use it at the community level. We do a lot of work also properly pricing the products, not trying to make it too expensive. Some of the models we open source. Some of the models, like Nach01, it's our natural and chemical languages foundation model, that's our true venture into multimodality.
One model can perform many tasks, that was kind of the premise of our MMAI Gym, where we realized that foundation models can now also perform some of those specialist tasks in drug discovery. We also put it on hyperscaler platforms, AWS and Microsoft, scale through there. Next slide. We've also managed to significantly progress our pipeline. Six PCCs were nominated, the year was very productive on the PCC front, giving us the ammo to go and partner with the pharmaceutical companies. Here, some of those targets are extremely popular, we actually had to decrease the novelty of the targets from time to time in order to satisfy the appetite of big pharmaceutical companies, that actually very often shy away from target novelty. They like to see clinical proof of concept most of the time.
There, we focus on the feats of molecular engineering, where we try to achieve absolutely unique molecular properties that are not possible with traditional approaches. You see that some of those PCCs are being partnered very successfully. We also have done pipelines on the clinical development front. We have multiple wholly owned pipelines, and Dr. Ren is going to cover it at length. That's why I'm going to be very brief. Rentosertib in 2025 progressed into phase IIa complete with the results published in "Nature Medicine." We're planning to go forward with this drug. I cannot comment too much about this yet, but there is a plan. PHD1/2 got restricted. We've initiated phase IIa trial in China and dosed our first patient. MAT2A, highly promising target. You've seen many business development activities around this target.
Our job is to demonstrate that it's safe. That will most likely lead to a partnership deal. We've completed already the first-in-person dosing in a global multi-center phase I trial. This is a hot target. Every pharmaceutical company that has PRMT5 is considering MAT2A. It's a very important companion for PRMT5, also for TROP2, and even for KRAS. TEAD. That's an extremely important target for us for BD. The idea is to demonstrate safety, but also hopefully some efficacy. We've completed the first inpatient dosing in a global multi-center phase I trial. We've also partnered multiple assets, and while we do not comment on our partner pipeline progress, it's been amazing. We started getting milestones from out-licensed and partnered assets. Of course, we have extended to new modalities. We're going after alpha-helical peptides and ADCs, nanobodies, antibodies, siRNA/PROTAC.
We have developed the software called Generative Biologics that a lot of people use and love. We are, of course, utilizing those same capabilities to concurrently evaluate those methods. I wouldn't be surprised if you see some biologics partnerships from our side. Next slide. On the business development front, we've managed to sign over $1.3 billion in total deal value in 2025, an increase of 39%. We've entered partnerships with Eli Lilly. That's a $100 million deal last year. That was actually a kind of a stepping stone after our partnership and software. I think Lilly is one of the top AI players in the world because they took most of our AI very early. Now many are catching up. We've partnered with Servier and announced the $888 million deal. Tenacia and achieved multiple milestones with [TheraSi], Eisai, and others.
Also achieved milestones with Menarini and TaiGen. We've entered multiple research collaborations, strategic collaborations. Now we've even announced our first partnership with a foundation model developer, with a very famous genius, Ramin, who is the CEO of Liquid AI. You probably remember some of the seminal papers in foundation modeling out of MIT in Boston, demonstrating that you can pack a lot of intelligence in a very small foundation model that is actually non-transformer based. These guys have progressed. Now we realize that those smaller models can perform at state-of-the-art performance after being reinforced in our MMAI Gym. We really hope that foundation model developers are going to notice that, read the papers. We just got a paper accepted for ICLR demonstrating the capabilities of MMAI Gym, and we believe that this could be a very significant market going forward. Next slide.
We've just partnered with Eli Lilly. You probably have seen the announcement over the weekend. This deal carries $150 million upfront, and about $2.75 billion in total deal value. It doesn't touch any of our clinical assets. This is early preclinical asset deal, and also collaboration. This is a testament to the high quality and novelty of these small molecules that are being generated by our Pharma.AI platform, and also gives you much more confidence in InSilico as a stable partner for research collaborations. We're also deeply proud to be a member of the Lilly Gateway Labs, which is one of the really top ecosystems in the world, I would say the top, that we've seen so far, that allows you to move R&D even faster. Next slide.
Now I'm going to pass the word to Alex Aliper, who is going to briefly go over AI platform. I've taken so much time. I'm sorry. I wanted to give you the gist of our strategy as well. Alex, please take the floor.
Thank you, Alex, and hi, everybody. At InSilico, we have been and remain committed to pushing the boundaries of our AI platform and advancing it forward. Our recent addition to AI platform is called Science MMAI Gym. As Alex explained, it's a framework for post-training and boosting the intelligence of frontier foundation models, frontier LLMs in science. Currently, it covers six scientific domains: chemistry, biology, clinical, longevity, material science, and agriculture. We plan to gradually expand its utility as we go forward. It integrates millions of proprietary data samples and over 1,000 different benchmarks that helps us evaluate the performance of every model in a specific task. The Gym also sits on top of strong foundation of reasoning datasets. We have over 300 reasoning datasets used to fine-tune and post-train frontier models and also many curated databases.
The way the framework work is that we provide a frontier model as an input, you put it through the Gym. First, we evaluate the baseline performance. We see which tasks the model performs well and which tasks the model lags behind. Then we put it through a post-training protocol in the MMAI Gym that combines proprietary reasoning datasets and experimentally validated reward models to boost its intelligence and do SFT, RFT post-training to reach up to 10x increase in performance compared to the baseline. Next slide. We have shown that this model, and Gym specifically, works on diverse set of tasks and benchmarks. One particular application is in target discovery. We showed that after post-training of one of the open source models, in the Gym, we're able to dramatically increase its utility in uncovering therapeutic targets.
Another application that we've just published, Alex referred to, is in collaboration with Liquid AI, where we showed that if we put their model, LFM 2.6 billion parameter model, through MMAI Gym, not only we outperform the frontier LLMs in the performance, but we're also achieving state-of-the-art performance among other models, among other specialist models. You can see the power of the Gym, where a single architecture can tackle many different modalities of data and domain-specific reasoning. It's capable pretty much of handling many different tasks across biology and chemistry within the drug discovery space very efficiently. Next slide. In terms of the business model for MMAI Gym, we tackle 2 main tracks. First track is a membership or licensing of MMAI Gym.
One way to license the Gym is provide us with your LLM model, we will do customized post-training of the foundation model and will provide the resulting model to the client. Second track is in providing teacher models or training data for post-training those frontier models. The second vertical of the business model is around commercialization of the resulting model that are coming out of MMAI Gym. Here we target 2 main tracks. One is annual licensing, where we provide the license to the resulting model to end user. The second track is pay-per-token licensing, where we put the models, we deploy them on the cloud and charge by token for every user's consumption of those models. It's a cloud consumption model for MMAI Gym models. Next slide. Another frontier that we're targeting from angle of AI platform is increasing efficiency of experimental validation.
Here, last year, we were proud to launch the next iteration of our automated robotics facilities, the lab called LifeStar 2. It's in full automated, AI-driven laboratory that accelerates drug discovery and development. It reaches different areas of drug discovery, including target discovery and validation, and also compound profiling, small molecule profiling within the same facility. It is supported by AI platform in the name of different modules such as PandaOmics and Chemistry42, which sit on top of it, and basically, we utilize LifeStar to seamlessly validate the AI platform hypothesis and also, in turn, make the AI platform smarter with every iteration. Next slide. In terms of the module capabilities, we have many different stations and utilities of the module. We have organoid and high-throughput safety profiling platform.
We have gene editing platform, HTS platform for high-throughput screening and validation, high-content imaging platform, and NGS multiomics platform for deep sequencing and gene expression data generation at scale. Next slide. In terms of utility for therapeutic areas, we're pretty much covering all our core target therapeutic areas, including oncology, metabolic disease, I&I, so autoimmune and inflammation, skin diseases, and longevity-specific assays, where we already demonstrated and published many of the case studies showcasing how we can integrate the AI platform, reasoning, and idea generation with seamless laboratory automation. Next slide. With that, let me turn over to Dr. Ren.
Thank you, Alex. Let's go to the next slide. I'm going to brief you with our strategy on the pipelines and also our current progress. For the strategy of pipelines, we have two strategies. One is so-called first-in-class. We use AI to discover novel targets and novel molecules get into a clinic. Second strategy is so-called best-in-class. For those programs we are pursuing the validated biology, but we develop using our AI to develop well-differentiated molecules to the clinic. On ratio of these two strategies, about 20% of internal programs are first-in-class and more than 80% are best-in-class. As Alex mentioned, we have 28 PCCs already till now, and that includes our own internal pipelines and also the collaboration pipelines. This is a highlight of what we have in the most advanced pipelines.
I want to give you updates on some key highlights. If you go to the next slide. Next slide. The first highlight is our TNIK program. This program is a first-in-class. We use AI to identify the novel target, TNIK, and also use AI to discover the novel molecule ISM001-055, also called Rentosertib. The targets, we have combined the literature and also our internal biology. We understand the targets regulating Wnt/TGF-beta/NF-κB/YAP pathway, which are all related to fibrosis. If you go to the next slide. We have completed the phase IIa trial in China. This is a comparison of our data with the available IPF data.
For all the commercial or approved IPF drugs, nintedanib, pirfenidone, and PDE4 inhibitors, their molecules in the phase III trial, they can only slow down the decrease of the patient's FVC, forced vital capacity, which is a golden standard for the clinical trial for IPF. You can see on the left-hand side, our molecule, it can revert. It can really improve the patient's FVC after three months of treatment. This is really remarkable, and if we can reproduce this result in the phase III trial, we can really bring through therapy for IPF patients. In the next slide. Based on this really impressive results, the CDE allow us to go direct to phase III without doing a phase IIb. You can see this is our phase III trial. We are going to enroll 320 patients with two groups.
One is a placebo, the other one is a 60 milligram qod dosing. We anticipate to have the first patient in the third quarter of this year, and the data will be expected in 2029. That's one highlight of our leading assets. If you go to the next slide, that's another highlight of our IBD program. The IBD is also use AI to identify prolyl hydroxylase domain 2 as a novel target for IBD, and we use our Chemistry42 tool to develop a novel molecule called Garutadustat. It can not only, it's an anti-inflammation, but also it drives the intestinal barrier protection. It's a dual mechanism of molecule, which is very rare comparing to all the available drugs for IBD and also the current clinical trials. If you go to the next slide.
Finished 2 phase I trials, one in Australia, the other one in China. The molecule shows very good tolerability, up to 1,000 mg. This is our phase II-A trial in China. We have 4 cohorts, placebo, 200 mg, 400 mg, and 800 mg QD. We have achieved the 1st patient in December last year. We are now very actively recruiting patients, and the top-line data is expected by next year. By next year, we will have the safety and also some preliminary efficacy result. Those are the 2 phase II trials. If you go to the next slide, we also have 2 oncology programs. I want to highlight one this time. We call the best-in-class pan-TEAD inhibitor, targeting the Hippo pathway. We are now the 2nd of the global for the pan-TEAD inhibitor in a clinic.
The 1st one is Vivace, but the molecule have very long half-life in human, 12-15 days. Because of that, their molecule have the accumulation in the clinical trial, so that they observe very severe side effects, including proteinuria and others. For us, our molecule has the half-life of 21 hours in human so that we don't observe those accumulation problem. For the Vivace's molecule, they have to take holiday dosing. For our molecule, we can do it once daily. If you go to next slide. Currently we are in the dose escalation stage. Even though it's a dose escalation, we did observe confirmed PRs. We observed at least 2 confirmed PRs already in the dose escalation, with the mesothelioma patients with already 4-5 prior lines of therapy. Not only that, we also observe very favorable safety data.
For our molecule, we didn't observe all the side effects are limited to grade 1, and also the proteinuria problem, which was observed very frequently in the Vivace's trial, but it's very rare in our trial. This is not only for mesothelioma. The TEAD can provide a basic as a backbone to combine with EGFR inhibitors, ADCs, KRAS inhibitors, et cetera. We are expecting to have some data available in 2nd half of this year, and we are going to start with part 2 dose-only in end of this year. If you go to next slide. In addition to that, we also have a lot of early programs. 1 is Pan-KRAS. The Pan-KRAS has a very selective comparing to the Pan-RAS. Our selectivity of the KRAS versus HRAS and NRAS is greater than 100-fold.
For that reason, we should observe a better safety profile. On the right-hand side, you can see that our molecule also showed very good comparable or even better efficacy comparing to the Revolution molecule, RMC-6236. Next slide, please. Our oncology strategy is we use a combination, right? We have the MAT2A in clinical trial. We also have PRMT5. If we combine MAT2A and PRMT5, it shows really good synergy. That's shown in the right-hand side bottom figure. Our Pan-KRAS inhibitor can also combine with SHP2. SHP2 is currently in the early enabling stage. For our oncology programs, we are doing a lot of combinations. If you go to the next slide. Oncology, I talk about fibrosis oncology. Now I want to share with you some of the early programs.
For us, we are targeting the program that can not only reduce fat mass, but also can preserve muscle, right? We have two programs. One is a GIPR antagonist. The other one is a bias APJ agonist. You can see on the right-hand side of the animal models that both molecules when it combine with semaglutide, it can further reduce the body weight by 10%-15%. At the same time, it can increase the lean mass to fat mass ratio. That means our molecule, when combined with semaglutide, can really preserve the muscle loss. That can also decrease the fat mass. That's our metabolic strategy. If you go to the next slide, we also have CNS programs. Targeting, focusing on pain, we have the best-in-class Nav1.8 inhibitor.
Our molecule shows better efficacy comparing to the Vertex VX-548. Also our molecule shows great solubility so that we can do the injection or infusion, IV dosing. That is our key differentiation. Another one is, we use AI, identify a novel target. We call it Target Z. Target Z is targeting different amyloid of Nav1.8. In the preclinical studies, we've shown that our molecule, it shows very good efficacy in multiple pain models. It shows better efficacy comparing with the SOCs pregabalin. We are very excited on that novel target, Target Z. If you go to the next slides. Alex has mentioned, this is our inflammation and autoimmune programs. We have the best-in-class STAT6, it shows much better potency comparing even to dupilumab. Also we have the VAV1, small degraders, protein degraders, and also we have IRAK4.
We use our protein degradation platform, developed quite a few of the anti-inflammation autoimmune programs. If you go to the next slides. As mentioned, the One Drug Pipeline, right? Our NOS3 inhibitor is the one typical One Drug Pipeline program. It can show efficacy in the MPTP model, which is a PD model. It also shows efficacy in the paracetamol hepatotoxicity model. Also, I want to mention that the foot ulcer model, our molecule shows great efficacy in that, which is a huge unmet medical need. Also it shows great activity in the dry AMD model. This molecule, this program is, we call it One Drug Pipeline, which means one molecule can cure a lot of diseases.
In the next slides, I also want to share with you another excitement is our AI-identified first-in-class program as a next generation of One Drug Pipeline. This target, we haven't disclosed the target yet, but we have already developed the molecule. You can show that this molecule also shows good efficacy in the Parkinson's disease model. On the left-hand side bottom figure, you can show that this molecule, when it combine with semaglutide, it further reduce the body weight by more than 10%. More importantly is once you retract the semaglutide, our molecule, the body weight rebounds. If you retract semaglutide but continue to take our molecule, the body weight remains, right? That means our molecule can prevent the body weight rebound, if you retract the semaglutide. This is very exciting.
Also on the right-hand side, the molecule shows the best result in the dry AMD model and also the mesothelioma model. This is a novel target, and we developed the molecule for this target for the One Drug Pipeline purpose. This will be the really big future program for us. I will stop there, and if you go to the next slide. I will stop there and pass the word to Leah.
Leah, we cannot hear you.
Sorry. Can you hear me now?
Of course.
Thank you. Can you go to the next slide, please? Yes. We run as a company a multi-pronged revenue generation business model. We have three main business models. One is our drug discovery and pipeline development. The second is our software solutions, and the third is our non-pharma sector business. The first one is our main focus, the drug line discovery and pipeline development. Out of that, we have generated so far, as Alex and Dr. Ren mentioned, 28 pipeline assets already. We have done more than 10 out-licensings and collaboration deals with a total contract value of, as of the annual report, $4.6 billion. But really after the Lilly deal, as we spoke about earlier, we have now $7.4 billion in total contract value. In 2025, we collaborated with more than 75 customers for the drug discovery business.
For the software solution part, as Alex and Alex mentioned before, we have our platform and we sell software to our business partners as a door opening for them to have deeper relationships with us. For that, we also came out with a new product, that is the MMAI Gym. Finally, for the non-discovery related non-pharma product, we actually collaborate with non-pharma sector leaders in their sector. For example, we have done a partnership in the past with Syngenta on agriculture and pesticides. We have collaborated with a major global energy company on carbon capture material and other new materials. Also we have developed next generation nutraceuticals and cosmetics products, all of which are generated by our proprietary generative AI platform. Next slide, please. As you can see, this again, is based on, at the cutoff of our annual report.
Now the total contract value is $4.7 billion, of which we have recognized $205 million in revenue by the cutoff of the report. We have more than actually $4.4 billion, or rather, we have more than $6.4 billion over in the future. Most of our business is actually based on our future gathering of our, receiving rather, of our milestone revenue and our profit-sharing revenue. We have more than 12 partnerships at this moment now. Next slide, please. This gives some details about our business and our milestones that we have achieved. These are the partnerships that we've formed in the past, and it shows you some of the major partners that we've worked with. Again, we have more than 75 partners in the entire system. Next slide, please. Our business, as I mentioned, not only is in our drug discovery business.
Our drug discovery business is the core validation of our AI platform, and it's the way that we have been able to improve and continuously build on our platform with validated data, as Alex mentioned before. However, our platform can be used for multiple different sectors. In going into the future, we will look for more collaborations in those sectors with major leaders in their fields. Next slide, please. Next, I will go over the financials with everyone. As Alex mentioned, we were the largest biotech IPO in Hong Kong in 2025, and subsequently, we were included in the Stock Connect on March 9th this year. Next slide, please. For the financials, our revenue declined in 2025 by $29.6 million to $56.2 million. The decline was due to the recognition of revenue generation and deal negotiation.
As you can see for now, we've already closed more than five deals as of this year in 2026. The revenue from software solutions increased by 23.8%, while the customer base only increased by a bit over 18%, which means that the per customer revenue actually increased from the software solutions product. The cost of revenue increased slightly by $2.1 million to $10.4 million, primarily attributable to the revenue composition of an increase in co-development revenue. Gross profit margin decreased slightly by similar reasons. Selling and marketing expense increased slightly by 0.8% due to the increase in share-based compensation expense. R&D expense decreased notably by $10.5 million. This is due to the enhanced efficiency for developing internal pipelines. Administrative expenses remained flat. Other income decreased by $2.6 million to $8 million, primarily due to the decrease in bank interest and income and subsidiary income.
The loss increased due to the change in fair value of financial liabilities as recorded a loss of $296.7 million. This is due to the conversion of preferred shares issued in previous financing series into ordinary shares based on the share price upon our listing. This is a common practice for newly listed company and is a financial basis non-cash item. The loss for the year increased by $335 million to $352 million due to similar reasons of mostly related to the fair value conversion. The loss for the year on non-IFRS measures increased by $21.2 million due to the decrease in revenue, which was partially offset by the decrease in R&D expenses. Our cash balance remains healthy and very strong at $393.3 Million at the time of the financial reporting.
That number has gone up dramatically now again, with the closing of several deals this year, including the Lilly deal, as just mentioned. Next, please. I will turn the mic back to Dr. Ren for closing remarks.
In summary, we will have more key milestones across both platform and therapeutic pipelines this year. In the first half of this year, we will have the preliminary safety efficacy biomarker data for the phase I trial of our TNIK inhibitor. We are having the inhalable Rentosertib and the approval for the IPF in China. For the NOS3, we are going to have the first subject in Australia trial, and also we are going to have the IND approval for phase I trial in China. The second half of the year, our phase III of Rentosertib for IPF will be started, and have the first patient in our NOS3, first subject in trial in China. Also we are going to have the MAT2A and TEAD for the close of the dose escalation for the phase I trial globally.
Also, we are going to have multiple new PCCs, and multiple INDs. In addition to that, we are going to have more BD deals to be announced in both first half and second half of the year, in addition to the Lilly deal. For the AI platform, we are continue to update, upgrade our Pharma.AI. I want to mention that we're going to continue to execute MMAI Gym collaborations. The Liquid AI is the first one, and we are going to have more of the MMAI Gym collaboration in the future. That will conclude our presentation, and we'd love to answer any question if you have.
Thank you, Dr. Ren. Now I will open it up to Q&A. If you'd like to ask a question, please click the Raise Your Hand button at the bottom of your window to queue up for live questions, or click on the Q&A button at the bottom of your Zoom window and type your question into the text box. Please include your name and organization so we know who the question is from. When I read your name, please unmute your microphone to ask the question. Thank you. The first question comes from Jack Lin, Morgan Stanley. Hi, Jack. Can you please unmute your phone? Jack?
Hello. Hi. Can you hear me?
Yes, we can hear you now.
Hi. Thanks again for taking my question, and congratulations again on the collaboration with Eli Lilly. This is Jack Lin from Morgan Stanley. I just have two quick questions for the company. First one is about kind of the PCC cadence and system-level learning, right? I was wondering if the company could update us a bit more on kind of the year-to-date progress in PCC nominations, and how you might see the cadence of PCC generation evolving as the discovery loop becomes more model-driven and more efficient. Also, in particular, how do we think about the kind of timeline in terms of further target discovery and disclosure timing, given that today we've shared a lot of developing pipelines and undisclosed assets. Just curious in terms of how we could expect the pipelines to be gradually disclosed to investors in the market.
The second question is about kind of foundational models and our kind of our overall platform setup, right? I think you shared a bit on the kind of the MMAI Gym. If I understand it correctly, it provides an environment to train and optimize foundational models against clinically relevant benchmark. I was wondering if the company could elaborate more on what role do you ultimately want foundational models to play within kind of the broader InSilico platform, and where they might have the most incremental leverage to kind of our existing already a very expansive suite of application modules. These two questions. Thank you.
Dr. Ren, maybe you go first on the pipeline, and I'll comment on the MMAI Gym.
For the PCC, since we got IPO, we've got more resources to support our early discovery programs. We are expecting much more PCCs comparing to the previous couple years. If you want me to say a number, we are targeting 8 to 10 PCCs every year since 2026. If you want to understand how many of those pieces will be progressed into the clinic by ourselves, I want to say that all depends on how many PDs we are going to do, right? Oftentimes, our partners will buy our molecules as a PCC or the early enabling stage. That if we sell more programs, that means we have less internal programs will push into the clinic.
We really want to push our first asset, TNIK inhibitor for IPF, into the phase III trial, and we want to achieve the clinical POC, get this compound approved. That will really complete our circle of the technology to drug discovery and also to the clinic and to the approval.
Yeah. Thank you, Jack, for the question. Love your work, by the way. For everybody on the call, please do follow the research. Also for everybody on the call, please do read the recent article that was just published on STAT on our deal with Eli Lilly, and it also talks a little bit about our platform. On the MMAI Gym, I just want to explain the entire framework there. About this time this last year, or maybe just a little bit earlier, we realized that foundation models could be trained to perform very complex drug discovery tasks. Not all, but many. That was our kind of our most revealing discovery last year.
I realized that we could turn many of the specialist models that we have into reasoning data, basically trying to explain what is happening to the model in text, in reasoning traces, and fine-tune foundation models to actually perform the same tasks almost at the same level. Think of it as a blind samurai We kind of got some inspiration from Zatoichi, if you know the movie, and also the story, where very often you lose a sensory system, like you cannot see, but you can still perform a very important function like samurai functions without seeing, by relying on different senses. We can describe what is happening to the model, and it would provide you with the expected result. That was one thing.
Another revelation was that if you later utilize a specialist model to reinforce the foundation model using reinforcement learning to basically an adversarial setting via reward and punishment, a foundation model achieves close to super intelligence in multiple tasks. It can perform multiple tasks simultaneously. At the same time, you can also measure some of the reasoning capabilities of the model. As you reinforce it in more and more AI drug discovery tasks, it actually starts reasoning between the tasks and acquires the ability to do zero-shot. Performing the tasks it has never seen before. That was a major revelation, we decided to actually even pause many of our specialist model development and focus on MMAI Gym.
This model, this approach took us maybe seven months to master, where we had to turn all of our smaller models, we have over 800 smaller models that perform close to or at state-of-the-art performance or surpass state-of-the-art performance. We think that we are state-of-the-art in multiple drug discovery tasks. You take the foundation model, for example, like an open source Qwen, we like to use it for demonstrations, give it to this reinforcement learning gym, where our smaller models act as teachers and teach whatever they know, whatever skills they acquired, to the foundation model. That's why it's called a gym. We also developed a proprietary set of benchmarks that nobody else has, only InSilico. That's actually the kind of purpose of those benchmarks.
Otherwise, if those foundation models see those benchmarks, they will hack the benchmarks and become state-of-the-art in the benchmark without actually learning the skill. We keep those benchmarks private. You can actually test the resulting foundation model in those benchmarks. In many of those benchmarks, those foundation models perform at state-of-the-art or above state-of-the-art. For us, it was a revelation. We actually believe that as time goes by and those foundation models become more capable in scale, also acquire different types of modalities, become highly multimodal, we will be able to achieve super intelligence in the pharmaceutical industry from the comfort of your prompt window. I'm not talking only about agentic AI.
Agentic AI is, of course, very important, kind of in the meantime, where you turn the foundation model into an agent by prompting it to do certain things, follow certain workflows. I think in the future, you'll be able to do the entire kind of prompt to drug. Basically prompt the foundation model to design a small molecule for a specific target or even for a biological process, and it will do it for you. It will take time, but we are going to get there. We published a paper together with Jie Shi, one of the top people in AI in the world. He currently leads AI drug discovery at Eli Lilly, I very often use him as a teacher because he understands kind of the future philosophy of AI drug discovery and taught us many directions of where to go in chemistry and biology.
We just published a paper, "From Prompt to Drug Towards Pharmaceutical Superintelligence." Please have a look at the paper describing this approach. We believe that you can monetize MMAI Gym. We already started doing so with the partnership with Liquid AI. Of course, we have multiple other partnerships with foundation model developers ongoing. For any investors on the call or analysts, please do query your foundation model developer if they have partnered with us, or what are their benchmarks in target discovery. Of course, the most important benchmark is going from zero to developmental candidate. So far, most of the foundation model developers, despite being around for several years, they have not resulted in a single PCC.
Some of the very overhyped AI companies with very popular AI tools on the market have made many claims and raised a lot of capital, but delivered zero PCCs and, of course, zero INDs. The ultimate test is a PCC. I think that in the future, you'll be able to actually deliver a PCC right from a foundation model with very little human intervention. That is our goal. In the meantime, we want to monetize that, and we think that the market of foundation models consuming our tools just from financially is larger than pharma. Pharma software market is actually genuinely small. You would be surprised. Many of the pharma companies prefer to develop software themselves. It's $ billions in software in AI pharma. It's a myth. Those kind of companies that just focus on software, they cannot scale too dramatically.
With the MMAI Gym and targeting the foundation model developers, we not only can scale, but also we make AI future friendly because at the end of the day, we need to enable more and more people to be able to discover drugs from the comfort of their prompt window, make it very simple, and then we just focus on operational efficiency. We are the SpaceX of drugs. Our job is to deliver zero to PCC and hopefully offload at that time to the pharmaceutical industry. In the meantime, we want to be a very significant player in AI. I think currently, no company that I know of has so many capabilities in drug discovery AI as we do, including smaller specialist models, but now also large foundation models that work with specialist models and use specialist models for training. MMAI Gym will be a very viable product.
It will also serve a very important role in the industry. As we all transition towards super intelligence, the world is never going to be the same. We want to be a very significant player in this process so that AI systems of the future remember us as a very significant player.
Got it. Thanks, Alex and doctor for the response. We'll definitely check out the super intelligence article and congratulations again from the collaboration. Thank you.
Thank you, Jack. In the interest of time, as we are already over time, we'll only take another question. The next question comes from Rui Han from CICC. Hi, Rui.
Hi, Leah. Can you hear me?
Yes, can hear you now. Thank you.
Thank you. Hi, this is Rui Han from CICC. Thanks for taking my questions and congratulations on solid set of results and the continued progress in your pipeline at BD. Actually, I have two questions. First question is about AI. We've seen two emerging technical framework in AI-driven drug discovery, like end-to-end foundation model platforms versus MOE systems. My question is about how do you think about your long-term architecture choice between these two approaches, especially in our early stage discovery? I'll stop here for the first question, I will have the second one. Thanks.
Sure. By the way, Leah, if we can allow some more time for other questions, that would be great. Thank you, Rui Han, for this wonderful question. I think that it actually doesn't matter what AI you use as long as you develop drugs. One of my favorite quotes in the world is actually a quote by one of the great leaders of the East that it doesn't matter whether your cat is black or white as long as it catches mice. As long as you can continuously deliver PCCs, and you can measure the R&D productivity in terms of the number of PCCs, cost per PCC, and the novelty of the PCC, and also potentially, like at the end of the day, we need to convert that into money and quality adjusted life years to the patients, to the real impact to the world.
You just need to measure zero to PCC. In the future, in terms of what is going to create more PCCs and higher quality PCCs, I think that foundation models. Currently we need to look at what's available today and what's available kind of in the near future. Currently, foundation models are very capable at reasoning, and we utilize them for program level reasoning, looking at the program from zero to approval. It's kind of like looking at human life from birth to death and reasoning within that, trying to identify the right path and also maybe to live longer and make more impact. We use foundation models for orchestration of the programs. We also even use them for business development now, expect some really cool announcements from us on the automated business development tools utilizing foundation models that allow you for end-to-end reasoning.
However, some of the very specialist applications, like physics, when you are doing molecular docking, folding, and some of the molecular property prediction, molecular generation, those tasks are still performed by specialist models with higher level of performance. In many cases, even those models are not reliable. You have to perform the experiment. Currently, we rely on hybrid systems where you have a very capable foundation model or actually a mixture of foundation models for general purpose reasoning, program-level reasoning, and smaller models for specialist tasks. Also the entire lab, the automated robotics lab that allows you to very rapidly validate some of the hypotheses and allow you to get the data and the answers that some of those models cannot produce at all.
It will take time, but in the future, it will all go into foundation models, in my opinion, where you would be able to loop them back on themselves, recreate the entire process, and reason to the point where you go from prompt to drug. Again, would love to refer you to that paper. InSilico, currently, we play in every category. We managed to develop many smaller models and put them on the market. We would cover biology, chemistry, medicine, and even science. We have even tools for environmental sustainability, and those same tools acquire more capabilities in small molecule design that we reuse for drug design. We also specialize in foundation models for end-to-end reasoning. We decided, again, to become a teacher rather than one of the fighters in the foundation model space.
Now we can actually evaluate which foundation model is better, utilizing over 1,000 benchmarks that we have, in order, as Alex has described already. With those benchmarks, we can see which one is performing better, evaluate, and then use it for our own programs. Preferably, those models must be reinforced by our MMAI Gym.
Thank you, Alex. That's very helpful. My second question is a quick question. How should we think about evaluating the real value of AI in our drug discovery today? I noticed on your slide, page seven, you gave us a formula. I was wondering for highly novel targets in particular, how to estimate the potential out-license value, as well as the underlying R&D cost in that framework?
You're absolutely correct. In that slide seven, I made an attempt to create a formula for how to evaluate pharma AI R&D productivity for the first time. We all need to think about it because otherwise, there's just too much hype, and everybody is telling the same story. At the end, there are no drugs. It's actually very easy to make a PCC, if you are going to do the same thing that has already been done before. A small iteration on already existing molecule. In China, many of the traditional pharmaceutical companies achieve greatness in time, cost, speed, and the quality of those iterations. However, many of them shy away from novelty.
If you are going for an ultra-novel target that has never been in a clinic before, pharmaceutical companies that are now rushing to replenish their pipelines, they are looking for something that has been already very proven to work on the target side and are willing to pay massive amounts of capital in order to get highly differentiated molecules with great clinical potential. However, novel targets are very difficult, and you need to push them to phase II or phase II complete in order to be able to license them out. Again, I don't think that us AI drug discovery companies should be taking drugs in the clinic at all.
If you want to prove the point that your AI is absolutely valuable and demonstrate many benchmarks and also build a bunch of benchmarks, you will need to push a few of those highly novel targets into the clinic. We believe that the potential of those targets, like for example, TNIK, that we are pushing into phase III, as Dr. Ren has described. In IPF, it is massive already, but I think that this target has a life way beyond IPF. It should work in multiple indications, because we found it using aging research. We just need to understand the dose and the molecular properties. We need to fully unlock the potential of this target. That is why we are working already on the second and third generation of the molecules. We want to own the space. In several other novel targets, it is the same thing.
You just have to grind, understand the biology that really affects your timelines, that affects your cost, that requires you to do many more experiments, take higher risk, get a lot of heat from the analysts and investors. At the end of the day, we believe we will prevail. In order to bet on those novel targets that have really high payoff potential also from the science side, we'd also need to go after low novelty targets to keep the lights running, lights on. That is why we bet on lower novelty targets with higher level of chemistry, higher level of novelty on the chemistry side, and create feats of molecular engineering that can sell for tens and hundreds of millions of dollars, as you have seen previously. That is done to be able to push the highly novel targets forward.
Those targets will create the real value for the company, for investors, and also for the patients and humanity. We are here to serve the patients, that's why novelty is extremely important.
Okay. Thanks, Alex. That's very helpful. Congrats again on having such a strong year. I'm really looking forward for your new achievements. Thanks.
Thank you.
Thank you, Rui.
Leah, can we take maybe one more?
We really can't, Alex. I'm sorry. We really need to jump on our next Chinese call, which starts at 10:30 A.M. Any last concluding remarks from you, Alex?
Thank you, everybody, for staying with us for this first earnings call. We'll try to do better next time. Better numbers, better performance. This is our first. We are here to serve the patients. We're here to serve humanity. When it comes to aging, every one of us is a patient. InSilico Medicine is a little bit unique in this regard because we are here to make money, but we are here to also advance science forward in order for us to all enjoy longer, healthier lives, preferably without losing too much function. At InSilico Medicine, we're really pushing the boundaries in both AI and drug discovery. We are competing with the most advanced pharmaceutical companies around the world, and mostly nowadays in China. Even though we're a global company, we come to the China gym to win. So far, we are doing extremely well.
We also rely on global infrastructure in order to move both AI and drug discovery forward. Please expect great results from us this year. You've already seen that just with one deal, we kind of doubled the revenue from last year. The revenue play should be there for this year because milestones are coming, and definitely I'm not stopping on the business development front. We should have many more great deals. We haven't licensed out any of our ultra-high-value clinical assets yet. But of course, everybody wants them now, especially the ones that are mid-stage clinical trials. We should do well. Of course, we're reinvesting heavily into AI. This ability to pivot into becoming a teacher, a trainer, a gym coach for a very large number of foundation model developers, I think it gives us a sense of sustainability that I have never seen before.
In the past, we were racing like a racehorse in a race. Now we can actually train a bunch of racehorses to run faster, while actually understanding which horse is the best for our purpose, and then pull the trigger on a very specific therapeutic disease target, and also on a molecule. You should expect us to scale, and you should expect more PCCs, right, Dr. Ren, this year? That's the most important metric.
Yes. Yes. Definitely.
Thank you. Of course, I would love to thank the team, would like to thank the investors. Some of the investors on the call believe in us as a longevity company that we hope we are. Please stay with us for as long as possible. We really value long-term friendships, commitments, and your feedback. We are actually very receptive to feedback, so please do send us notes. Thank you so much. Here, we would love to conclude. Please stay with us. Even though now we are kind of in the hands of the macro environment, it's kind of like weather that we cannot control, we want to be the best company in the industry and deliver great success for you. Thank you.
Thank you.
Thank you.