This article is a part of Poland Unpacked. Weekly intelligence for decision-makers
Molecule.one is developing Maria, a platform that combines AI models, an autonomous laboratory, and chemical data. The startup worked with OpenAI to optimize a reaction used in drug manufacturing, and in 2025 increased its revenue to more than PLN 10 million.
Gathering information, automating processes, and generating content are among the most commonplace uses of artificial intelligence. But the technology is also powering projects with truly transformative potential. One such initiative, focused on quantum chemistry, is the autonomous system El Agente Q, developed in part by Nvidia.
Yet there is no need to wait years to see AI deliver breakthrough scientific results. Today, an "AI chemist" can complete in just three months as many experiments as a human conducting three reactions a day, without a single day off, would finish in a decade – more than 10,000 experiments. That is what Polish startup Molecule.one demonstrated in collaboration with OpenAI as it works to accelerate both drug discovery and pharmaceutical manufacturing.
"For us, this is an absolute milestone. In June 2024, we created a product development roadmap. One of our assumptions was that our AI would make its first scientific discovery in the first quarter of 2026. We were off by just one month. It illustrates that the most effective way to predict the future of technology is to create it. That mindset has guided us from the very beginning," says Piotr Byrski, CEO and co-founder of Molecule.one.
Good to know
Work on AI for chemistry is gathering pace
Technology giants, leading universities, and major pharmaceutical companies are already developing AI agents specialized in chemistry. Among the most notable are Co-Scientist from Google DeepMind, El Agente Q, developed by Nvidia, the University of Toronto, and the Vector Institute, Cactus from the Pacific Northwest National Laboratory, SheEPhERD from the Massachusetts Institute of Technology, and Robin from FutureHouse.
Specialized companies applying AI to drug discovery are also emerging, including Owkin, which works closely with pharmaceutical giant Sanofi. Poland has joined this technological race as well, thanks to startup Molecule.one.
Solving a scientific problem in an unconventional way
For its first scientific discovery, Maria – Molecule.one's proprietary AI agent – used the GPT-5.4 model. The breakthrough focused on improving the efficiency of the Chan–Lam reaction, a widely used process in the pharmaceutical industry that typically joins a carbon atom with a nitrogen atom. Chemists regard the reaction as notoriously "temperamental," often failing on the first attempt.
"Our agent proposed several ways to solve the problem, including unconventional approaches that chemists would generally consider unlikely. It then carried out the experiments and identified an oxidizing agent known as TEMPO, which significantly increased the probability of success. Across 80% of the substrates tested – the starting materials for the reaction – the average yield was higher, while for more than half of them it more than doubled," says Piotr Byrski.
In a publicly available scientific paper, the team has described its methodology in detail, allowing any chemist to apply it. The aim was to provide researchers with a tool for synthesizing challenging compounds that are important, among other things, in research into cancer and diabetes.
"We see this as a demonstration of our capabilities. From a business perspective, we have showcased a 'scientific problem-solving engine' that we can commercialize in partnership with individual companies," Piotr Byrski explains.
Expert's perspective
A breakthrough on the scale of AlphaGo
The proposal to use TEMPO as the oxidizing agent in the Chan–Lam coupling reaction is as surprising as Move 37, the famous play made by AlphaGo. It is further evidence that AI models are not merely capable of reproducing existing knowledge – they can also generate original ideas.
The scale of the experimental work is equally striking: 10,080 reactions. Confirming the research hypothesis across such a large number of experiments shows that the finding is not a statistical fluke but a rigorously validated scientific discovery. It is an excellent example of the value created by combining synthetic chemistry with a robust, large-scale experimental approach. It also highlights the Molecule.one team's hands-on expertise not only in building AI models, but also in practical drug discovery.
OpenAI and Molecule.one share the same ambition
The partnership between OpenAI and Molecule.one grew out of a shared vision for the future of artificial intelligence in science. Both teams are pursuing the same objective: to dramatically accelerate scientific discovery, particularly in drug discovery and pharmaceutical manufacturing.
"Our Polish roots helped us at the beginning of the project. We had the opportunity to meet Jakub Pachocki, OpenAI's Chief Scientist, through mutual friends from the University of Warsaw. It quickly became clear that we shared the same vision of how AI could transform scientific research," says Piotr Byrski.
He admits that, while he is naturally optimistic – otherwise he would never have founded a startup – he was still surprised by how quickly the first discovery came. Just two and a half months elapsed between submitting the first prompt to the model and achieving the breakthrough.
"During that time, Maria-GPT evaluated four hypotheses addressing different scientific problems. It initially disproved one, confirmed another – which we have already published – and a third remains undisclosed. As for the fourth, we still need a few more data points before we can confirm the result with complete confidence. Each of these hypotheses has the potential to improve the way new drugs are discovered and manufactured," the entrepreneur says.
Good to know
AI refutes Erdős’ hypothesis after 80 years
In May 2026, an OpenAI model disproved a 1946 hypothesis by Paul Erdős concerning the distances between points. For decades, mathematicians had assumed that the best results were achieved through arrangements based on regular grids. The model, however, discovered an entire family of superior solutions, and the resulting proof was later verified by mathematicians. It is one of the most significant examples to date of AI making a genuine contribution to solving an open scientific problem.
AI as the laboratory operator, not the human
The startup was founded in 2018 by Piotr Byrski, Paweł Włodarczyk-Pruszyński, and Dr Stanisław Jastrzębski. Its original goal has remained unchanged – the only thing that has evolved is the scale at which it applies generative artificial intelligence (GenAI).
"We recognized the potential of AI agents, which have recently become a global trend, long before they gained widespread attention. In 2022, we launched an AI chemistry laboratory designed to generate data tailored to the needs of AI agents. Artificial intelligence is its primary operator, while humans play a supporting role. This approach is the opposite of how AI is implemented in most cases," explains Piotr Byrski.
The opening of the facility in Dziekanów Leśny near Warsaw was an earlier milestone. The company had already developed software for chemists to plan experiments related to chemical synthesis. The bottleneck, however, was the quality of available data, which was insufficient compared with the capabilities of modern AI models. This led to the creation of the Maria platform – named after Maria Skłodowska-Curie, still the only woman to have received two Nobel Prizes.
"The system consists of a laboratory, a data repository, and an AI model. It is designed not only to replicate known processes, but also to discover new ones. The data we generate is used to further train the models in subsequent iterations. The number of experiments we conduct is an order of magnitude greater than what is possible in a traditional human-operated laboratory. Over time, we have built the largest publicly known dataset of its kind – more than 400,000 experiments, meaning chemical reactions conducted at the microliter scale [using very small volumes – editor's note]," says Piotr Byrski.
Expert's perspective
Better not to separate humans from AI
Nevertheless, the presence of humans and their role in decision-making appear to be essential. What matters is not only the final outcome, but also preserving human expertise. Dependence on algorithms and the loss of the ability to conduct processes without them could prove as dangerous as the economic and raw material dependencies seen today in some sectors of the economy, including pharmaceuticals. The withdrawal of a supplier or a disruption in deliveries can bring an entire production process to a halt if alternative pathways have previously been eliminated.
Equally dangerous is superficial human analysis of results. This point should be strongly emphasized: we should use modern tools, but we must remember that human processes of analysis and decision-making have not accelerated – and should not be artificially accelerated. For this reason, I do not find the term "synthetic chemists" for algorithms entirely appropriate, as it suggests competition with "protein-based" chemists. People should not feel pressured that because AI accelerates work, they must also work faster. The process of acquiring knowledge and conducting analysis requires time.
Another important issue is the security of these solutions and the need for international legal regulations, rather than relying solely on internal rules established by organizations implementing such tools. AI agents, like any technology, can be used in harmful ways – and by a much broader group of people than before, no longer limited to specialists.
From the perspective of scientists and businesses using the results of their work, intellectual property is another crucial issue. Various electronic tools have supported researchers for years, for example in statistical analysis. However, assessing "scientific novelty" or the uniqueness of a solution – including when granting patent protection – has so far remained a human responsibility. Today, we need to consider how to define the "authorial" role of an algorithm within this framework. Intellectual property law will likely need to be adapted quickly to match the current level of technological development.
High-profile investors and a U.S. funding round in the pipeline
Molecule.one has recently seen growing interest from potential customers. Even the laboratory’s enormous capacity by market standards – 12,000 experiments per week – is beginning to become a constraint. The company is therefore planning, potentially as early as this year, to expand the capabilities of its facility in Poland and launch a similar operation in the United States.
Each of these investments will amount to millions of złoty. To date, the startup has raised approximately PLN 40 million (about EUR 9.4 million) through grants and investor funding. It has already secured the backing of an international group of established venture capital funds, including Sunfish Partners, Expeditions Fund, Atmos Ventures, Possible Ventures, AME Cloud Ventures, Cherubic Ventures, Hanover, and Outsized Ventures. Key business angels include Mati Staniszewski (ElevenLabs), Marcin Żukowski (Snowflake), Piotr Romanowski (Selvita), and Sebastian Guth (Bayer).
"Following the success of our collaboration with OpenAI, we are working on our next funding round, this time in the United States. The country is home to many outstanding projects transforming science through AI, and they can count on significantly greater financial support. I believe our success comes from taking a different approach. The dominant narrative is still based on the idea of building systems around humans, with artificial intelligence serving only as an assistant. From the beginning, we have put AI at the center. This is the only way to improve efficiency by an order of magnitude, rather than by a few or several dozen percent," says Molecule.one’s CEO.
AI, chemistry and biology specialists wanted
The significantly greater availability of capital in the United States – where funding can be raised more easily "locally" – was one of the factors behind Piotr Byrski’s move to the US this year. Equally important was the desire to accelerate growth in the company’s key American market. Frequent direct meetings with partners and investors make that easier.
"The pace of AI development is also crucial. In the United States, and specifically in California, it is the fastest in the world. If we want to push the boundaries of how artificial intelligence is used, there is no better place to do it than the San Francisco Bay Area," the entrepreneur explains.
The vast majority of the company’s 30 employees still work in Poland, however. That number is expected to increase significantly. Molecule.one is currently recruiting for its machine learning team, and within a year the company plans to hire several dozen experts in AI, chemistry, biology, sales, and operations.
"At the same time, we have started building a team of around a dozen people in the US. It will be responsible for developing relationships with local partners, as well as launching and growing our future laboratory," Piotr Byrski says.
First profit and multi-million revenues
The startup is not limiting itself to research and development activities. It has been generating revenue for several years. In 2025, it tripled its revenue to more than PLN 10 million (approximately EUR 2.3 million) and plans another threefold increase this year. Most of its projects are carried out for US customers, although interest is also growing in Europe. The company benefits from maintaining its main operational hub in Poland.
"Since last year, we have already been generating positive EBITDA and net profit. We could focus on commercializing our existing technology, which would significantly improve profitability. But at this stage, that is simply not our priority. We still have many achievements ahead of us, so it is more valuable for us to invest. We will discuss margin levels in a few years," says Piotr Byrski.
The company is simultaneously acquiring new customers and expanding cooperation with existing ones. One example is a project carried out for global company W.R. Grace, which Molecule.one secured after winning a competition with a USD 1 million prize. The two companies are currently focused on discovering new methods for synthesizing compounds used in the production of drugs such as Ozempic.
"We have entered a crucial stage of development. We have successful collaborations with renowned customers behind us, as well as a high-profile scientific achievement. The key remaining barrier is market readiness – in terms of standards and processes – to adopt what we offer. Interest is strong. Now we need to turn that interest into new large-scale projects," says Piotr Byrski.
Good to know
A scientific breakthrough just around the corner?
The founders of Molecule.one describe their views as contrarian. They rarely change their minds based on external opinions alone. Instead, they always seek to understand the essence of a problem, and only new evidence can persuade them to revise their position.
"This means our conclusions may sometimes appear radical. But in the case of the results from our project with OpenAI, we were right. We achieved the highest level of AI autonomy in science to date in terms of proposing and experimentally testing new ideas in a laboratory setting. Moreover, for the first time, this was achieved through collaboration between a language model and a laboratory without a critical role for humans," says Piotr Byrski.
Together with his co-founders, he believes the scientific world is on the verge of a fundamental shift in how knowledge is created. Scientific progress typically requires expensive and time-consuming experiments. The recent refutation by an OpenAI model of Paul Erdős’ 1946 mathematical hypothesis, as well as the joint project with Molecule.one, demonstrate that artificial intelligence could remove one of the key barriers to scientific advancement: the limited number of hypotheses that researchers are able to test.
"Progress in chemistry is highly nonlinear. Significant discoveries are extremely rare, but they usually have a huge impact – for example, on the new medicines humanity will be able to create. The number of possible substances and combinations is difficult to comprehend. We can compare it to the scale of our knowledge of Earth set against the size of the entire universe. Essentially, almost everything still lies ahead of us," says Piotr Byrski.
Key Takeaways
- Purpose and technology. Startup Molecule.one was founded in 2018 by Piotr Byrski, Paweł Włodarczyk-Pruszyński, and Dr Stanisław Jastrzębski. Its goal is to radically accelerate scientific discoveries, including drug discovery and manufacturing, through the use of artificial intelligence. In 2022, the company created the Maria platform, which consists of an AI laboratory, a data repository, and an AI model. The laboratory’s primary operator is an AI agent, while humans play a supporting role. The facility has a capacity of 12,000 experiments per week.
- Key project. In collaboration with OpenAI, the Polish startup made a widely reported scientific breakthrough this year. Its Maria platform used the GPT-5.4 model to improve the efficiency of the Chan–Lam reaction. This is a widely used process in the pharmaceutical industry that typically involves linking a carbon atom with a nitrogen atom and often fails on the first attempt. The model proposed solutions to the problem – including approaches considered unlikely by chemists – conducted experiments, and identified an oxidizing agent that increased the probability of success. In less than three months, it carried out more than 10,000 reactions, a scale of work that could take a human chemist up to a decade to complete.
- Investors and results. To date, Molecule.one has received approximately PLN 40 million (about EUR 9.4 million) in grants and investor funding. It has attracted backing from an international group of renowned venture capital funds, including Sunfish Partners, Expeditions Fund, AME Cloud Ventures, and Outsized Ventures. Key business angels include Mati Staniszewski (ElevenLabs) and Marcin Żukowski (Snowflake). The startup has already become profitable, but it continues to focus on investment and revenue growth. In 2025, it tripled its revenue to more than PLN 10 million (about EUR 2.3 million) and plans another threefold increase this year.