Flagship founder: The next step for AI for Science is Multi-Agent

01.

Flagship Creation

The creation of Flagship stems from Noubar Afeyan’s personal experience and deep thinking. Born in Lebanon, Afeyan comes from an Armenian diaspora family. His family has been displaced for generations, and Afeyan himself immigrated to Canada as a political refugee. In 1983, Afeyan entered MIT to pursue a PhD in biochemical engineering and became the first graduate of the program. During his studies, an exchange with David Packard, the founder of HP, at an international conference pointed him in the right direction.

In 1987, at a time when venture capital in the United States was primarily going to people like former Merck or IBM executives, Afeyan, a 24-year-old immigrant with no background, obtained funding to start his first company, PerSeptive Biosystems (a biotechnology instrument company that was acquired by Applied Biosystems in 1998).

In the process of starting a business, Afeyan began to think: Why is entrepreneurship seen as a random, improvisational, and even gamified process rather than a professional activity? Afeyan believes that startups are one of the greatest inventions of mankind, and he does not understand why such an important cause is seen as a random gamified activity. As a scientist with an engineering background, he firmly believes that entrepreneurship in key areas such as healthcare and climate change cannot be simply seen as "like shots on goal and this and that."

Meta-question: How to professionalize entrepreneurship

Afeyan then began to think systematically about how to professionalize entrepreneurship. He believed that one of the criteria for judging whether an activity is professional is whether it can carry out multiple projects at the same time. VCs can invest in multiple projects at the same time, while entrepreneurs are expected to focus on a single project. This cognitive difference prompted him to explore the possibility of "parallel entrepreneurship."

To test this concept, he started his first company in the late 1990s while co-founding other companies. But he soon realized that he couldn’t achieve this goal alone, so he created a company dedicated to incubating companies—originally called “newcogen” (New Company Generation), and later renamed Flagship Pioneering.

Flagship's mission is to transform entrepreneurship into a professional and institutionalized activity, just as investment has been institutionalized, operating in a team form, setting clear goals, and creating value more effectively. Its screening and incubation process includes four key stages:

•  Explorations: In the initial stage, Flagship Labs proposes seemingly far-fetched hypotheses and explores the question “What if…?” The company’s team of scientists evolves these hypotheses through mutation and screening, and collaborates with external experts to test the pros and cons of new concepts. Flagship conducts 80-100 such explorations each year to find potential breakthrough innovations. These explorations are not simple creative brainstorming, but systematic explorations based on rigorous scientific methods.

•  ProtoCos: Potential Explorations enter the Prototype Company (ProtoCos) stage. Each ProtoCo is numbered according to its position in the Flagship Labs sequence (such as FL1, FL2...FL63, etc.). At this stage, Flagship's founding team tests the concept, and ProtoCos that cannot verify the scientific principles in the laboratory will be terminated. Flagship calls this process "origination", which means conceiving, iterating and launching new companies around unique breakthroughs. The company creates an average of 8-10 ProtoCos each year.

•  NewCos: When the founding team is able to turn the question “What if...?” into a definitive answer to “It turns out...” , the ProtoCo is upgraded to a new company (NewCo), which receives a formal name and a large capital commitment from the Flagship. Each NewCo focuses on developing a proprietary platform that can continue to produce innovative products in the coming years. At this stage, a formal board of directors, CEO, and leadership team are formed. Flagship forms 6-8 NewCos each year.

•  GrowthCos: At the final stage, NewCo spins off into a GrowthCo. The leadership team of a GrowthCo is responsible for attracting external investors, establishing strategic partnerships, and building a business model that creates long-term value. Many GrowthCos eventually become public companies. Since 2013, Flagship has had 25 GrowthCos complete IPOs, most notably biotech companies such as Moderna.

This systematic approach enables Flagship to systematically innovate in areas of high uncertainty and transform scientific breakthroughs into commercial value.

Unlike the traditional venture capital model, Flagship does not look for existing entrepreneurial teams to invest in, but instead identifies opportunities on its own and builds companies from scratch. This "company-creation" model has become a unique paradigm in the biotechnology field.

Emergent Innovation

Afeyan summarized his experience in entrepreneurial practice over the years as "emergent innovation." In 2021, he and Gary P. Pisano published an article in Harvard Business Review to explain this concept.

Emergent innovation challenges traditional goal-oriented design thinking. Afeyan found that human design is usually based on clear goals, but in fact, if you ask what the goal of creating NVIDIA is, it is difficult to give an exact definition. NVIDIA initially just thought that it could create valuable gaming products, and the opportunity of AI was completely natural and not a foreseeable result in the initial business plan or five-year plan.

Afeyan seeks inspiration from nature, where this unpredictable novelty is characteristic. In nature, variation, selection, and iteration create incredibly complex systems, such as life, which we have yet to artificially create. This is called "emergence."

As for whether "variation" or "selection" occurs first, the answer is that in nature, the selection pressure comes from the sum of all things in nature. The same is true in the business world. Consumer preferences will react to products, and these reactions will form new consumer preferences. Therefore, the innovation process needs to create an environment where the two can interact to produce new products and services.

Afeyan believes that applying these three principles in any field will lead to emergent innovation - in the field of ideas, revolutionary concepts, political ideas and religions; in the field of products, iconic products such as Air Jordan. Although participants often claim that they came up with these innovations, Afeyan believes this is a manifestation of human arrogance. "People tend to describe everything in an arrogance way. We have our own language, and now we have LLMs, which are very good at translating reality into a narrative of 'how humans dominated this.' It's like they often say that those who win the war will write history."

In his 38 years of entrepreneurial career, Afeyan has come to realize that anything he has created is not entirely the product of personal work, but the result of emergence. Most successful entrepreneurs superficially attribute "success" to "hard work", but in fact secretly worship the "God of Opportunity". The core work of Flagship is not some genius technology or super intelligence, but to create an environment that promotes emergence.

02.

How to realize scientific research automation

Flagship started exploring AI 25 years ago 

Looking back over the past 25 years, Flagship, founded in 1999-2000, faced a unique market environment at the time. The Internet and e-commerce boom was in full swing, and most of the funding went to websites such as sunglasses.com and diapers.com, while it was extremely difficult to obtain funding for life science and medical projects. Despite this, Afeyan saw the huge market demand for products such as drugs and decided to focus on the intersection of biology and technology.

Flagship has been committed to systematically conceiving and creating companies from the beginning, but it did not bet on being able to systematically achieve breakthrough innovations in the early days. This is exactly the ability they gradually mastered in the following years - the company has evolved from a simple entrepreneurial ideation platform to a systematic engine that can continuously produce disruptive scientific breakthroughs.

Seven years ago, Flagship had only about 50 employees. Today, the company has 550 employees, of which about 200 are scientists, engineers and doctors, and applies for 600-700 patents each year. More importantly, Flagship has internalized its company-building capabilities and established an internal engine capable of creating multiple companies in parallel, greatly accelerating the learning cycle.

Contrary to popular belief, Flagship's exploration in the field of AI dates back 25 years. As early as 2001, they founded Affinnova, which used machine learning evolutionary algorithms to develop consumer products online. Today, their 100th company, FL100, is using Generative AI to develop related technologies. Flagship believes that Generative AI's ability to come up with hypotheses and conceptual ideas is amazing and is improving every day.

In the future, Flagship will place more emphasis on creating breakthrough technologies that can benefit the entire industry. To expand the impact of innovation, they have established large partnerships with pharmaceutical companies such as Pfizer, Novo Nordisk, and GlaxoSmithKline, as well as technology companies such as Thermo Fisher, Analog Devices, and Samsung.

Multi-agents system

Enables emergent innovation

Afeyan believes that the goal of his work is to create a new type of platform and establish an autonomous scientific discovery method - capable of generating hypotheses, designing experiments, executing experiments, collecting data, interpreting data and iterating hypotheses, realizing scientific research automation similar to Waymo's self-driving cars.

Although there is still a gap compared to fully automated scientific research, Afeyan said AI is already good enough to demonstrate its key elements and integrate them. In certain areas of biological research, AI has achieved breakthroughs similar to those in chess and Go.

Flagship is most interested in multi-agent systems that enable emergent performance. They are studying these technologies in the product space, developing new brands and products, and applying them in the field of mental health to design agent-based early intervention measures. This is not simply training AI to imitate doctor behavior, but letting the system interact on its own and learn from the dynamics between different types of agent models.

For example, Lila Science, the latest incubator of Flagship. Lila is an AI research platform, a "scientific superintelligence platform". Lila has built an AI research platform that combines AI and automated experimental equipment, which can shorten the research that originally took several years to 6 months. It has made breakthroughs in the fields of genetic drug design, new catalysts and carbon capture materials. The company adopts a business model of opening the platform to industry partners. It does not promote clinical trials on its own, but achieves commercialization through cooperation or spin-offs. Flagship's GP Geoffrey von Maltzahn serves as CEO, and Harvard geneticist George Church joins as CSO.

Strategic preference: betting on platform companies

Generally speaking, bio-tech companies can be divided into platform-based and asset-based, which focus on one or several pipelines. Afeyan believes that Flagship's strategy is to explore undeveloped frontiers. It is logically unreasonable to take risks just to acquire a single asset. If it is asset-based, it is wiser to bet on proven technology routes and make minor improvements.

Afeyan stressed that when companies venture into new areas such as RNA, DNA, gene editing or computational protein design, diversification strategies are crucial, as project failures often stem from factors unrelated to the underlying technology. It is based on this concept that all 110 companies supported by Flagship in the past 25 years have adopted a platform model without exception, a strategy that allows them to "go beyond adjacent fields, beyond the reasonable scope, and into the 'unreasonable' field."

Although most companies have this desire, not all are able to adopt a platform strategy. Afeyan analyzes three key obstacles here:

• Significant challenges in funding requirements. The development of a single project already requires a lot of money, and building a platform to support multiple projects requires an extremely large investment;

• Investors generally have difficulty accurately assessing the value of platforms. They often fail to understand the synergies between projects within the platform, i.e. how one project is de-risked by the success of other projects. Conversely, platform companies are often disadvantaged in valuations because they are perceived as "too expensive" and "too complex"

•  Investors are concerned about the management team's ability to execute. Managing a single project requires completely different skills than coordinating multiple projects simultaneously, and this fragmentation of skills can lead to poor execution. Afeyan is concerned that this limitation may create "doomed companies," which is essentially a matter of probability.

Afeyan observed that in the current market environment, the biotechnology industry is undergoing a transformation, especially for "single-asset" companies. The scale of "fast-followers" has made the market competition more intense. Companies in places such as India are entering the same field at a lower cost, and the clinical data access barriers faced by these places are more advantageous.

Faced with this global competitive landscape, Afeyan would question how single-asset biotech companies can remain competitive in commercial development. While pharmaceutical giants can profit from this, biotech startups face the risk of being commoditized, especially under the higher cost structure in the West.

These factors combined make Afeyan believe in the value of the platform strategy. The platform model at least provides companies with the opportunity to establish diversified partnerships and find ways to survive. Although there is no guarantee that all platform companies will succeed, it provides greater survival space and development opportunities than the single asset model, especially for entrepreneurs and visionary investors who truly own platform technology.

03.

Investment layout

Emphasis on "experimental spirit"

Flagship adopts an "emergent" thinking mode in field selection. However, they are clearly aware that not everything requires a scientific leap, and not all fields are suitable for this type of activity, so they maintain an experimental spirit.

In its early attempts, Flagship used a variety of technologies in addition to deep neural networks on Moderna, using data from the entire mRNA field to guide the development and production of a new generation of products. At the same time, they also began to view AI as an innovative way to design proteins.

About six or seven years ago, Afeyan's team launched a forward-looking project to explore the possibility of computationally designing proteins with specific functions. Although AlphaFold and quantum folding models are now available, they took a different path at the time. They wanted to explore whether learning algorithms could generate completely new proteins by analyzing a large number of functional examples and their corresponding DNA sequences. Although many people are skeptical about this and believe that it is necessary to first understand the DNA sequence, protein sequence, and folding structure, Afeyan insists that the fact that DNA is passed down from generation to generation in nature without instructions, but can achieve functional transmission, means that this information must be encoded in the DNA in some way.

This hypothesis drove them to jump right into experiments, since technological advances had made it much cheaper to ask such questions. Within a few years, they had demonstrated that, at least for antibodies and their binding to targets, computational methods could achieve breakthroughs that were impossible with traditional experimental methods, and this achievement gave rise to Generate: Biomedicines, one of NVIDIA’s first large partnerships in biology. Today, the company has more than 15 computationally designed antibody programs, some of which have entered clinical trials.

Afeyan's team then expanded this computational design approach to cell models, DNA, RNA and various molecules, as well as lipid nanoparticles (LNP) design, and has now expanded to multiple cutting-edge research fields. Abiologics is an important part of this expansion strategy. By integrating proprietary generative AI algorithms and peptide synthesis technology, it creates Synteins™ with strong and ideal pharmacological properties, breaking through the central law of traditional biopharmaceuticals and creating a new paradigm for drug design.

AI + Clinical Trials

There are significant challenges in converting AI-driven innovative drug candidates into marketable therapies. From a regulatory process perspective, Afeyan analyzed this path in reverse: the final step is to submit a biologics license application (BLA) or new drug application (NDA) to the FDA for approval, and before that, phase III clinical trials need to be completed to prove that the drug has statistically significant superiority and no toxicity in a large enough population.

Afeyan believes that while AI and computational methods have significantly increased the number and quality of candidates in the early stages of drug discovery, these later stages are still strictly regulated and require expensive and time-consuming traditional clinical trials. He believes that in the future we may be able to create models based on data to predict outcomes, but before that, the industry still needs to wait for large trials and the hundreds of millions of dollars they require.

On the question of when this shift will come, Afeyan feels that this change should have happened sooner, especially given the impact of the current situation on patients. He cited Operation Warp Speed ​​as an example of how the private sector, public sector, and regulators were able to work together effectively during the COVID-19 emergency, not taking shortcuts, but re-prioritizing, recognizing that the goal was to find solutions and avoid delays caused by excessive caution.

Afeyan believes the problem with the current system is that chronic diseases like cancer and neurological disorders are not considered urgent threats despite their huge impact, resulting in slow development of innovative treatments.

As a solution, Afeyan proposed a core idea: to redesign the clinical trial process using a data-driven approach. The core of this idea is to classify and understand the disease more accurately. He proposed the concept of "biological staging" that goes beyond the traditional disease staging system. This method does not simply divide the disease into four stages, but may be subdivided into tens of thousands of different stages, which can capture subtle changes in disease development at the molecular level.

Based on this fine classification, Afeyan believes that clinical trial participants can be selected more accurately, reducing the interference of the heterogeneity of the test population on the trial results. This approach allows for smaller but more targeted clinical trials, first obtaining a smaller range of approved indicators, and then expanding the scope of application.

Afeyan stressed that the key to achieving this vision is to make full use of existing patient data. Under the premise of complying with privacy regulations such as HIPAA, this data can be used to train models to identify the unique mechanisms of specific disease subtypes. For example, this approach can help identify molecular targets that need to be targeted for a specific subtype of Parkinson's disease.

He acknowledged that this precision medicine approach may conflict with the business model of large pharmaceutical companies that want to make drugs available to a wide range of people, but it could be of great value to biotech startups with limited resources and patients who are in urgent need of new treatments.

Beyond drug development

Flagship has an extremely wide range of investment portfolios, not only involved in the development of therapeutic drugs, but also in different fields such as nutrition, agriculture and climate. In fact, Flagship's business scope is wider than the outside world imagines. As early as the early 2000s, Flagship began to get involved in supercomputing companies and also invested in network companies. Now, they have also begun to make new attempts in the field of materials, such as semiconductor materials, carbon capture materials, etc.

Flagship is usually highly cautious about entering completely new fields unless it has a core advantage, whether it is intellectual property or a fearless spirit of innovation arising from a lack of knowledge of the field.

For Flagship, the first attempt in a field will influence the next five projects. If these five projects are not successful, they will think, "Maybe we can't get the expected return from innovation in this field, and quickly give up and move on to more promising projects."

Flagship’s core strategy is to look for areas that can “accelerate the realization of future technologies,” but Afeyan also pointed out that not all advanced technologies have commercial value at the moment.

Take the experience in renewable energy as an example. Flagship has worked in this field for many years and invented an advanced method to produce carbon-neutral liquid fuels by modifying photosynthetic bacteria. These genetically engineered bacteria can directly absorb carbon dioxide and then produce and secrete diesel fuel, which is a huge technological breakthrough. Between 2008 and 2012, they founded Joule to promote this technology.

However, changes in the market environment made it difficult for this innovation to obtain the expected returns. When Flagship started the project, the price of carbon was $50 per ton, and it dropped to $5 per ton by the end of the project. When the project started, the United States relied on energy imports, but by the time the project was completed, the United States had abundant energy. In the end, they realized that no matter what innovations they made in this field, they could not get a premium, so they decided not to invest anymore.

04.

Investment philosophy

Managing uncertainty

In the real business world, when opening up a new field, a company not only needs to solve technical challenges, but also needs to create an entire ecosystem (including regulatory frameworks, market awareness, public acceptance, etc.). Regarding how to manage the many uncertainties in unknown areas, Afeyan shared Flagship's thinking and response strategies.

Afeyan first distinguished between the two key concepts of "risk" and "uncertainty". He used a vivid metaphor to explain the area where innovation occurs: imagine everything that is currently known and exists as a circle, and the area immediately outside this circle is the area that will be known in the next period of time. In this "adjacent circle", most innovations occur, and people can make reasonable estimates of risks and returns. This is exactly the meaning of traditional due diligence - consulting field experts (key opinion leaders), aggregating information and making investment decisions.

But when you move away from the known range, the situation changes qualitatively. Afeyan believes that at this point people can no longer estimate the probability of success and the rate of return. This situation should not be called risk, but "uncertainty." He disagrees with the idea promoted by Wall Street and other institutions that everything can be put into the risk matrix. Taking the current nuclear fusion technology as an example, Afeyan believes that this is not a risk but an uncertainty because no one can really estimate its probability of success.

Afeyan admitted that opening up unfamiliar territory does add more layers of uncertainty. But he also raised a paradox: although the risk in "adjacent areas" seems lower, there is actually an overlooked problem, namely commoditization risk. Because all players - whether startups, academic labs or large companies - are paying attention to these adjacent innovation areas, competition is extremely fierce and they will eventually face commoditization risks. Even if the innovation is successful, it is difficult to obtain high returns because multiple similar solutions will appear on the market. Therefore, Flagship prefers to choose to embrace uncertainty cautiously.

As for how to manage uncertainty, Flagship's way of dealing with it is to conduct experiments. Afeyan believes that for things whose value and feasibility are still uncertain, the key is to design appropriate experiments, turn them into reality, verify their feasibility, and at least control the controllable parts.

Moderna, the 18th company of Flagship, is a prime example of this strategy. Before Moderna was founded, the market for mRNA drugs or vaccines was basically non-existent and almost no one was involved. This meant that they needed to fight for regulatory changes or approvals, establish market pricing mechanisms, and solve production and manufacturing problems that were completely unknown at the beginning. According to Afeyan, Moderna's returns were foreseeable even without the epidemic, and in fact they had created a lot of value before the epidemic.

The core way of thinking at Flagship is to accept uncertainty and manage it systematically. Afeyan emphasized that if you are not willing to do this, you can only work in the "Me Too value pools". Flagship does not think that they are smarter, harder-working or more connected than others. Their core competitiveness lies in their willingness to bear uncertainty.

Polyintelligence

In Flagship’s 2025 annual letter, Afeyan proposed the concept of “polyintelligence,” the idea and relevance of human intuition in the future development of the world.

Afeyan first proposed a rethinking of the nature of "intuition". Afeyan believes that human intuition can essentially be understood as a cognitive model - a model that everyone generates and simplifies based on their own experience.

From this perspective, intuition and LLM have conceptual similarities, the difference lies in the scale of the data: LLM is trained on data from millions of people, while personal intuition is built only on the basis of an individual's limited experience.

So what aspects of humanity will retain their unique value in the age of AI? Inspired by the PBS documentary Leonardo da Vinci , Afeyan took a more macro perspective: Current discussions about AI often focus on the binary relationship between humans and machines, ignoring a fundamental fact: the core of science has always been the interaction between humans and nature.

In Afeyan's framework, the future intelligent ecology is not a simple opposition or integration between man and machine, but a dynamic system consisting of three parties: human intelligence, machine intelligence and natural intelligence. These three forms of intelligence interact and adapt to each other to form an evolving whole.

Afeyan emphasized that in this triangle relationship, the role of humans is still irreplaceable, because human thinking and behavior patterns are fundamentally different from computers and other forces in nature. The value of humans does not lie in competing with machines for information processing, but in participating in this tripartite system in a unique way and contributing human-specific thinking patterns, creativity and goal orientation.

Polyintelligence is now a key guiding principle of Flagship’s mission: to not only understand nature, but to innovate beyond its current capabilities. Similar to how Leonardo da Vinci mastered perspective and color theory before breaking traditional boundaries, Flagship hopes to come up with new solutions by having AI learn and encode the fundamental “rules” by which nature works. Just as researchers used machine learning to analyze the complex click patterns of sperm whales and revealed a structured form of communication that challenged our understanding of language and intelligence, Quotient Therapeutics is applying AI to analyze somatic cell genomics, linking mutations to functional outcomes and revealing insights that were inaccessible to traditional genetics.

The "tripartite relationship" in Polyintelligence is described by Afeyan as "a beautiful new axis of emergence", suggesting that the integration of multiple intelligences will drive the future development of life forms. This framework goes beyond technological determinism and places humans, machines and nature in a mutually beneficial and co-evolving system.