From Generate to Lila to Expedition, Molly Gibson has spent her career building the next generation of scientific infrastructure - ventures that merge large-scale structural biology, lab automation, AI, and molecular design. As an Origination Partner at Flagship Pioneering, she’s helped shape how new scientific platforms are conceived and built from the ground up. We wanted to learn from Molly how she thinks about creating, building and scaling companies at the forefront AI and science.
In this conversation, we cover:
Creating chemistry’s ‘AlphaFold’ moment
Lessons from scaling an AI drug discovery company
How to make the lab part of the algorithm
Emerging business models in AI for science
And what’s really needed for the next leap forward
Background
Expedition Medicines
Founded after three years of incubation within Flagship Pioneering, Expedition Medicines is rethinking how small molecules are designed - using chemoproteomics as a data engine for AI and applying generative, quantum-informed models to targets once considered “undruggable.” Officially launched on October 22, 2025, the company aims to open a new chapter in covalent drug discovery by fusing quantum chemistry, machine learning, and large-scale experimental data generation.
Financing: Backed by $50 million in initial funding from Flagship Pioneering, following three years of in-house incubation. Expedition has also initiated a multi-target research collaboration with Pfizer focused on prostate cancer, under the broader Flagship–Pfizer strategic partnership.
Platform: Expedition is developing a generative AI platform for covalent drug discovery, combining quantum chemistry, AI, and chemoproteomics to design small molecules for previously “undruggable” targets.
CEO: Molly Gibson
Lila Sciences
Founded in 2023 by Flagship Pioneering, Lila Sciences is building what it calls a “scientific superintelligence” — an AI-driven research engine that unites generative models, robotics, and automated labs to accelerate discovery across biology, chemistry, and materials science. Having only emerged from stealth in early 2025, Lila has become a central player in the growing movement to make the scientific method itself programmable..
Financing: Raised ~$550 million total; Series A ~$350 million (incl. ~$115 million extension) bringing valuation to > $1.3 billion.
Platform: Lila’s “AI Science Factory” connects large-scale generative AI with autonomous, data-rich laboratories. The platform will allow AI to design, execute and learn from experiments in real time, effectively bringing the laboratory into the algorithm.
CEO: Geoffrey von Maltzahn
Generate Biomedicines
Founded in 2018 by Flagship Pioneering (with co-founder Molly Gibson) and headquartered in Somerville MA, Generate:Biomedicines is a clinical-stage company advancing a new era of programmable biology to engineer better medicines for patients, faster. Decoding Bio covered Generate in detail here.
Financing: Raised nearly ~$700 million since 2020; Series C ~$273 million closed Sept 2023
Platform: Generative biology platform using machine learning and biological engineering to design novel protein therapeutics at scale.
CEO: Mike Nally
Molly Gibson
Molly is an experienced founder, technologist, and builder of AI-native platforms that expand the boundaries of scientific discovery. Prior to Expedition, as Co-Founder and President at Lila Sciences, she led the development of the AI Science Factories - a platform designed to build ‘Scientific Superintelligence’ by allowing AI to run the scientific method through interaction with a physical lab. As Co-Founder and Chief Strategy & Innovation Officer at Generate: Biomedicines, she helped lead the development of one of the world’s first transformer-based models for protein design, pioneering a new approach to therapeutic creation through generative AI.
At Flagship, Molly in an Origination Partner at Flagship Pioneering, where she leads a Pioneering Business Unit focused on founding, building, and growing companies at the intersection of artificial intelligence and science. Her work has led to numerous patents and peer-reviewed publications in Science and Nature and underpins programs now advancing through clinical development.
A computer scientist by training, she earned her PhD in computational biology at Washington University in St. Louis.
A Conversation with Molly Gibson on Scaling Scientific Discovery
Expedition Medicines | Small Molecules Meet AI
Zahra: Congratulations on the launch! Tell us, what was the initial spark behind Expedition?
Molly: We have a good understanding of many biological targets - but often, we just don’t have the chemical tools to solve them. Recent technologies, such as protein-based technologies, aren’t capable of reaching all targets, especially those deep within cells.
Small molecules should be the answer, but discovery remains too empirical and trial-and-error. We wanted to bring the AI revolution from protein design into chemistry, so we could design small molecules even for the so-called “undruggable” targets.
Zahra: How did that idea become a platform?
Molly: We approached this problem from three directions:
First, learning from nature. In proteins, evolution leaves patterns AI can learn. Small molecules lack those evolutionary signatures, but nature still solves binding via catalysis. We treat the proteome as a surface full of latent catalytic micro-environments and view binding as a catalysis-first, covalent reaction driving precise bonds between a molecule and its target.
Second, building the missing dataset. The PDB underpinned the AI-protein revolution; chemistry has no equivalent for protein-small-molecule interactions. So we built a high-throughput chemoproteomics platform that screens ~20,000 sites across the proteome in parallel (typical assays do 100 at a time!). We read covalent engagement by mass spectrometry at single-amino-acid, atom-level resolution. It’s the high-quality, high-content, generalisable data that chemistry has been missing.
Third, fusing AI with quantum chemistry. For proteins, AI and atomistic models have taken us a long way. For small molecules, electronic structure dominates behaviour. To model and design at that level, we’ve combined deep learning with quantum chemistry, particularly Density Functional Theory (DFT). This allows us to simulate electron overlap between ligands and their micro-environments, capturing how new covalent bonds could form.
Zahra: What allowed you to scale the chemoproteomics from 100 to 20,000 sites in parallel?
Molly: We’ve reimagined chemoproteomics not just as a discovery method but as a data-generation engine for AI. That required innovation across both hardware and software.
We’ve partnered closely with Thermo Fisher, who are equity investors and collaborators in advancing mass spectrometry technology for our needs - and they’ve noted that Expedition is operating at some of the highest experimental throughputs in the field. We’ve paired those gains with novel probe design and cell-based expression systems to achieve unprecedented diversity and coverage across hard-to-reach targets - essential both for building robust models and for launching drug discovery programs against first-in-class targets.
Zahra: What about the quantum chemistry side, how does Expedition go beyond what companies like DeepMind or Microsoft are doing?
Molly: We’re not focused on developing proprietary DFT models ourselves. Instead, we collaborate with leading groups advancing those methods, and integrate their approaches with new deep learning architectures trained on what we believe is a uniquely rich experimental dataset. Our strength lies in data alignment: we bring large-scale, high-quality experimental data that allows these quantum models to actually learn chemistry - capturing bond formation and reactivity at a mechanistic level. Here we are combining the power of synthetic quantum data with real-world chemistry.
Zahra: And once you have a hit, how does optimisation work?
Molly: Every SAR campaign runs against the entire platform, so we can monitor potency and off-target effects across the proteome in real time.
The next frontier, I think, is letting large language models help run the medicinal chemistry loop: suggesting modifications, exploring chemical space, and finding ideas humans might not think of.
Zahra: And what about novelty, are you finding new chemistry or new biology?
Molly: Both. We’ve found some very novel cases - in one instance, we bound to a featureless surface on a transcription factor and, as we optimised the compound, it actually induced the formation of a pocket. Binding reshaped the protein. We’re also seeing a much higher proportion of functional binders than standard approaches. And if something binds but isn’t functional, we can often convert it into a degrader.
Zahra: How do you pick targets? Do you let the platform guide you or do you guide the platform?
Molly: We start with a curated set of hundreds of genetically validated, high-value targets where the biology is solid but no drugs exist - all first-in-class or first-in-mechanism opportunities.
Right now, we use our high-throughput data-generation platform to screen across those targets and generate generalisable data across the proteome. But increasingly, we’re moving toward the approach you described - saying, “this is the target we want,” and then using machine learning to generate the initial hit compound if we don’t find one experimentally. That first binder is usually the hardest part of tackling these tough targets - now we can start to design it computationally.
Zahra: What do you think is the biggest challenge for Expedition?
Molly: Honestly, it’s the same challenge across AI-driven drug discovery - the translation into humans.
We’ve made so much progress in early discovery, but it still takes capital and time to bring these programs into the clinic and learn from human data. People sometimes say, “there’s no proof that AI helps in drug discovery,” and I get that skepticism -but it’s really a matter of time horizons.
It’s already transformed how protein drugs are designed - I can’t imagine anyone developing a new biologic today without using AI somewhere in the process. We just haven’t seen enough clinical proof points yet. Those will come.
Generate Biomedicines | Reflections
Zahra: Looking back on Generate, what were the biggest lessons, and what would you approach differently now?
Molly: I learned so much in building Generate, but probably the biggest lessons I learned had to do with how to build and scale teams. How to motivate teams, how to inspire teams to do seemingly impossible things, how to hire great people and part ways when needed (with humility and respect), how to resolve conflict, build enduring culture. I really grew as a leader at Generate and it’s stayed with me.
One thing I’d approach differently now was how fast we scaled - it’s so hard to scale companies at the pace Generate grew and maintain an amazing culture and alignment across the team.
Overall, I feel Generate built a sustainable platform that’s producing a pretty amazing pipeline, which I’m proud of.
Zahra: On that theme, do you see a world where biotech platforms can become a core and substantial source of revenue in themselves? Or will dependence on a pipeline be inevitable?
Molly: I think we have to figure out platform monetisation. Right now, value doesn’t flow back to early discovery technologies until you have proven they change outcomes downstream.
Once we have more proof points, once we have shown repeatedly that AI can improve success rates or move the needle in meaningful ways - I hope these value inflection points shift earlier, and platforms become more licensable and properly valued. Today it’s still hard to avoid building a pipeline.
Zahra: What was the effect of open-sourcing Generate’s Chroma model - e.g. scientifically and culturally?
Molly: In my view, publishing Chroma was absolutely a net positive. It gave the scientific community visibility into the frontier work happening inside Generate and helped advance the field by sharing some of the unique insights the team had created into scalable, generalisable, generative models for proteins.
It also shaped culture: internally it reinforced a spirit of openness and scientific contribution, and externally it built trust and connection with the broader community. For us, it was about being part of the ecosystem that’s moving the whole field forward.
Lila Sciences | The AI Science Factory
Zahra: With Lila, you’re facing all of these challenges but on a much bigger scale: platform monetisation, determining what to build when… how do you prioritise what to build first across that stack?
Molly: One big goal for Lila is to create an experimental verifier for AI - essentially scientific self-play: ask a question, run the experiment, and learn from the result. To do that, you need the whole tech stack: hardware, APIs, orchestration software, and AI - all working together.
We’re building a generalised, end-to-end system. Historically, lab automation has been trapped in fixed workflows; we’re breaking it down into unit operations so AI can design any experiment it wants - or even the instruments it needs. That means innovating across every layer simultaneously, building what we call reinforcement learning for science, with both experimental and simulated data feeding the self-improving loop.
Zahra: We touched earlier on platform monetisation being difficult in drug discovery, but you mentioned that Lila might be different. How do you see Lila capturing value - through partnerships, licensing, or something else?
Molly: Lila is building an intelligence operating system for science - a unified stack that connects a central AI with data, simulation, and automated labs across biology, chemistry, and materials. It’s designed to reason, design, and experiment across domains, creating a common layer of intelligence that compounds with every use.
Instead of capturing value in a single asset, we’re monetising that intelligence itself - as shared infrastructure others can build on, use, and benefit from. That breadth should allow new business models to emerge in unique ways in science - from platform licensing to joint ventures to usage-based models that reflect how science actually scales.
Zahra: When people talk about “lab-in-the-loop” systems, what does that actually look like today? Are there examples where models already influence live experiments, or is it still largely human-guided? And what’s the hardest part of getting to full autonomy?
Molly: Lab-in-the-loop is really about how we train AI models - not from static, curated datasets, but by letting AI design and learn directly from experiments. It’s active learning in the physical world.
While there are still humans in the loop today, we’re bringing the lab much closer to the AI. Full autonomy is a longer-term goal, but active learning with real-time lab feedback is already happening.
The challenge with full autonomy in the lab is a lot like self-driving cars: it’s not the easy cases that are hard, it’s the edge cases. Experiments fail for subtle, context-specific reasons, and the system needs to understand and adapt to that. In some ways, it might be even harder than driving, because science isn’t fully observable - you can’t always see what’s happening inside a cell or a reaction in real time. Getting to full autonomy means building systems that can reason through that uncertainty. I think we can get there, but there’s a lot of value between here and full autonomy as well.
Zahra: Where are we on the stack today - especially the “AI orchestrates the lab” vision? Do we need to reinvent robots, or can we use what already exists?
Molly: It’s a bit of both. For a good portion of the lab, we use existing instruments, but we build new APIs and drivers so AI can interact with them. In some areas - especially on the physical sciences side - we’ve had to invent entirely new instruments for high-throughput scientific verification.
So when an instrument exists, we use it; when it doesn’t, we build it. We tried different ways to use existing automation tech to get general capabilities, and we found it’s better to own the full vertical integration - hardware, software, and AI - end to end is what enables generality and scale.
Zahra: How do you think about “product” at Lila - is it therapeutics, delivery vehicles, coatings, catalysts?
Molly: All of those are on the table , but for partners, not for us directly. Lila isn’t using its capital to build traditional products like therapeutics. We want to help others create those.
Think of Lila as the infrastructure and intelligence layer of science. Others can come to the platform and build their products. Our job is to make that layer generalisable and impactful across the industry.
Zahra: Lila mentioned in a press release that it has made thousands of discoveries already - are these product discoveries? Where is my invisibility cloak?
Molly: Too early! But those discoveries span a range of domains: from improved mRNA stability to new materials formulations. The point isn’t the count, it’s that the rate of discovery is accelerating as the system learns. Each new result strengthens the AI’s ability to reason about science itself.
Zahra: How do you see China’s role in this new hardware-heavy science?
Molly: China is scaling lab automation and scientific manufacturing at extraordinary speed. I think there’s a lot to learn from that → especially how they integrate robotics, data systems, and physical infrastructure.
For Lila, hardware innovation is global - we collaborate with groups in North America, Europe, and Asia - I think the future will be collaborative, not competitive in this space.
The Future of AI and Science
Zahra: Impossible question: what’s actually holding back the next frontier of AI in biology - compute, data, or algorithms?
Molly: All three matter, but if I had to pick its data.
If you had the data, compute would follow and the algorithms are already remarkably robust. We have come very far with transformers. The challenge is that science will never have internet-scale data - or at least, not soon.
Most experiments are only points in time, not full narratives - we don’t record the conditions, failures, or context in ways that machines can learn from. Metadata matters enormously, and much of science simply isn’t written down in a machine-readable way. The PDB is the exception, not the rule, and it took decades to build.
Zahra: So is the challenge less about collecting more data and more about finding entirely new ways of observing biology?
Molly: It’s both! here’s a distinction between general foundation models and problem-specific data. What Expedition is doing and what Generate did is create large datasets that try to learn generalisable rules: for Generate, protein structure; for Expedition, how small molecules form covalent bonds with proteins. Those general rules let you generalise across many targets.
But I’m also excited about platforms or machines that can create the specific datasets you need, on demand. Imagine taking cell-based assays and making them automatable: I have a new target; I want a new cell-based assay; I don’t want six months of method development, transfer, and engineering. What if an AI could design the assay, optimise it autonomously on a lab, and learn from that data? Then you can explore specific hypotheses at high throughput.
I even imagine AI that can say what instruments it needs: like a robot saying, “to cut a sandwich, I need a knife.” One day AI could look at a lab and say, “to test this hypothesis, I need this instrument”, this instrument could well be a new way of observing biology that us humans have not thought of! It’s down the road, but we’re just scratching the surface of the assays we’ll build and the kinds of biology we’ll observe.
Zahra: How do you see tech giants like OpenAI or Anthropic entering biology?
Molly: Science is a tough business. Training a model is one thing; running labs, generating experimental data, and connecting that feedback loop is another.
Flagship’s advantage is that we’ve done it over and over and we know how to bridge AI reasoning with physical experimentation.
There’s no reason not to collaborate. AI companies bring scale in reasoning; we bring grounding in biology and infrastructure. The combination could be powerful.
Zahra: What’s most overhyped right now in AI-driven biotech?
Molly: Foundational cell models. We don’t even agree on what a “cell model” means yet. It’s easy to over-promise. The potential impact is huge, but it will take years of hard work.
Zahra: And what’s under-appreciated?
Molly: The hard work itself - the grind of making these systems actually produce impact. It’s not about the next flashy model; it’s about doing the experiments, generating the data, and seeing them through.
Zahra: What topic or idea is fascinating you most right now?
Molly: Understanding how the brain itself works, and how that might inspire the next architectures for AI.
Maybe transformers aren’t the end, maybe we’ll find new neural designs closer to biology. We’re still learning how to build machines that learn the way nature does. That’s the next horizon.
Zahra: Thank you Molly, for the generous conversation, and for reminding us how much discovery still depends on curiosity, persistence, and the courage to build what doesn’t yet exist!
Would you like to contribute to Decoding Bio by writing a guest post? Drop us a note here or chat with us on Twitter: @decodingbio.