Where the Alpha Lives in AI x Bio
TLDR: we are looking for 4 defensible ingredients: deterministic, proprietary, integrated, and infra.
27/03/2026
Luka Nićin
Artificial intelligence in drug discovery moved from niche computational biology to the biggest buzzword in venture capital. Everyone has a deck, everyone is fine-tuning a transformer, and everyone promises to "revolutionize the discovery pipeline."
But let’s cut through the noise.
As investors, we are seeing dozens of pitches that fundamentally misunderstand where the value accrues in AI-driven biology. We are rapidly entering an era where certain AI capabilities are becoming table stakes, while others are creating significant moats. We have to know the difference between a commoditized wrapper and a true generational leap.
Here is a breakdown of what we believe is overcrowded, what is commoditized, and where we see the absolute largest opportunities in AI for biology.
The Commoditized Trap: What We Are Avoiding
Basic Structure Prediction Predicting a protein structure is no longer a business model, rather a utility. AlphaFold proved it was possible, and now open-source models have democratized it. "We predict 3D structures better" is playing a losing game against open-source momentum and deep-pocketed tech giants.
"Screening" vs True "Design" Historically, computational antibody design has been brutally inefficient, with experimental success rates routinely hovering below 0.1%. Startups that use AI simply to narrow down a massive library of millions of designs so a wet lab can run high-throughput screening are not truly designing drugs.. That is an overcrowded space with thin margins.
Generalist LLM Wrappers Using off-the-shelf Large Language Models to summarize PubMed articles or write basic Python scripts for bioinformatics is a feature, not a company. The barrier to entry is zero, and the defensibility is non-existent.
The Deep Alpha: Where the Largest Opportunities Lie
If the above is the noise, here is the signal. At Pace Ventures, we are aggressively looking at founders building in the following four spaces.
1. Deterministic Engineering Over Probabilistic Screening
We are moving away from the era of "shots on goal" and entering the era of deterministic molecular engineering. The real opportunity lies in rational design, models that generate drug candidates, skipping several steps of the in-vitro playbook.
That’s why we backed Ternary Therapeutics, a team of drug hunters from GSK and AI for drug discovery experts from BenevolentAI building a platform for rational design of molecular glues and other modalities inducing proximity.
Additionally, look at Chai Discovery, recently raised a $130M Series B at a $1.3B valuation. Their Chai-2 model is not predicting structures. It is designing full-length, drug-like monoclonal antibodies with state-of-the-art double-digit hit rates. They are successfully generating binders for historically "undruggable" membrane proteins like GPCRs (which account for a third of approved drugs, but almost zero antibodies).
Similarly, Boltz (backed by a $28M seed) has solved a massive bottleneck: binding affinity. Their Boltz-2 model predicts protein-ligand binding affinity at an accuracy approaching expensive, physics-based free-energy perturbation (FEP) simulations, but operates over 1,000x faster. Companies that can simultaneously optimize for structure, binding affinity, and manufacturability (developability) right out of the gate are the ones that will win.
Nevertheless, although both Chai and Boltz entered the stage with a blast, it’s early days and both companies need to demonstrate that they can deliver on those promises.
2. Proprietary, Multimodal Data Generation
Foundation models are only as good as the underlying training data. You cannot build a defensible AI biotech company solely by scraping the public Protein Data Bank (PDB). The next massive winners are building physical data factories to generate bespoke, multimodal human datasets that no one else has.
Take Noetik as a prime example. They realized that cancer's complexity exceeds human interpretation, so they built a wet-lab engine specifically to generate training data for foundation models. They are generating massive tissue microarrays capturing clinical H&E, spatial transcriptomics, custom protein panels, and DNA sequencing, spanning 1.5 petabytes of data across 250 million cells. Data generation at this scale is an impenetrable moat.
However, data accessibility is a major bottleneck, given that pharma is reluctant to share proprietary data to the outside world, even trusted partners. That’s why we’re convinced that novel digital infra is necessary to allow for federated learning environments, as being pioneered by our portfolio company Bitfount.
3. The Integrated Biology Environment
Biology has a severe fragmentation problem. Scientists lose countless hours bouncing between tissue culture rooms, outdated Excel templates, databases, and disparate software tools. There is a massive opportunity to build the biological equivalent of an IDE (Integrated Development Environment) for software engineers.
Phylo, backed by a16z and Menlo, is pioneering what they call the Integrated Biology Environment (IBE). Through their Biomni Lab, they deploy collaborative AI agents that sit seamlessly alongside biologists. Instead of a scientist executing tedious data plumbing, an AI-native biologist simply states a hypothesis, and the agent queries databases, runs multi-omics analyses, and can even dispatch instructions to robotic labs.
4. New Biotech Business Models: "Infrastructure as an Asset"
For decades, AI-biotech deals with Big Pharma were structured as service collaborations or co-development on single targets. That is changing. We are witnessing the birth of a new asset class: the licensing of biological models as enterprise infrastructure.
In January, Noetik signed a landmark $50M upfront deal with GSK not just for a drug asset, but for a direct, non-exclusive license to their virtual cell foundation models. Simultaneously, Boltz announced a strategic collaboration to integrate its agents and foundation models directly into Pfizer’s preclinical discovery workflows.
Companies that can monetize their AI infrastructure OR agentic workflows via SaaS-like subscriptions while retaining their own internal drug pipelines will command massive, tech-like multiples.
The Bottom Line
At Pace Ventures, we are ignoring the wrappers, the incremental computational screeners, and the generalist LLMs. We want to talk to founders who are building proprietary physical data moats, designing drugs with zero-shot atomic precision, automating the wet-lab/dry-lab workflow with autonomous agents, and rewriting the business model of biotech.
If you're building the deterministic future of biology, we want to hear from you. Let's engineer the future.
Artificial intelligence in drug discovery moved from niche computational biology to the biggest buzzword in venture capital. Everyone has a deck, everyone is fine-tuning a transformer, and everyone promises to "revolutionize the discovery pipeline."
But let’s cut through the noise.
As investors, we are seeing dozens of pitches that fundamentally misunderstand where the value accrues in AI-driven biology. We are rapidly entering an era where certain AI capabilities are becoming table stakes, while others are creating significant moats. We have to know the difference between a commoditized wrapper and a true generational leap.
Here is a breakdown of what we believe is overcrowded, what is commoditized, and where we see the absolute largest opportunities in AI for biology.
The Commoditized Trap: What We Are Avoiding
Basic Structure Prediction Predicting a protein structure is no longer a business model, rather a utility. AlphaFold proved it was possible, and now open-source models have democratized it. "We predict 3D structures better" is playing a losing game against open-source momentum and deep-pocketed tech giants.
"Screening" vs True "Design" Historically, computational antibody design has been brutally inefficient, with experimental success rates routinely hovering below 0.1%. Startups that use AI simply to narrow down a massive library of millions of designs so a wet lab can run high-throughput screening are not truly designing drugs.. That is an overcrowded space with thin margins.
Generalist LLM Wrappers Using off-the-shelf Large Language Models to summarize PubMed articles or write basic Python scripts for bioinformatics is a feature, not a company. The barrier to entry is zero, and the defensibility is non-existent.
The Deep Alpha: Where the Largest Opportunities Lie
If the above is the noise, here is the signal. At Pace Ventures, we are aggressively looking at founders building in the following four spaces.
1. Deterministic Engineering Over Probabilistic Screening
We are moving away from the era of "shots on goal" and entering the era of deterministic molecular engineering. The real opportunity lies in rational design, models that generate drug candidates, skipping several steps of the in-vitro playbook.
That’s why we backed Ternary Therapeutics, a team of drug hunters from GSK and AI for drug discovery experts from BenevolentAI building a platform for rational design of molecular glues and other modalities inducing proximity.
Additionally, look at Chai Discovery, recently raised a $130M Series B at a $1.3B valuation. Their Chai-2 model is not predicting structures. It is designing full-length, drug-like monoclonal antibodies with state-of-the-art double-digit hit rates. They are successfully generating binders for historically "undruggable" membrane proteins like GPCRs (which account for a third of approved drugs, but almost zero antibodies).
Similarly, Boltz (backed by a $28M seed) has solved a massive bottleneck: binding affinity. Their Boltz-2 model predicts protein-ligand binding affinity at an accuracy approaching expensive, physics-based free-energy perturbation (FEP) simulations, but operates over 1,000x faster. Companies that can simultaneously optimize for structure, binding affinity, and manufacturability (developability) right out of the gate are the ones that will win.
Nevertheless, although both Chai and Boltz entered the stage with a blast, it’s early days and both companies need to demonstrate that they can deliver on those promises.
2. Proprietary, Multimodal Data Generation
Foundation models are only as good as the underlying training data. You cannot build a defensible AI biotech company solely by scraping the public Protein Data Bank (PDB). The next massive winners are building physical data factories to generate bespoke, multimodal human datasets that no one else has.
Take Noetik as a prime example. They realized that cancer's complexity exceeds human interpretation, so they built a wet-lab engine specifically to generate training data for foundation models. They are generating massive tissue microarrays capturing clinical H&E, spatial transcriptomics, custom protein panels, and DNA sequencing, spanning 1.5 petabytes of data across 250 million cells. Data generation at this scale is an impenetrable moat.
However, data accessibility is a major bottleneck, given that pharma is reluctant to share proprietary data to the outside world, even trusted partners. That’s why we’re convinced that novel digital infra is necessary to allow for federated learning environments, as being pioneered by our portfolio company Bitfount.
3. The Integrated Biology Environment
Biology has a severe fragmentation problem. Scientists lose countless hours bouncing between tissue culture rooms, outdated Excel templates, databases, and disparate software tools. There is a massive opportunity to build the biological equivalent of an IDE (Integrated Development Environment) for software engineers.
Phylo, backed by a16z and Menlo, is pioneering what they call the Integrated Biology Environment (IBE). Through their Biomni Lab, they deploy collaborative AI agents that sit seamlessly alongside biologists. Instead of a scientist executing tedious data plumbing, an AI-native biologist simply states a hypothesis, and the agent queries databases, runs multi-omics analyses, and can even dispatch instructions to robotic labs.
4. New Biotech Business Models: "Infrastructure as an Asset"
For decades, AI-biotech deals with Big Pharma were structured as service collaborations or co-development on single targets. That is changing. We are witnessing the birth of a new asset class: the licensing of biological models as enterprise infrastructure.
In January, Noetik signed a landmark $50M upfront deal with GSK not just for a drug asset, but for a direct, non-exclusive license to their virtual cell foundation models. Simultaneously, Boltz announced a strategic collaboration to integrate its agents and foundation models directly into Pfizer’s preclinical discovery workflows.
Companies that can monetize their AI infrastructure OR agentic workflows via SaaS-like subscriptions while retaining their own internal drug pipelines will command massive, tech-like multiples.
The Bottom Line
At Pace Ventures, we are ignoring the wrappers, the incremental computational screeners, and the generalist LLMs. We want to talk to founders who are building proprietary physical data moats, designing drugs with zero-shot atomic precision, automating the wet-lab/dry-lab workflow with autonomous agents, and rewriting the business model of biotech.
If you're building the deterministic future of biology, we want to hear from you. Let's engineer the future.
Artificial intelligence in drug discovery moved from niche computational biology to the biggest buzzword in venture capital. Everyone has a deck, everyone is fine-tuning a transformer, and everyone promises to "revolutionize the discovery pipeline."
But let’s cut through the noise.
As investors, we are seeing dozens of pitches that fundamentally misunderstand where the value accrues in AI-driven biology. We are rapidly entering an era where certain AI capabilities are becoming table stakes, while others are creating significant moats. We have to know the difference between a commoditized wrapper and a true generational leap.
Here is a breakdown of what we believe is overcrowded, what is commoditized, and where we see the absolute largest opportunities in AI for biology.
The Commoditized Trap: What We Are Avoiding
Basic Structure Prediction Predicting a protein structure is no longer a business model, rather a utility. AlphaFold proved it was possible, and now open-source models have democratized it. "We predict 3D structures better" is playing a losing game against open-source momentum and deep-pocketed tech giants.
"Screening" vs True "Design" Historically, computational antibody design has been brutally inefficient, with experimental success rates routinely hovering below 0.1%. Startups that use AI simply to narrow down a massive library of millions of designs so a wet lab can run high-throughput screening are not truly designing drugs.. That is an overcrowded space with thin margins.
Generalist LLM Wrappers Using off-the-shelf Large Language Models to summarize PubMed articles or write basic Python scripts for bioinformatics is a feature, not a company. The barrier to entry is zero, and the defensibility is non-existent.
The Deep Alpha: Where the Largest Opportunities Lie
If the above is the noise, here is the signal. At Pace Ventures, we are aggressively looking at founders building in the following four spaces.
1. Deterministic Engineering Over Probabilistic Screening
We are moving away from the era of "shots on goal" and entering the era of deterministic molecular engineering. The real opportunity lies in rational design, models that generate drug candidates, skipping several steps of the in-vitro playbook.
That’s why we backed Ternary Therapeutics, a team of drug hunters from GSK and AI for drug discovery experts from BenevolentAI building a platform for rational design of molecular glues and other modalities inducing proximity.
Additionally, look at Chai Discovery, recently raised a $130M Series B at a $1.3B valuation. Their Chai-2 model is not predicting structures. It is designing full-length, drug-like monoclonal antibodies with state-of-the-art double-digit hit rates. They are successfully generating binders for historically "undruggable" membrane proteins like GPCRs (which account for a third of approved drugs, but almost zero antibodies).
Similarly, Boltz (backed by a $28M seed) has solved a massive bottleneck: binding affinity. Their Boltz-2 model predicts protein-ligand binding affinity at an accuracy approaching expensive, physics-based free-energy perturbation (FEP) simulations, but operates over 1,000x faster. Companies that can simultaneously optimize for structure, binding affinity, and manufacturability (developability) right out of the gate are the ones that will win.
Nevertheless, although both Chai and Boltz entered the stage with a blast, it’s early days and both companies need to demonstrate that they can deliver on those promises.
2. Proprietary, Multimodal Data Generation
Foundation models are only as good as the underlying training data. You cannot build a defensible AI biotech company solely by scraping the public Protein Data Bank (PDB). The next massive winners are building physical data factories to generate bespoke, multimodal human datasets that no one else has.
Take Noetik as a prime example. They realized that cancer's complexity exceeds human interpretation, so they built a wet-lab engine specifically to generate training data for foundation models. They are generating massive tissue microarrays capturing clinical H&E, spatial transcriptomics, custom protein panels, and DNA sequencing, spanning 1.5 petabytes of data across 250 million cells. Data generation at this scale is an impenetrable moat.
However, data accessibility is a major bottleneck, given that pharma is reluctant to share proprietary data to the outside world, even trusted partners. That’s why we’re convinced that novel digital infra is necessary to allow for federated learning environments, as being pioneered by our portfolio company Bitfount.
3. The Integrated Biology Environment
Biology has a severe fragmentation problem. Scientists lose countless hours bouncing between tissue culture rooms, outdated Excel templates, databases, and disparate software tools. There is a massive opportunity to build the biological equivalent of an IDE (Integrated Development Environment) for software engineers.
Phylo, backed by a16z and Menlo, is pioneering what they call the Integrated Biology Environment (IBE). Through their Biomni Lab, they deploy collaborative AI agents that sit seamlessly alongside biologists. Instead of a scientist executing tedious data plumbing, an AI-native biologist simply states a hypothesis, and the agent queries databases, runs multi-omics analyses, and can even dispatch instructions to robotic labs.
4. New Biotech Business Models: "Infrastructure as an Asset"
For decades, AI-biotech deals with Big Pharma were structured as service collaborations or co-development on single targets. That is changing. We are witnessing the birth of a new asset class: the licensing of biological models as enterprise infrastructure.
In January, Noetik signed a landmark $50M upfront deal with GSK not just for a drug asset, but for a direct, non-exclusive license to their virtual cell foundation models. Simultaneously, Boltz announced a strategic collaboration to integrate its agents and foundation models directly into Pfizer’s preclinical discovery workflows.
Companies that can monetize their AI infrastructure OR agentic workflows via SaaS-like subscriptions while retaining their own internal drug pipelines will command massive, tech-like multiples.
The Bottom Line
At Pace Ventures, we are ignoring the wrappers, the incremental computational screeners, and the generalist LLMs. We want to talk to founders who are building proprietary physical data moats, designing drugs with zero-shot atomic precision, automating the wet-lab/dry-lab workflow with autonomous agents, and rewriting the business model of biotech.
If you're building the deterministic future of biology, we want to hear from you. Let's engineer the future.