Drug discovery has a target problem.
Finding a molecule that binds to a protein is difficult. Designing a drug that is safe, stable and easy to manufacture is difficult too. But before any of that happens, scientists have to make a much more basic decision: what part of the disease should they try to change?
That choice can determine almost everything that follows.
Choose the wrong biological target and a company can spend years making an excellent drug for a mechanism that does not matter enough in real patients. Choose a target that looks important in a simplified lab model but is not important inside the human body, and even beautiful preclinical results may fail once clinical testing begins.
This is one reason drug development remains so difficult. A widely cited review estimated that roughly 90% of drugs entering clinical development fail, with weak understanding of disease biology and target selection among the major problems the industry still needs to solve.
Artificial intelligence is now being pushed much further upstream to attack this problem.
And Greater Boston may be one of the most interesting places to watch it happen.
Across Boston, Cambridge, Somerville, Lexington and Watertown, a group of biotechnology companies is using AI, machine learning, single-cell data, human patient records, imaging, structural biology and large-scale experiments to rethink how drug targets are found and validated.
Some are trying to identify previously hidden molecular targets.
Others are going further.
They are treating a cell state, a protein’s movement, a biomolecular condensate, an immune-cell interaction, or even the way a medicine travels through a living organism as the thing that needs to be understood and controlled.
That is a much bigger change than simply replacing a scientist’s spreadsheet with an AI model.
It changes what the word target can mean.
The Short Version: AI Is Expanding the Definition of a Drug Target
The traditional target-discovery process often starts with a fairly simple question: which gene or protein appears to be connected to a disease?
Scientists then gather evidence, test the biology, decide whether the target can be affected with a drug and begin searching for molecules.
AI can speed up that process by processing far more data than a person could examine manually. But Boston’s most interesting companies are using computation for something more ambitious.
They are asking whether the industry’s basic picture of disease is too simple.
Cellarity, for example, argues that some diseases cannot be understood by looking at one protein at a time. Its platform uses single-cell and multi-omics data to model dysfunctional cellular states and then predict interventions capable of moving those cells toward healthier behavior. The company says this approach led to the discovery of a previously unknown target involved in the globin-switching mechanism behind its clinical-stage sickle cell candidate CLY-124.
Valo Health approaches the problem from another direction. It starts with human data, including access to more than 17 million de-identified patient records, and uses AI, knowledge graphs, causal inference and statistical genetics to identify patient subgroups, biological pathways and potential therapeutic targets before moving into chemistry.
Dewpoint Therapeutics is trying to identify disease-driving biomolecular condensates. Its target-discovery process combines multi-omics, knowledge graphs, bioinformatics, AI and experimental assays to nominate and validate condensate-associated targets.
Relay Therapeutics looks at protein motion. Its Dynamo platform combines computational and experimental methods to study the moving shapes of proteins and use those insights to attack targets that were previously considered difficult to drug.
The result is a new target-discovery landscape.
The question is no longer only:
Which protein should we block?
The emerging questions are:
Which cell state should we change?
Which disease-driving molecular interaction matters most?
Which form of a protein should we hit?
Which condensate should we disrupt?
Which immune-cell relationship should we control?
Which tissue should a medicine reach?
And, perhaps most important, which of these mechanisms actually matters inside a human being?
Why Traditional Target Discovery Has Been So Hard
To understand why AI matters here, it helps to understand the old problem.
Biology produces an enormous amount of correlation.
A gene may be highly active in diseased tissue.
A protein may be present in much larger amounts in patients with a condition.
A mutation may appear more frequently in people who become sick.
A cell population may expand during disease.
Those signals are useful, but they do not automatically prove that changing the signal will improve the disease.
A fire produces smoke. Removing the smoke does not necessarily put out the fire.
Drug discovery faces the same problem.
Scientists have to distinguish a biological change that causes or maintains disease from one that simply appears because the disease already exists.
A Good Target Must Pass Several Tests
An attractive signal is not enough.
A useful drug target usually needs to sit at the intersection of several questions. Does changing it meaningfully affect disease? Is the effect large enough to matter? Can a drug reach it? Can the target be changed without unacceptable side effects? Is the biology similar enough between the experimental model and real patients?
AI can help connect these questions, but only when the model has access to meaningful data.
This is why Boston’s advantage may be less about having the biggest general-purpose AI models and more about its access to many different forms of biology.
The region combines hospitals, academic laboratories, sequencing centers, biotechnology companies, pharmaceutical research groups and major research institutions. The Broad Institute, for example, has helped develop large-scale single-cell genomics, perturbation technologies and resources such as the Human Cell Atlas, all of which make it easier to connect genes with specific cell types and disease states rather than studying biology only in bulk averages.
Perturb-seq, developed by researchers including scientists associated with the Broad, provides a good example of why this matters. The technique combines genetic perturbation with single-cell RNA sequencing, allowing researchers to change genes and then measure the resulting cellular effects across thousands of individual cells.
That transition—from observing biology to deliberately perturbing it—is extremely important for AI target discovery.
Observation tells a model what is associated with disease.
Perturbation begins to tell it what might actually matter.
BostonBizScene Original Research: Mapping Boston’s AI Target-Discovery Stack
To understand what is actually happening in Greater Boston, BostonBizScene reviewed eight biotechnology companies with substantial Boston-area operations and publicly described AI or machine-learning-enabled therapeutic platforms.
The sample includes Valo Health, Cellarity, Dewpoint Therapeutics, Relay Therapeutics, Montai Therapeutics, Manifold Bio, Generate and Seismic Therapeutic.
The goal was not to rank these companies.
Instead, we coded how each platform thinks about the basic unit of biological intervention.
Methodology
Companies were included if they met three conditions.
First, they needed headquarters or a major research presence in Greater Boston.
Second, AI, machine learning or advanced computation needed to be a meaningful part of the drug-discovery platform rather than simply a support tool.
Third, there had to be enough public technical information to understand the company’s approach.
We reviewed current company platform pages, pipeline disclosures and recent public updates available through October 6, 2026.
For each company, we coded five questions:
- What biological object does the platform primarily try to understand or control?
- What type of data appears central to the platform?
- Is computation connected to experimental feedback?
- Does the company move beyond the idea of a static single-protein target?
- Has a platform-derived or platform-supported therapeutic reached human clinical development?
The results are striking.
Table: Eight Boston-Area AI Biotech Platforms and What They Actually Target
| Company | Primary discovery or intervention object | Major data or experimental input | Role of AI/computation | Clinical evidence by Oct. 6, 2026 |
|---|---|---|---|---|
| Valo Health | Human disease pathways and therapeutic targets | 17M+ de-identified patient records, genetics, longitudinal health data | Patient stratification, knowledge graphs, causal inference, target prioritization | Platform has produced licensed preclinical programs |
| Cellarity | Diseased cell states and gene networks | Single-cell transcriptomics, multi-omics, human-relevant models | Models disease states and predicts interventions | CLY-124 in Phase 1 |
| Dewpoint Therapeutics | Disease-driving biomolecular condensates | Multi-omics, knowledge graphs, high-content imaging, cellular assays | Target nomination, phenotypic analysis, molecule discovery | DPTX3186 in Phase 1a/2a |
| Relay Therapeutics | Dynamic protein conformations | Structural biology, cryo-EM, molecular dynamics, experiments | Maps protein motion to reveal druggable opportunities | Zovegalisib in Phase 3 |
| Montai Therapeutics | Hard-to-drug biological pathways | Large chemical space, biological data, structural and molecular information | Finds chemistry for difficult validated mechanisms | MTAI-1025 in Phase 1 |
| Manifold Bio | Tissue targeting and in-vivo molecular behavior | Massively multiplexed living-system experiments | AI-guided protein design using direct in-vivo feedback | Preclinical platform/pipeline |
| Generate | Therapeutic protein sequence and function | Protein sequences, structures and proprietary experimental measurements | Generative protein design | GB-0895 in Phase 3 |
| Seismic Therapeutic | Immune mechanisms and multi-receptor interactions | Structural biology, protein engineering, translational immunology | Multi-property biologic design and optimization | Multiple clinical programs |
Sources: Valo, Cellarity, Dewpoint, Relay, Montai, Manifold Bio, Generate and Seismic Therapeutic.
Original Finding #1: Boston’s AI Target-Discovery Market Is No Longer Mainly About Finding Protein Names
The most important finding from our review is that five of the eight platforms studied—62.5%—meaningfully broaden the intervention unit beyond the traditional idea of a static single protein.
That does not mean proteins are disappearing.
Proteins remain central to medicine.
What is changing is the level at which companies choose to understand the disease.
Chart: Boston Platforms Expanding the Meaning of a Target
| Target concept in BostonBizScene sample | Companies | Share |
|---|---|---|
| Broader than a static single-protein target | 5 | 62.5% |
| Primarily conventional target or molecule-centered | 3 | 37.5% |
Broader intervention models identified: cell states, condensates, protein conformations, tissue-level behavior and multi-cell immune mechanisms.
That is a significant shift.
Cellarity is not simply asking which gene is abnormal. It models the wider network state of the cell.
Dewpoint is not simply asking whether a protein exists in disease. It asks whether groups of molecules are forming abnormal condensates that alter cell behavior.
Relay is not treating a protein as a frozen object. It studies how the protein moves and changes shape.
Manifold Bio pushes the problem out another level by asking how engineered proteins behave inside an entire living organism.
Seismic designs molecules around coordinated immune mechanisms rather than treating optimization as a simple one-target, one-property exercise.
Taken together, this suggests that one of Boston’s most important contributions to AI drug discovery may be target abstraction.
AI allows researchers to model disease at a higher level than was practical before.
Cellarity: What If the Disease State Is the Target?
Cellarity offers one of the clearest examples.
Traditional drug discovery often starts by identifying a protein believed to cause disease and then looking for a molecule that changes that protein.
Cellarity starts with the behavior of the cell.
Its platform uses high-dimensional transcriptomic and multi-omics data to build a picture of how genes interact inside healthy and diseased cells. AI models then attempt to understand which mechanisms distinguish the unhealthy state and predict chemical interventions capable of shifting the system toward healthier function.
This is an important conceptual change.
A disease can involve hundreds or thousands of connected biological changes.
Selecting one target too early can force scientists to simplify that network before they fully understand it.
A cell-state model tries to preserve more of that complexity.
CLY-124 Shows Why This Matters
Cellarity says its discovery platform identified an undiscovered target controlling globin switching, leading to CLY-124, an oral medicine being studied for sickle cell disease. The candidate entered Phase 1 testing in 2025, and Cellarity’s 2026 priorities included advancing the study and reporting patient data.
That example is more important than the words “AI-powered platform.”
It gives us a measurable test of the concept.
The model was not simply used to optimize a molecule after a human scientist chose the mechanism.
The computational system helped reveal the mechanism.
That is what genuine AI-powered target discovery should look like.
Valo Health: Start With Patients and Work Backward
Valo takes a different route.
Its approach begins with human disease data rather than beginning with a molecule.
The company says it has access to more than 17 million de-identified patient records, with some patient histories spanning two or three decades. AI and knowledge graphs are used to identify patient subgroups and map those groups to biological pathways. Causal inference and statistical genetics are then used to identify possible intervention targets.
The distinction between correlation and causality matters here.
A model can easily find thousands of patterns in a large healthcare dataset.
That does not mean all of those patterns should become drug programs.
Valo’s stated approach is therefore particularly interesting because it explicitly combines machine learning with techniques intended to get closer to causal biology.
Human Data Can Change the Order of Drug Discovery
The traditional sequence often looks something like this:
laboratory hypothesis → animal model → drug candidate → human testing
Human-first discovery attempts to reverse part of that flow:
human disease evidence → causal hypothesis → target → experiment → candidate
That may reduce one of the most expensive forms of uncertainty: whether the mechanism that looked important in the lab actually matters in human disease.
Valo’s collaboration with Novo Nordisk offers evidence that this approach has moved beyond theory. Their agreement included discovery work using Valo’s human dataset and AI capabilities, while Novo licensed three preclinical cardiovascular programs developed through the platform.
For Boston founders, this points toward an important strategic lesson.
The most valuable dataset may not be the biggest dataset.
It may be the dataset closest to the final biological question.
Dewpoint Therapeutics: Finding Targets Inside Condensates
Dewpoint is working on a different layer of biology.
Cells are not simply bags containing freely floating proteins.
Many molecules organize into temporary compartments known as biomolecular condensates. These structures help control important functions such as gene regulation and cellular signaling.
When condensate behavior goes wrong, disease can follow.
Dewpoint’s thesis is that abnormal condensates may expose therapeutic opportunities that older target-discovery methods miss.
Its platform combines disease-relevant cellular models, high-content imaging, automation and computational analysis to discover and optimize molecules capable of changing condensate behavior.
More importantly for target discovery, Dewpoint says its nomination process integrates multi-omics, bioinformatics, AI and knowledge graphs containing disease, sequence and mechanistic information. Candidate condensate targets are then tested experimentally in disease-relevant models.
The Target Is Becoming a Biological Context
This illustrates another change AI is enabling.
A protein can behave differently depending on what it is interacting with, where it is located and what physical state surrounds it.
The “target” therefore becomes more than the protein’s name.
The target may be a specific behavior of that protein inside a disease-linked molecular environment.
Dewpoint’s lead program DPTX3186 entered a Phase 1a/2a study in 2026. The drug is designed to modulate beta-catenin signaling by changing its behavior in biomolecular condensates, providing an early clinical test of the approach.
Relay Therapeutics: A Protein Is Not a Photograph
Relay’s Dynamo platform attacks yet another blind spot.
Proteins move.
They bend, rotate, open, close and shift between different shapes.
Traditional structural drug discovery has often relied heavily on static pictures of those proteins. Relay’s approach combines experiments with computational modeling to understand the different conformations a protein can adopt.
Why does that matter for target discovery?
Because a target that appears difficult to drug in one shape may expose a useful pocket in another.
The opportunity may not be “PI3Kα.”
The opportunity may be a particular moving state of PI3Kα that is present in mutant disease biology but less important in the healthy form.
Relay’s zovegalisib program shows how this can translate into drug design. The company says it solved the full-length cryo-EM structure of PI3Kα and used long-timescale molecular dynamics simulations to study conformational differences between mutant and normal versions of the protein. Those insights supported the design of zovegalisib, which is now being evaluated in Phase 3 breast-cancer testing.
This turns computational biology into a search tool for hidden druggability.
That may be one of the largest commercial opportunities in AI drug discovery.
Instead of competing only for the targets everyone already knows, companies can attempt to reveal new ways of attacking biology that previously appeared unusable.
Original Finding #2: Boston Is Building Systems That Link Prediction With Experiment
There is a second pattern across the dataset.
Every company in our eight-company sample publicly describes some form of connection between computational prediction and experimental evidence.
The exact loops are very different.
But the underlying architecture is remarkably consistent.
Chart: Computational Models Connected to Experimental Feedback
| Platform architecture | Companies in sample | Share |
|---|---|---|
| Computation tied to experimental or biological validation | 8 | 100% |
| Computation presented mainly as a stand-alone prediction layer | 0 | 0% |
This should not be read as proof that every platform is equally strong.
It tells us something different.
Within this Boston sample, serious AI-biotech companies are not betting on software alone.
They are building learning systems.
A prediction creates an experiment.
The experiment creates new data.
That data changes the model.
The improved model chooses the next experiment.
This loop may be much harder to copy than the model itself.
Generate Makes the Loop Explicit
Generate describes its platform as a cycle of generate, build, measure and learn.
Machine-learning models design protein sequences. Those proteins are physically created. Experimental systems measure their properties. The resulting data is fed back into the platform to improve future designs.
Generate says its models have learned from hundreds of millions of known proteins and can create proteins with chosen therapeutic properties.
This company is better described as a generative medicine platform than a pure target-discovery company.
That difference is useful.
It shows why target discovery and molecule design are beginning to merge.
Once a system can rapidly produce many molecules against a biological hypothesis, scientists can use those molecules as experimental probes.
The model is not just building a drug.
It is helping scientists ask biology questions.
Manifold Bio Pushes the Learning Loop Into Living Systems
Manifold Bio takes that idea even further.
Many AI models are trained using experiments performed in test tubes or cell systems. Those experiments are easier to run at scale, but the human body is far more complicated.
A molecule that looks excellent in vitro can behave badly once it encounters metabolism, blood flow, tissue barriers and the immune system.
Manifold’s approach is therefore built around direct in-vivo measurement.
The Boston company combines AI-guided protein design with molecular barcoding and massively parallel testing inside living systems. It says its newer infrastructure can measure more than 500,000 binders in a single in-vivo experiment.
That number deserves attention.
Chart: Selected Public Scale Signals From Boston AI-Biotech Platforms
| Company | Public scale signal | What it represents |
|---|---|---|
| Montai | 3B+ compounds | Diverse chemical search library |
| Generate | Hundreds of millions of known proteins analyzed | Protein sequence/structure learning space |
| Valo Health | 17M+ patient records | Human disease evidence |
| Manifold Bio | 500,000+ binders measured in one in-vivo experiment | Experimental throughput |
These numbers are not directly comparable because they measure completely different things. They are included to show the scale at which modern Boston platforms can now explore chemistry, human biology, proteins and living-system behavior.
Manifold’s model is particularly interesting for target discovery because each engineered binder can become both a possible therapeutic component and an experiment.
A binder that unexpectedly travels to a certain tissue may reveal useful biology.
A molecule that binds but fails to distribute may expose a delivery constraint.
A successful tissue-specific binder may reveal a new path for reaching biology that had previously been difficult to access.
The boundary between discovering a target and discovering how to reach the target starts to disappear.
Original Finding #3: Experimental Throughput May Matter More Than Model Size
Much of the public conversation around AI focuses on model size.
That may be the wrong metric for biotech.
A drug-discovery model ultimately needs to learn things about the physical world.
Biology can provide that feedback, but experiments are expensive.
That means a company’s long-term advantage may come from how rapidly it can produce reliable biological measurements.
Consider three examples.
Generate continuously builds and tests AI-designed proteins.
Cellarity describes a lab-in-the-loop approach in which predictions are tested in human-relevant experimental models.
Manifold can test massive numbers of engineered molecules directly in living systems.
The strategic implication is important.
A competitor may eventually gain access to similar algorithms.
It is much harder to reproduce years of proprietary experiments created by a tightly integrated AI-and-lab system.
The Real Moat Can Be the Data-Producing Machine
This changes how investors and founders should evaluate AI-biotech companies.
Instead of asking only, “How sophisticated is the AI?”
Ask:
How quickly can the company test its predictions?
How close are those experiments to real human biology?
How much proprietary information does each experiment generate?
Can the experimental results be fed back into the system automatically?
Does each new program make the platform smarter for future programs?
Those questions get much closer to whether an AI-biotech company is building a durable advantage.
Montai Therapeutics: Sometimes the Target Is Known but Still Unusable
AI-powered target discovery does not always mean discovering a completely new biological target.
Sometimes scientists already know the mechanism matters.
The problem is that nobody has found a practical way to drug it.
Montai Therapeutics is a useful example.
The Cambridge company focuses on difficult biology in inflammation and immunology. Its platform searches a very large and diverse chemical space, with the goal of finding oral small molecules capable of acting on pathways that have historically been difficult to reach with conventional chemistry. The company says its library spans more than three billion compounds and reports a ninefold hit rate compared with leading public machine-learning models in that chemical space.
Its first clinical candidate, MTAI-1025, entered Phase 1 in 2026 and targets NRF2 in ulcerative colitis.
That is not target discovery in the narrowest sense.
NRF2 was already known.
But it represents something equally valuable: target enablement.
Target Discovery and Target Tractability Are Converging
In the old model, companies could divide work into neat stages.
First discover a target.
Then determine whether it is druggable.
Then find chemistry.
AI platforms increasingly examine those questions together.
A biological mechanism that looks promising but has no workable chemistry may rank lower.
A target that initially seems difficult may become attractive if AI uncovers a new binding strategy.
This means future target-discovery systems may optimize for three things simultaneously:
Is the biology important?
Can it be changed?
Can it be changed with a medicine that patients can actually use?
That is a far more useful definition of a target.
Seismic Therapeutic: Designing Around the Immune System as a Network
Seismic’s work shows the same shift in biologics.
The company’s IMPACT platform combines machine learning with structural biology, protein engineering and translational immunology. Rather than optimizing one property at a time, the company uses computational methods to optimize multiple features of a therapeutic protein together.
That includes biological function, binding behavior, manufacturability and the risk that the immune system itself reacts to the medicine.
Its S-4321 program is particularly interesting because the molecule is designed to engage both PD-1 and FcγRIIb inhibitory pathways, affecting different parts of the immune system through one therapeutic design. Seismic’s current pipeline lists S-4321 in Phase 1b development.
Another program, S-1117, entered Phase 1 in 2025 after being engineered through Seismic’s machine-learning platform to reduce potentially immunogenic epitopes while maintaining its biological function.
Once again, the AI is not simply searching for an isolated target.
It is helping design an intervention around a biological system.
Original Finding #4: 75% of the Boston Platforms Studied Have Reached Human Testing
AI drug discovery has often faced an obvious criticism.
A beautiful model is not a medicine.
The real test begins when platform-generated ideas move into humans.
In our eight-company sample, six companies—75%—had at least one publicly disclosed clinical-stage program by October 6, 2026.
Chart: Clinical Translation in the BostonBizScene Sample
| Status | Companies | Share |
|---|---|---|
| At least one clinical-stage program | 6 | 75% |
| No publicly identified clinical-stage program in our review | 2 | 25% |
The six clinical-stage companies in this sample are Cellarity, Dewpoint Therapeutics, Relay Therapeutics, Montai Therapeutics, Generate and Seismic Therapeutic.
This does not prove that AI target discovery works.
Clinical-stage drugs still fail.
Some of these targets were known before the companies’ platforms were developed.
And the role of AI differs significantly from company to company.
But the data does show something important.
Boston’s AI-biotech ecosystem is moving beyond platform demonstrations.
Real medicines are now being used to test whether these computational discovery systems produce better biological decisions.
That is where the next several years become extremely important.
The Broad Institute Helps Explain Why This Is Happening in Boston
Boston’s biotech AI story cannot be explained only by startups.
The research infrastructure underneath them matters.
The Broad Institute has been deeply involved in areas such as large-scale genomics, single-cell biology, human genetics, perturbation biology, machine learning and biological data infrastructure.
The Human Cell Atlas is a useful example.
Traditional tissue measurements often average many different cells together. That can hide rare but important cell populations.
Single-cell methods allow scientists to see which genes are active in specific cell types and states. The Human Cell Atlas aims to map these cells across the human body, creating a reference that can help researchers identify disease-linked cell types and therapeutic opportunities.
The Broad’s Single Cell Portal currently contains large collections of these datasets, including Human Cell Atlas studies spanning hundreds of thousands of cells in individual projects.
AI Gets More Useful When Biology Becomes More Measurable
This may sound obvious, but it is one of the most important points in the entire AI-biotech story.
AI cannot discover biology that nobody has measured.
Every new experimental technology expands the information available to the model.
Single-cell sequencing allows models to study individual cell types.
Spatial biology adds information about where those cells are located.
Proteomics measures proteins.
High-content imaging captures cellular shape and behavior.
Perturbation experiments show what happens when specific genes or pathways are changed.
Human medical records reveal disease progression.
Genetics provides naturally occurring experiments across populations.
Structural tools reveal the physical forms of proteins.
In-vivo screening shows what actually happens inside an organism.
Boston has unusually dense access to many of these layers.
That is why the region can compete even if the largest general AI companies are located elsewhere.
Original Finding #5: Boston’s Advantage May Be Biological Data Density, Not Raw AI Talent Alone
Boston will probably never beat Silicon Valley by trying to become Silicon Valley.
It does not need to.
The region’s advantage is the density of biological systems that can generate useful evidence.
Our eight-company sample contains platforms built around:
human longitudinal records,
single-cell biology,
multi-omics,
molecular imaging,
structural biology,
protein dynamics,
large chemical libraries,
protein sequence data,
high-throughput laboratory experimentation,
and direct in-vivo screening.
The important point is not the length of that list.
It is that many of these data types exist within a small geographic cluster where biotechnology companies, hospitals, academic research groups and pharmaceutical scientists regularly exchange talent and ideas.
More Data Is Not the Same as Better Data
A company can collect billions of low-value measurements.
That does not guarantee useful target discovery.
The best target datasets answer a decision.
For example:
Does changing this gene move a diseased cell toward a healthy state?
Does this target matter in real patients?
Does a molecule bind only the disease-driving form of the protein?
Does the intervention work inside living tissue?
Does the effect reproduce across different biological models?
These questions generate much more valuable information than simply adding more rows to a database.
Boston’s next biotech advantage may therefore come from building decision-grade biological datasets.
Why Causal Biology Will Become More Important
The growth of AI will make correlation cheaper.
That sounds positive, but it creates a new problem.
If a model can identify millions of possible relationships, researchers still need to know which relationships are worth acting on.
That increases the value of causal evidence.
Human genetics can help.
Longitudinal patient records can help.
Carefully designed perturbation experiments can help.
Interventional studies can help.
The strongest platforms will combine several types of evidence.
From Prediction to Intervention
A useful mental model is to divide AI target discovery into four levels.
Level One: Association
The model finds a connection between a biological feature and disease.
This is useful for generating hypotheses, but it is also where false leads can multiply quickly.
Level Two: Mechanism
The model helps explain how the biological feature may influence disease.
Now the evidence becomes stronger.
Level Three: Perturbation
Researchers deliberately change the suspected mechanism and observe the result.
This is where target validation becomes much more convincing.
Level Four: Human Translation
The intervention produces the expected biological effect in people.
This is the hardest and most valuable test.
The AI-biotech winners will not simply generate the largest number of Level One predictions.
They will build systems that move ideas through all four levels more reliably.
What Boston Biotechs Should Measure Instead of “Targets Discovered”
AI companies often need simple numbers for investor decks.
“10,000 targets identified” sounds impressive.
It may mean almost nothing.
A target list is not a pipeline.
A much better set of performance measures would follow the probability that a target survives deeper testing.
Table: Better KPIs for AI-Powered Target Discovery
| Weak metric | Better metric |
|---|---|
| Number of targets generated | Percentage surviving independent biological validation |
| Number of model predictions | Prospective prediction success rate |
| Model benchmark score | Accuracy on unseen experimental outcomes |
| Size of dataset | Amount of decision-relevant proprietary data created |
| Number of compounds screened | Validated hits per experiment |
| Number of targets nominated | Percentage showing desired phenotype after perturbation |
| Speed of computation | Time from target hypothesis to experimental answer |
| Preclinical programs started | Programs reaching human proof of mechanism |
This is especially important because retrospective AI analysis can look impressive.
A model may perform well when scientists already know the answer.
Prospective prediction is harder.
The model has to make a decision before the experiment is run.
That is the test Boston companies should increasingly report.
What Investors Should Ask AI Target-Discovery Startups
The target-discovery market is crowded with companies that use similar language.
AI-powered.
Data-driven.
Platform.
Foundation model.
Closed loop.
Those words no longer provide enough information.
Investors need to understand what the system actually does.
Start with the data.
Where does it come from? Is the company buying it, licensing it or generating it internally? Does the company have exclusive access? Does every experiment create information competitors cannot easily obtain?
Then examine validation.
Has the system made prospective predictions that were later confirmed experimentally?
Next, look at biology.
Is the platform discovering new mechanisms, or merely optimizing molecules against known targets?
Both can create value. They are simply different businesses.
Finally, look at clinical translation.
A platform that can repeatedly produce human-validated mechanisms may become dramatically more valuable than one producing large volumes of preclinical candidates.
What Founders Should Learn From the Boston Model
The strongest message from Boston is that an AI biotech company should not begin by asking, “Where can we use AI?”
That is backwards.
Start with the biological bottleneck.
What important drug-discovery decision is currently slow, expensive or unreliable?
Then determine what evidence would make that decision better.
Then build the data-generation system.
Only after that should the team decide what type of AI is required.
This order matters.
Build Around the Experiment, Not the Demo
A model demo can be built quickly.
A biological learning system takes years.
But the learning system creates the stronger company.
If every experiment produces proprietary information, each program improves the next.
If the platform only depends on public data and generally available models, competitors can close the gap much faster.
The long-term value is not simply the software.
It is the system that continually produces better biological evidence.
Where the Next Generation of Targets May Come From
The next major wave of drug targets may not resemble the targets that built the previous generation of pharmaceutical companies.
Several areas look particularly important.
Cell States
Instead of attacking one molecule, companies can attempt to move whole cellular programs from disease toward health.
Cellarity is the clearest Boston example.
Protein Conformations
AI and simulation can reveal temporary protein shapes and hidden pockets that traditional structural methods miss.
Relay demonstrates how this can expand the druggable space.
Biomolecular Condensates
Disease-relevant condensates create a new level of biological organization that older target frameworks did not capture.
Dewpoint is pushing this category toward clinical testing.
Multi-Protein and Multi-Cell Mechanisms
Immune disease often emerges from networks of interacting receptors and cells.
Seismic’s multi-functional biologics show how a therapeutic can be designed around these relationships.
Tissue-Level Targeting
A perfect drug is useless if it cannot reach the right cells.
Manifold’s direct-to-vivo platform suggests that delivery behavior itself can become a computational discovery problem.
Causal Human Biology
Valo’s approach points toward another major shift: using real human evidence to decide which biological mechanisms deserve investment before large chemistry programs begin.
These categories are not mutually exclusive.
The most powerful future systems may combine several of them.
The Next Big Opportunity: Joining Target Discovery With Target Validation
Historically, target discovery and validation have often been treated as separate stages.
AI may collapse them into one process.
A model proposes a mechanism.
A perturbation experiment tests it.
Single-cell measurements show which cells changed.
Imaging measures phenotype.
Multi-omics reveals the downstream response.
The model updates.
Another experiment begins.
Instead of producing a list of targets once per year, the platform continuously learns which biological hypotheses survive contact with reality.
That is a very different operating model for biotech.
Original Finding #6: Boston’s Real Target-Discovery Moat May Be Intervention Density
There is another way to understand the Boston ecosystem.
Think about the number of ways scientists in the region can experimentally interfere with biology.
They can knock genes out.
Turn genes up.
Change protein sequences.
Create small molecules.
Modify protein interactions.
Alter cell states.
Measure single cells.
Image cellular structures.
Screen molecules.
Test engineered proteins in animals.
Compare predictions with human medical records.
Each method creates a different type of intervention.
We call this intervention density.
It is not a standard industry measure.
But it may become useful.
Why Intervention Density Matters
An AI system trained only on observations learns what biology looks like.
A system connected to many interventions begins to learn how biology responds when something changes.
That difference is fundamental.
If Boston can combine its computational talent with unusually dense access to laboratories, hospitals, sequencing infrastructure and biological engineering, it could build models that are not merely descriptive.
They could become increasingly predictive about what scientists should do next.
That is much closer to the real goal of drug discovery.
The Biggest Risk: Confusing Better Prediction With Better Biology
AI can still fail badly.
The model may learn bias in the dataset.
The experimental system may not reproduce human biology.
The company may optimize a measurable laboratory endpoint that has little connection to clinical benefit.
The algorithm may become excellent at predicting an assay that itself is misleading.
This is why target discovery cannot become a purely computational discipline.
Every prediction must eventually collide with biology.
The best companies will design those collisions deliberately.
Human Validation Remains the Final Judge
Clinical testing will remain slow.
It will remain expensive.
It will remain uncertain.
AI will not change that overnight.
But if AI improves the quality of decisions made before clinical testing begins, the impact could still be enormous.
A modest increase in the probability that a chosen target is biologically correct can create significant value when the alternative is spending hundreds of millions of dollars developing the wrong mechanism.
What Pharmaceutical Companies Should Do Differently
Large drug companies should not treat AI target discovery as a software purchase.
The greater opportunity is to redesign discovery around learning loops.
That means connecting human data, target nomination, experiments, chemistry and clinical information instead of keeping them in separate systems.
The question should not be:
“Which AI tool should our researchers use?”
A better question is:
“Where does information currently get lost between biological observation and the next experiment?”
Fixing those gaps may create more value than purchasing another model.
Boston’s AI Target-Discovery Stack Is Starting to Look Like a New R&D Architecture
The eight companies examined in our analysis are very different.
Valo begins with human disease.
Cellarity begins with cellular state.
Dewpoint focuses on condensates.
Relay focuses on molecular motion.
Montai explores chemistry around difficult biology.
Generate programs proteins.
Manifold measures engineered molecules inside living systems.
Seismic integrates machine learning with complex immunology and biologic design.
At first glance, they may look like separate categories.
Together, however, they form something that resembles an emerging R&D stack.
Layer 1: Human Disease Evidence
Which mechanisms appear to matter in real people?
Layer 2: Cellular and Molecular State
Where does biology diverge from health?
Layer 3: Causal Perturbation
What happens when scientists deliberately change that biology?
Layer 4: Target Tractability
Can the mechanism be controlled with a useful medicine?
Layer 5: Molecule Design
What sequence or chemical structure can create the desired effect?
Layer 6: Living-System Validation
Does the molecule reach the right place and behave correctly inside an organism?
Layer 7: Clinical Proof
Does changing the mechanism help people?
No single Boston company currently owns every layer.
That creates opportunity.
The most valuable partnerships may increasingly occur between companies controlling complementary pieces of this stack.
Final Takeaway: The Next Great Drug Targets May Not Look Like Traditional Targets
AI-powered target discovery is often described as a search problem.
Feed large amounts of biology into a model.
Generate promising proteins.
Rank them.
Choose the top candidates.
That description is becoming outdated.
Boston’s biotech ecosystem suggests a much bigger transformation.
The next generation of drug discovery may not simply produce better lists of proteins.
It may reveal new levels of biology that can be controlled.
A diseased cell state can become the target.
A temporary protein conformation can become the target.
A biomolecular condensate can become the target.
A relationship between immune cells can become the target.
A tissue-specific delivery route can become the target.
A causal pathway identified in millions of patient records can become the target.
That shift matters because biology is not built from isolated parts.
Disease emerges from systems.
AI gives scientists new ways to measure those systems, connect them and decide where intervention is most likely to work.
But computation alone will not determine the winners.
The companies with the strongest position will be those that connect prediction to experiment, experiment to proprietary data, proprietary data to better models and better models to medicines that work in real patients.
That is why Boston is such an important place to watch.
Its competitive advantage is not simply that it has AI researchers.
It is that AI sits next to some of the world’s densest infrastructure for measuring and changing biology.
And in drug discovery, the ability to run the next meaningful experiment may ultimately matter far more than the ability to generate the next million predictions.
