Artificial intelligence has become very good at finding patterns.
Give a modern AI system enough molecules, proteins, patient records, genetic data, or microscope images, and it can often find relationships that would take scientists much longer to see. It can predict which molecule may bind to a protein. It can separate patients into groups. It can suggest new proteins or chemical structures. It can even find biological signals hidden inside millions of data points.
But drug discovery has a problem that pattern recognition alone cannot solve.
A pattern tells you that two things appear together. A drug company needs to know what happens when it changes something.
If a gene is highly active in people with a disease, does changing that gene improve the disease? If a protein appears alongside tumor growth, is the protein driving the cancer or simply responding to it? If patients with one biological marker respond better to a treatment, did the marker affect the response, or is another hidden factor responsible?
Those are causal questions.
And they matter because making a medicine is ultimately an intervention.
A company does not get paid for predicting disease accurately. It needs to change biology in a useful way.
That difference is pushing one of the most important shifts happening in AI-powered biotech today: the move from systems that mainly recognize patterns toward systems designed to understand interventions, mechanisms, experiments, and cause-and-effect relationships.
The term often used for this emerging field is causal AI.
The idea is still young. There is no single technical approach that deserves the label, and companies use different language to describe similar ideas. Some talk directly about causal inference. Others talk about perturbations, cell-state changes, human genetics, mechanistic models, active learning, lab-in-the-loop systems, or autonomous experimentation.
Boston may be one of the best places to watch this transition.
Companies across Cambridge, Somerville, Lexington, Watertown, and Boston are connecting AI models to human data, biological experiments, protein engineering, chemistry, animal studies, and increasingly clinical development. Some are explicitly trying to uncover causal biology. Others are building experimental systems that let AI propose an intervention, observe what happens, and learn from the result.
That distinction matters.
Our analysis of seven Boston-area AI-biotech companies found that all seven publicly describe platforms centered on designing or testing interventions rather than simply predicting labels or outcomes. Five explicitly describe an iterative experimental feedback system. Six already show some form of external translation through clinical programs, licensing, or major pharmaceutical partnerships.
Yet only one of the seven clearly and repeatedly describes its platform using explicit causal-inference language.
That is the bigger story.
Boston biotech may be becoming causal in practice faster than it is becoming “causal AI” in name.
The Short Version: Drug Discovery Does Not Need Better Correlations Alone
Traditional machine learning is extremely useful when the question looks like this:
What is likely to happen?
Causal methods become important when the question changes to:
What would happen if we changed something?
That difference sounds small. In drug discovery, it is enormous.
A predictive model might learn that patients with high levels of Protein X tend to have severe disease. But that does not prove that reducing Protein X will improve their condition. Protein X could be a consequence of the disease rather than its cause. Or both Protein X and the disease could be driven by another biological process.
Causal reasoning asks a harder question.
If we intervene on Protein X while other factors are accounted for, does the disease actually change?
Medical researchers have dealt with this problem long before modern AI existed. Randomized controlled trials are powerful largely because random assignment helps separate treatment effects from hidden differences between groups. Researchers also use genetics, carefully designed laboratory experiments, natural experiments, longitudinal data, and statistical methods to make stronger causal claims.
Modern AI does not make those principles obsolete.
It gives scientists new tools for applying them to far larger and more complex datasets.
A major review of causal machine learning in healthcare explains the difference clearly: conventional machine-learning systems often learn associations, while causal methods try to reason about interventions and counterfactual outcomes. The authors divide the field into areas such as causal representation learning, causal discovery, and causal reasoning.
A 2024 Nature Medicine perspective similarly argues that causal machine learning can help estimate treatment outcomes and individual treatment effects, while warning that biased data and incorrect assumptions can still lead to incorrect conclusions.
For drug discovery, this means the next AI advantage may not come from predicting existing data slightly better.
It may come from knowing which experiment to run, which biological lever to pull, and what can reasonably be expected to happen when it is pulled.
Pattern Recognition Was a Necessary First Step
It would be a mistake to talk about pattern recognition as if it failed.
It has been enormously useful.
Machine learning can search chemical space, predict molecular properties, analyze images, detect biomarkers, estimate protein structure, prioritize targets, find drug-target interactions, and generate new molecules. Reviews of AI in drug discovery show that ML is now used across target discovery, screening, molecular design, toxicity prediction, repurposing, and other stages of research.
The limitation appears when prediction is mistaken for biological understanding.
A strong prediction can still be scientifically misleading
Imagine a dataset containing thousands of cancer samples.
A model discovers that the activity of Gene A is strongly linked with aggressive tumors.
That could mean Gene A causes aggressive growth.
It could also mean aggressive tumors cause Gene A to become active.
Or Gene A and tumor growth could both be responding to Gene B.
A predictive model does not automatically know the difference.
It may not need to know the difference if its job is simply to classify tumors. But a pharmaceutical company considering Gene A as a drug target absolutely needs to know.
The question has moved from:
Can Gene A predict aggressive cancer?
to:
Will changing Gene A change aggressive cancer?
That second question is closer to what matters commercially.
Drug discovery is an intervention business
A therapy changes something.
It blocks an enzyme. Activates a receptor. Replaces a missing protein. Changes gene expression. Reprograms an immune cell. Delivers genetic material. Alters a disease pathway.
This is why causality matters so much in pharmaceutical research.
A review of machine learning in drug development makes the point directly: target identification ultimately requires establishing that changing a biological target affects disease, whether through genetic evidence or experimental intervention.
A model that only predicts what biology currently looks like can be valuable.
A model that helps predict how biology responds when scientists deliberately change it could be much more valuable.
What Causal AI Actually Means
Causal AI can sound more mysterious than it is.
At a practical level, the goal is to separate association from intervention.
Suppose:
A → B
If changing A reliably changes B, A may have a causal effect on B.
Now imagine:
C → A
C → B
A and B may move together simply because C affects both.
That is called confounding.
Healthcare data is full of this problem.
Patients receiving one treatment may be older. People receiving another may have more severe disease. Hospitals may follow different treatment rules. Genetic ancestry, income, geography, prior treatments, lifestyle, and dozens of other variables can affect outcomes.
AI can find patterns inside this mess.
But without careful causal thinking, it can also learn the mess.
Causal AI has several forms
Researchers are working on several related problems.
Causal discovery tries to learn which variables may cause others.
Treatment-effect estimation asks how much an intervention changes an outcome.
Counterfactual reasoning asks what might have happened to the same person or biological system under a different intervention.
Causal representation learning tries to find meaningful causal variables inside high-dimensional data such as images or molecular measurements.
Experimental design and active learning choose the next experiment partly to reduce uncertainty about the system.
These approaches should not be treated as interchangeable.
And causal AI does not magically prove causation.
Experiments remain extremely important. A recent review focused specifically on methodological and regulatory issues around causal AI in drug development stresses that causal inference requires clear assumptions, careful uncertainty estimates, and rigorous validation.
That is especially important in biotech, where a false causal conclusion can lead to years of work on the wrong target.
Why Boston Is a Natural Testing Ground
Boston has an unusual combination of ingredients needed for causal AI.
It has large research hospitals, human genetics programs, academic biology, drug companies, computational researchers, wet labs, robotics, clinical development teams, and large pools of venture capital within a relatively small geographic area.
More importantly, many Boston AI-biotech companies are not building software that sits outside biology.
They are building AI directly into the experimental process.
That changes the economics of causal learning.
The best way to learn whether an intervention causes a biological effect is often to perform the intervention.
If experiments are slow and expensive, AI must depend heavily on historical datasets.
If experiments become faster, more automated, more parallel, and cheaper, models can learn from their own interventions.
That creates a very different kind of AI company.
Instead of:
data → model → prediction
the system becomes:
data → model → hypothesis → intervention → experiment → result → improved model → next intervention
That loop may be one of Boston’s strongest long-term advantages.
BostonBizScene Original Research: Measuring the Move Toward Causal Drug Discovery
To test this idea, BostonBizScene created an original dataset covering seven Boston-area companies whose public materials show substantial use of AI in biological or therapeutic discovery.
The companies were:
Valo Health
Cellarity
Lila Sciences
Manifold Bio
Generate
Relay Therapeutics
Dyno Therapeutics
The analysis was conducted using publicly available company materials and current information available through early October 2026.
This is not meant to be an exhaustive list of Boston AI-biotech companies. We deliberately used a smaller sample where enough primary information was publicly available to apply the same framework consistently.
Our five-factor methodology
Each company received one point for each of five observable features.
| Factor | Test used in our analysis |
|---|---|
| Explicit causal methods | Does the company publicly state that it uses causal inference, causal AI, or equivalent formal causal methods? |
| Intervention-first output | Does the AI system propose, design, or optimize an intervention rather than only classify or predict existing data? |
| Iterative experiment loop | Does the company explicitly describe experimental results feeding back into model improvement or subsequent experimental decisions? |
| Biological intervention evidence | Is there public evidence of real experimental testing in cells, tissue, animals, or humans? |
| External translation | Has the platform produced a clinical-stage asset, licensed program, or major external therapeutic collaboration? |
This scoring system measures causal-discovery readiness, not proof that every company uses formal causal inference.
That qualification is important.
A company can score highly because its operating system creates the conditions needed for causal learning even if it does not describe its algorithms as causal AI.
Chart 1: How Common Are the Building Blocks of Causal Discovery?
Share of companies in our seven-company Boston sample
| Capability | Companies | Share |
|---|---|---|
| Intervention-first AI output | 7 of 7 | 100% |
| Biological experimental validation | 7 of 7 | 100% |
| External clinical/partner translation | 6 of 7 | 86% |
| Explicit iterative experimental loop | 5 of 7 | 71% |
| Explicit causal-AI/inference language | 1 of 7 | 14% |
Visual view
Intervention-first systems
████████████████████ 100%
Experimental biological validation
████████████████████ 100%
External translation
█████████████████ 86%
Explicit model-experiment feedback
██████████████ 71%
Explicit causal methods
███ 14%
BostonBizScene analysis of publicly available company information, October 2026.
The gap is striking.
Only 14% of the sample clearly meets our strict test for explicit causal-method language.
Yet 100% is already focused on interventions.
That suggests something important.
Original Finding #1: Biotech Is Becoming Causal Before It Calls Itself Causal AI
The causal AI market may be bigger than the phrase “causal AI” suggests.
Searching company websites for those exact words would miss much of the transition.
Boston companies increasingly describe systems that change variables, measure responses, and feed results into later decisions. That is much closer to causal scientific reasoning than simply training a model on static historical data.
Valo Health is the clearest explicit example.
Valo says its platform combines real-world data, AI, advanced causal inference, statistical genetics, human tissue validation, and closed-loop chemistry. Its system is designed to uncover causal relationships in human data and identify targets for therapeutic intervention.
In April 2026, Valo went even further by appointing a Chief Causal AI Officer. The company said the role would advance its human causal biology platform using large-scale longitudinal data that now includes information from more than 20 million patients.
That is explicit causal AI.
But other Boston companies are approaching the same larger problem from different directions.
Cellarity tries to predict chemical interventions that move unhealthy cell states toward healthier ones. Its platform combines high-dimensional transcriptomics, AI modeling, human-relevant experiments, and an iterative lab-in-the-loop system.
Manifold Bio generates protein-design hypotheses and then obtains massive experimental measurements in living systems. Those measurements become labels that can improve predictive design.
Generate describes a continuous feedback loop in which machine learning drives protein design, experiments test the proteins, and the results refine the system. The company says it has already generated, built, and tested more than 42,000 proteins.
These companies are not identical.
But they share a larger direction.
They are connecting prediction with intervention.
Chart 2: Boston Causal-Discovery Readiness Score
Using the five-factor framework above, we calculated the following scores.
| Company | Score out of 5 |
|---|---|
| Valo Health | 5 |
| Cellarity | 4 |
| Manifold Bio | 4 |
| Generate | 4 |
| Lila Sciences | 3 |
| Relay Therapeutics | 3 |
| Dyno Therapeutics | 3 |
Average score: 3.71
Median score: 4
These numbers should not be read as rankings of company quality.
A score of five does not mean Valo has a better drug-discovery platform than a company scoring three. They operate in different areas, use different modalities, and disclose different amounts of technical information.
The useful insight is structural.
Even with a deliberately demanding framework, the typical company in the sample already shows roughly four of the five building blocks associated with intervention-driven learning.
Original Finding #2: The Wet Lab Is Becoming Part of the AI Model
The usual story about AI says data goes into a computer and predictions come out.
That story is increasingly incomplete in biotech.
The laboratory itself is becoming part of the learning system.
This may ultimately matter more than the size of any single model.
Generate shows what a continuous learning loop can look like
Generate says its platform removes the traditional separation between computation and biology. Machine learning generates protein designs, experiments test them, and the new information feeds back into the system.
By 2026, Generate described itself as a clinical-stage company with a platform connecting computational design and large-scale biological experimentation.
This matters because biological models face a basic problem.
There is far more possible biology than experimental data.
A model may suggest one million proteins. Scientists cannot take all one million through normal development.
The competitive advantage therefore becomes deciding which experiments generate the most useful information.
That is where active learning and causal thinking begin to overlap.
The model does not simply ask:
Which design looks best?
It can increasingly ask:
Which experiment would teach us the most?
Manifold pushes that idea directly into living biology
Manifold Bio may offer one of the clearest examples of why experimentation can become a data moat.
In March 2026, the Boston company said it tested one million binder designs against 127 targets in one multiplexed experiment, measuring more than 100 million protein-protein interactions. Specific binders were identified for 68% of targets in that study.
Its earlier mBER work also described using massively parallel testing to produce biological labels that feed model training and predictive design.
That changes the usual AI equation.
Many models may eventually gain access to similar public protein sequences and structures.
Far fewer companies will be able to produce enormous proprietary datasets showing what millions of designed interventions actually do inside biological systems.
That data is much harder to copy.
Original Finding #3: Experimental Throughput Could Become a Causal-AI Moat
This leads to a counterintuitive conclusion.
The best causal-AI company may not necessarily be the company with the largest AI model.
It could be the company capable of performing the most informative interventions.
Consider two companies.
Company A has an enormous model trained on public data but runs only a few hundred experiments each year.
Company B has a slightly smaller model but can perform hundreds of thousands or millions of carefully designed experiments, record what changed, and retrain its system.
Over time, Company B may learn things that simply did not exist in any historical dataset.
Its data is created through intervention.
That is a powerful moat.
Lila Sciences is pushing this idea even further. The company is building what it calls AI Science Factories: automated laboratories where AI systems can propose hypotheses, design experiments, run them, interpret results, and repeat the process.
Lila had raised $550 million in total capital by October 2025 to expand this model.
The strategic idea is bigger than drug discovery.
But for biotech, it points toward a future where the AI model and experimental machine become one system.
Valo Health Is the Clearest Boston Example of Explicit Causal AI
Among the companies we analyzed, Valo is the most direct example of the causal-AI thesis.
The company starts with human data.
It uses large longitudinal datasets, genetics, AI, and causal-inference approaches to identify biological drivers that may matter to disease. It then connects those insights to human tissue models and drug design.
That sequence is strategically important.
Human data comes before chemistry
Traditional drug discovery often begins with a biological hypothesis and tries to turn it into a drug.
Valo is attempting to move more human evidence toward the beginning.
Its expanded partnership with Novo Nordisk is a useful example.
In January 2025, the companies expanded their collaboration to cover up to 20 potential drug programs in obesity, type 2 diabetes, and cardiovascular disease. The agreement included near-term payments of up to $190 million and potential milestones of about $4.6 billion.
The companies said their earlier work had already produced several potential novel targets and multiple active small-molecule discovery programs.
In November 2025, Valo also announced a collaboration with Merck KGaA focused on Parkinson’s disease and related disorders. That program is designed to use Valo’s human causal biology platform for target discovery and its closed-loop platform for compound generation.
This is causal AI becoming an operating model rather than an academic exercise.
Cellarity Shows Why Causality May Need to Be Learned at the Cell-State Level
Traditional target discovery often simplifies disease into a small number of proteins.
Cellarity is making a different bet.
The company focuses on the state of the entire cell.
Its platform uses high-dimensional transcriptomic information and AI models to map disease states and predict chemical interventions capable of shifting dysfunctional cells toward healthier behavior.
That framing is inherently intervention-focused.
The question is not simply:
What does a sick cell look like?
It becomes:
What change would move this cell toward a healthy state?
Cellarity also describes a lab-in-the-loop approach where experimental results are used during optimization.
Importantly, this work has moved beyond research.
Its CLY-124 program entered a Phase 1 clinical study in 2025 after the platform helped identify a previously undiscovered mechanism connected to globin switching in sickle cell disease.
That is precisely the kind of translational evidence causal AI eventually needs.
It is easy to build an impressive biological model.
The harder test is whether insights from that model produce useful interventions.
Relay Therapeutics Shows Why Mechanistic Models Still Matter
Causal AI should not be confused with purely statistical causal inference.
Mechanistic understanding matters too.
Relay Therapeutics was built around understanding protein motion rather than treating proteins as static objects.
Its Dynamo platform combines experimental structural biology, biophysics, computation, chemistry, and genomics to understand how changes in protein conformation affect function.
That approach has already produced clinical programs.
Relay’s zovegalisib program, previously known as RLY-2608, emerged from work that included cryo-electron microscopy and long-timescale molecular simulations designed to understand differences between mutant and normal PI3Kα.
By 2026, the drug had entered Phase 3 development in metastatic breast cancer.
Relay does not need to market this work as causal AI for it to matter to the causal story.
It is trying to understand how physical changes in a biological system alter function, then use those insights to create an intervention.
That is deeper than recognizing a statistical pattern.
Dyno Therapeutics Shows Why In-Vivo Evidence Changes the Quality of AI Data
Dyno Therapeutics applies AI to designing AAV capsids used for gene delivery.
The difference between predicting a good vector and proving that a vector works inside a living system is enormous.
Dyno has increasingly moved toward real in-vivo validation.
In 2025, the company reported AI-designed capsids tested in non-human primates for delivery to the eye, muscle, and central nervous system.
In 2026, it disclosed further capsids and updates to its AI platform.
Industry partners have also exercised licensing options.
Roche exercised an option in 2025 for an AI-designed neurological gene-delivery capsid, while Astellas exercised another in 2026 for skeletal-muscle delivery.
Again, the larger lesson is about evidence.
An AI model predicting that a vector should work is useful.
A living organism showing where the vector travels, how strongly it expresses, and what side effects appear is far more informative.
Causal AI Changes Target Discovery First
The most obvious near-term application may be target selection.
Target failure remains one of the great sources of waste in pharmaceutical research.
Researchers can spend years developing an excellent molecule against a target that turns out not to matter enough in human disease.
Causal evidence can reduce that risk.
Human genetics is especially useful because genetic variation can sometimes act as a natural experiment.
If people carrying variants that reduce the activity of a biological pathway consistently show lower disease risk, that can strengthen the case for targeting the pathway.
It is not perfect evidence.
But it is often stronger than correlation alone.
The AI opportunity is combining many weak clues
A target might have:
genetic evidence,
patient-record evidence,
single-cell evidence,
protein-network evidence,
experimental perturbation evidence,
animal-model evidence,
and clinical biomarker evidence.
Traditional teams may examine these datasets separately.
A causal-AI system could increasingly connect them into one structured hypothesis.
The goal is not merely to produce a target score.
It is to explain why manipulating that target should produce a desired effect, under which conditions, for which patients, and with what uncertainty.
That would be far more useful.
Causal AI Could Also Change Biomarker Discovery
Predictive biomarkers tell researchers who is likely to experience an outcome.
Causal biomarkers can potentially help explain why.
That distinction matters for precision medicine.
Suppose patients with Biomarker A respond unusually well to Drug B.
A predictive model may use Biomarker A to select patients.
A causal model asks whether Biomarker A is part of the mechanism that creates the response.
Those are different questions.
The second question can help researchers design better drugs, identify resistance mechanisms, choose combinations, and discover new targets.
This is one reason causal machine learning has received attention for treatment-effect estimation.
The Nature Medicine causal-ML framework highlights the potential to estimate how treatment effects differ between patients, although it also emphasizes that strong assumptions and validation are needed.
The Biggest Opportunity May Be Choosing the Next Experiment
There is another use case that deserves more attention.
Causal AI can help determine what scientists should test next.
Biology has too many possibilities.
A lab cannot test every gene, molecule, dose, combination, cell type, disease model, and patient subgroup.
So the quality of experimental selection becomes critical.
Prediction asks which candidate is best
Suppose a model has 10,000 possible drug candidates.
A conventional optimization system may rank them and choose the highest-scoring candidate.
Experimental intelligence asks which experiment teaches us the most
The most informative experiment may not involve the candidate with the highest predicted score.
It may involve a candidate where the model is uncertain.
The result could reveal whether the model’s theory of the biology is correct.
This approach can create faster scientific learning.
Instead of optimizing only for:
best molecule now
the system can partly optimize for:
best information for the next round
That is how closed-loop discovery becomes more powerful over time.
Original Finding #4: Boston’s Competitive Advantage May Be Intervention Density
Silicon Valley has extraordinary AI talent and capital.
Boston’s advantage may be different.
It has unusually high intervention density.
By that we mean the concentration of institutions capable of actually changing biological systems and measuring what happens next.
The region contains universities, hospitals, biotech companies, contract research infrastructure, clinical researchers, robotic labs, sequencing capabilities, medicinal chemistry teams, and translational scientists within a relatively tight geographic network.
For causal AI, that matters.
Training data tells a model what has happened before.
Experiments create new evidence about what happens when the world is changed.
Boston can generate both.
This could become more valuable as foundation models commoditize
Powerful biological foundation models will probably become widely available.
Open models are improving.
Cloud computing is becoming easier to access.
Protein structure and sequence models are spreading quickly.
If model access becomes cheaper, proprietary advantage moves elsewhere.
One potential moat is experimental data.
Another is clinical data.
Another is the ability to repeatedly connect model predictions to physical interventions.
Boston is unusually strong in all three.
A Better Way to Think About the Causal Drug-Discovery Stack
Instead of treating causal AI as one algorithm, biotech leaders should think about it as a stack.
Layer 1: Human and biological observations
This includes genetics, patient records, omics, pathology, imaging, protein data, disease progression, and clinical outcomes.
The goal is to understand the system before intervention.
Layer 2: Causal and mechanistic hypotheses
AI and statistics help identify possible drivers rather than mere associations.
This layer should make assumptions explicit.
What causes what?
What might be a confounder?
What evidence would prove the current theory wrong?
Layer 3: Experimental intervention
Scientists perturb genes, proteins, cells, organisms, molecules, or pathways.
This is where hypotheses collide with reality.
Layer 4: Feedback and model updating
Experimental results are fed back into the system.
Failed predictions are especially valuable.
They reveal where the model’s understanding is incomplete.
Layer 5: Human translation
Eventually, hypotheses must survive the hardest environment of all: patients.
Clinical data then becomes another source of evidence that can improve the next generation of discovery.
The strongest causal-AI companies may eventually operate across most of this stack.
Why More Data Is Not Enough
Biotech has spent years assuming that larger datasets will automatically create better models.
Often they do.
But quantity and causal value are not the same thing.
Ten million observational measurements showing the same correlation may be less useful than a carefully designed intervention that separates two competing explanations.
This is why experiment design matters.
Data has different levels of information.
Observational data tells researchers what occurred.
Longitudinal data shows how things changed over time.
Genetics can provide useful natural variation.
Perturbation studies show what happens when scientists deliberately change something.
Randomized experiments provide stronger control over confounding.
The future AI-biotech data strategy should therefore ask:
What type of evidence are we collecting?
not simply:
How many terabytes do we have?
The Biggest Mistake: Calling Every Knowledge Graph “Causal”
Causal AI will inevitably become a marketing term.
Biotech leaders should be skeptical.
A graph showing relationships between genes, proteins, diseases, and compounds is not automatically a causal graph.
A machine-learning model identifying important features is not automatically identifying causal drivers.
A model explaining why it produced a prediction is not the same as proving that the variables it highlights cause the outcome.
And a biological mechanism suggested by AI is not established until evidence supports it.
Ask what intervention supports the claim
When evaluating causal-AI technology, investors and pharmaceutical partners should ask:
What interventions have been performed?
Did the predicted effect occur?
Were competing explanations tested?
Did the model work on unseen experimental conditions?
Was the finding validated in human-relevant systems?
Did the result survive prospective testing?
Those questions cut through marketing quickly.
Causal AI Does Not Replace Randomized Trials
This is one of the most important points in the entire discussion.
Causal AI is not a machine for turning observational data into certainty.
Randomized experiments remain extremely valuable precisely because randomization helps remove confounding.
Causal methods can make better use of existing data. They can help researchers design trials, estimate heterogeneous treatment effects, identify important covariates, create hypotheses, and sometimes support decisions where perfect experiments are unavailable.
But assumptions still matter.
If hidden confounders are ignored, causal models can still be wrong.
The healthcare causal-machine-learning literature repeatedly makes this limitation clear.
Good causal AI should therefore encourage more rigorous experiments.
It should not become an excuse to avoid them.
What Boston Biotech Companies Should Build Now
The practical opportunity is not to add the phrase “causal AI” to a company presentation.
It is to redesign discovery around intervention.
Start saving failed experiments
Successful results are attractive.
Failed predictions may be more useful for model improvement.
A model predicted that changing Target A would affect Pathway B.
Nothing happened.
That result helps narrow the space of plausible biological explanations.
Companies should structure, label, and preserve negative experimental results so they can train future models.
Record experimental context
Biological results depend on context.
Cell line.
Dose.
Time.
Temperature.
Disease state.
Assay conditions.
Patient background.
Animal model.
Combination treatment.
A result without context can become dangerous training data.
Causal models need to know the circumstances under which an intervention produced an effect.
Design experiments for information gain
Do not always test the candidate expected to perform best.
Some experiments should deliberately test areas where competing biological explanations disagree.
That can improve the model faster.
Connect computational teams with experimental teams
A weekly handoff between AI scientists and wet-lab scientists is not enough.
The feedback loop should be designed into the platform.
Generate’s public description of computation and experimentation as one continuous system is directionally important here.
The companies that reduce the delay between prediction and experiment from months to days—or eventually hours—could compound knowledge much faster.
What Investors Should Measure Instead of Model Size
Model size is an easy metric.
It may not be the most useful one.
A better diligence framework would include several operational measures.
| Metric | Why it matters |
|---|---|
| Prediction-to-experiment time | Shows how quickly computational ideas meet reality |
| Experiments per model cycle | Measures learning throughput |
| Percentage of experiments feeding training data | Shows whether the loop truly closes |
| Prospective validation rate | Tests performance outside historical data |
| In-vivo validation rate | Measures movement toward biological reality |
| Human-data anchoring | Shows whether early discovery is connected to human disease |
| Clinical translation | Measures whether platform insights survive development |
These numbers will often tell investors more than a benchmark showing a model is 3% more accurate than another algorithm.
What Pharma Should Ask Causal-AI Partners
Pharmaceutical companies evaluating platforms should insist on separating retrospective performance from prospective evidence.
A system can perform beautifully on historical datasets because similar examples were present during training.
The harder test is what happens next.
Can it propose a new target?
Can researchers perturb the target?
Does the expected biological change occur?
Can the result be repeated?
Does it hold in a more realistic model?
Can a medicine be built around it?
Does the mechanism survive in humans?
That progression should be the real benchmark.
Original Finding #5: Translation Is Already Becoming Part of the Boston AI Story
One of the strongest results in our sample is that six of seven companies already have substantial external translation through clinical programs or major therapeutic partnerships.
That equals roughly 86% of the sample.
This matters because the AI-biotech sector has historically faced an obvious criticism:
Where are the drugs?
That question is gradually becoming easier to answer.
Generate has a clinical pipeline enabled by a platform combining machine learning with large-scale biological experimentation.
Relay has progressed a platform-derived drug into Phase 3 development.
Cellarity has moved CLY-124 into Phase 1 development.
Dyno has seen major partners exercise licensing options on AI-engineered vectors.
Manifold signed a Roche collaboration in 2025 involving $55 million upfront and more than $2 billion in potential milestones for programs built around its AI-guided direct-to-vivo technology.
Valo’s Novo Nordisk collaboration now reaches up to 20 possible drug programs, while its Merck KGaA collaboration is applying human causal biology to Parkinson’s discovery.
These outcomes do not prove that AI has solved drug discovery.
They show something more modest but still important.
The platforms are leaving the demo stage.
Where Causal AI Will Probably Struggle
There are serious limits.
Biology is not a neat spreadsheet.
Systems contain feedback loops. Multiple causes can produce the same result. The same intervention can behave differently across tissues. Diseases evolve. Patients differ. Drugs hit several pathways. Biological measurements are noisy.
Causal models are only as trustworthy as their assumptions and evidence.
Hidden confounding will remain difficult
Researchers can adjust for variables they measure.
Unknown variables are harder.
This is especially dangerous when causal claims come from observational healthcare data.
Biological causality can change with context
A pathway may drive disease in one genetic background and barely matter in another.
A drug may work during one disease stage and fail later.
A protein may have different functions in different tissues.
There may not be one universal causal graph.
Counterfactuals cannot always be observed
For one patient, we observe what happened under the treatment they actually received.
We cannot usually observe what would have happened to that exact same patient at the same moment under a different treatment.
That fundamental problem does not disappear because AI becomes more powerful.
Why Causal AI Could Favor Biotech Over Pure Software Companies
The next phase of biological AI may reward companies that own both computation and experimentation.
A pure software company can build impressive models.
But if valuable causal information increasingly comes from interventions, access to laboratories becomes strategically important.
This does not mean every AI company needs to own a giant wet lab.
Partnerships can work.
Cloud labs can work.
Contract research can work.
Hospital collaborations can work.
But the feedback loop must be fast and structured.
A company that waits six months for experimental validation will learn differently from one that generates data every day.
That is another reason Boston is worth watching.
The region’s biotech infrastructure makes it easier to connect computation to physical science.
What to Watch Through 2027
Several signals will tell us whether causal AI is becoming a real drug-discovery advantage rather than another AI label.
First, watch for prospective target-validation results.
Companies should begin showing that targets selected through causal approaches succeed experimentally at higher rates than targets selected using conventional pipelines.
Second, watch for closed-loop experiment counts.
Statements about AI models will matter less than evidence about how quickly those models create and learn from new experiments.
Third, watch human translation.
The industry ultimately needs platform-derived drugs to perform in clinical trials.
Fourth, watch causal models move deeper into multimodal biology.
Genetics, single-cell data, imaging, proteomics, longitudinal patient records, and perturbation experiments become much more useful when analyzed together.
Fifth, watch AI begin choosing experiments autonomously.
The moment systems consistently decide which experiment will resolve the most important scientific uncertainty, the competitive landscape could change quickly.
The Bigger Story: AI Is Moving From Seeing Biology to Changing Biology
The first generation of biological AI became remarkably good at seeing patterns.
It could look across millions of observations and identify structures humans missed.
The next generation is being asked to do something harder.
Understand what happens when biology changes.
That means distinguishing drivers from passengers.
Cause from correlation.
Mechanism from coincidence.
Useful intervention from impressive prediction.
This is not simply a new algorithmic trend.
It changes how biotech companies should build their entire discovery process.
Data collection becomes experimental strategy.
Labs become learning infrastructure.
Negative results become training data.
Clinical translation becomes part of model evaluation.
And the boundary between computation and biology begins to disappear.
Boston already has companies operating along much of that path.
Final Takeaway: The Winning AI May Be the One That Knows What Experiment to Run Next
The last decade of AI was largely about learning from data that already existed.
The next era of drug discovery may be about creating the right new data.
That is the core reason causal AI matters.
Finding a pattern in biology can produce a hypothesis.
A medicine requires much more.
Scientists have to intervene, observe the consequence, understand why it happened, revise their assumptions, and repeat the process until the biology becomes clear enough to build a therapy around it.
Our BostonBizScene analysis suggests that Boston’s AI-biotech industry is already moving in that direction.
Across the seven-company sample, 100% centered their platforms on interventions. 100% showed experimental biological validation. 71% explicitly described a recurring experiment-and-learning loop. 86% had reached meaningful external translation through clinical development, licensing, or pharmaceutical collaboration.
Yet only 14% clearly marketed the technology around explicit causal methods.
That may tell us where the market is going.
The causal-AI revolution in biotech may not arrive under one label.
It may arrive quietly through better experiments, richer perturbation data, human genetics, automated labs, mechanistic models, closed-loop chemistry, direct-to-vivo screening, and systems that learn not only from what biology looks like—but from what happens when they deliberately change it.
For Boston, that could be a major advantage.
The city already has AI.
It already has biotech.
It already has world-class experimental infrastructure.
The next opportunity is connecting them into systems that do more than predict biology.
They learn how to control it.
