For most of the past few years, the biggest story in AI and biotechnology has been the model itself. Researchers trained larger models on proteins, startups showed that AI could predict structures or generate new sequences, investors poured money into the field, and major pharmaceutical companies started building dedicated AI teams. At almost every biotech conference, the same broad question appeared in different forms: what happens when generative AI finally becomes part of drug discovery?
Converge Bio represents a different and potentially more important stage of that story because the company is not only trying to build powerful biological models. It is trying to move generative AI into the normal day-to-day work of biotech researchers. Instead of asking scientists to learn how to run protein language models, build machine-learning pipelines, connect multiple prediction tools, or hire large computational teams, Converge is building systems around specific scientific jobs that researchers already need to perform.
Those jobs include finding new targets, designing antibodies, improving protein production, discovering biomarkers, and analyzing difficult biological systems. Converge says its models are trained on the core biological languages of DNA, RNA, and protein sequences, but its broader goal is not simply to produce better models. The company wants to turn those models into experimentally tested systems that can be used inside existing drug-development programs. (converge-bio.com)
That distinction matters because the next stage of AI in biotech may not be defined by one spectacular model that discovers a medicine by itself. It may instead be defined by thousands of smaller decisions that become faster and better every week. Which target should a team investigate? Which antibody sequences should be synthesized? Which molecule should move into the next assay? Which protein sequence might produce a higher manufacturing yield? Which experimental result deserves another round of testing? These are the kinds of decisions that fill ordinary biotech R&D, and they are exactly where AI may create practical value.
Converge Bio is particularly interesting because it is building directly for that world. For Boston, this model may be especially important because Greater Boston already has the wet labs, translational scientists, antibody teams, cell biologists, manufacturing specialists, research hospitals, universities, CRO relationships, and experienced drug developers required to test AI-generated ideas in the physical world. The question is increasingly not whether Boston can train another biological foundation model. The more important question is whether AI can be inserted into the thousands of scientific decisions already being made across Kendall Square and the wider biotechnology ecosystem.
The Short Version: Converge Bio Is Trying to Turn Generative AI Into a Normal Lab Tool
The easiest way to understand Converge Bio is to avoid thinking of it as a version of ChatGPT for scientists, because that comparison is too simple. Converge’s system is closer to a computational research department that sits beside a biotechnology company’s existing experimental team. A scientist defines a problem, AI generates or ranks possible biological answers, additional computational models filter those answers, and a small number of candidates move into physical experiments.
Converge publicly describes systems for target discovery, antibody design, protein manufacturing optimization, and biomarker discovery. The company says these systems combine proprietary models with experimentally validated workflows rather than simply delivering raw model predictions. (converge-bio.com) That makes the company interesting because the useful product is not necessarily the neural network itself. The useful product is the decision that comes out of the system.
A scientist does not need an impressive model merely for the sake of having an impressive model. The scientist needs to know whether a particular antibody should be built, whether one sequence should be rejected, whether a protein can be produced more efficiently, whether a gene deserves more investigation, or whether one design is more likely to improve a molecule without creating another problem. This is where generative AI starts moving from an impressive research demo into everyday biotechnology.
Converge Bio at a Glance
| Metric | Publicly reported position |
|---|---|
| Founded | 2024 |
| Main locations described publicly | Boston and Tel Aviv |
| Series A announced | $25 million |
| Total funding stated at Series A announcement | $30 million |
| Additional Gates Foundation grant announced in June 2026 | $2.5 million |
| Pharma and biotech customers reported in January 2026 | More than 12 |
| Completed pharma and biotech programs reported in 2026 | More than 40 |
| Public platform areas | Antibody design, target discovery, biomarker discovery, protein manufacturing optimization |
| Publicly listed case studies reviewed by BostonBizScene | 9 |
Converge announced its $25 million Series A in January 2026, led by Bessemer Venture Partners with participation from TLV Partners, Vintage Investment Partners, Saras Capital, and investors connected with companies including Meta, OpenAI, and Wiz. The company said the financing brought total capital raised to roughly $30 million, and it also reported more than a dozen pharmaceutical and biotechnology customers at that point. (prnewswire.com)
By March 2026, public material connected to a collaboration with Purple Biotech stated that Converge had already completed more than 40 programs with pharmaceutical and biotechnology companies. (markets.financialcontent.com) Those numbers do not make Converge one of the largest AI-biotech companies in the market, but they make it something potentially more revealing: a relatively young company already being used repeatedly on real development problems.
Why Converge Bio Is Different From the First Wave of AI Biotech
The first generation of AI-biotech companies often started with a technological breakthrough. One company might build a new structure-prediction system, another would generate small molecules, another would build a biological knowledge graph, and another would train on cell images. In many cases, the technology itself became the story, while the everyday scientific workflow remained secondary.
Converge appears to be trying to invert that relationship. Instead of beginning with the question of which model should be built, the company is starting with the job a scientist needs to complete and then assembling the models and tools required to perform that job more efficiently.
The Model Sits Underneath the Workflow
Converge says its systems can combine several computational methods rather than depending on one universal model. For antibody design, for example, CEO Dov Gertz told TechCrunch that the system includes a generative component that creates potential antibodies, predictive models that filter candidates based on molecular properties, and a physics-based docking system that examines how the molecules may interact with their targets. (techcrunch.com)
That architecture tells us something important about where AI drug discovery is heading. A pharmaceutical researcher does not care whether every stage of the process uses the newest transformer architecture. The researcher cares whether the final sequence works. If generative AI is useful for exploring sequence space, it should be used there. If a predictive model is better for checking solubility, it should be used for that. If physics-based simulation is useful for another filter, it belongs in the workflow as well.
The product therefore becomes a system of models rather than a single model. That is much closer to how laboratories already work because real drug research is already made up of many specialized tools. One assay measures binding, another looks at aggregation, another examines expression, and another tests activity in cells. Scientists combine all of those signals before making a decision, and AI systems are beginning to develop the same multi-step shape.
Converge explicitly says it prioritizes experimental validation instead of relying only on model benchmarks. The company describes wet-lab results and real program impact as the standard by which its systems should ultimately be measured. (converge-bio.com) That sounds obvious, but it is one of the most important changes happening across AI biotech because a model that performs well on a benchmark is not automatically useful in a laboratory.
Original BostonBizScene Analysis: 56% of Converge’s Public Case Studies Are Already About Antibody Work
To understand where Converge is actually putting visible product energy, BostonBizScene reviewed every case study listed on the company’s public case-study page as of October 7, 2026. There were nine public case studies in the sample, and we manually classified each one according to its main research and development task rather than the marketing language used in the title.
Our dataset placed each case study into one of three categories: antibody design and engineering, biological target or biomarker discovery, or protein production optimization. Converge currently lists five antibody-focused studies, two target or biological-discovery studies, and two protein-yield studies. (converge-bio.com)
Public Case-Study Distribution
| Workflow | Public case studies | Share |
|---|---|---|
| Antibody design and engineering | 5 | 55.6% |
| Target, biomarker, or cell discovery | 2 | 22.2% |
| Protein yield and manufacturing | 2 | 22.2% |
| Total | 9 | 100% |
The distribution is clear. Antibody design accounts for 55.6% of Converge’s visible case-study library, while target discovery and protein-manufacturing work each account for 22.2%. This does not tell us the exact mix of Converge’s private customer programs because those engagements are not publicly visible, so the internal business mix may look very different.
Even with that limitation, the public evidence is useful because more than half of the company’s published case studies concentrate on antibodies. That makes strategic sense. Antibody engineering gives generative AI a clear design object, measurable molecular properties, large sequence datasets, established laboratory assays, significant commercial value, and a relatively straightforward path from digital prediction to physical validation.
For that reason, antibody design is becoming one of the strongest proving grounds for biological generative AI. Companies can generate sequences computationally, synthesize a limited number of them, measure what happens in the laboratory, and compare those outcomes directly with the original molecules.
ConvergeAB Is Where the Strategy Becomes Easy to See
ConvergeAB is the company’s antibody design and engineering system, and its public case studies show a clear progression in the kinds of problems the company is trying to solve. The early question in AI antibody design was often simple: can an algorithm improve an antibody? The more useful question now is whether AI can improve several competing properties at the same time without damaging the characteristics that already make the antibody valuable.
That second problem is much closer to real pharmaceutical development. An antibody can bind extremely well and still be a poor drug candidate because it may aggregate, become unstable, create unwanted immune responses, prove difficult to manufacture, or bind too strongly for the biological job it is supposed to perform. Researchers rarely need the strongest binder in isolation. They need the best overall compromise across several properties.
The FMC63 Experiment Shows Why Multi-Objective AI Matters
One of Converge’s newer public case studies focuses on FMC63, the antibody binder used in several approved CD19 CAR-T products. Converge did not simply ask its system to increase binding. Instead, the platform was given several goals at once: humanize the antibody, remove predicted immune-related liabilities, improve developability, and preserve its existing CD19 affinity.
The resulting lead contained 45 amino-acid changes. Converge reported that framework humanness increased from 77.8% to 93.3%, two predicted CD4 T-cell epitopes fell to zero, thermal stability improved by about 19°C, and monomer content improved by roughly 13 percentage points. Importantly, the final binding strength remained close to the parent antibody’s level. (converge-bio.com)
A different generated candidate actually bound CD19 roughly four times more tightly than the original molecule, yet Converge did not select that candidate as the preferred lead. The chosen design had a reported KD of 0.184 nM versus 0.197 nM for the parent, while another generated candidate reached 0.048 nM. Converge deliberately selected the candidate that kept binding near the clinically established range while improving other properties. (converge-bio.com)
That is where generative AI begins to look much more like engineering than simple search. The goal is no longer to maximize one number. The goal is to satisfy several biological and development constraints at the same time, which is much closer to the way real medicines are designed.
From Single-Property Optimization to Multi-Objective Design
The older optimization process usually begins by finding a molecule, improving one property, testing it, discovering that another property became worse, and then starting another round of engineering. This can continue through many iterations, especially when affinity, stability, immunogenicity, solubility, and manufacturability are all interacting with one another.
The emerging AI approach is different because researchers can define several desired characteristics before generation begins. The system can generate candidates inside those limits, score them across multiple properties, reject most of them computationally, and send only a small group into the wet lab. This does not remove experiments from the process. It attempts to make each experiment more informative and more likely to produce something useful.
Original Finding #1: Converge’s Antibody Funnel Discards More Than 99.9% of Generated Designs Before Final Wet-Lab Selection
One of the most useful public datasets Converge has released comes from a four-target antibody experiment using ConvergeAB. The company started with four antibody-antigen pairs covering different starting conditions and used its platform to generate roughly 30,000 candidate sequences for each starting antibody. That means the total initial design space contained approximately 120,000 candidates across four programs.
Only around 5% moved into deeper structure simulation and docking. The workflow eventually narrowed the complete candidate pool down to 42 molecules that were physically tested. (converge-bio.com)
The ConvergeAB Funnel
| Stage | Approx. candidates | Share of initial pool |
|---|---|---|
| Initial AI-generated candidate pool | ~120,000 | 100% |
| Approx. top 5% moving deeper | ~6,000 | 5.0% |
| Wet-lab candidates | 42 | 0.035% |
| Wet-lab candidates with improved affinity | 18 | 0.015% |
The 120,000 figure is approximate because Converge describes the initial candidate pool as roughly 30,000 per target. Even so, the shape of the funnel is striking because the value of the system is clearly not just that it can produce 120,000 sequences. Modern biological models can generate very large numbers of potential sequences. The difficult and commercially important part is deciding which tiny fraction deserves to become a physical experiment.
BostonBizScene’s calculation shows that the workflow reduced the original design space to only 0.035% before wet-lab testing. Put another way, roughly 99.965% of the original candidates were removed before physical experiments began.
The Filtering Ratio Is Roughly 2,857 to 1
For every candidate that reached the wet lab, approximately 2,857 candidates were removed computationally. That may be one of the most useful ways to think about generative AI in biotechnology because the system is not necessarily replacing experiments. It is changing which experiments get run.
This matters because physical experiments are expensive, slow, and capacity constrained. Generating 100,000 sequences on a computer is relatively cheap compared with synthesizing and testing 100,000 physical molecules. The value appears when computation prevents the laboratory from spending time on weak candidates.
Original Finding #2: Roughly 69% of Measurable Candidates Improved Affinity
Converge reported that 42 antibody candidates were experimentally tested and 26 produced measurable binding. Eighteen of those 26 measurable binders improved on their respective parent antibodies. (converge-bio.com)
BostonBizScene calculated several useful conversion rates from those reported figures. Around 61.9% of all tested candidates produced measurable binding, 69.2% of measurable binders improved affinity, and approximately 42.9% of all tested candidates represented an affinity improvement over the parent molecule.
| Measure | Result |
|---|---|
| Measurable binders among all tested candidates | 26 / 42 = 61.9% |
| Improved candidates among measurable binders | 18 / 26 = 69.2% |
| Improved candidates among all tested candidates | 18 / 42 = 42.9% |
These numbers come from a Converge-run case study, so they should not be treated as independent third-party validation. Company-run experiments can contain selection effects, and broader replication across different laboratories would make the evidence stronger.
Even with that limitation, the data illustrates what a commercially useful AI system could look like. The model does not need every design to succeed. It needs the small group entering expensive experiments to be meaningfully enriched for useful molecules compared with a much less selective process.
Four Targets Produced Four Very Different Levels of Improvement
The same experiment covered PD-L1, SARS-CoV-2 Spike, HER2, and CD19. Converge reported very different best-case improvements across those four targets, which is useful because real biological systems rarely produce the same result in every context.
| Target | Parent KD | Best reported AI-designed KD | Improvement |
|---|---|---|---|
| PD-L1 | 22.5 nM | 0.04 nM | 556× |
| Spike | 7.5 nM | 0.29 nM | 26× |
| HER2 | 23.9 nM | 9.28 nM | 2.6× |
| CD19 / tafasitamab | 0.56 nM | 0.21 nM | 2.7× |
Converge says it used the same weights and general procedure across the four targets without target-specific fine-tuning. It also reported that PD-L1 and Spike were held out of the model’s training set for the experiment. (converge-bio.com)
The wide difference between a 556-fold improvement and a roughly 2.6-fold improvement should not automatically be viewed as a weakness. It reflects a basic reality of molecular engineering: different starting molecules have different amounts of room to improve. A highly optimized therapeutic antibody may have far fewer obvious opportunities for dramatic improvement than an earlier discovery-stage binder.
That is why biotech teams should be cautious when evaluating AI platforms that imply the same dramatic gains should be expected across every target. Real biology is uneven, and strong platforms need to work within that uneven reality rather than hide it.
Why the Cetuximab Experiment Got So Much Attention
Another Converge experiment focused on cetuximab, an established antibody targeting EGFR. According to the company’s reported study, ConvergeAB received the antibody sequence and receptor sequence, generated 100,000 candidates, selected ten sequences for laboratory testing, and identified a candidate with an average binding affinity 2.1 times stronger than cetuximab.
The company said the computational part of the project was completed in eight hours without target-specific training or a dedicated research campaign. (prnewswire.com) That does not mean a new cancer treatment was created in eight hours, because everything downstream still matters. Further biological testing, pharmacology, safety work, manufacturing, toxicology, clinical development, regulation, and commercial production would all remain necessary.
What the experiment suggests is that one design step may be compressed substantially. Drug development is made up of many such cycles, and reducing the time required for enough of those cycles could begin to change the economics of research.
Generative AI’s Biggest Near-Term Win May Be Fewer Bad Experiments
AI-biotech discussions often jump immediately to the idea of creating billion-dollar medicines, but that skips an important middle layer. Most biotechnology research is not the final discovery of a drug. It is the everyday process of deciding what to test next, which experiments deserve resources, and which ideas should be removed from the pipeline.
A protein scientist may need to choose 20 sequences from tens of thousands of possible designs. An antibody engineer may need to decide which mutations deserve synthesis. A target-discovery group may need to reduce hundreds of genes to a short list of ten. A process-development team may need to determine which expression sequence is most likely to improve output.
These are repetitive and expensive decisions. AI can create meaningful economic value long before it independently “discovers a drug” if it can consistently improve which candidates receive expensive laboratory resources.
The Economic Logic Is Simple
Imagine that a team can physically test only 20 candidates during one development cycle. Without strong computational filtering, perhaps only one or two of those candidates prove useful enough to advance. If AI can increase the proportion of useful candidates in that same group of 20, the laboratory gets more information and more progress from the same budget.
That creates several possible benefits at once. The same experimental budget may produce more usable results, fewer design-test cycles may be needed before finding a strong lead, and scientists may spend less time studying candidates that were weak from the beginning. This is less dramatic than an autonomous AI scientist, but it is much closer to the way productivity improvements normally spread through an industry.
ConvergeGEO Shows AI Moving Downstream Into Manufacturing Problems
Drug discovery receives most of the attention, but generating a useful therapeutic sequence is only part of the development process. Biotechnology companies must eventually produce the molecule at the required quality, consistency, and cost, which makes manufacturing characteristics extremely important.
A strong therapeutic protein can become commercially difficult if it expresses poorly or requires inefficient production systems. Converge has therefore built ConvergeGEO around protein yield optimization, and its public case-study directory includes two studies centered on protein manufacturing. One describes a 4.5-fold increase in yield, while another looks at zero-shot therapeutic protein yield optimization across a broader group of examples. (converge-bio.com)
This is strategically important because AI-biotech companies have traditionally concentrated heavily on the earliest stages of discovery. Manufacturing optimization is much closer to operational biology. If generative AI can move successfully into that part of the workflow, the addressable market becomes much larger because AI can support not only discovery but also the process of turning biological designs into real products.
The Sequence Can Influence the Factory
Proteins are produced by biological systems, and small changes in the genetic instructions used to produce a therapeutic protein can influence expression efficiency without necessarily changing the desired final protein. That creates another search problem where there may be many possible sequences but only a limited number that can be tested economically.
The same overall pattern therefore appears again. AI generates or explores a large possibility space, computational tools rank and filter the options, and a much smaller set moves into physical testing. This is how generative AI begins to spread horizontally through R&D even when the scientific problem changes from molecule design to manufacturing efficiency.
ConvergeCELL Pushes in the Other Direction: From Molecules Back Toward Disease Biology
Converge is also working on systems that sit much earlier in the discovery process. The company’s public material describes target and biomarker discovery, and its case-study library includes work using ConvergeCELL on multiple myeloma as well as a broader system for disease-agnostic target and biomarker discovery. (converge-bio.com)
This introduces a very different problem from antibody optimization. Instead of asking how an existing molecule should be redesigned, the scientist is asking what part of the biology should be targeted in the first place.
That can require integrating information about cells, genes, disease states, tissue differences, molecular pathways, and interactions between many biological systems. It is a more open-ended question than improving a known molecule, and it can also be more valuable because a biotechnology company that chooses the wrong biological target may design an excellent drug that ultimately does not matter.
The Virtual Cell Could Become One of the Most Important Long-Term Pieces
In June 2026, Converge announced a $2.5 million multi-year grant from the Gates Foundation to extend its modeling approach into crop genomics. The project involves a multimodal, long-context foundation model designed to identify genetic variants associated with characteristics such as crop yield and climate resilience.
Converge described a system capable of examining millions of DNA base pairs together, combining sequence and gene-activity information, ranking genes or mutations, and identifying the genomic regions most responsible for a prediction. (prnewswire.com)
Although the immediate application is agriculture, the technical direction matters for biotechnology because the model is moving beyond designing a single protein sequence and toward understanding biological systems in context. That is central to the wider “virtual cell” idea, where AI attempts to model the behavior of biological systems rather than only isolated molecular objects.
Why Long Context Matters in Biology
A gene rarely acts by itself because its effect depends on other genes, regulatory regions, cell state, environmental signals, and the wider biological system around it. A mutation that looks important when studied in isolation may become much less important once its surrounding context is considered.
Older computational tools often simplify these relationships because handling extremely large biological contexts is difficult. Long-context biological models attempt to examine more of the system at the same time. If those systems become accurate enough, target discovery may gradually move beyond simple correlation and toward deeper questions about biological cause.
For Boston’s drug-development ecosystem, that would be significant because the region already contains many of the laboratories and translational research groups needed to challenge those predictions experimentally.
Original Finding #3: Converge’s Public Product Map Now Touches Both Ends of Preclinical R&D
BostonBizScene created a simplified drug-development workflow and mapped Converge’s publicly documented systems against it. The goal is not to claim that Converge has solved every task inside each stage, because it clearly has not. The purpose is to understand how far the company’s strategy has expanded beyond one narrow AI use case.
| R&D stage | Public Converge capability visible? | Example |
|---|---|---|
| Disease biology | Yes | ConvergeCELL and virtual-cell direction |
| Biomarker discovery | Yes | Biomarker discovery system |
| Target identification | Yes | Target discovery |
| Molecule generation | Yes | ConvergeAB |
| Lead optimization | Yes | Affinity, stability, humanness, solubility |
| Developability optimization | Yes | Multi-property antibody engineering |
| Protein production optimization | Yes | ConvergeGEO |
| Clinical trial execution | No major public system identified | — |
| Regulatory execution | No major public system identified | — |
The table does not show that Converge owns the full drug-development stack. What it does show is a clear direction of travel. The company is expanding horizontally from molecular design into the scientific decisions that surround molecular design.
That may ultimately be more important than building one enormous foundation model because drug development contains many bottlenecks. A platform that can support several of those bottlenecks may become more deeply embedded in everyday R&D than a system built around only one spectacular capability.
Original Finding #4: The Business Is Expanding Faster Than a Pure Research Project Usually Would
Converge’s reported team growth also provides useful context. TechCrunch reported that the company’s headcount increased from nine people in November 2024 to 34 people by January 2026. (techcrunch.com)
By March 2026, a public announcement connected with Purple Biotech described Converge as having a 40-person team. (markets.financialcontent.com)
Team Growth
| Date | Reported team size | Growth vs. Nov. 2024 |
|---|---|---|
| Nov. 2024 | 9 | Baseline |
| Jan. 2026 | 34 | +278% |
| Mar. 2026 | 40 | +344% |
Moving from nine employees to 40 means the reported team became roughly 4.4 times larger in about sixteen months. Team growth alone does not prove product-market fit, and fast hiring can create its own operational risks. However, when that growth is considered alongside more than a dozen reported customers and more than 40 completed programs, it looks different from a company that is still mainly conducting internal AI research.
Converge appears to be building an organization designed around repeated delivery into external biotech and pharmaceutical programs. That matters because the operational requirements of serving customers are very different from those of publishing impressive model results.
More Than 40 Programs May Be the Most Important Number in the Story
Funding announcements get headlines, model benchmarks impress researchers, and dramatic antibody results are easy to market. Yet the claim that Converge has completed more than 40 programs may tell us more about where AI biotech is going than any one technical result.
Repeated programs create operational learning. A company discovers where customers actually struggle, which outputs scientists trust, which predictions routinely fail, how experimental data arrives, where software needs more explanation, and which model errors are harmless compared with errors that waste several weeks of laboratory work.
That experience can become a powerful competitive advantage because biotechnology contains enormous numbers of edge cases. A team that has worked through dozens of messy real-world customer programs may learn practical lessons that are difficult to obtain from public datasets or academic benchmarks alone.
Data Is Not the Only Moat
AI companies often describe proprietary training data as their primary moat, and data certainly matters. However, workflow memory can become equally important because it reflects repeated knowledge about how AI is actually used inside scientific work.
Consider two companies with models of roughly similar quality. One has completed five tightly controlled internal demonstrations, while the other has worked through 100 customer programs with different targets, assay types, experimental constraints, data formats, and research teams. The second company may understand much more about how to make AI useful in practice even if its core model is not dramatically better.
That type of operational knowledge can be difficult to reproduce because it accumulates slowly through real use.
Why “No Code” Matters More in Biotech Than It Sounds
Converge says its systems are designed to deliver useful outputs to biologists without requiring them to write code or build AI infrastructure. (prnewswire.com)
At first glance, that sounds like a normal software feature, but it is central to adoption because most pharmaceutical scientists are not machine-learning engineers and should not be required to become machine-learning engineers. A protein scientist’s scarce skill is understanding proteins. An immunologist’s scarce skill is understanding immune biology. A process-development scientist’s scarce skill is understanding manufacturing.
Forcing those people to spend months assembling AI pipelines reduces the productivity gains that AI is supposed to deliver. The winning interface may therefore be the one that hides most of the machine learning and lets scientists describe the biological problem directly.
The Scientist Should Define the Biological Goal
A future workflow might begin with a scientist providing an existing antibody, its target, and a desired product profile. The scientist could specify that affinity should remain within a defined range, request better thermal stability, require lower predicted immunogenicity, and lock specific residues that cannot be changed.
The AI system could then generate a large candidate space, remove obviously poor designs, evaluate the remaining sequences across several computational models, and return a small group that deserves testing. The scientist would not need to know which neural network produced candidate 17. The scientist would need to understand why candidate 17 deserves an expensive physical experiment.
That is a much more useful way to design scientific software because it keeps the researcher focused on biology rather than software infrastructure.
Explainability Becomes More Important When Experiments Are Expensive
A bad consumer chatbot answer may cost a few seconds. A bad biological prediction can trigger weeks of work, consume laboratory capacity, and waste expensive materials. That changes the standard AI systems need to meet inside biotechnology.
Converge’s virtual-cell work, for example, is described as not only ranking possible genes or mutations but also showing which parts of a sequence influenced those predictions. (prnewswire.com)
That kind of explanation matters because scientists must decide whether a prediction makes biological sense before assigning resources to it. Researchers need to know what changed, why one candidate is ranked higher, which property improved, what may have become worse, how uncertain the prediction is, and what experiment would most efficiently test whether the model is wrong.
Those questions will increasingly separate serious scientific platforms from generic AI tools because scientific trust depends on understanding both the prediction and its limitations.
The Patent Extractor Is Another Clue About Everyday AI
Converge’s public research material also includes a patent-extraction tool focused on antibody sequences. Its resource library lists “Mine Antibody Sequences from Patents” among its 2026 releases. (converge-bio.com)
This may sound less exciting than designing a new therapeutic molecule, but it is one of the reasons Converge’s direction is interesting. Biotechnology researchers spend large amounts of time gathering, cleaning, and organizing information before they design anything. Patent documents can contain valuable sequence information, yet extracting those sequences can be slow and tedious.
AI that reduces this type of work does not need to make a scientific breakthrough to create value. It only needs to return hours to researchers, and that is how enterprise software becomes embedded in a profession. A platform may attract attention through a spectacular capability, but long-term adoption often comes from removing dozens of smaller frustrations that scientists experience every week.
The Real Product May Become a Generative Lab, Not an AI Model
Converge increasingly describes itself as a “Generative AI Lab for the Life Sciences.” That positioning may ultimately be more strategic than calling itself an AI drug-discovery company because the two models imply very different businesses.
A traditional drug-discovery company is expected to create and advance its own therapeutic assets. A generative lab can instead provide capabilities across many customer programs while allowing biotech companies to retain control of their underlying drug-development strategy.
Two Very Different Business Models
The traditional biotech model involves raising capital, choosing targets, building proprietary medicines, running experiments, advancing a relatively small number of internal programs, and carrying substantial clinical-development risk. The company may create enormous value if one of those medicines succeeds, but the feedback loop can be slow because clinical development takes years.
A generative-lab model can work across many external programs, reuse AI systems across different customers, learn from repeated engagements, and potentially generate much more frequent feedback about what works and what does not. That feedback may be particularly valuable for AI because model performance improves when systems receive reliable information about both successful and unsuccessful predictions.
A generative lab connected to dozens of programs could therefore build a much broader learning network than a company working on only a small internal drug pipeline.
Original Finding #5: Converge’s Public Evidence Suggests AI Is Becoming a Funnel, Not a Replacement Scientist
Across Converge’s published examples, the same broad pattern appears repeatedly. It shows up in antibody design, protein optimization, target ranking, and genomics, even though the underlying scientific problems are very different.
The pattern begins with a large possibility space. AI generates or identifies potential options, computational systems filter those options, scientists review the strongest candidates, a small subset moves into physical experiments, and the results of those experiments inform the next decision.
This is a much more realistic model of AI in biotechnology than the popular idea that an AI system will simply replace the scientist. The old mental model is that AI produces the answer. The emerging model is that AI reshapes the search space so scientists can spend more time on the most promising parts of it.
That difference may sound small, but it changes how biotech teams should evaluate AI. Biology is too complex for most important problems to have one obvious computational answer, yet AI can still create enormous value by changing which possibilities researchers examine first.
Why Boston Is an Especially Good Place for This Model
Boston has spent decades building the physical side of biotechnology. The region contains major research hospitals, universities, pharmaceutical companies, venture-backed startups, specialized laboratory infrastructure, CROs, equipment suppliers, manufacturing expertise, experienced biotech executives, and scientists who have taken medicines from early discovery into clinical development.
Generative AI needs exactly that environment because a protein model becomes more valuable when someone can test the protein. A target-ranking system becomes more valuable when a disease-biology team can perturb the target. A virtual-cell model becomes more useful when experimental teams can challenge its predictions. An antibody generator becomes more valuable when candidates can quickly be expressed, purified, and measured.
Boston therefore does not need to beat Silicon Valley at every layer of AI infrastructure. Its advantage can come from combining computational systems with one of the densest experimental biotechnology ecosystems in the world.
Boston Can Compete Through the Experimental Loop
The highest-value biological AI company may not be the one with the largest model. It may be the company that moves fastest from prediction to experiment, from experiment to measurement, and from measurement back into the next prediction.
Boston already contains much of the physical infrastructure needed for that cycle. Converge’s Boston connection is therefore strategically interesting even though its technical team and operations span multiple locations.
Public information does not provide a reliable city-by-city breakdown of Converge employees, so it would be inaccurate to describe the company as purely Boston-built. It is better understood as part of a new class of globally distributed AI-biotech companies whose commercial and scientific activity intersects with Boston’s unusually dense drug-development ecosystem.
What Boston Biotechs Should Learn From Converge
The biggest lesson is not that every Boston biotech company should immediately buy Converge. The more useful lesson is that AI adoption should begin with a workflow rather than with a vague technology strategy.
Too many companies begin by asking what their “AI strategy” should be. That question is usually too broad to produce useful action because it begins with the technology rather than the scientific bottleneck.
A better starting point is to ask which repeated decision consumes the most time, experiments, or scientific attention. For an antibody company, that may be candidate selection. For a protein company, it may be expression. For a genomics team, it may be variant prioritization. For an oncology company, it may be target selection. For a platform biotech, the bottleneck may be literature or patent mining.
Once the bottleneck is identified, the team can ask whether AI can make that particular decision faster, cheaper, or more accurate.
Do Not Measure AI by How Many Candidates It Generates
Generation is becoming cheap, while filtering is becoming valuable. A model that produces one million sequences may sound impressive, but if scientists still need to test 5,000 of them, the system has not solved the most expensive part of the problem.
A much better KPI is experiment compression. Companies should measure how effectively AI reduces a huge computational possibility space into a small and experimentally useful shortlist.
A Practical AI-R&D Dashboard
| Metric | Why it matters |
|---|---|
| Generated candidates | Measures search breadth |
| Candidates surviving computational filters | Measures selectivity |
| Candidates entering wet-lab tests | Measures experimental burden |
| Experimentally valid candidates | Measures enrichment |
| Improved candidates | Measures useful hit rate |
| Number of design-test cycles | Measures iteration speed |
| Cost per successful candidate | Measures economic value |
| Time per useful design | Measures workflow improvement |
This is one of the strongest lessons from Converge’s public antibody work. The headline number may be 120,000 generated sequences, but the business number is 42 experiments. That difference explains where much of the practical value comes from.
Wet-Lab Success Rate Should Become a Standard AI-Biotech Metric
AI companies frequently report computational benchmarks, and those benchmarks remain useful for comparing models. However, customers ultimately care about what happens after synthesis, expression, purification, or testing begins.
Converge’s four-target antibody study provides the type of metric that more AI-biotech platforms should publish: how many candidates were tested and how many actually improved in the laboratory. That does not completely solve the validation problem because company-run studies can still contain selection bias, but it moves the industry toward a more useful standard.
A biological AI system should increasingly be judged by practical questions. How many designs express correctly? How many bind? How many improve the desired property? How many retain the other properties that matter? How often does the result reproduce? How often does the model fail? Those numbers are far more useful to a biotech buyer than an abstract machine-learning score.
Where Converge Still Has to Prove Itself
Converge has published interesting data, but there are important limitations that should be considered. Much of the detailed validation available publicly comes from Converge itself. That is common for an emerging platform, but independent benchmarking and third-party replication would make the evidence much stronger.
Better affinity or improved developability also does not guarantee a better medicine because clinical biology is far more complicated than an in-vitro binding measurement. A strong molecular result can still fail because of safety, pharmacology, biology, manufacturing, dosing, or clinical factors.
The company’s broad product expansion also creates execution risk. Antibody optimization, genomics, cell-state modeling, biomarkers, and manufacturing are different scientific problems, and building excellent products across all of them is much harder than producing isolated demonstrations.
Finally, the economic advantage still needs to be proven repeatedly. Customers do not adopt AI because a model is technically elegant. They adopt it because programs move faster, experiments become cheaper, molecules improve, or success rates increase. Those outcomes can take time to measure.
AI Cannot Remove Biology From Biotechnology
This point is worth emphasizing because AI marketing can sometimes blur the limits. Drug development still requires physical experiments, clinical trials remain necessary, manufacturing must work, regulators require evidence, patients are biologically diverse, models can be wrong, datasets can contain errors, and assays can vary between laboratories.
Converge itself has published work questioning the reproducibility and reliability of public antibody data, including studies involving antibody-antigen affinity and developability datasets. (converge-bio.com)
That is significant because a model trained on weak or inconsistent biological measurements does not become reliable simply because it contains billions of parameters. AI makes data quality more important because the model can scale whatever signal is present in the data, whether that signal is useful or misleading.
Failed Experiments May Become Some of the Most Valuable Data
Traditional scientific publishing creates a strange bias because successful experiments are more likely to become visible while failures often disappear. AI changes the economics of failure data because an unsuccessful design can still teach the model where not to search next.
If a company generates 100 antibody sequences and 85 perform poorly, those 85 failures contain information. One sequence may bind strongly but aggregate. Another may express well but lose affinity. A third may be stable but show a predicted immunogenicity concern. A fourth may perform well across several properties but fail somewhere else in the workflow.
Over thousands of experiments, those negative and mixed results can become extremely valuable. This creates a strong reason for AI companies to operate close to physical laboratories because the system does not simply receive the final winning sequence. It can learn from the entire landscape of what worked, what failed, and why.
The Bigger Moat Could Be Experimental Memory
This leads to one of the most important strategic questions around Converge and other AI-biotech companies: what becomes more valuable as the company operates for several years?
The obvious answer is the model, but models spread quickly. Research papers are published, open-source implementations appear, computing capacity becomes more widely available, and competitors recruit machine-learning scientists with similar backgrounds.
Experimental memory is harder to reproduce because it contains repeated knowledge about how computational design choices behave in real biological systems. A company that repeatedly learns which generated sequences succeed, fail, aggregate, express poorly, remain stable, or lose biological activity can accumulate a rich map between digital design decisions and physical outcomes.
If those lessons can be used lawfully and contractually while respecting customer confidentiality and ownership, they may become a powerful competitive advantage. Converge says customers can use proprietary data in private fine-tuned model instances while retaining ownership of their data. (prnewswire.com)
How AI companies structure those data rights will likely become one of the most important commercial questions in the sector.
What Biotech Leaders Should Ask Before Adopting Generative AI
The buying process should not begin with a model demonstration. It should begin with experimental economics.
A biotech leader should calculate how many candidates normally enter a development cycle, how many experiments are required to reach a useful result, how long the cycle takes, and how much the process costs. The team can then run the same workflow using an AI system and compare the difference.
The most useful questions are practical. Has the AI system been validated experimentally on problems similar to yours? What percentage of generated candidates usually survive into wet-lab testing? How does the system handle competing objectives? Can critical parts of an antibody or protein be locked so they do not change? How is uncertainty communicated? How is proprietary data stored, used, and protected? What happens when the model is wrong?
The most important question should be asked before the project begins: what metric will define success? If the answer is simply “better model performance,” the team should keep asking until the metric becomes connected to real scientific or economic outcomes.
Boston Companies Should Build Closed Loops, Not Isolated AI Departments
Large biotechs may be tempted to create separate AI organizations that sit beside the rest of the research operation. That structure can make sense for fundamental platform work, but it can also create organizational distance between computational predictions and physical experiments.
Computational scientists may produce predictions, experimental scientists may receive spreadsheets, several weeks may pass, and results may return in a format that is difficult to reuse. In that structure, nobody has built a truly efficient learning loop.
A stronger model organizes computational and experimental scientists around the same scientific objective. The unit of progress becomes the complete cycle rather than the model or the assay by itself.
Scientists define a product profile, AI generates candidates, computational systems score them, researchers review the shortlist, the laboratory tests them, and the results return in a structured form that improves the next round. This cycle can then repeat quickly.
That approach is much closer to Converge’s “AI lab” idea, and it is also more likely to create durable productivity improvements because every part of the organization is connected to the same feedback system.
Generative AI May Eventually Become Invisible Inside Biotech
The most successful technologies often disappear into normal work because people stop talking about the tool and start talking about the task. Scientists no longer announce that they are using a computer every time they analyze data because computers have become part of the standard research environment.
The same may eventually happen with AI. An antibody researcher may not say that they are starting a generative AI workflow. They may simply open the design system, enter the constraints, and review the strongest candidates.
A cell biologist may upload a dataset and receive ranked targets. A process-development scientist may request optimized expression sequences. A discovery scientist may ask the system to compare hundreds of candidate variants before selecting the next assay.
At that point, AI stops being a special project and becomes infrastructure. Converge appears to be building toward that stage.
The Real Competition Will Be Over Trust
Models will continue improving, computing costs will continue falling, and more biological models will become open source. As those capabilities become easier to access, another scarce resource becomes increasingly important: scientific trust.
A researcher has to believe that using the system is worth sacrificing an experimental slot. That is a much higher standard than convincing a consumer to try a chatbot.
Trust will come from reproducible results, clear methods, transparent limitations, strong data handling, reliable security, intellectual-property protection, and repeated scientific success. The platforms that earn that trust may become deeply embedded in research workflows because once a system helps determine which sequence gets synthesized or which target receives millions of dollars in development resources, replacing it becomes a significant decision.
What This Means for Boston’s AI-Biotech Ecosystem
Boston has spent the last decade producing AI drug-discovery companies built around increasingly powerful scientific models. The next generation may become more operational and more deeply connected to everyday research.
Some companies will continue building massive biological foundation models. Others will develop autonomous laboratories. Some will design and own their own medicines. Another important category is emerging between traditional biotechnology and software: companies that provide reusable AI capabilities directly inside existing drug-development programs.
Converge Bio fits that pattern. Its public work now reaches from biological target discovery through antibody engineering and protein-production optimization. It combines generation with prediction, filtering, and experimental validation. It has grown from a nine-person team in late 2024 to dozens of employees, raised significant venture financing, worked with more than a dozen biotechnology and pharmaceutical customers, and reported more than 40 completed programs. (techcrunch.com)
The story is therefore larger than one startup. It gives Boston a useful view into what practical AI adoption may look like inside biotechnology.
The future may not be one robot replacing a scientist or one model discovering every medicine. It may instead be thousands of search, ranking, filtering, and optimization decisions becoming partially computational.
That could be far more transformative because it changes everyday R&D rather than waiting for one spectacular breakthrough.
Final Takeaway: The Biggest AI-Biotech Shift May Be From Generating Molecules to Choosing Better Experiments
Converge Bio’s most important contribution may ultimately have less to do with the phrase “generative AI” than it appears at first glance. Generation is only the beginning of the workflow. The real challenge in biotechnology is deciding which possibility deserves scarce experimental time.
Converge’s public antibody data makes that point especially clear. In one four-target program, roughly 120,000 computational candidates were reduced to 42 wet-lab tests. Only around 0.035% of the original design space reached physical experiments, and 18 candidates ultimately showed improved affinity. (converge-bio.com)
That is a more realistic picture of AI’s future in biotechnology. Models explore a possibility space that humans could never search manually, prediction systems narrow the field, scientists decide what matters, laboratories test the remaining candidates, and the resulting data improves the next decision.
When that loop becomes faster and more reliable across enough programs, biotechnology R&D starts to change in a fundamental way. The advantage no longer comes from simply generating more ideas. It comes from learning which ideas deserve expensive physical validation.
For Boston, this may become one of the most important opportunities created by AI biology. The region already has extraordinary experimental capacity, experienced scientific talent, world-class hospitals, major universities, large pharmaceutical companies, and a dense network of biotechnology startups.
Boston does not necessarily need to build the world’s largest biological model to win. It can build the world’s strongest connection between models and real experiments.
Converge Bio provides an early view of what that future may look like. Generative AI is moving beyond impressive demonstrations and gradually becoming part of the everyday machinery of biotech research and development.
