Converge Bio’s $25M round says something useful about AI drug discovery

Converge Bio has raised $25 million, with Bessemer leading and executives from Meta, OpenAI, and Wiz backing the company. Plenty of AI biotech startups can still raise money. The more useful signal is what Converge says it's selling.

The company says it's already shipping production systems for three jobs pharma teams actually pay for: antibody design, protein yield optimization, and biomarker or target discovery. It has run more than 40 programs for over a dozen pharma and biotech customers across the U.S., Canada, Europe, and Israel, and it's expanding into Asia. Since its $5.5 million seed in November 2024, the team has grown from 9 people to 34.

None of that means Converge has cracked AI drug discovery. Nobody has. It does suggest buyers are moving away from broad "foundation model for biology" pitches and toward narrower systems that fit existing R&D pipelines and can be judged on output.

That's good for the category.

The stack looks familiar, which is a good sign

Converge's setup is straightforward in concept, even if building it well is hard.

For antibody design, the company combines:

• a generative model that proposes candidates
• predictive models that score those candidates on relevant properties
• a physics-based docking layer that simulates 3D interactions and estimates binding

That's a lot closer to how serious teams actually work than the usual pitch about one giant model designing molecules end to end. In practice, drug discovery systems are staged. One model generates. Another filters. Physics checks whether the proposal survives contact with structure and chemistry. Wet lab work decides what mattered.

Converge's reported results fit that pattern. The company says some antibodies reached single-nanomolar affinity, and one protein yield program delivered a 4x to 4.5x improvement in a single computational iteration. If those numbers hold across programs, they're strong. They also point to where the value is. These systems don't have to discover new biology from scratch. If they narrow the search space, cut dead ends, and improve candidate quality before expensive experiments start, customers will care.

That's enough to build a business.

This matters because it isn't centered on LLMs

Converge CEO Dov Gertz has been direct about the limits of text LLMs for molecular work. Fair enough. A chatbot hallucination is annoying. A bad sequence suggestion can waste weeks of lab time and a lot of money.

So Converge reportedly uses LLMs mostly for literature exploration, summarization, and reference linking, not for core molecular reasoning. For the harder scientific work, it leans on sequence models, predictive models, and physics constraints.

That choice makes sense.

Drug discovery data doesn't behave like internet text. Protein and antibody sequences come with structure, evolutionary bias, assay noise, and hard physical limits. Small changes can swing binding, stability, viscosity, manufacturability, or immunogenicity in ways a text-trained model won't reliably capture. Anyone still selling general-purpose LLMs as the engine of molecular design is either talking to nontechnical buyers or dodging validation questions.

The stronger teams now treat language models as a supporting tool, usually for retrieval or annotation. The decision layer sits elsewhere.

The pipeline matters more than novelty

A technical reader will notice that little in Converge's stack looks exotic on paper. Protein language models, graph-based predictors, docking, active learning, Bayesian optimization, assay feedback loops. Versions of this already exist across the field.

That's why execution matters.

A molecular design platform usually lives or dies on a few unglamorous details:

Data contracts

Sequence, structure, assay, and literature data all carry messy metadata. Buffer conditions, temperature, cell line, assay protocol, expression system, readout type. If that isn't encoded consistently, the model learns junk associations and validation gets shaky fast. "Data quality" sounds vague until one mislabeled assay condition sends the whole optimization loop in the wrong direction.

Uncertainty gating

If the system is proposing expensive wet lab experiments, calibrated uncertainty matters as much as top-line accuracy. Out-of-distribution detection, conformal prediction, ensemble variance, simple thresholding on known liabilities. These aren't extras. They decide whether the platform is a useful filter or a budget burner.

Multi-stage compute

Physics helps, but it's expensive. Docking every generated candidate at high fidelity is a bad use of compute. The practical pattern is tiered scoring: fast triage first, heavier docking or molecular dynamics for the shortlist. Obvious on paper. Still where plenty of teams run into trouble when a research prototype meets production budgets.

Auditability

If the software plugs into regulated workflows, every generated candidate needs lineage. Dataset version, model version, features, prompts if LLMs are involved, scoring outputs, simulation runs, and links to wet lab results. Reproducibility is already hard in standard ML. In pharma, weak provenance turns into a trust problem quickly.

That seems to be the product Converge is selling: a controlled pipeline with measurable checkpoints.

Why the timing fits

The market helps explain why this round happened now.

AlphaFold turned structure prediction into standard infrastructure instead of a moonshot. Eli Lilly and Nvidia are building dedicated AI compute for pharma. At the same time, hundreds of startups are trying to sell some piece of AI-enabled drug discovery, and buyers are harder to impress.

That changes what counts as differentiation.

A few years ago, saying "we trained a biology foundation model" could open doors. Now customers want proof that the system fits target discovery, hit identification, developability, and scale-up. They want case studies. They want integration work. They want someone to own the ugly handoff between model output and lab workflow.

There's still plenty of hype here. The vendor bar is getting higher, and that's healthy.

Where I'd stay skeptical

Converge's numbers are promising, but the caveats are obvious.

First, 40-plus programs sounds solid, but "program" is a slippery unit. It could mean a substantial long-running engagement or a smaller pilot. Without win rates, baseline comparisons, and wet lab follow-through, the topline count shows demand, not necessarily durable product advantage.

Second, reported affinity and yield gains don't automatically turn into approved drugs or even viable candidates. Antibody design has plenty of traps after initial binding. Developability problems can kill a strong binder. Protein expression gains can disappear when conditions, systems, or scale change. Early metrics matter. They're still early.

Third, biology data is biased and sparse in irritating ways. Public sequence databases overrepresent what researchers chose to study. Assay data is fragmented and noisy. Private customer data helps, but it also makes outside evaluation of generalization harder.

So yes, this looks technically credible. No, the problem isn't solved.

What senior engineers should take from it

If you're building in this space, or buying from vendors like Converge, a few priorities stand out.

Treat architecture like a systems problem

You'll probably need multiple models doing different jobs, not one monolith. Generation, property prediction, structural scoring, experimental prioritization, and literature retrieval need clear interfaces and handoff criteria. Clean boundaries beat model sprawl.

Build feedback loops early

Wet lab data has to flow back into the system quickly enough to matter. That means versioned datasets, schema validation, experiment tracking, and retraining paths that aren't held together by notebooks and luck.

Optimize for decision quality

The useful output isn't "novel sequence generated." It's "candidate we're willing to spend money testing." That puts the focus on calibration, ranking quality, safety filters, and reproducibility. Fancy generation with weak gating is a bad product.

Watch infrastructure cost

Long-sequence transformers, structure-aware models, and simulation workloads can burn through GPUs quickly. Efficient attention kernels like FlashAttention, mixed precision, batching discipline, and a tiered compute plan matter. So does knowing when to skip the expensive step.

Treat security and governance as product features

Pharma buyers care about IP isolation, access controls, audit trails, and where data is stored. Startups that treat those as enterprise extras usually find out otherwise during procurement.

What this round signals

The investor list helps. The more interesting part is that Converge is pitching AI drug discovery in a way that sounds grounded: domain models instead of chat models for core science, constrained generation, physics paired with prediction, and software built to survive real R&D workflows.

That's a better direction than broad claims about foundation models solving biology. It's also harder to build, which is why this corner of the market is finally getting interesting.