BenchSci vs Edison

A new kind of AI has arrived in drug discovery: autonomous agents that read the literature, run the analysis, and propose the next experiment on their own. Several vendors now describe these systems as an AI scientist that works alongside a research team rather than a tool the team operates by hand. For a preclinical organization, the real test is how much of that autonomy holds up inside its own biology, across a program that runs from an early target call to an IND filing.

EMET by BenchSci and Edison both come up in that evaluation. Edison Scientific is the commercial spinout of FutureHouse, the nonprofit lab, and its product Kosmos is an "AI Scientist for R&D" that reads literature, analyzes data, generates hypotheses, and runs scientific workflows autonomously from target identification through clinical development and filing. EMET is narrower by design, built only for biopharma R&D and its preclinical core.

This comparison runs EMET and Kosmos through the five criteria that decide whether an AI platform earns a lasting place in preclinical R&D: the data each system reasons over, the depth of its biology-specific workflows, the service that drives adoption, the freedom to use the best model for each task, and where the vendor stands on developing drugs of its own. The two land in the same place on model flexibility and on neutrality, since both are AI vendors rather than drug developers. On the other three, BenchSci pulls ahead.

What to look for in AI for preclinical R&D

When you evaluate AI for preclinical R&D, the demo rarely tells you what you need to know. The traits that decide whether a platform is still earning its place a year later are structural, and they are easy to miss when the output on screen looks polished. Five of them do most of the work.

• Proprietary, science-specific evidence to reason over. A model can only reason as well as the evidence beneath it. Curated, structured biology that a competitor cannot reach beats a general index of the same public sources, and the gap is widest on the questions that decide a program.
• Depth of biology-grounded workflow and integration. Preclinical work runs from a literature question to a sequencing run to a decision at the bench, and it is specific to biology at every step. Value comes from a platform that acts across that whole span in biology, not one that spreads thin across every field.
• A service model that drives adoption. Buying software changes nothing on its own about how a lab works. People on the ground who build workflows for specific teams are what decide whether the tool is still in use a year later.
• The freedom to use the best model for each task. No single model wins every scientific job, and the frontier moves constantly. A platform welded to one lab's models inherits that lab's roadmap, pricing, and outages.
• A clear position on whether the vendor develops its own drugs. You are handing this system your targets, your assays, and the programs that define your pipeline. Whether the vendor also develops medicines of its own is worth understanding before that IP leaves your walls.

Few platforms carry all five with equal weight. A capable general-science agent can read, reason, and run analyses well while leaving the deepest biology data, the biology-specific execution, and the rollout lighter than a preclinical team needs. The sections below take EMET and Edison through each criterion in turn.

Criterion 1: The data and evidence the platform reasons over

Evidence sets the ceiling on any answer. Two systems can use the same reasoning models and still diverge sharply, because the one reasoning over deeper, harder-to-reach biology reaches conclusions the other cannot.

Edison reasons over the literature and your own data

Edison's research lineage is real, and its agents are good at reading and reasoning over the literature. Kosmos reads literature, analyzes data, and generates hypotheses, and through its Incyte partnership it now integrates with an organization's ELNs, assay data, biomarker data, and unstructured notes to reason across proprietary internal knowledge.

That is a strong design for turning a customer's own data into insight. The evidence Kosmos reasons over is the literature plus whatever each customer brings. Edison builds internal scientific world models from those inputs, and it doesn't describe a proprietary, licensed data asset of its own.

BenchSci reasons over closed-access science and experimental data no one else holds

A decade of BenchSci's data work sits under EMET. The platform reasons over a proprietary Biological Evidence Knowledge Graph of 858M nodes and 2.2B relationship edges, built from 38M+ publications. That graph is one of the largest structured maps of disease and target biology built for drug discovery. Beneath it sit 16M closed-access papers, opened through eight years of licensing with publishers like Elsevier, Springer Nature, Wiley, and Oxford University Press, and reachable in EMET and nowhere else.

The part that is hardest to match is not the literature. It is the reagent and model-systems layer: 16M antibodies, 22M RNAi entries, 18M CRISPR records, 500K cell lines, and 700K animal models, curated by a team of 60 PhD scientists. That experimental layer answers the practical questions that decide a bench experiment: which antibody, which cell line, which model.

A neuro-symbolic evaluation loop ties every generated output back to that verified graph, reaching 95%+ accuracy on biological questions across 600+ tests and 8+ benchmarks, two to four times better than frontier models alone. Both systems cite their sources and read your internal data. The difference is the depth and exclusivity of the biology each one starts from.

Criterion 2: Depth of biology-grounded workflow and integration

Discovery does not live inside a chat box. It moves from a literature question to an omics dataset to a decision at the bench, and it is specific to biology at every step. The question is how much of that work the platform can actually do, and how deep it goes in biology specifically.

Edison runs autonomous workflows across the R&D lifecycle

Kosmos is built to act, not only to answer. It steers investigations, branches hypotheses, and executes scientific workflows autonomously, provisions compute for GPU-heavy tasks such as AlphaFold and molecular docking, and Edison says it has rearchitected Kosmos to span the full R&D lifecycle from discovery through clinical development and filing. It is a capable autonomous scientist, and that breadth is part of the pitch.

That breadth is also the trade-off. Kosmos is a general-purpose AI scientist that stretches from discovery to filing, so its depth is spread across the lifecycle rather than concentrated in preclinical biology.

EMET runs the biology, deeply and connected to your data

Biology has fragmented across thousands of sources and hundreds of tools, and scientists lose 40% to 60% of their time to finding and cleaning data. EMET carries 200+ proprietary scientific skills and more than 250 agents and workflows that move a program across the full preclinical journey, from target identification and validation through lead optimization, safety, and IND-enabling studies.

It also does the biology-specific analysis itself. EMET runs bioinformatics pipelines such as FastQC and RNA-seq differential expression from inside the conversation, fetches and visualizes protein structures, and retrieves ADMET data. It connects to your internal and proprietary data through connectors BenchSci's scientists build for each account, and acts in the lab through a lab-in-the-loop design.

Fusing those pieces into one biology-grounded system is what makes EMET a platform rather than a marketplace, where the value comes from binding the pieces together in biology, not from the pieces themselves. Each new connector and workflow raises the value of the next. A general-science agent can run a wide range of analyses; EMET concentrates that same energy on the preclinical biology stack and the systems a lab already runs on.

Criterion 3: The service model that drives adoption

Good software still dies if nobody adopts it. Whether a platform survives its first year rarely comes down to the interface. It comes down to who shows up to bend the tool around each team's real work. Both vendors staff that work with scientists, and here the two are closer than on the criteria above.

Edison embeds forward-deployed scientists

Edison does not hand customers a login and walk away. It embeds scientists rather than only shipping software, placing forward-deployed experts alongside a customer's own team and building custom world models on the customer's data. For an enterprise standing up AI across R&D, that hands-on model is a real strength and worth weighing.

BenchSci embeds a dedicated scientific team per account

BenchSci structures the same idea into a standing commitment. Every deployment comes with a dedicated team of PhD scientists fluent in your biology and your therapeutic areas, who work alongside your researchers and build to your stack.

That team leads training, steers the change, and builds the connectors and workflows that fit how your people actually operate, tailored per organization rather than pulled off a shelf. It stays involved as an ongoing scientific partner rather than a launch-week deployment. Both companies agree that science adoption takes scientists; BenchSci commits a dedicated, standing team to each account to make it stick.

Criterion 4: Model flexibility and vendor independence

No lab holds the frontier for long, so the platforms worth committing to are the ones free to reach for a better model whenever one appears. This is the criterion where BenchSci and Edison land in the same place.

Edison stays model-agnostic as new models ship

Edison builds no frontier model of its own to lock customers into. Kosmos remains model-agnostic and adopts new state-of-the-art models as they emerge, and Edison can also fine-tune the models underlying Kosmos on partner data. A customer's research is not tied to one model vendor's roadmap.

BenchSci is model-agnostic and routes every task to the best model

Because EMET builds no model of its own, every question can go to whatever performs best on it. The platform selects per step across 40+ models: a frontier model for open reasoning, a specialized biomedical model such as ESM-2 or AbLang2 where sequence or structure is the question, and each new frontier model folded in as it ships.

On the question that decides lock-in, whether your research is tied to one lab's models, both platforms answer the same way. Neither welds you to a single provider, and both stay free to adopt the next model that beats the last.

Criterion 5: Whether your AI partner also develops drugs

In drug discovery the vendor relationship is unusually intimate. You are trusting a partner with your targets, your assays, and the programs that define your pipeline, so it is fair to ask whether that partner might develop drugs of its own from what it learns. This is the second criterion where the two stand on similar ground, because both are AI vendors rather than drug developers.

Edison is an AI company, not a drug developer

Edison sells Kosmos to pharma R&D teams, as its partnership with Incyte shows, and it runs no drug pipeline of its own. Its stated mission is to accelerate the full pipeline of new medicines, and it co-builds new biotechs alongside partners such as Population Health Partners, but those drug programs sit inside the partner companies rather than inside Edison. For a preclinical team judging whether the vendor itself will turn into a drug competitor, Edison stays on the tooling side.

BenchSci has committed never to develop its own drugs

BenchSci makes that same posture a categorical, permanent commitment: it will never discover its own drugs. The commitment is structural, since BenchSci's only path to winning is to make your scientists more effective, so it has no program of its own that your data could feed. It is built into how EMET runs, inside your own cloud, with your data isolated and never used to train a shared model. BenchSci is SOC 2 Type II attested and GDPR and CCPA compliant.

So neutrality is not where these two separate. Both are AI vendors rather than drug developers, and a preclinical team weighing IP risk finds little daylight between them. The real separation runs through the other three criteria: the data each platform reasons over, the depth of the biology-specific work it does, and the people who make it stick.

Which platform belongs in preclinical R&D

This was never an argument about whether Edison is good. Kosmos is a capable autonomous scientist, it reasons well over the literature and a customer's own data, and Edison embeds real scientists to make it land. Across the five criteria that decide a preclinical platform, EMET and Edison are evenly matched on model flexibility and on neutrality, since both are AI vendors rather than drug developers. On proprietary data, biology-specific workflow depth, and the service model, the separation runs in BenchSci's favor.

EMET reasons over 16M closed-access papers and a PhD-curated reagent and model-systems database that exist nowhere else, where Kosmos reasons over the literature and the data you bring. EMET concentrates its execution on the preclinical biology stack and connects to the data your lab already holds, where Kosmos spreads across the whole R&D lifecycle. And BenchSci embeds a dedicated, standing team of PhD scientists in every account.

Edison is a strong choice for teams that want a broad, autonomous agent reading and reasoning from discovery through filing. EMET by BenchSci is built to go deeper into the biology itself: grounded in proprietary data no one else holds, executing the preclinical analysis, deployed inside your walls, and staffed by scientists. When the question is where your programs and your data will live for the next several years, that depth is the difference that lasts.

See what EMET does on your own targets. Book a demo with the BenchSci team.