COMPARE

Onkydra vs Tempus, Charles River, BPGbio, Claude Science.

The same problem class, four different depths of field. Tempus, Charles River, and BPGbio are decision-grade if you're top-30 pharma. Anthropic Claude Science is the general-purpose AI workbench, launched June 30, and is now the interaction-quality bar every AI-for-science tool is compared to. Onkydra is what happens when you take that interaction quality and specialise it to the one vertical where curated anchor cohorts and evidence-size honesty actually decide whether the report gets peer-reviewed: rare paediatric cancer.

SHORT VERDICT

Different depth of field. Different buyer. Same interaction quality.

Tempus, Charles River and BPGbio are the right answer for top-30 pharma with a hundred-million pipeline. Claude Science is the right answer if the buyer is a general-purpose scientist across ten disciplines who wants a workbench, not a vertical. withZeta.ai (Lantern Pharma) is a vertical rare-cancer co-scientist that shares our buyer profile but ships as a byproduct of an internal oncology pipeline. Onkydra is the right answer when the buyer is a rare-cancer researcher, a Seed-A biotech R&D lead, or a Series-A founder who needs the indication library already curated, the CIs computed over the real anchor n, and the disagreement between paradigms surfaced as the report headline. The interaction quality bar is Claude Science's. The vertical depth is ours.

SIDE BY SIDE

The matrix.

Twelve axes, five points of comparison, one challenger. The blue column is us.

Axis	Onkydra	Tempus	Charles River	BPGbio	Claude Science	withZeta.ai
Target buyer	PhD researchers, Seed-A biotech R&D, Series-A founders	Top-30 pharma, large clinical networks	Mid-large pharma, biotechs with $20M+ pre-clinical budgets	Pharma partnership deals, public-market investors	General-purpose scientists · enterprise pharma R&D	Rare-cancer researchers, biotechs, and academic labs (same buyer as us)
Entry price	€300 / month per academic lab, unlimited people, 7-day free trial. €800 / month per commercial seat. Provisional: set against published prices in computational drug discovery software	Six-figure annual contracts (after procurement)	$150k+ per pre-clinical engagement	Multi-year partnership deals (undisclosed)	Bundled into Claude Pro ($20 / mo) / Team ($25-30 / seat) / Enterprise	Introductory / Academic / Commercial tiers (public pricing not fully disclosed)
Time from question to answer	Minutes, not weeks. Not yet measured across real runs.	Weeks to months (depending on workflow)	Months (study scoping → execution → report)	Weeks to months (project-based)	Minutes to hours (per-run, BYO compute cost)	Minutes to hours per agentic query
Indication focus	Rare and paediatric cancers. One indication wired: H3 K27M DMG, running within a stated scope. Seven scoped, not shipped.	Broad: adult oncology, growing rare	Broad: all therapeutic areas	Multi-disease (oncology, immune, CNS)	Horizontal · any life-sciences discipline	438 cancer types, chemistry/CNS-permeability-anchored generalist
Curated anchor cohorts	1 selectable: DMG, n=60 anchor. Ewing (n=222) is curated and held behind the coverage gate until DMG is proven out. Six more scoped, not curated.	Yes, proprietary patient databases	Yes, per-engagement	Yes, mass-spec + proprietary	None. User assembles ad hoc from PubMed / public databases per session.	Ontology-level (Orphanet 6,500 rare diseases, NCI-t 100K cancer concepts). No public per-indication anchor cohort.
Perturbation model	CellOracle, permissively licensed and running: in-silico knockouts precomputed over a DMG regulatory network built from 4 real scATAC-seq samples unioned with CellOracle's published human promoter network, covering 45 transcription factors, plus ACVR1 through an explicit ID1/ID3 bridge. It abstains rather than guessing for every other target, and the readout is an uncalibrated ordering signal. The reference it simulates on is a mixed cohort: about half its cells are midline high-grade glioma records and the rest are hemispheric, site-unstated or ependymoma comparators, with H3 K27M stated on no record. Alongside it, two independent stated rule sets score the same strata so their disagreement is visible.	Proprietary	Proprietary + engagement-specific	Proprietary	BioNeMo bundle: Evo 2, Boltz-2, OpenFold3. No single-cell perturbation model.	PredictBBB (blood-brain barrier, 16-model ensemble, 7,000+ compounds) + ETHER0 (chemistry). No transcriptomic perturbation model.
Methodology transparency	Methods public · preprint in draft, CC-BY on release · implementation proprietary	Proprietary	Proprietary	Proprietary (NeurOme / proprietary biology platform)	Reproducible artifacts with code + env history · reviewer is same model	Corporate roadmap disclosed (ZetaSwarm, ZetaOmics). RADR platform is internal.
Evidence size disclosed in output	Yes, bootstrap CIs over real anchor n in every report	Standard biostatistics (CI conventions vary by product)	Custom per engagement	Not customer-facing	Not enforced · depends on the user's prompt	Not enforced at the output level (ontology-and-literature-anchored)
Adversarial honesty layer	Honesty Critic, an evidence class on every layer and on every row of the predictions export, and disagreements reported rather than resolved	Internal QA	Internal QA + sponsor review	Internal QA	Built-in reviewer: same underlying model checking itself	Multi-agent agentic loop. No publicly documented different-model second opinion.
Reproducible / re-runnable	Yes, every run audit-logged, reproducible from public data	Per-engagement reports	Per-engagement reports	Per-engagement reports	Yes, code + environment snapshot embedded on every artifact	Yes (agentic loop is queryable per session)
Procurement gate at entry	None by design. Checkout is not open yet, so nothing can be bought today.	Yes, enterprise procurement	Yes, master service agreement	Yes, partnership negotiation	None, Claude Pro subscription	Introductory tier is self-serve; Commercial requires contact
Vendor is also a drug developer	No · we are a tool vendor	No	No	Partially · discovery partnerships	Yes · Anthropic runs its own preclinical programs (2026)	Yes · Lantern Pharma runs an internal oncology pipeline (RADR)

Incumbent details derived from each company's public disclosures + standard CRO contracting conventions, last reviewed 2026-07-05 and not re-read since: treat every column but ours as a July snapshot.

ONE BY ONE

Each comparison in turn.

Tempus

AI-enabled precision oncology · public (NASDAQ: TEM)

The big platform. Massive patient database, broad oncology coverage, deep pharma integrations. Enterprise pricing, procurement-gated, multi-week onboarding. If you need a stack the FDA already knows, this is it. We don't compete with Tempus on enterprise oncology. We compete with the gap they leave behind for everyone with a budget under €100k a year.

Differentiator · Onkydra ships a result in minutes, not weeks, from €300 a month. Tempus ships a result in weeks for six figures. Different products.

Charles River Laboratories

Contract research organisation · public (NYSE: CRL)

The CRO incumbent for pre-clinical drug development. In vivo and in silico mixed under a master service agreement. The bottom-end engagement starts at $150k and runs three-to-six months. A founder running a $1M seed cannot afford a single Charles River study, much less the three or four it takes to triangulate a target.

Differentiator · Charles River sells one project at a time. We sell a workspace you can re-run yourself any time. Automatic re-runs as new data lands are in build, not yet live.

BPGbio

AI-driven biology platform · pre-IPO, partnership deals

Mass-spec + AI for biomarker and target discovery. Big partnerships with pharma; closer to a discovery engine than a workspace. Strong on multi-omic integration. If you have a partnership-style deal flow and patient timelines, this is in scope. For a Series-A founder doing weekly board check-ins, it isn't.

Differentiator · BPGbio is closed-source and partnership-priced. Onkydra is open-methodology and subscription-priced.

Recursion · Cellarity · Lantern · Genialis · Aitia

Adjacent in-silico drug-discovery platforms

All real, all interesting, none of them are answering "what would CDK7 inhibition do in H3 K27M DMG?" in minutes at a price a single lab can approve. Most are phenotypic-screen platforms, multi-omic biomarker platforms, or causal-AI platforms, sold as enterprise or research collaborations. The closest thing to Onkydra in pricing model and buyer is Genialis, and Genialis is biomarker-discovery, not target-prioritisation.

Differentiator · These are valid choices for the partnership-style buyer. We sit in the gap below them.

Lantern Pharma · withZeta.ai

Rare-cancer AI co-scientist · public (NASDAQ: LTRN) · launched 2026-04-14

The closest structural match to Onkydra's buyer story. Positioned as "the world's first multi-agentic AI co-scientist for rare cancer drug discovery." Three tiers (Introductory / Academic / Commercial), 438 cancer types, and a connector list that reads like an ontology roundup: Orphanet, NCI Thesaurus, HPO, OpenFDA, OLS, ChEMBL/PubChem, Cellosaurus, ClinicalTrials.gov via AACT. Their purpose-built AI modules are PredictBBB (blood-brain-barrier permeability, 16-model ensemble on 7,000+ compounds, 94.1% accuracy) and ETHER0 (chemistry reasoning, Apache 2.0, built by FutureHouse). Real distribution: NASDAQ ticker, AACR 2026 launch momentum, an internal RADR platform they can market against. Real weakness: they are chemistry-and-CNS-permeability-anchored generalists across 438 cancers, and their agentic loop does not publicly ship a different-model second opinion.

Differentiator · withZeta is broad and chemistry-forward across 438 cancers. Onkydra is deep and transcriptomic, with a DMG anchor cohort of 60 real cases, bootstrap CIs over the real n, and an evidence class on every layer and on every exported row: what was measured, what a model computed, and what is an assumption we stated. One indication runs, within a stated scope, and it is the only one: the other seven are on the roadmap, not built, and available to nobody. Both of us have Lantern-scale distribution to beat; ours is via foundations + patient advocacy, theirs is via a public-company channel.

Anthropic Claude Science

General-purpose AI workbench for scientists · launched 2026-06-30

The one that raised the interaction-quality bar. Multi-agent subagents that fan out in parallel, reproducible artifacts with full code and environment history, plain-language editing of figures, native rendering of 3D protein structures, chemical structures, genome tracks. Bundled into Claude Pro / Team / Enterprise, so a subscriber pays nothing extra to use it. Horizontal: the demo hero was PKU, one autosomal-recessive monogenic disease. No pre-curated H3 K27M DMG cohort, no ATRT cohort, no Ewing cohort. BioNeMo ships Evo 2, Boltz-2, and OpenFold3, none of which predict a transcriptomic perturbation, so single-cell perturbation is a capability gap. The reviewer agent is the same underlying model checking itself. TechCrunch and Hacker News both flagged this as a scientific-rigor limitation. Anthropic also announced they are now running their own preclinical drug programs on neglected diseases, which turns them into a competitor to any pharma customer they sell to.

Differentiator · Claude Science is horizontal · Onkydra is vertical. Their reviewer is the same model checking itself · ours separates a mechanistic model run from a curated registry from an assumption we wrote, and says which is which on every number rather than letting one settle another. Their perturbation stack is BioNeMo · ours is CellOracle, real and precomputed over a DMG network. Their default anchor cohort is whatever the user pulls from PubMed · ours is n=60 K27M cases from DKFZ and CPTAC, pre-loaded, and tested against cases held out of the fit with the anchor sample ids excluded by construction, so the test is leakage-free (r=0.92 within the same study, r=0.59 across institutions). We do not call that externally validated, and our own expansion gate records it as unmet: the validation script computes no permutation, no shuffled-label arm and no baseline, so neither correlation has a zero to be read against, and both vectors are dominated by the same few high-frequency drivers. Quote the lower figure, which is the transportability one. We are shipping an Onkydra MCP server so a Claude Science user can call Onkydra from inside their session. Distribution over ownership.

COMMON OBJECTIONS

What about… ?

“Isn't this just withZeta.ai from Lantern Pharma?”

Closest vertical match. withZeta covers 438 cancer types with an ontology-heavy connector layer (Orphanet, NCI-t, HPO, OpenFDA, ChEMBL, Cellosaurus) and two purpose-built modules, PredictBBB and ETHER0. Onkydra ships one indication curated end to end (DMG, n=60 anchor) with seven more scoped, every number labelled as a model run, a proxy or a stated assumption, and disagreements between layers reported rather than resolved. Same buyer. Deeper science per indication. Weaker distribution than a NASDAQ ticker. That's the gap we close through foundations, methodology preprint, and the free-DMG-cohort lead magnet.

“Isn't this just Claude Science?”

Different shape. Claude Science is horizontal: one workbench for every scientific discipline, with the anchor cohort assembled ad hoc from PubMed at run time, and a reviewer agent that is the same underlying model checking itself. Onkydra ships one rare-cancer indication library curated end to end (H3 K27M DMG, n=60 anchor) with seven more scoped, separates a mechanistic model run from a curated registry from an assumption we stated, and runs a transcriptomic perturbation model where BioNeMo bundles none (Evo 2, Boltz-2 and OpenFold3 predict sequence, structure and folding, not a perturbation response). CellOracle is real today: in-silico knockouts precomputed over a DMG regulatory network built from four real scATAC-seq samples unioned with CellOracle's published human promoter network and served as an artifact, covering 45 transcription factors, with a documented bridge from ACVR1 to its downstream TFs and abstention for every other non-TF target. The cell-state reference behind it is a mixed cohort, roughly half midline high-grade glioma records and the rest hemispheric, site-unstated or ependymoma comparators, and we say so on the atlas rather than calling it a clean DMG reference. Different depth of field, same interaction quality bar.

“Can you really do what Tempus does at 1/100th the price?”

Not the same product. We do the in-silico target-validation slice (cohort sampling, dependency anchoring, resistance ranking, citation-backed reports) from €300 to €1,600 a month, with a quoted programme tier above it. We don't do Tempus's clinical-data integration or tissue testing.

“Is the methodology actually defensible?”

Yes. The methodology page is the current public write-up; the preprint is in draft and has not been posted yet. Monte Carlo Gaussian-copula sampler, Ledoit-Wolf shrinkage, co-occurrence audit, bootstrap CIs over the real anchor n, and a linear baseline fitted on measured expression and held out by sample. See /methodology.

“Why should I trust a single-founder shop?”

Don't yet. Join the waitlist; read the methodology page; check that the system flags its own knowledge gaps in every report header. Trust is something we earn run by run, not claim on a homepage.

Read the methodology See pricing Full FAQ