The evidence layer for biology.
We judge every experiment your program is built on, so your decisions stand on solid ground, with every gap in view.
PartnersbioRxiveLifeWileyPublished inEMBO ReportsIn the pressNatureScienceScientists atHarvardYaleOxfordCornellUC BerkeleySalkWeizmannTuftsTel Aviv University
PartnersbioRxiveLifeWileyPublished inEMBO ReportsIn the pressNatureScienceScientists atHarvardYaleOxfordCornellUC BerkeleySalkWeizmannTuftsTel Aviv University
Every go/no-go is a bet on evidence you can’t fully validate.
A weak link surfaces only when it’s too late.
A weak experiment, hard to validate
Discovery
Strong data added on top
Preclinical
Still resting on it
Phase I
It surfaces here
Phase II
Never reached
Phase III
~71%
of Phase II programs never reach Phase III
BIO, Informa, QLS$28B
a year spent on irreproducible preclinical research, in the US alone
Freedman et al., PLoS Biol 2015$1.3B
average cost to bring one drug to market, failures included
Wouters et al., JAMA 2020The best AI won’t fix it. The flaw is in the evidence, not the reasoning.
Today’s AI tools weigh every experiment the same. With QED, we weigh every experiment before reasoning on top.
Without a judged evidence layer, each source counts at face value. Solid lines support adding C; dashed lines argue against. Illustrative.
The same sources weighed, plus QED’s own patient-level re-analysis. Block size and line weight: how far you can count on each one. Illustrative weights.
“I had a more rigorous, attentive, constructive and useful experience with QED than with many of my peer reviewers in my 35+ years submitting papers.”
ML
Prof. Michael Levin
Tufts
“I’m impressed, the AI was very accurate at catching real gaps and suggesting logical (and mostly doable) experiments.”
BL
Dr. Bluma Lesch
Yale
“It gave me Nature reviewers-level feedback on two different papers I uploaded.”
NI
Dr. Netanella Illouz Eliaz
Salk Institute
“This is exceptionally powerful. It can create a new level of scientific integrity.”
RF
Asst. Prof. Ryan Flynn
Harvard
“This is just what I was missing, it’s already highly valuable to the lab.”
MO
Dr. Meital Oren-Suissa
Weizmann Institute
“I was surprised by the system’s precision, it accurately captured the essence of my work.”
AR
Prof. Asya Rolls
Tel Aviv University
“This should be in every research lab!”
TS
Prof. Tom Schonberg
Tel Aviv University
“I had a more rigorous, attentive, constructive and useful experience with QED than with many of my peer reviewers in my 35+ years submitting papers.”
ML
Prof. Michael Levin
Tufts
“I’m impressed, the AI was very accurate at catching real gaps and suggesting logical (and mostly doable) experiments.”
BL
Dr. Bluma Lesch
Yale
“It gave me Nature reviewers-level feedback on two different papers I uploaded.”
NI
Dr. Netanella Illouz Eliaz
Salk Institute
“This is exceptionally powerful. It can create a new level of scientific integrity.”
RF
Asst. Prof. Ryan Flynn
Harvard
“This is just what I was missing, it’s already highly valuable to the lab.”
MO
Dr. Meital Oren-Suissa
Weizmann Institute
“I was surprised by the system’s precision, it accurately captured the essence of my work.”
AR
Prof. Asya Rolls
Tel Aviv University
“This should be in every research lab!”
TS
Prof. Tom Schonberg
Tel Aviv University
The largest curated, experiment-driven dataset in biology. A new infrastructure layer for reasoning.
>500K
experiments validated
for customers and users
>200K
scientific gaps found
across our portfolio
25K
principal investigators
judging the evidence in their own fields
>2K
institutes and organizations
in our network
From your question to a decision, in five steps.
?
01
Your question
The decision your program hangs on
→
02
Gather
Every source: papers, trials, patents, datasets, your data
→
03
Curate
Every experiment judged and validated
→
04
Augment
New analyses and experiments fill the missing pieces
→
Go
05
Reason
On what holds, and only on that
Holds upUnconfirmedDidn’t holdAdded by QED
You get
The decision
With its confidence, and the one experiment that would change it
The knowledge layer
Every experiment behind it, judged and linked. Yours to keep and query
With QED, an answer becomes a decision.
General AIOn evidence it can’t weigh
[1][2][3][4][5][6][7][8][9][10][11][12][13][14]
More to read. The same uncertainty.
QEDOn evidence it has weighed
Node size and line weight: how far you can count on it. Dashed: points against.
Less to read. A decision you can trace to every experiment.
You get the decision, and the evidence layer behind it.
Every experiment weighed and linked. Yours to query, augment with your own data, and build every next decision on.
1,080 experiments · 9 datasets · 3 new analyses · every link mapped
IDTypeFindingWeightLinks
EXP-0412PaperTarget knockdown reduces disease marker↑
14 links
DS-GWAS-07GWASVariant near target linked to disease risk↑
31 links
EXP-0877TrialPhase II, randomized: primary endpoint missed↓
9 links
DS-PROT-02ProteomicsTarget protein elevated in patient tissue↑
22 links
QED-RA-01Re-analysisEffect not seen with the patient as the unit↓
17 links
EXP-1033PatentBinder affinity, no replication reported↑
3 links
EXP-0156PreprintOrganoid screen supports mechanism↑
6 links
DS-SC-04scRNA-seqPathway active in two of five cell types↑
19 links
EXP-0703FilingSafety signal at high dose↓
8 links
EXP-0288PaperMouse model, n = 6, never repeated↑
2 links
EXP-0961PaperEffect holds in second species↑
11 links
DS-COH-03CohortBiomarker tracks response in 212 patients↑
26 links
+ 1,068 more experiments, each weighed and linked
Yours to build on
Interrogate it
Ask the next question. Every answer stays anchored on the same judged evidence.
Query the data
GWAS to proteomics, linked to every experiment that touches them.
Run new analyses
Ad hoc re-analyses of raw data wherever the literature is thin.
Deploy it your way
On our platform, through an API, or on your own premises.
Add your own data. Secured for you, never trained on.
One real engagement’s counts; rows and map illustrative.
The same frontier model, alone and on QED’s data layer, against an expert panel.
Models reason well. What they reason on decides the answer.
What an R&D decision needs
40%Agreement with expert panel100%
Model alone+ QED layerΔData coverage
found the decision-relevant evidence
58%†
94%†
+36
Evidence validation
judged whether each experiment holds
61%
87%
+26
Gaps found
identified gaps in the R&D logic
76%
92%
+16
Suggested experiments
proposed the experiment experts would run
70%
85%
+15
Harmonization
reconciled conflicts across sources and models
49%†
83%†
+34
Grounding
every claim traceable to a primary experiment
64%†
99%†
+35
Mechanistic precision
kept distinct mechanisms apart
52%†
88%†
+36
Recency
included evidence from the last 24 months
45%†
96%†
+51
Frontier model alone Same model on QED’s data layer Expert-to-expert agreement (ceiling)
Error profile · gaps found
PrecisionRecallFP / qFN / q
Model alone0.58†0.47†2.1†3.4†
+ QED layer0.91†0.86†0.4†0.9†
Calibration. When QED reports 80% confidence, the panel agreed 78%† of the time.
False positive: asserted, but rejected by the panel majority. Cost: a wasted experiment or a wrongly stopped program.
False negative: identified by the panel, missed by the system. Cost: a liability found late.
False negative: identified by the panel, missed by the system. Cost: a liability found late.
† illustrative, pending benchmark run · evidence validation, gaps and experiments from QED’s head-to-head
Pharma R&D teams already bring us their hardest calls.
Chief Strategy Officer
Top-20 pharma
Combination design
Does a third arm earn its place in our bispecific?
Uses: QED Review · program models
Director of Bioinformatics
Global pharma and life-science group
Indication selection
Which disease should this asset go after first?
Uses: program models
SVP
Clinical-stage biotech
Portfolio prioritization
Which few of 30–50 targets deserve investment?
Uses: program models
Director of [Clinical Development]
Top-10 pharma
Scientific review
QED Review, built into their internal publishing workflow.
Uses: QED Review · API
Real engagements · names withheld · ranked by revenueEach square: one judged experiment behind the decision
Case studyCombination design · Global pharmaAssessment delivered October 2026
Does a third arm, C, earn its place on a validated A × B?
What we did
1Judged 1,080 experiments: 411 papers, 107 trials, 43 patents, 123 regulatory records
2Re-analyzed raw patient data where the published evidence was thin
3Ran the same question through best-in-class AI agents
Fig. 1 | The published signal for C, re-tested per patient. a, C-pathway signal in treated vs untreated lesions at the same body site (6 vs 5 patients). Filled: per-cell Wilcoxon test with Bonferroni correction, as published. Open: the same comparison with the patient as the unit (Benjamini–Hochberg). b, The per-cell test treats ~129,000 correlated cells as independent observations.
What changed
Published signal for C
P = 8 × 10⁻⁴⁸→not significant per patient
Evidence base
2 primary papers→107 trials, 1,080 experiments
AI agents
lean no→open: one experiment decides
Indications
12 candidates→lead with 1; 2 needs no third arm
46
orders of magnitude: the published signal, re-tested per patient
429
trial participants missing from the AI answer
0
errors found in the client’s review
1
decisive experiment, before the big spend
“QED found evidence our own searches and the AI tools had missed, and turned an open debate into one experiment we can run.”Head of R&D Strategy, global pharma
We are a team of scientists and builders sharing a vision.
Physicians, scientists and AI leaders, from Google, Teva, MeMed and SparkBeyond.
Niv Samuel Mastboim
MD, Technion
Oded Rechavi
PhD, Tel Aviv University
Prof. @ Tel Aviv University
Enav Weinreb
PhD, Ben Gurion University
Formerly @ Google, Thomson Reuters
Meir Maor
BS, Computer Science, Technion
Formerly @ SparkBeyond
Guy Ronen
BS, Computer Science, Hebrew University
Formerly @ Google
Orit Shaked
Formerly @ Teva, CytoReason
Lior Onn
PhD, Ben Gurion University
Mor Sela
PhD, Tel Aviv University
Efrat Goffer
PhD, Harvard-MIT HST
Lilach Bien
MSc, Computer Science, Tel Aviv University
Sagi Hammer
BS, Computer Science, The College of Management
The four questions every R&D head asks.
“Why not just use an LLM agent?”
It reads every source the same. We weigh each one, with the experts in its field.
“We already have evidence platforms.”
They find the papers. We tell you which experiments hold, and what’s missing.
“How deep is the coverage?”
400,000 experiments judged, 200,000 gaps tagged, growing every day.
“Is our data safe?”
Never trained on. SOC 2 Type II, ISO 27001, and your data stays in your control.
Bring us one decision. We’ll show you what it really stands on.
Book a demo →Never trained on your data · SOC 2 Type II · ISO 27001