Gyri for the life sciences

The biomedical
knowledge graph
your AI walks
for you.

A persistent, queryable graph of papers, trials, targets, drugs, and SEC filings — wired into your AI through MCP. Ask in natural language. Gyri executes the multi-hop traversals that take a postdoc a week, returns the answer in seconds, and remembers what it learned for next time.

workspace://biomed · live conversation live
i. You ask your AI.
The user asks a natural-language question. No GraphQL, no schema, no field names — just the question they'd ask a senior colleague.
"Which protein targets in IPF have strong genetic evidence and a confidently-folded structure I could actually drug — and where's the IP and recent science already pointing?"
ii. Your AI sends a query to Gyri.
Your AI requests structured data from Gyri through MCP, translating the question into a multi-hop graph traversal. One round trip across every relevant source.
// your AI translates the question into an 8-hop graph traversalquery DeepTargetTriage {disease(id: "EFO_0000768") { // IPFassociatedTargets {geneSymbol score evidenceCountentity {alphafold { globalPlddt }mousePhenotype { nPhenotypes }proteinDrugs { summary { nDrugs } }}} } }// + cross-reference IP and recent literaturepatents: googlePatentsSearch(query: "TERT telomerase fibrosis") {hits { title patentId assignee publicationDate }}papers: openalexSearch(query: "TERT IPF", publicationYearMin: 2023) {hits { openalexId title citedByCount primarySourceName }}}
8 hops7 sources~3,200 tokens
iii. Gyri responds to your AI.
Gyri responds with typed, citable, hydratable entities — every claim bound to a source the next model can re-fetch and verify. Persisted to the workspace by default.
// 25 targets · 17 patents · 470 recent papers · joined in one round tripdisease.name: "idiopathic pulmonary fibrosis"associatedTargets[]: 25 · patents[]: 17 · papers[]: 470 TERT · score 0.86 · 24,865 evidence · pLDDT 80.2 DSP · score 0.84 · 20,975 evidence · pLDDT 75.3 RTEL1 · score 0.84 · 10,260 evidence · pLDDT 71.7… 22 more
expand full response · patents + literature
TERT · score 0.86 · 24,865 ev · pLDDT 80.2 · 33 mouse phenotypes DSP · score 0.84 · 20,975 ev · pLDDT 75.3 · 41 mouse phenotypes RTEL1 · score 0.84 · 10,260 ev · pLDDT 71.7 · 29 mouse phenotypes PARN · score 0.79 · 2,533 ev · pLDDT 81.1 · 2 mouse phenotypes MUC5B · score 0.65 · 2,548 ev · 23 mouse phenotypes PDGFRB · score 0.59 · 15,800 ev · pLDDT 72.4 · 91 mouse phenotypes FGFR1 · score 0.59 · 16,441 ev · pLDDT 73.9 · 202 mouse phenotypes FGFR2 · score 0.59 · 13,456 ev · pLDDT 73.9 · 279 mouse phenotypes… 17 more (FAM13A, ABCA3, SFTPC, EDNRB, EDNRA, TOLLIP, IFNGR1, FLT1, FLT4, FGFR4, ACTRT3, DPP9, AKAP13, MUC5AC, IFNGR2, TNFRSF6B, PTGIR)
"Compounds, compositions, methods, and kits relating to telomere extension"patentId: US20240277752A1 · filed 2024-08-22assignee: Stanford University "Telomerase-containing exosomes for treatment of aging-associated diseases"patentId: EP3920889B1 · filed 2024-11-13assignee: UT System "PAPD5 inhibitors and methods of use thereof"patentId: US12171755B2 · filed 2024-12-24assignee: Children's Medical Center · PAPD5 inhibition restores telomere length "Risk prediction, diagnosis, prognosis for IPF"patentId: US20230313307A1 · National Jewish Health… 13 more across telomere-maintenance, fibroblast modulators, kinase inhibitors
"KIF15 missense variant is associated with the early onset of IPF"W4387216921 · Respiratory Research · 2023 · 6 citations "Mapping the genetic architecture of IPF: meta-analysis"W4388714602 · Gene · 2023 · 7 citations "Clinical, radiological and histopathological features of familial PF"W4399571280 · Respiratory Research · 2024 "Radiographic Lung Abnormalities in First-Degree Relatives of PF Patients"W4316127164 · CHEST Journal · 2023 · 10 citations… 466 more, all with hydratable refs
412ms+25 targets · 17 patents · 470 paperspersisted
iv. Your AI answers you.
Your AI weaves the structured response back into plain language, with the underlying refs available the moment you want them.
"Three stand out: TERT, DSP, and RTEL1. Each clears 0.83 on Open Targets and folds confidently in AlphaFold. The telomere story is where the field is converging — Stanford filed US20240277752A1 on telomere extension last August, Children's Medical Center patented PAPD5 inhibitors that restore telomere length, and 470 papers since 2023 corroborate. RTEL1 is the under-appreciated one: same evidence tier as DSP, far less crowded patent space."
v. Saved for later — with full hydration and citations.
The whole exchange is persisted as a typed insight. Every claim binds to a citable source the next model can re-fetch. Every reference is hydratable on demand. Tomorrow's session — yours, your colleague's, or a model that doesn't exist yet — picks up exactly where this one ended.
// stored as
insight:ipf-targets-2026-q2
"IPF target triage with structural and IP context (8-hop)" — pinned to workspace://biomed
// citations · all hydratable
312 refs · 25 targets · 17 patents · 470 papers
Each pulls back to openalex, opentargets, alphafold, impc, chembl, google_patents — re-fetchable on demand.
// resolvable later by
any future session, any model, any colleague
Tomorrow's model reads from the same substrate. Your reasoning compounds, instead of evaporating with the chat window.
// connected graph
linked to 3 existing IPF insights
insight-c9895498-f40 · insight-30e1f386-937 · insight-563b0b2c-0c7
auto · 9s
1 question · 5 steps · persisted 412ms · 8 hops · 7 sources joined
OpenAlex · 250M works ClinicalTrials.gov Open Targets FDA · DailyMed · FAERS ChEMBL Bioactivity UniProt AlphaFold MONDO · EFO · MeSH SEC EDGAR Filings Google Patents DepMap Essentiality IMPC Mouse Phenotypes Gene Ontology ClinVar · GWAS OpenAlex · 250M works ClinicalTrials.gov Open Targets FDA · DailyMed · FAERS ChEMBL Bioactivity UniProt AlphaFold MONDO · EFO · MeSH SEC EDGAR Filings Google Patents
Trust roots · 14 federated sources

Anchored to the literature, by default.

// CORPUS
OpenAlex
250M+ scholarly works, resolved at work-ID level.
// TARGETS
Open Targets
Genetic + clinical evidence for target-disease pairs.
// REGULATORY
FDA
DailyMed labels, FAERS adverse events, approvals.
// TRIALS
ClinicalTrials.gov
Active + historical trials with outcome data.
// CHEMISTRY
ChEMBL
Bioactive molecules with assay-level provenance.
// STRUCTURE
UniProt
Protein sequence, function, cross-references.
// PREDICTION
AlphaFold
Structural predictions with confidence anchors.
// FUNCTIONAL
DepMap
Cell-line essentiality across the cancer proteome.
// ONTOLOGY
MONDO / EFO
Disease ontologies for cross-source disambiguation.
// PHENOTYPE
IMPC
Mouse phenotyping, knockout signal, reverse genetics.
// FUNCTION
Gene Ontology
Functional annotations across the proteome.
// MARKETS
SEC EDGAR
8-K, 10-Q, 10-K, S-1 — full-text searchable.
// IP
Google Patents
Patent filings, assignees, priority dates.
// VARIANTS
ClinVar · GWAS
Variant interpretations + association catalogs.
// + YOURS
Your data
Internal claims, syntheses, lab notebooks — federated.
Capabilities

A knowledge graph
your AI doesn't forget.

01 / Persistence

Every claim, refutation, and synthesis becomes a typed artifact.

Conversations end. Their reasoning doesn't. Each assertion your AI makes is recorded as a structured object with its evidence chain intact — addressable across sessions, models, and providers. Tomorrow's session inherits today's work.

02 / Portability

Carry your reasoning between any frontier model.

MCP-native. Claude today, GPT tomorrow, the next thing after that — all read from the same substrate.

03 / Multi-hop traversal

Cross-source joins that took a postdoc a week.

Targets to trials to sponsors to SEC filings. Diseases to targets to drugs to adverse events. One GraphQL query. One round trip. Provenance preserved at every edge.

04 / Verifiability

Every claim binds to a specific anchor.

Not a footnote — a typed reference the next model can re-fetch and re-evaluate. Disagreement gets localized to the exact binding that broke, not lost in prose.

05 / Federation

Your data, queried at the edge.

No customer-data ingestion. Gyri queries against your sources at runtime, with your trust roots, your epistemology.

06 / Compounding intelligence

Quality goes up with every query.

Each model invocation against your workspace deepens the graph: emits new claims, hydrates old ones, surfaces contradictions, accrues citations. The substrate is non-decreasing in value as your team uses it.

Multi-hop traversals

One question.
One query. Six sources.

// traversal · i · 4 hops
Triage every druggable target in a disease.
"For idiopathic pulmonary fibrosis, give me every protein target with strong genetic evidence — and for each, the drugs in development and how their trials are reading out."
the path
i.
Disease Resolve MONDO:0007739 (IPF) and pull associated targets.
ii.
→ Targets Filter by Open Targets maxScoreDirect ≥ 0.7 — strong genetic evidence.
iii.
→ Drugs For each target, pull indicationDrugs with phase + sponsor.
iv.
→ Trials Hydrate trial outcomes, readouts, stop reasons.

A six-source join that would take a research associate the better part of a week. Returns in 412ms with provenance intact.

POST /workspace/biomed/graphql 412ms · 4 hops
query TargetTriage($id: ID!) { disease(id: $id) { name aliases associatedTargets { name otAssociation { summary { maxScoreDirect totalEvidenceDirect } } depmapEssentiality { isEssential pctStrong } indicationDrugs { name phase mechanism trials { nctId recruitmentStatus whyStopped } } } } } // variables: { "id": "EFO_0000768" }
response · disease resolved · 14 targets
live · verified data
disease.name: "idiopathic pulmonary fibrosis"
disease.aliases: ["IPF", "UIP", "cryptogenic fibrosing alveolitis", …25 more]
TGFB1 · otScore 0.80 · 133,024 evidence pieces consensus
indicationDrugs: pirfenidone approved, nintedanib approved
CCN2 (CTGF) · otScore 0.46 · pamrevlumab trial-fail
14 targets surfaced for IPF. TGFB1 holds the strongest direct evidence (0.80, 133K evidence pieces) and anchors both approved therapies. CCN2's pamrevlumab missed primary in NCT03955146 — worth knowing before you cite the case for it.
expand full response
TGFB1 · otAssociation.summary: {
maxScoreDirect: 0.8005, nDiseasesDirect: 3209
nDiseasesScoreAbove0_7Direct: 2, totalEvidenceDirect: 133024
}
depmapEssentiality: { isEssential: false, nLines: 1183, nTissues: 29 }
indicationDrugs: 8
PIRFENIDONE · approved · 30+ indications
NINTEDANIB · approved · tyrosine kinase inhibitor

JAK2 · otScore 0.80 · nDiseasesScoreAbove0_7: 8
predominantly hematology indications · indirect IPF relevance
CCN2 (CTGF) · otScore 0.456 · nDiseasesDirect: 1101
pamrevlumab · NCT03955146 terminated · did not meet primary FVC endpoint
NCT04419506 also terminated for futility
PDGFRA · otScore 0.71 trial-fail
LOXL2 · otScore 0.68 simtuzumab failed
… 9 more (full list available)

_meta: {
sourcesJoined: ["open_targets", "clinicaltrials", "depmap", "chembl"]
latencyMs: 412, refsHydrated: 312
}
// traversal · ii · 4 hops
Catch a safety signal before the conference call.
"Pull pirfenidone's FAERS adverse-event profile, then cross-reference recent SEC filings from any sponsor with an active IPF program. Are competitors disclosing tolerability stories that line up with what FAERS is actually showing?"
the path
i.
Drug entity Lookup by name → resolve drug:<slug> with mechanism + sponsor.
ii.
→ FAERS Pull faersAdverseEvents rollup: maxLlr, totalReports, top events.
iii.
→ Active trials Trials in same indication or class with overlapping AE profile.
iv.
→ SEC filings 8-K/10-Q text-search the sponsor for tolerability or label risk language.

The query an analyst runs the morning before the earnings call. The graph crosses signal databases no chat tool has access to in one shot.

POST /workspace/biomed/graphql 638ms · 4 hops
query SafetySignal($drugName: String!) { drug(name: $drugName) { name phase indications faersAdverseEvents { summary { nEvents totalReports maxLlr medianLlr maxEvent maxCountEvent } } } competitorFilings: secFilings( ticker: "PLRX", since: "2026-01-01", limit: 10 ) { accession formType filingDate companyName } } // variables: { "drugName": "pirfenidone" }
response · safety profile + filings cross-referenced
live · verified data
drug.name: "PIRFENIDONE" · phase: approved
faersAdverseEvents.summary: {
nEvents: 33, totalReports: 9,476, maxLlr: 1263.2 strong signal
maxEvent: "decreased appetite", maxCountEvent: "nausea"
}
9,476 FAERS reports, median LLR of 196 — strong tolerability signal. The labelled SAE ("decreased appetite") matches what the data says; competitor sponsors filing IPF programs need a differentiated tolerability story to displace pirfenidone, not equivalent efficacy.
expand full response
indications[]: 30 · primary: "idiopathic pulmonary fibrosis"
also: systemic scleroderma, hypersensitivity pneumonitis, interstitial lung disease, …

faersAdverseEvents.summary: {
medianLlr: 196.12, maxLlr: 1263.23
// LLR > 4.0 = signal threshold; pirfenidone shows broad signal across 33 events
}

competitorFilings: 5 recent · ticker PLRX (Pliant Therapeutics)
8-K · 2026-04-17 · acc 0001746473-26-000050
10-K · 2026-03-11 · acc 0001746473-26-000037
8-K · 2026-03-30 · acc 0001746473-26-000046
8-K · 2026-03-11, 8-K · 2026-03-03

competitorFilings: 3 recent · ticker UTHR (United Therapeutics)
10-K · 2026-02-25 · acc 0001082554-26-000006

_meta.sourcesJoined: ["chembl", "faers", "clinicaltrials", "sec_edgar"]
_meta.latencyMs: 638
// traversal · iii · 4 hops
Find undisclosed competition on a target.
"Tell me everything about JAK2 as a target — genetic evidence, druggability, every disclosed program, and where the IP is concentrated. I want to know if there's whitespace before I write the proposal."
the path
i.
Protein target Resolve UniProt accession → pull rollups.
ii.
→ DepMap + GeneBurden depmapEssentiality, geneBurden, geneticConstraint for druggability scoring.
iii.
→ Patents googlePatentsSearch by gene symbol + mechanism class.
iv.
→ Sponsors Resolve assignees to sponsor:<slug>; cross-check pipeline + trials.

Surfaces the lab that filed an IP claim but hasn't filed an IND yet. Cannot be reached by trial registries alone.

POST /workspace/biomed/graphql 523ms · 4 hops
query CompetitiveTargetScan($accession: ID!) { protein(accession: $accession) { name depmapEssentiality { isEssential pctStrong meanEffect } geneBurden { summary } otTarget { tractability } proteinDrugs { drug { name phase sponsorRef } } } patents: googlePatentsSearch( query: "$accession AND inhibitor AND therapeutic" ) { results { title assignee filingDate publicationNumber } } }
response · target druggability + IP map
live · verified data
protein.name: "JAK2" · accession: O60674
otAssociation.summary: {
maxScoreDirect: 0.801, nDiseasesDirect: 1685
nDiseasesScoreAbove0_7: 8 strong evidence in 8 indications
}
JAK2 has strong evidence in 8 indications and broad clinical history. The graph shows where it's been tried, where it succeeded, and what's currently in late stage — without you reading the literature.
expand full response
depmapEssentiality: {
isEssential: false, nLines: 1183
pctStrong: 0.25%, meanEffect: 0.092
// not pan-essential — selective inhibition is tolerated
}

otAssociation.summary: {
nDiseasesDirect: 1685, nDiseasesIndirect: 5,200+
nDiseasesScoreAbove0_7Direct: 8
topTherapeuticAreas: ["hematology", "immunology", "oncology"]
}

proteinDrugs: · known clinical assets
RUXOLITINIB · approved · myelofibrosis, polycythemia vera
FEDRATINIB · approved · myelofibrosis
PACRITINIB · approved · myelofibrosis with thrombocytopenia
MOMELOTINIB · approved · myelofibrosis with anemia

patentScan: 17 active patent families
"Selective JAK2 inhibitors" · multiple assignees
5 filings in last 24 months — emerging selectivity refinements

_meta.sourcesJoined: ["uniprot", "open_targets", "depmap", "chembl", "google_patents"]
_meta.latencyMs: 523
// traversal · iv · 4 hops
Find repurposing candidates for a pathway.
"Find approved drugs whose primary or off-target activity hits any node in this pathway. Filter to clean tox profiles and patent runway past 2030."
the path
i.
Pathway nodes Resolve workspace pathway synthesis → list of UniProt targets.
ii.
→ Drugs chemblTargets + proteinDrugs for primary and off-target hits.
iii.
→ Tox + patents faersAdverseEvents filter + patent-expiry cross-check.
iv.
→ Indication history Cross-check whether the indication has been tried before, by whom, and why it stopped.

The BD scout query. Three weeks of manual work returned in a single round trip with the failed-prior-attempts column already filled in.

POST /workspace/biomed/graphql 487ms · 4 hops
query RepurposeScan($pathwayId: ID!) { synthesis(synthesisId: $pathwayId) { title linkedTargets { name proteinDrugs { drug { name phase mechanism indications drugWarnings faersAdverseEvents { summary { maxLlr } } drugProfile { chemistry { patentExpiry } } trials(indication: "NASH") { nctId whyStopped outcomeSummary } } } } } }
response · NASH repurposing surface
live · verified data
disease.name: "non-alcoholic steatohepatitis"
indicationDrugs: 9 disclosed across NR1H4, THRB, ACACA
RESMETIROM · approved 2024 · first NASH approval de-risked
OBETICHOLIC ACID · approved PBC · 14 indications including NASH attempts
FIRSOCOSTAT · ACC1 inhibitor · phase 2 mixed
Resmetirom's 2024 approval validates the THRB axis. Obeticholic has 14 indications on file but never closed NASH; it's a cautionary tale, not a candidate. ACC1 inhibitors are the next live mechanism worth tracking.
expand full response
target: NR1H4 (FXR)
proteinDrugs.summary.nDrugs: 9
OBETICHOLIC ACID · approved
indications: 14 · primary biliary cirrhosis, NASH, NAFLD, alcoholic hepatitis, …
FDA delayed 2020 NASH approval — REGENERATE outcomes ambiguous

target: THRB (thyroid hormone receptor β)
RESMETIROM · approved 2024
indications: ["non-alcoholic steatohepatitis", "NAFLD", "hypercholesterolemia"]
first FDA-approved NASH therapy · de-risks the mechanism class

target: ACACA (ACC1)
proteinDrugs.summary.nDrugs: 4
FIRSOCOSTAT · phase 2 (Gilead) · mixed efficacy CV signal concerns
CLESACOSTAT · phase 2 (Pfizer) · liver fat reduction confirmed

patentRunway: resmetirom composition-of-matter expires 2034
→ first-mover window approximately 10 years; expect IP-around filings to accelerate

_meta.sourcesJoined: ["chembl", "open_targets", "clinicaltrials", "fda_dailymed"]
_meta.latencyMs: 487
// traversal · v · 4 hops
Profile a sponsor's pipeline + risk surface.
"Profile Pliant Therapeutics (PLRX): every recent SEC filing, what their disclosure cadence looks like, and how their pipeline lines up against fibrosis-space peers I should compare them to."
the path
i.
Sponsor Resolve sponsor:<slug> — entity with aliases + ticker.
ii.
→ Pipeline All drugs by sponsor with phase, mechanism, target, indication.
iii.
→ Trials Active + recent readouts per drug.
iv.
→ SEC filings 8-K / 10-Q / S-1 corpus filtered by ticker and timeframe.

A diligence pass that connects the science to the disclosure. The asset doing the most work in the deck is usually not the one doing the most work in the pipeline.

POST /workspace/biomed/graphql 714ms · 4 hops
query SponsorProfile($ticker: String!) { filings: secFilings( ticker: $ticker, since: "2026-01-01", limit: 10 ) { accession formType filingDate companyName sponsorRef } peerSet: secFilings( formType: "10-K", since: "2026-01-01", limit: 5 ) { ticker companyName filingDate } } // variables: { "ticker": "PLRX" }
response · sponsor profile · PLRX
live · verified data
sponsor: "Pliant Therapeutics, Inc." · sponsor:pliant-therapeutics
filings[]: 5 since 2026-01-01
10-K · 2026-03-11 · acc 0001746473-26-000037 annual report
8-K · 2026-04-17 · acc 0001746473-26-000050 most recent material event
8-K · 2026-03-30 · acc 0001746473-26-000046
5 SEC filings since January including the just-filed 10-K. Three 8-K material events in March alone — material disclosure cadence is high. Read the 10-K runway language before next investor update.
expand full response
filings[]: 5 total since 2026-01-01
8-K · 2026-04-17
accession: 0001746473-26-000050
company: "PLIANT THERAPEUTICS, INC."
8-K · 2026-03-30
accession: 0001746473-26-000046
10-K · 2026-03-11 annual report
accession: 0001746473-26-000037
// hydrate sections to read R&D allocation, runway language, risk factors
8-K · 2026-03-11
accession: 0001746473-26-000039
// concurrent with 10-K filing — typically earnings or guidance
8-K · 2026-03-03
accession: 0001746473-26-000024

peerSet: recent 10-K filings · IPF / fibrosis space
CTNM · Contineum Therapeutics · 10-K filed 2026-03-05
UTHR · United Therapeutics · 10-K filed 2026-02-25
ABBV · AbbVie · 10-K filed 2026-02-20

_meta.sourcesJoined: ["sec_edgar", "clinicaltrials", "chembl"]
_meta.latencyMs: 714
// traversal · vi · 4 hops
From a single paper to the whole landscape.
"I just read the 2024 'TGF-β signaling in health, disease and therapeutics' review. Walk the citation graph and tell me what's been published downstream — and which of those papers connect to active drug programs."
the path
i.
Paper Resolve OpenAlex work ID → extract mentioned proteins / diseases.
ii.
→ Targets Hydrate each entity with full graph context.
iii.
→ Drugs Every disclosed compound across ChEMBL + OT mechanisms.
iv.
→ Trial outcomes Hydrate stop reasons, readouts, current sponsor activity.

The query that turns "interesting paper" into "what does this mean for the field" without a week of literature work.

POST /workspace/biomed/graphql 589ms · 4 hops
query PaperToLandscape($query: String!) { openalexSearch( query: $query, publicationYearMin: 2024, perPage: 5 ) { total hits { openalexId title doi publicationYear citedByCount primarySourceName isOpenAccess } } } // variables: { "query": "TGFB1 pulmonary fibrosis" }
response · paper grounded in landscape
live · verified data
openalexWork.title: "TGF-β signaling in health, disease and therapeutics"
id: W4394747404 · year: 2024 · citedBy: 873 high impact
primarySource: "Signal Transduction and Targeted Therapy"
doi: 10.1038/s41392-024-01764-w
total: 1,074 related works since 2024
A single review (873 citations) anchors a literature surface of 1,074 papers since 2024. The query lets you walk from any one of them into the full target / drug / trial graph in a single hop.
expand full response
openalexSearch.total: 1,074 · publicationYearMin 2024

[1] "TGF-β signaling in health, disease and therapeutics"
openalexId: W4394747404 · year 2024 · citedByCount 873
doi: 10.1038/s41392-024-01764-w
primarySourceName: "Signal Transduction and Targeted Therapy"

[2] "Idiopathic pulmonary fibrosis-specific Bayesian network
integrating extracellular vesicle proteome and clinical info"
W4390898360 · year 2024 · citedByCount 15
primarySourceName: "Scientific Reports"

[3] "Targeting mTORC1/TGFB1 signaling with a novel
Bergapten-Esculetin combination in IPF"
W4416239903 · year 2025 · citedByCount 2 emerging
doi: 10.1007/s11030-025-11401-5

followGraph: paper → mentioned proteins → indicationDrugs
TGFB1 → pirfenidone, nintedanib (both approved IPF)
SMAD2/3 → no clinical assets · 12 patents · pre-IND space
CCN2 → pamrevlumab terminated; 3 follow-ons phase 1

_meta.sourcesJoined: ["openalex", "open_targets", "chembl", "clinicaltrials"]
_meta.latencyMs: 589
The other way

Deep research fakes it.
Gyri actually knows.

Hand a frontier model the same question. It will spend an afternoon scrolling the open web, burn hundreds of thousands of tokens making the same lookup over and over, hand you a wall of paraphrase, and then forget every word of it the moment you close the tab. A typed graph and a one-shot query do something completely different.

"Which protein targets in IPF have strong genetic evidence and a confidently-folded structure I could actually drug — and where's the IP and recent science already pointing?"
// the other way
Frontier model + deep research
scraping the open web · 47 minutes · still going
001web_search("IPF drug targets 2024")+4,212
002web_fetch("nature.com/articles/s41392...")+18,440
003web_search("TGFB1 pulmonary fibrosis evidence")+3,884
004web_fetch("opentargets.org/target/ENSG...")+22,103
005web_fetch("clinicaltrials.gov/study/NCT0...")+9,772
006web_search("pirfenidone trial outcomes")+5,318
007web_fetch("alphafold.ebi.ac.uk/entry/AF-...")+14,201
008web_search("TERT IPF telomerase mechanism")+4,902
009web_fetch("pubmed.ncbi.nlm.nih.gov/...")+11,544
010web_fetch("impc.mousephenotype.org/g...")+8,129
011web_search("TGFB1 drug targets 2024") // duplicate?+4,212
012web_fetch("reactome.org/PathwayBrowser...")+16,773
013web_search("DSP desmoplakin pulmonary")+3,991
014web_fetch("chembl.org/target/CHEMBL2...")+12,309
015web_fetch("sec.gov/edgar/0001746473-...")+24,118
016web_search("RTEL1 telomere fibrosis")+4,402
017web_fetch("uniprot.org/uniprotkb/O95...")+13,887
018web_search("FAERS pirfenidone adverse")+5,902
019web_fetch("fda.gov/safety/medwatch/...")+19,234
020web_fetch("opentargets.org/target/ENSG...") // rate limit+22,103
021web_search("MUC5B promoter polymorphism")+4,121
022web_fetch("nejm.org/doi/10.1056/NEJMo...")+17,556
023web_search("PARN poly(A) ribonuclease IPF")+4,318
024web_fetch("jamanetwork.com/journals/...")+11,873
025web_search("IPF drug targets 2024") // re-issued+4,212
026web_fetch("globalgenes.org/rare-disea...")+8,902
027web_search("FGFR2 fibroblast growth factor")+4,884
028web_fetch("depmap.org/portal/gene/FGFR...")+15,402
// 213 more tool calls · still scrolling open-web pages
001web_search("IPF drug targets 2024")+4,212
002web_fetch("nature.com/articles/s41392...")+18,440
003web_search("TGFB1 pulmonary fibrosis evidence")+3,884
004web_fetch("opentargets.org/target/ENSG...")+22,103
005web_fetch("clinicaltrials.gov/study/NCT0...")+9,772
006web_search("pirfenidone trial outcomes")+5,318
007web_fetch("alphafold.ebi.ac.uk/entry/AF-...")+14,201
008web_search("TERT IPF telomerase mechanism")+4,902
009web_fetch("pubmed.ncbi.nlm.nih.gov/...")+11,544
010web_fetch("impc.mousephenotype.org/g...")+8,129
tool calls241
tokens consumed847,402
wall-clock latency47 min
code execution against output✕ none
persistence after the chat ends✕ none
verifiable provenance✕ paraphrase only
vs.
// gyri
Frontier model + typed graph
one query · 412ms · 6 sources joined
step 1 · query
query DeepTargetTriage {
disease("EFO_0000768") { // IPF
associatedTargets { geneSymbol score entity { alphafold mousePhenotype proteinDrugs } } } }
step 2 · response · 25 targets · 412ms
TERT · score 0.86 · 24,865 evidence · pLDDT 80.2
DSP · score 0.84 · 20,975 evidence · pLDDT 75.3
RTEL1 · score 0.84 · 10,260 evidence · pLDDT 71.7
… 22 more
step 3 · run code against the output
# classify targets by structural confidence × genetic evidence
import pandas as pd, matplotlib.pyplot as plt
df = pd.DataFrame(response["associatedTargets"])
df["tier"] = pd.cut(df.plddt, [0,70,85,100], labels=["C","B","A"])
tier_a = df.query("tier == 'A' and score > 0.5")
# → 4 confidently-folded high-evidence targets
# top 6 by score · evidence pieces TERT ████████████████████████ 24,865 DSP ████████████████████ 20,975 FGFR1 ████████████████ 16,441 PDGFRB ███████████████ 15,800 FGFR2 █████████████ 13,456 RTEL1 ██████████ 10,260
# chart written to /workspace/biomed/charts/ipf-tier-A.png
step 4 · persist as a typed insight
insight:ipf-targets-2026-q2 · 25 targets · 4 tier-A workspace://biomed
citations 312 · refs 25 hydrated · pinnedAt 2026-05-02 queryable forever
stored · queryable by any future model
+25 entities · 412ms
// tool calls
deep research241
gyri1
// tokens consumed
deep research847K+
gyri~3K
// code execution on output
deep researchnone
gyrinative
// persists after chat ends
deep researchnever
gyrialways

"Models are only going to get smarter — but whether they preserve what you taught them is entirely on you."

Stop teaching the same thing
to every new session.

Gyri's biomedical workspace is in private access for research groups, biotech teams, and infrastructure builders working at the frontier. Tell us what you're researching.