GeneCore is now live in the API and MCP. Read more

Amass

Building an AI Scientist · London · Jul 3–5, 2026

Build your AI scientist on cited life-science data.

Amass is a proud sponsor of the hackathon. Every participant gets $500 in Amass API credits. It is one REST API across five cross-linked Cores: biomedical literature, clinical trials, drugs, genes, and FDA + EMA regulatory.

Full records, typed schemas, every answer cited to source. Find below how to claim your credits, and a copy-paste API recipe for each track.

Claim $500 in credits →Jump to your track ↓

Applied automatically within minutes when you sign up with your hackathon email. Valid until 12 July 2026. Questions? hello@amass.tech.

your agent · amass api
you ▸Is Tagrisso's target druggable, and is the drug approved in the US and the EU?
  1. genecoreEGFRAMGC_GZzP…
    tyrosine-protein kinase · small-molecule tractable (Approved Drug) · LOEUF 0.475 · 7p11.2
  2. drugcoreOsimertinib (Tagrisso)AMDC_VmrG…
    small molecule · irreversible EGFR inhibitor · max stage APPROVAL
  3. trialcoreFLAURA2AMTC_zlXK…
    NCT04035486 · Phase 3 · AstraZeneca · N=587 · 21 countries · has results
  4. regulatorycoreTagrissoAMRC_ZeWz…
    FDA active 2015-11-13 ⇄ EMA active 2016-02-01 · orphan
✓

Yes. EGFR is a small-molecule-tractable kinase, and Osimertinib (Tagrisso) targets it irreversibly — with active FDA and EMA authorizations for EGFR-mutant NSCLC.

gene → drug → trial → approval — four calls, every hop cited.

The offer

Claim your $500 in API credits

Sign up with the same email you registered for the hackathon with, and your $500 is applied automatically, usually within a few minutes.

01

Sign up with your hackathon email

Create an account at platform.amass.tech using the email you signed up for the hackathon with.

02

$500 lands automatically

Matched on your registration email, your org is credited $500 automatically, usually within a few minutes during the hackathon. Valid until 12 July 2026; check the balance via GET /v1/credits/api-credits.

03

Create a key & build

Open API Keys, click Create API Key (starts with amass_, shown once), and make your first call.

Credits not showing yet?

Credits apply automatically within a few minutes. If yours have not appeared, check you used your registration email, then email hello@amass.tech or find Alex on-site.

What $500 buys

≈ 10,000 searches or 50,000 record fetches. Search is $0.05 per 20 results; get-by-id and lookup are $0.01 each. Plenty of headroom for a weekend.

Get started

From key to first response in a minute

Everything lives at platform.amass.tech. One base URL, one auth header, JSON back.

bash
# base: https://api.amass.tech/api/v1
# auth: Authorization: Bearer amass_YOUR_KEY  (every request)
curl "https://api.amass.tech/api/v1/cores/trialcore/records\
?query=cancer+immunotherapy&phase=PHASE3\
&overallStatus=RECRUITING&limit=5" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK
{
  "data": [
    {
      "amassId": "AMTC_…",
      "briefTitle": "…",
      "phase": "PHASE3",
      "overallStatus": "RECRUITING",
      "sponsorName": "…",
      "enrollment": 900
    }
  ]
}

Platform & keys

Sign up, create API keys, check your credit balance.

Open

API reference

Interactive docs for all five Cores, every param and field.

Open

Quickstart

Your first cited response in under five minutes.

Open

LLM quick reference

Self-contained, paste into your agent’s system prompt.

Open

Starter agent · Python

A runnable natural-language agent with the API loop built in.

Open

Starter agent · TypeScript

The same agentic loop in TypeScript, ready for a Node or web app.

Open

OpenAPI spec

Machine-readable: code-gen, Postman, type inference.

Open

The data

Five cross-linked Cores

Same auth, same error format, same rate limits across all of them. Every record cross-links to the others, so a paper, trial, drug, gene, and approval all resolve to the same entity.

BioMedCore

40M+

/cores/biomedcore/records

Peer-reviewed biomedical literature from PubMed and PubMed Central. Every record includes MeSH terms and IDs, publication types (RCT, meta-analysis, systematic review), JuFo 0–3 journal quality tier, citation count, DOI, PMID, and PMCID. Optional: full text for PMC articles, ORCID-disambiguated authors with ROR-mapped affiliations, and links to the trials each paper describes in TrialCore.

TrialCore

1.2M+

/cores/trialcore/records

Clinical trial records from ClinicalTrials.gov and international registries. Includes protocol data, eligibility criteria, primary and secondary endpoints, sponsor details, arm groups, recruitment status, NCT ID, and the full lifecycle — start, completion, results posting, and why stopped. Outcome measurements are available as structured values per arm, not PDF blobs. Records link back to the publications that describe them in BioMedCore.

DrugCore

22K+

/cores/drugcore/records

ChEMBL-derived drugs and molecules, harmonized. Names, trade names, and synonyms; chemical structures with InChIKey and SMILES; modality classification; and highest clinical stage from preclinical through approval. Each drug cross-links to the trials, papers, and regulatory authorizations it appears in.

RegulatoryCore

FDA + EMA

/cores/regulatorycore/records

FDA and EMA regulatory data on one unified schema. FDA application records, drug labels, review documents, and approval letters; EMA EPARs, assessment reports, and SmPCs. Each record carries a unified authorization status, designations (Breakthrough Therapy, PRIME, Accelerated Approval), and orphan status — and cross-links to the drug it covers in DrugCore.

GeneCore

43K+

/cores/genecore/records

43,000+ harmonized human gene records from HGNC, NCBI, UniProt, and Open Targets. Druggability and tractability assessments, target safety data, genetic constraint via gnomAD v4.0 (pLI, LOEUF), and cellular essentiality from DepMap CRISPR screens. Each gene cross-links to the drugs that target it in DrugCore, the trials that study it in TrialCore, and the papers that cite it in BioMedCore — from gene to clinical evidence in one call.

Tracks

An API recipe for every track

Inspiration, not coverage. Each command below was run against the live API and the responses are real (trimmed for width). Jump to your track:

Experimental dataSafetyPre-clinical strategyIPTarget identificationStructure-based designLigand modelsOpen stream
Track

Experimental data

Proteomics / MS ingestion & visualisation · Curve copilot (GraphPad ↔ Claude) · Transform IC50/EC50 → Ki/Kd, then train a QSAR model on old vs. corrected values.

BioMedCoreDrugCore

Ground an IC50→Ki agent in the actual methodology, ranked by journal quality and citations, cited to source.

bash
curl "https://api.amass.tech/api/v1/cores/biomedcore/records\
?query=Cheng-Prusoff+IC50+Ki+inhibition+constant\
&minJournalQualityJufo=1&limit=4" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · 4 records
TITLE                                                   JOURNAL                       YR   CITED
The power issue: determination of KB or Ki from IC50…   J Pharmacol Toxicol Methods   2002   171
An exact correction to the "Cheng-Prusoff" correction   Journal of Receptor Research  1988   109
Development of a Quantum Chemical Method for IC50        Molecular Informatics         2016    18
Binding Curve Viewer: equilibrium & kinetics of…        J Chem Inf Model              2024     4

Also try

For the QSAR half: DrugCore returns canonicalSmiles + mechanismsOfAction per ChEMBL compound, a ready feature set to compare models trained on raw vs. Ki/Kd-corrected labels.

Track

Safety

Evidence-backed toxicity report (e.g. “CRBN in atopic dermatitis”) · Zero-shot hERG de-risking: predict cardiac liability from structure and design an optimisation loop that drops hERG binding while preserving potency.

GeneCoreRegulatoryCoreTrialCoreBioMedCore

hERG in a single call: curated cardiac liabilities, channel class, and loss-of-function constraint, the backbone of a de-risking loop.

bash
curl "https://api.amass.tech/api/v1/cores/genecore/records\
?query=KCNH2&limit=1" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · GeneCore KCNH2 (hERG)
symbol            KCNH2  (hERG)
location          7q36.1          LOEUF  0.568
targetClass       Ion channel › Voltage-gated potassium channel
safetyLiabilities
  • Torsades de Pointes              ClinPGx
  • Prolongation of QT interval      Bowes et al. (2012)
  • Increased mortality              AOP-Wiki
  • Receptor binding (in vitro)      ToxCast

Sweep the full text of every FDA label & EMA SmPC for a safety phrase. Matched sections come back with the exact excerpt and an addressable id.

bash
curl "https://api.amass.tech/api/v1/cores/regulatorycore/records\
?query=torsades+de+pointes&limit=4" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · full-text label sweep
Apokyn   (FDA)  →  §5.11  QTc Prolongation and Potential for Proarrhythmic Effects
Geodon   (FDA)  →  label: "…greater capacity to prolong the QT/QTc interval…"
Floxin   (FDA)  →  review: Precautions · Animal Pharmacology
Dacogen  (EMA)  →  matched on active substance

Also try

For the CRBN report: GeneCore returns CRBN with small-molecule “Approved Drug” tractability and PROTAC predictive buckets (the classic molecular-glue hub, LOEUF 1.15, LoF-tolerant). Cross-link to atopic-dermatitis trials with TrialCore and the literature with BioMedCore.

Track

Pre-clinical strategy

Assay optimisation (clever DoE / Bayesian) · Protein & antibody finder (literature use-cases, not datasheets) · Disease-model planner · Cross-species translational alignment.

TrialCoreBioMedCoreGeneCore

Benchmark a study design before you run it: allocation, masking, model, and sample size across completed trials in your indication.

bash
curl "https://api.amass.tech/api/v1/cores/trialcore/records\
?query=atopic+dermatitis&phase=PHASE2\
&overallStatus=COMPLETED&limit=5" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · completed Phase 2
TRIAL                                   N     ALLOCATION   MASKING   MODEL
Secukinumab for Atopic Dermatitis       41    RANDOMIZED   TRIPLE    PARALLEL
ILV-094 (anti-IL-22) proof-of-concept   60    RANDOMIZED   TRIPLE    PARALLEL
S-777469 dose-finding                  209    RANDOMIZED   DOUBLE    PARALLEL
Apremilast, moderate-to-severe AD      191    RANDOMIZED   TRIPLE    PARALLEL
CM310 open-label                       127    NA           NONE      SINGLE_GROUP

Also try

For the protein/antibody finder & disease-model planner: run BioMedCore with include=authorsMetadata to pull real use-cases, institutions (ROR), and ORCIDs straight from the literature, the in-domain evidence a datasheet leaves out.

Track

IP

Float the hidden “key compound” in a 500-molecule patent · Markush infringement testing · Claims drafting with WIPO feedback loops · Automated freedom-to-operate (FTO) search.

DrugCorePatentCore · live Fri Jul 3

Pull clean, canonical structures (SMILES + InChIKey) to feed RDKit for Tanimoto similarity, Markush membership, or an FTO risk score.

bash
curl "https://api.amass.tech/api/v1/cores/drugcore/records\
?query=EGFR&drugType=SMALL_MOLECULE&limit=4" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · structures + stage + MoA
NAME            CHEMBL          STAGE     INCHIKEY
Erlotinib HCl   CHEMBL1079742   APPROVAL  GTTBEUCJPZQMDZ-UHFFFAOYSA-N
Aumolertinib    CHEMBL4761468   APPROVAL  DOEOECWDNSEFDN-UHFFFAOYSA-N
Icotinib        CHEMBL2087361   APPROVAL  QQLKULDARVNMAL-UHFFFAOYSA-N
Mavelertinib    CHEMBL3989970   PHASE1    JYIUNVOCEFIUIU-GHMZBOCLSA-N
# each record also carries canonicalSmiles + mechanismsOfAction (INHIBITOR → EGFR)

Also try

PatentCore goes live on Friday 3 July, the first day of the hackathon: 170M+ patents from USPTO, EPO, and WIPO, with claims, families, and citations. Pair it with DrugCore’s ChEMBL-derived structures and mechanisms for a grounded chemical layer under your FTO and Markush work.

Track

Target identification

Predict clinical-trial attrition and classify it (toxicity vs. lack of efficacy) · The “undruggable” protein classifier: spot hidden cryptic pockets from sequence / structure.

GeneCoreTrialCoreBioMedCore

Prioritise targets in one query: druggable + essential + loss-of-function-constrained, each with the numbers to defend it.

bash
curl "https://api.amass.tech/api/v1/cores/genecore/records\
?query=kinase&isDruggable=true&isEssential=true\
&targetClass=ENZYME&maxConstraintLoeuf=0.6&limit=5" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · Open Targets intelligence
SYMBOL  NAME                        LOEUF  DEPMAP dep/tested  SM TRACTABILITY
WEE1    WEE1 G2 checkpoint kinase   0.41   1183 / 1183        Advanced Clinical
CDK2    cyclin dependent kinase 2   0.39    770 / 1183        Advanced Clinical
PLK1    polo like kinase 1          0.21   1183 / 1183        Advanced Clinical
SMG1    SMG1 (PI3K-related)         0.10    340 /  351        —
PI4KA   PI4K alpha                  0.55    767 / 1183        —

Learn why programs die: filter TrialCore to terminated trials and read whyStopped, labelled training data for an attrition model.

bash
curl "https://api.amass.tech/api/v1/cores/trialcore/records\
?query=amyloid+Alzheimer&overallStatus=TERMINATED&limit=6" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · attrition signal
SPONSOR             PHASE   N      WHY STOPPED
Hoffmann-La Roche   PHASE3  1382   "…terminate development of gantenerumab following a
                                    pre-planned safety & efficacy analysis (GRADUATE I & II)…"
Hoffmann-La Roche   PHASE3  1053    (same program, prodromal / mild AD)
Inst. Neurodeg.     PHASE1     3   "…tracer did not evidence as a marker of disease"
Track

Structure-based design

Assess co-folding training-set similarity (how novel is my prediction?) · Target deconvolution for compounds with unknown targets · Visualise selectivity differences between targets.

GeneCore (protein)DrugCore

Get the structural context for a target: 3D availability, every PDB id, and Pfam domains, a novelty prior for co-folding predictions.

bash
curl "https://api.amass.tech/api/v1/cores/genecore/records\
?query=EGFR&include=protein&limit=1" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · EGFR protein block
symbol           EGFR    ENSG00000146648
has3dStructure   true    ·    354 PDB structures
pdbIds           1IVO, 1M14, 1M17, 1MOX, 1NQL, 1XKK, …
pfam             PF00757 · PF07714 (PK_Tyr_Ser-Thr) · PF01030 · PF14843 · PF21314
sequence         1210 aa    ·    134,277 Da

Also try

Pair the PDB count as a training-set-coverage prior with DrugCore’s canonicalSmiles for the ligand side of a co-fold. (The Boltz team is offering API credits at this event too, a natural pairing for the co-folding problems.)

Track

Ligand models

Few-shot bioactivity from a handful of actives · Conquering activity cliffs (predict them, do not smooth them) · Multi-objective generative design against 3–4 conflicting properties.

DrugCoreGeneCore

Assemble a few-shot actives set for a target: SMILES plus clinical stage as a weak relevance label.

bash
curl "https://api.amass.tech/api/v1/cores/drugcore/records\
?query=EGFR&drugType=SMALL_MOLECULE&limit=4" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · canonicalSmiles + stage
Erlotinib HCl  APPROVAL  C#Cc1cccc(Nc2ncnc3cc(OCCOC)c(OCCOC)cc23)c1.Cl
Icotinib       APPROVAL  C#Cc1cccc(Nc2ncnc3cc4c(cc23)OCCOCCOCCO4)c1
Aumolertinib   APPROVAL  C=CC(=O)Nc1cc(Nc2nccc(-c3cn(C4CC4)c4ccccc34)n2)c(OC)cc1N(C)CCN(C)C
Mavelertinib   PHASE1    C=CC(=O)N[C@@H]1CN(c2nc(Nc3cn(C)nc3OC)c3ncn(C)c3n2)C[C@H]1F

Also try

GeneCore EGFR → include=referencesDrugCore returns 87 targeting drugs in one hop, a ready-made labelled set. Contrast approved vs. Phase 1 scaffolds to mine activity cliffs.

Track

Open stream

By definition you already know the problem. Bring your own; the whole engine is open. The strongest entries usually chain Cores together.

All five CoresCross-Core links

Cross-Core links let one entity resolve across all five Cores, every hop cited to source.

bash
curl "https://api.amass.tech/api/v1/cores/genecore/records\
?query=EGFR&include=referencesDrugCore&limit=1" \
  -H "Authorization: Bearer amass_YOUR_KEY"
response · 200 OK · one gene → 87 drugs → …
EGFR  →  referencesDrugCore: 87 drugs
             │
             ├─ each drug → referencesTrialCore       (the trials that ran it)
             ├─ each drug → referencesBiomedCore      (the papers behind it)
             └─ each drug → referencesRegulatoryCore  (FDA / EMA labels)

gene → drug → trial → paper → approval, in one thread. all cited.

Playbook

Nine ways to get further, faster

Patterns that tend to separate a demo that lands from one that fights its own plumbing.

01

Build an agent, not a script.

Let the model call the API in a loop: search broadly, filter, drill into a record, repeat. Our open-source starter agents (Python and TypeScript) ship exactly this loop; fork one and go.

02

Ground every claim in a record.

The whole point is answers you can trust: each result carries a stable Amass id and links back to its source. No hallucinated PMIDs or NCT numbers.

03

Cross-Core is your edge.

A gene resolves to the drugs that target it, the trials that ran them, the papers behind them, and the FDA/EMA labels that cover them. Follow the links, one call at a time. Treat an empty reference array as “none recorded,” not “no evidence.”

04

Query, do not scrape.

PubMed, ClinicalTrials.gov, ChEMBL, HGNC, Open Targets, FDA & EMA are already normalised and cross-linked. Spend your hackathon on the idea, not the ETL.

05

Narrow with filters.

journalQualityJufo, phase, maxConstraintLoeuf, tractability, agency… tighter queries mean higher signal and fewer credits burned. Skip include=fulltext unless you truly need it.

06

Be nice to the rate limit.

60 requests / 60s, keyed by user + org. On a 429, read Retry-After and back off exponentially. Every response carries X-RateLimit-Remaining and X-Amass-Credit-Cost.

07

Read the data envelope.

Every response is wrapped in { "data": … }; errors come back in a different shape under "error". Read results from data, and check for error before you parse. A batch lookup can even fail item-by-item.

08

Convert external IDs first.

Get-by-id endpoints take Amass IDs only (AMBC_, AMTC_, AMDC_, AMRC_, AMGC_). Holding PMIDs, DOIs, NCTs, ChEMBL, or gene IDs? Batch them through the /lookup endpoints, then fetch.

09

Cache, and mind pagination.

Same query, same result. Memoise locally so a weekend of iterating does not re-spend credits. And there is no pagination: 300 results max per search, so narrow with filters instead of trying to page.

Prototype in Claude

Sketch your agent in Claude before you write app code.

Connect the Amass MCP in Claude.ai, Claude Code, Cursor, ChatGPT, or Codex and query all five Cores in natural language. Explore the data, find the filters that work, then drop the same queries into the REST API in your build.

In Claude Code, install the Amass SKILL.md and let the model wire up your data layer for you — a quick path from “what data is even in here?” to a working prototype.

Connect the MCPStarter agent:PythonTypeScript

Go further

More ways to put the credits to work

Not a coder on your team, or just want to move faster? Pick a lane.

No-code builds

Paste your key into Lovable, Streamlit, or an n8n HTTP node and ship a filterable trial tracker or literature app in an afternoon.

See examples

Claude Code + SKILL.md

Install the Amass skill and let Claude Code build your Amass data layer, then iterate on the science instead of the SDK.

Get the skill

Demo it live in Claude

With the MCP connected, your agent’s answers come back cited in the chat, a clean, judge-friendly demo with zero deploy.

How it works

Chain the Cores

Strong entries fan out: gene → drug → trial → paper → label. Follow the reference fields and cite every hop.

See open stream

Share on Discord

Show what you are building with life-science AI builders, and tag the Amass team for a boost.

Join now
Alex Junge, Amass co-founder and CTO

On the ground

Meet the Amass team on-site

Alex Junge, Amass co-founder and CTO, is here Friday and Saturday until 14:00. Come find him for credits, API help, or to talk through your build. Meet the full Amass team →

LinkedInGitHubXEmail

Go build your AI scientist

Claim your $500 in credits, wire up a Core, and let an agent do the reading. We cannot wait to see what you make.

Claim $500 in credits →API referenceConnect the MCP

Credits apply automatically within minutes · valid until 12 July 2026 · questions? hello@amass.tech