One question in.
Your scientific stack answers together.

How Perslis integrates PubMed

PubMed is the National Library of Medicine's index of biomedical literature — abstracts, MeSH terms, DOIs and citation metadata across MEDLINE. It is the first place most research questions land.

On the console, pubmed_search runs a relevance- or date-sorted query and returns structured records; pubmed_abstracts fetches the full abstract text by PMID. Both hand back the identifiers a claim can be traced by.

The failure mode this fixes is the fabricated citation — the plausible paper that does not exist. Because Lois can only report what PubMed returned, every reference on the console carries a live PMID that resolves. There is no path for a made-up DOI to enter the record.

PubMed is also the hub of the literature mesh: a PMID that carries a PMCID hands straight to pmc_fulltext for open-access full text, and the same question fans out to OpenAlex and Semantic Scholar so a search is a triangulation, not a single opinion.

How Perslis integrates PubMed Central

PubMed Central (PMC) is the NIH free full-text archive of biomedical literature. Where a paper is open access, PMC has the complete article — introduction, methods, figures legends, results and references.

On the console, pmc_fulltext takes a PMCID and returns the article body. It is the second half of a PubMed hit: search finds the paper, PMC reads it.

The handoff is automatic. When pubmed_search returns a record that carries a PMCID, that identifier is the key straight into full text — no second search, no guessing a URL. The claim and the paragraph it came from stay one click apart.

This is what makes grounded reading possible offline too: full text pulled through PMC can be harvested into a local topic pack, so a review still has the source article when the cable is pulled.

How Perslis integrates arXiv

arXiv is the open preprint server for physics, mathematics, computer science and quantitative biology (q-bio). A paper is here before, and often instead of, peer review.

On the console, arxiv_search runs a query, optionally scoped to a category such as q-bio.GN, and returns titles, authors, abstracts and stable arXiv identifiers that link straight to the PDF.

Preprints are where the risk of hallucination is highest — the work is new, so a model is most tempted to invent a plausible-sounding one. Routing through arxiv_search removes the temptation: every preprint the console cites has an arXiv ID that resolves to a real abstract and PDF.

arXiv sits alongside bioRxiv on the preprint bench and beside OpenAlex and Semantic Scholar on the scholarly graph, so a frontier question is answered from preprints and the published record at once.

How Perslis integrates bioRxiv & medRxiv

bioRxiv (biology) and medRxiv (clinical/health) are the life-science preprint servers. Their API is date-window based — there is no keyword search — so you browse what posted, newest first, and filter by category.

On the console, biorxiv_recent returns a window's preprints (optionally filtered to a category like genomics), and biorxiv_fetch retrieves a specific DOI, following through to the published article when one exists.

The runtime is honest about the API's shape. bioRxiv has no keyword search, so the instrument does not pretend to: it returns exactly the date window asked for, filters category client-side, and states plainly when a window is empty rather than inventing preprints to fill it.

When a preprint has been published, biorxiv_fetch follows the DOI to the version of record — so a citation points at the peer-reviewed paper, not a superseded draft, the moment one exists.

How Perslis integrates OpenAlex

OpenAlex is a free, open replacement for the old proprietary citation databases. It covers 250 million-plus works with authorship, venue, year and citation metadata — biomedicine and everything beyond it.

On the console, openalex_search queries the corpus and openalex_work resolves a single work by ID, returning the metadata that lets an agent weigh how established an idea is.

OpenAlex is the runtime's discipline-agnostic index: PubMed covers biomedicine, but a materials or physics question needs a broader map, and OpenAlex provides it without a paywall or a key. Every work carries an OpenAlex ID and, where known, a DOI — both resolve.

It answers next to Semantic Scholar, whose forward-citation graph complements OpenAlex's breadth, so a single question yields both how much a paper is cited and who built on it.

How Perslis integrates Semantic Scholar

Semantic Scholar is the Allen Institute's scholarly graph — papers, abstracts, citation counts and, distinctively, the forward citation edges that let you trace a line of work forward in time.

On the console, semantic_scholar_search queries papers and semantic_scholar_citations returns the works that cite a given paper — the shortest path to what happened next.

Its unique value on the console is direction. Given a landmark paper — say AlphaFold's Nature article, DOI 10.1038/s41586-021-03819-2 — the citation instrument walks forward to the work that built on it, turning a single reference into a live research front an agent can follow.

The unauthenticated tier is rate-limited, and the runtime treats that honestly: a throttled call is reported as unverified, never silently dropped or filled with a guessed number. It pairs with OpenAlex, whose open breadth backs up any gap.

How Perslis integrates UniProt

UniProt (UniProtKB) is the expert-curated protein knowledgebase — reviewed entries with function, subcellular location, disease associations, sequence features and cross-references to structure and genome databases.

On the console, uniprot_search resolves a name or gene to accessions, and uniprot_entry returns the full record — function, disease, domains and its list of PDB structures — pinned to the accession.

UniProt is the join key of the protein bench. Ask about hemoglobin and Lois resolves it to P69905, then that one accession fans out: its PDB cross-references become structure lookups, its sequence becomes an AlphaFold model, its features become InterPro domains. The instruments talk to each other because they share the UniProt anchor.

Curated disease text — alpha-thalassemia, Heinz body anemia — is reported verbatim from the entry, never paraphrased into something a model finds tidier. If UniProt says it, the console says it; if it does not, the console does not invent it.

How Perslis integrates RCSB PDB

RCSB PDB is the worldwide repository of macromolecular structures solved by X-ray crystallography, cryo-EM and NMR. Each entry records how it was determined and at what resolution.

On the console, pdb_search finds structures and pdb_entry returns a single one — method, resolution, molecular weight, primary citation (with DOI and PMID) and direct .pdb/.cif download links.

PDB is where a protein's structure claim becomes checkable. A UniProt entry lists its cross-referenced PDB IDs; the console follows one straight to pdb_entry, and the resolution and primary citation come back with it — so a structural statement always carries the evidence and the paper behind it.

Experimental structure sits deliberately beside AlphaFold's prediction on the same bench: an agent can compare a measured structure against a predicted one and see, from the metadata, which is which.

How Perslis integrates AlphaFold DB

AlphaFold DB (EMBL-EBI) hosts the DeepMind-predicted 3D structure for a UniProt protein, with per-residue confidence (pLDDT) and predicted aligned error (PAE) that quantify how reliable each region is.

On the console, alphafold_structure takes a UniProt accession and returns the model version, mean pLDDT, and download URLs for the CIF/PDB coordinates and the PAE.

Prediction is only useful with its confidence attached, so the instrument never returns a fold without its pLDDT. A high score (hemoglobin's is 98.06) says the model is reliable; a low one is a warning the console passes through unedited rather than hiding.

Because AlphaFold keys on the same UniProt accession as everything else on the protein bench, a predicted model and an experimental PDB structure line up for one protein automatically — the anchor is what lets the console show measured and predicted side by side.

How Perslis integrates InterPro

InterPro (EMBL-EBI) integrates a dozen protein-signature databases into one classification — families, domains, homologous superfamilies and sites — each with a stable InterPro accession.

On the console, interpro_domains returns the entries matching a UniProt protein, and interpro_entry resolves a single InterPro accession. A count of zero is a real result, reported as such.

The honest zero is the whole point. A model asked what domains a protein has will, unprompted, offer a confident-sounding list. InterPro replaces that guess with a lookup: hemoglobin returns the Globin domain and its superfamilies; an unclassified protein returns nothing, and the console reports nothing rather than filling the silence.

Keyed on the UniProt accession, InterPro completes the protein picture the bench builds — sequence and disease from UniProt, structure from PDB and AlphaFold, and composition from InterPro, all for one anchor.

How Perslis integrates Ensembl

Ensembl (EMBL-EBI) is the annotated reference genome for human and other species — genes, transcripts, coordinates and cross-references to protein and variant databases, all on a stated assembly like GRCh38.

On the console, ensembl_gene resolves a symbol to its record, ensembl_sequence returns FASTA, and ensembl_xrefs lists cross-references — the links that carry a gene to its protein and its variants.

Ensembl is the genome anchor, the way UniProt is the protein anchor. A gene resolved here — HBA1 to ENSG00000206172 on GRCh38 — carries the assembly with it, so a coordinate always means one place, and the cross-references hand off cleanly to VEP for variant effects, dbSNP for known variants, and NCBI for sequence records.

Every field is pinned: the gene ID, the region, the canonical transcript, the assembly. A genomic statement on the console names its coordinate system, because a position without an assembly is a position that could be wrong.

How Perslis integrates Ensembl VEP

Ensembl VEP annotates a variant with its predicted molecular consequences across all overlapping transcripts, the most severe consequence, and pathogenicity predictions from SIFT and PolyPhen.

On the console, vep_consequences accepts an rsID or HGVS notation and returns per-transcript consequence terms, impact levels, and SIFT/PolyPhen scores — a prediction, clearly labelled as such.

VEP is prediction, and the console keeps that boundary sharp. It returns SIFT and PolyPhen scores as computed estimates, distinct from ClinVar's expert clinical classification — an agent sees the algorithmic call and the curated call as two different kinds of evidence, never merged into one confident verdict.

Keyed to the same coordinates as dbSNP and Ensembl, VEP completes the variant triangle: dbSNP says the variant exists, VEP predicts its effect, ClinVar reports what clinicians concluded. Every call is stamped research use only.

How Perslis integrates NCBI

NCBI hosts the primary nucleotide (nuccore) and protein sequence databases — RefSeq curated records, GenBank submissions, RefSeqGene entries — the authoritative store for a sequence and its provenance.

On the console, ncbi_sequence_search queries either database and returns accessions with titles and lengths; ncbi_fetch_sequence retrieves the FASTA or GenBank record for one.

NCBI is where a raw sequence keeps its identity. A search for hemoglobin's gene returns NG_059186.1 — a RefSeqGene with an accession, a length, an organism — not an anonymous string of bases a model could have hallucinated. The accession is the receipt.

It sits beside Ensembl, which anchors the annotated genome, and UniProt, which anchors the protein: NCBI provides the underlying sequence records that back both, so a nucleotide or protein sequence on the console can always be traced to a submitted, accessioned entry.

How Perslis integrates ClinVar

ClinVar (NCBI) aggregates submitted interpretations of the clinical significance of genetic variants, with the review status that says how much scrutiny each classification received.

On the console, clinvar_search finds variants and clinvar_variant returns one — the classification, review status, associated condition and canonical SPDI — each field carried through untouched.

This is the sharpest edge of the whole runtime. A language model asked to judge a variant will happily produce a clinical opinion — which, for a real patient's variant, is dangerous. The console forbids it: ClinVar's classification passes through verbatim, review status attached, and the model is structurally barred from upgrading, downgrading or summarising it into a verdict.

ClinVar's clinical call sits beside VEP's algorithmic prediction and dbSNP's record so an agent sees them as separate evidence — the curated human classification never blurred with a computed guess. Every result is stamped research use only, not medical advice.

How Perslis integrates dbSNP

dbSNP (NCBI) is the reference archive of short genetic variation — SNPs and small indels — each with a stable rsID, mapped location, alleles, functional class and any aggregated clinical significance.

On the console, dbsnp_variant takes an rsID (with or without the 'rs' prefix) and returns its position on a named assembly, its genes, alleles and the HGVS strings that other instruments consume.

dbSNP is the front door of the variant bench. An rsID resolved here — rs6025, Factor V Leiden, in the F5 gene at 1:169,549,811 — carries the assembly and the HGVS notations, which hand straight to VEP for predicted effect and to ClinVar for clinical classification. One identifier, three instruments.

Where dbSNP aggregates clinical significance, it is reported verbatim — the same never-re-graded rule ClinVar follows — and stamped research use only. A variant on the console always knows which genome build its coordinates belong to.

How Perslis integrates PubChem

PubChem (NCBI) is the largest open chemistry database — compounds with computed and curated properties: formula, molecular weight, SMILES, InChIKey, XLogP and more, each under a stable CID.

On the console, pubchem_compound resolves a name to its record and pubchem_compound_by_cid fetches one by CID — the canonical identity every downstream chemistry step keys on.

SMILES strings are exactly where a language model's confidence outruns its correctness — a plausible-looking structure can be subtly wrong. PubChem replaces generation with resolution: ask for aspirin and the console returns CID 2244 with the real InChIKey, not a string a model assembled. The identity is looked up, never invented.

That verified identity is the key for the chemistry bench: the same compound resolves into ChEMBL for measured bioactivity and into KEGG and Reactome for the pathways it acts on — one molecule, correctly identified, across four databases.

How Perslis integrates ChEMBL

ChEMBL (EMBL-EBI) is the manually-curated database of bioactive molecules — structures, calculated properties, clinical development phase, and measured activities (IC50, Ki, EC50) against biological targets, drawn from the literature.

On the console, chembl_search finds molecules and chembl_bioactivity returns the measured activities for one — each an experimental value with a source, not an estimate.

A potency number is the kind of specific-sounding fact a language model invents most convincingly. ChEMBL makes that unnecessary: aspirin resolves to CHEMBL25 with its real molecular weight and max clinical phase, and any activity value comes back as a measured datapoint tied to its assay — never a plausible figure with no experiment behind it.

ChEMBL joins the chemistry bench through the compound's identity: PubChem fixes what the molecule is, ChEMBL says what it does, and its target links reach back to UniProt proteins — the same anchor the structure bench is built on.

How Perslis integrates KEGG

KEGG (Kyoto Encyclopedia of Genes and Genomes) links genomic and molecular information to higher-order functions — pathway maps and the enzymes, compounds, diseases and drugs that populate them, each with a stable KEGG identifier.

On the console, kegg_search finds entries and kegg_entry fetches the full flat-file record for one — a pathway with its gene list, compounds, modules and associated drugs.

KEGG turns an isolated molecule into a position in a process. A compound identified in PubChem or a target from ChEMBL can be placed on a KEGG pathway — glycolysis, hsa00010, with its enzyme genes, its compounds from glucose to pyruvate, and the drugs (mitapivat, etavopivat) that act on it — every element carrying a KEGG ID.

KEGG shares the pathways bench with Reactome; the two curate biological process differently, so the console offers both rather than collapsing them into one view. Every entry is pinned, so a pathway claim resolves to a KEGG record.

How Perslis integrates Reactome

Reactome is an open, manually-curated and peer-reviewed database of human pathways and reactions — molecular events organised into a navigable hierarchy, each with a stable identifier like R-HSA-70171.

On the console, reactome_search queries pathways, reactions and proteins by name or species and returns entries with their stable IDs — the anchor a pathway claim resolves against.

Reactome and KEGG both map biological process, but they curate it differently — reaction-level detail versus pathway maps — so the console keeps both rather than picking a winner. A question about glycolysis returns Reactome's R-HSA-70171 and KEGG's hsa00010, and an agent sees two independent curations of the same biology.

Because Reactome's proteins carry UniProt accessions, a pathway found here reaches back to the protein bench — the same anchor that ties structure, prediction and domains together, now placing a protein inside the reactions it participates in.

Compute on what the instruments retrieved.

Biopython, pandas, SciPy, NumPy and RDKit sit beside the source lanes, so sequence and molecule work can run without copying evidence into another application. It is labeled honestly: process isolation, not a security sandbox.

Experiments that outlive the session.

WetHands runs managed background work for up to 24 hours with saved progress, monitoring and cancellation. Its file operations stay inside the assigned workspace. The same job contract is the integration surface for LIMS, electronic lab notebooks and instrument automation APIs; no autonomous laboratory is claimed.