> For the complete documentation index, see [llms.txt](https://docs.neurolabai.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.neurolabai.io/reference/glossary.md). # Glossary ## Abstention An intentional decision to withhold or de-emphasize an output when route-specific evidence, applicability, or directional support is insufficient. ## Account preferences Account-owned presentation and workflow defaults, such as evidence posture, response detail, starting view, and personalization choices. Preferences do not alter the underlying model result. ## Applicability domain The region of input space for which a route has relevant training, retrieval, or evaluation support. In the structure route, chemical similarity is one diagnostic of applicability but is not a correctness probability. ## ATC The World Health Organization Anatomical Therapeutic Chemical classification system. Neurolab’s ATC head ranks classes; the ranking is not a regulatory classification decision. ## B0 The production known-compound GNN architecture: an edge-feature heterogeneous `HeteroConv(TransformerConv)` model with two layers, hidden width 128, and four attention heads. ## Brain channel A separately labeled brain-related data source or method, such as a raw GNN parcel ranking, receptor-density hypothesis, circuit-localization hypothesis, literature-evidence map, measured map, or legacy compatibility output. ## Caller-provided assumption Input supplied by the user—such as receptor values, affinities, actions, or dose context—that conditions downstream output without being independently verified by Neurolab. ## CircuitFramework A typed mechanism layer connecting receptors, regions, tracts, source nuclei, and cognitive functions. It supports symbolic composition and fine-cognition hypotheses; it is not a validated dynamical brain simulation. ## Composed hypothesis An output created by explicit rules, retrieval, or symbolic relationships rather than directly by one learned prediction head. ## Confidence interval An interval produced by a specified resampling or statistical procedure. Historical raw-head intervals bootstrap compounds, not model seeds or chemistry components, and are population-level development evidence rather than per-prediction confidence. Current evaluation tooling can use chemistry-component cluster resampling for future split-aware studies. ## DE-LEAK The v43.262 training and serving line in which relations used as supervision targets are excluded from GNN message passing, preventing held-out compound embeddings from seeing answer-like edges. ## Data Studio A developing analysis surface for structured comparison, visualization, and natural-language operations over research artifacts while preserving provenance and source values. ## Evidence artifact A response-safe bundle of sources, evidence records, support links, nodes, edges, paths, scope, filters, release identity, and caveats assembled for a report or section. ## Facts-and-evidence release The versioned set of governed identities, records, provenance, evidence coverage, and support artifacts used to compose and explain a report. It can advance independently of the predictive-model release. ## Evidence state The origin and epistemic status of a report value: model-ranked prediction, source-backed record, composed hypothesis, reference context, caller-provided assumption, or unavailable/withheld material. ## Field-specific support Evidence bound to an exact report pointer or ranked item. It is narrower and more directly applicable than section-level or contextual evidence. ## GraphRAG Graph- and retrieval-augmented assembly used by the novel route to retrieve chemical-neighbor and KG context. Retrieval does not make an analog’s facts true of the query. ## Heterogeneous graph A graph with multiple node and relation types. Neurolab’s KG includes compounds, receptors, genes, regions, concepts, sources, classifications, and other typed entities. ## Hit\@K A ranking metric that asks whether at least one gold label occurs in the first K predictions. It does not measure calibration, effect size, causality, or clinical value. ## InfoNCE The contrastive objective used for the five categorical known-route heads, with temperature 0.3 in the frozen v43.262 recipe. ## Knowledge graph (KG) Neurolab’s governed, typed graph substrate. Different projections support GNN message passing, novel-route artifacts, evidence retrieval, and reference context. ## Known-compound route The primary beta path for a compound already represented in the governed KG and deployed transductive model. It runs the three-seed B0 ensemble. ## Legacy channel An output retained for compatibility with earlier serving behavior. “Legacy” does not imply current validation; the novel-route legacy brain channel is explicitly unvalidated. ## Mechanism of action (MoA) A class or explanation of how a compound may produce effects through molecular targets or pathways. Neurolab has both a ranked MoA head and mechanism-oriented composed material; their evidence states can differ. ## Message passing The propagation of information along allowed graph relations inside a GNN. Only the checkpoint KEEP projection is message-passed; the presence of a relation elsewhere in the KG does not imply model use. ## Model-ranked prediction A label or value ordered by a learned model component. Its score is not automatically a calibrated probability. ## Morgan fingerprint A fixed-length chemical fingerprint used for compound features and structure-route similarity/retrieval. ## Molecular Portfolio A developing workflow for organizing saved forecasts and comparing selected fields across a small, version-compatible set of molecules. ## Needs SMILES A routing state returned when a compound name does not resolve to the known route and no usable structure is available. It is not a scientific report. ## Novel-structure route The conditional beta path for a supplied SMILES outside the known route. It uses structure-to-receptor estimation, chemical-neighbor retrieval, and mechanism reasoning; it does not execute B0. ## Pooled development cohort The validation-plus-test list used by frozen known-head reporting. It is development evidence, and a later audit found four train/test scaffold-family overlaps; it is not scaffold-disjoint or an untouched independent test. ## Provenance Information that identifies where data or a claim came from, how it was transformed, and which release or artifact produced it. ## Raw payload The exact machine-readable report returned to the application, including fields, nulls, metadata, reliability, and optional evidence bundles. ## Receptor-conditioned route A structure workflow in which the user supplies receptor rows. Downstream outputs are conditional on those rows and must remain attributed to the caller. ## Reference-only context Material available for interpretation or exploration but not admitted as model input or query-specific support. Examples include some dosage, atlas, population, pathway, nucleus, and tract layers. ## Release identity The versions and content identifiers that bind an output to an application, predictive model, facts-and-evidence layer, knowledge graph, schema/report, and generation time. ## Research Workbench A developing account-owned workspace for preserving a research goal, compounds, targets, filters, model settings, graph state, and versioned report references. ## Report pointer A path to an exact location in the structured report. Support links use pointers to express the scope of evidence. ## Scaffold-disjoint A chemistry split intended to keep molecular scaffold families separated between fitting and evaluation cohorts. Its exact construction and leakage controls must be documented for any claim. ## Chemistry-component split A deterministic split that groups compounds by scaffold and connects related chemistry at a declared fingerprint-similarity threshold before assigning whole components to train, validation, or test. ## Section Inspector The application surface for reading MoA, cognition, and side-effect rankings with reliability, report evidence, and separately labeled external literature. ## Data Explorer A bounded, read-only interface over existing report rankings. Natural language selects a structured action; it does not execute arbitrary queries or modify model output. ## Evidence Constellation The Knowledge Graph Explorer’s 3D rendering of report-local abstract topology. Its geometry is not anatomical space. ## SMILES Simplified Molecular Input Line Entry System, a text representation of molecular structure. A pasted SMILES still requires identity, stereochemistry, and chemistry review. ## Source-backed record A report item retrieved from a named, governed source. It represents the source’s assertion under its scope; it does not validate every nearby prediction. ## Structure-to-receptor (struct2rec) The novel-route component that combines a Morgan-to-receptor model with a Tanimoto-based KG prior to produce a receptor profile hypothesis. ## Support link A typed link between a report pointer and evidence records, graph elements, or sources. Its scope determines which claim it can support. ## Surfaceability A policy state that determines whether source material can appear directly, in redacted form, only under a restricted permitted use, or not at all. It is separate from scientific quality. ## Tanimoto similarity A similarity measure commonly applied to chemical fingerprints. It describes fingerprint overlap and is used in retrieval/applicability logic; it is not biological confidence. ## Top-K The K highest-ranked items returned by a head or route component. Top-K membership does not imply completeness or clinical importance. ## Transductive model A model whose inference operates on entities represented in its trained graph. The known-compound B0 model cannot accept an arbitrary new compound node at request time. *** Return to [Welcome to Neurolab](/readme.md). --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://docs.neurolabai.io/reference/glossary.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.