> 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/engineering-and-scientific-assurance/model-inference-overview.md).

# Model Inference

Neurolab’s serving architecture is best understood as three cooperating engines: a transductive predictor for known compounds, a structure retriever for novel chemistry, and a mechanism reasoner. It is not one universal model.

Predictive checkpoints and facts-and-evidence data have separate release identities. The v43.262 predictive family remains the current served model line. v44.16 identifies the latest candidate facts-and-evidence package; each report should disclose the package actually served rather than treating a candidate as an active release.

## Known-compound predictor

The known route serves the **v43.262 DE-LEAK B0** model family: an edge-feature `HeteroConv(TransformerConv)` heterogeneous GNN with two layers, hidden dimension 128, four attention heads, and dropout 0.3.

The route uses a three-seed ensemble trained with seeds 42, 123, and 777. It produces six raw output heads:

| Head                | Output space                              |
| ------------------- | ----------------------------------------- |
| Cognition           | Top-K over a coarse 30-concept vocabulary |
| Reported usage      | Top-K over a fixed 50-label vocabulary    |
| Side effects        | Top-K over a fixed 57-label vocabulary    |
| ATC                 | Top-K over 50 therapeutic classes         |
| Mechanism of action | Top-K over 100 MoA classes                |
| Brain               | Dense ranking over 392 parcels            |

Cognition, MoA, and side effects use per-head receptor attention. Reported usage, ATC, and brain use direct compound readouts.

The heads are parallel objectives. They are not steps in a single biological causal chain, and their scores do not share one calibrated scale.

The current service adds human-readable atlas metadata to the top parcel rows without changing the 392-dimensional brain output. See [Brain Localization and Atlas Labels](/engineering-and-scientific-assurance/brain-localization.md).

## What DE-LEAK means

Earlier model vintages allowed relations used as supervision targets for cognition, usage, side effects, and ATC to participate in message passing. That let held-out compound embeddings see answer-like edges. In v43.262 those supervision-target relations are excluded from the GNN feed.

The change retired the older leakage-inflated headline metrics. Current documentation uses the content-bound deployment validation report instead.

## Novel-structure retriever and reasoner

A new SMILES string is not a node in the transductive GNN. The structure route therefore does not execute or load B0.

```mermaid
flowchart LR
    SMILES["Submitted SMILES"] --> Morgan["Morgan fingerprint"]
    Morgan --> S2R["Structure-to-receptor MLP + Tanimoto KG prior"]
    Morgan --> Retrieval["GraphRAG chemical-neighbor retrieval"]
    Receptors["Inferred or caller-supplied receptors"] --> Reasoner["CircuitFramework mechanism reasoner"]
    S2R --> Receptors
    Retrieval --> Compose["Route-specific composition"]
    Reasoner --> Compose
    Compose --> Output["Hypotheses + provenance + applicability + abstention"]
```

The route combines:

* **struct2rec:** a Morgan-to-receptor MLP blended with a Tanimoto-based KG prior;
* **GraphRAG retrieval:** chemical-neighbor retrieval used to assemble analog context and compatible head outputs;
* **mechanism reasoning:** receptor-to-CircuitFramework composition for fine-cognition and multiscale mechanism hypotheses; and
* **route policy:** applicability gates, unresolved-direction withholding, provenance, and abstention.

Neighbor facts remain facts about the neighbors unless explicitly supported for the query.

## Relationship safety

Downstream receptor evidence uses typed relationship policy rather than treating every nearby graph edge as positive support. Negative-binding, reverse-traversal, and noncausal relations are excluded from positive receptor evidence. Relation direction and action are resolved conservatively; unsupported sign remains uncertain. The relation policy also distinguishes inhibitor and antagonist analog records. These controls prevent a negative or navigation-only relation from being silently promoted into mechanistic support.

## Caller-supplied receptors

When receptor rows are supplied in Advanced input, structure-to-receptor inference is bypassed or conditioned according to the request contract. The output must retain `receptor_source` and indicate that downstream reasoning depends on caller input. Model-seed gating cannot validate a receptor action supplied by the caller.

## Brain output channels

The system preserves multiple brain-related channels rather than merging them into one confidence map:

* the known-route raw GNN parcel output;
* a compatibility-only legacy structure channel;
* direct or expanded receptor-density hypotheses;
* circuit-localization hypotheses;
* literature-evidence channels when independently loaded; and
* measured channels only when actual measurements are supplied.

Reference anatomy and atlas density are not functional activation. A circuit composition is not a measured propagation trace. The legacy structure channel remains unvalidated and must be labeled accordingly.

## Report composition

After route inference, deterministic composition can add reliability, provenance, reference records, mechanism explanations, and optional KG support. Some fields are direct learned outputs; others are retrieved or composed. The report’s field-level evidence state is therefore as important as the route itself.

***

**Current predictive release:** v43.262 DE-LEAK, three-seed ensemble. Facts-and-evidence updates do not imply a new predictive checkpoint.

Next: [Knowledge Graph](/engineering-and-scientific-assurance/knowledge-graph.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/engineering-and-scientific-assurance/model-inference-overview.md?ask=<question>&goal=<endgoal>
```

`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.