How FastScience! Cuts Research Time with Graph AI

The instinct in 2026 is to throw a bigger language model at any research problem. That instinct is wrong, and it costs you weeks of compute.

Here is the stack we actually use, and why each layer earns its place.

Layer 1: Neo4j as the causal backbone

We do not store papers — we store entities and causal arcs extracted from papers. A typical kwiKBio knowledge graph for a single domain holds on the order of tens of thousands of entities and hundreds of thousands of typed arcs (see the microbial consortia case study for exact figures from one production graph). Arc types matter: regulates, inhibits, co-occurs-with, and causally-precedes are not interchangeable.

Neo4j is the right database here because path queries are first-class. The questions we ask — "what regulates X two hops away from Y?" — are a few-line Cypher query on Neo4j and a sprawling SQL join on Postgres.

Layer 2: Sub-linear association mining

Sub-Linear Association Mining (SLAM) is a body of academic work — notably advanced by Evan Steeg — designed to find higher-order correlations in multidimensional data that pairwise statistics miss. We use a SLAM-style approach to surface candidate causal arcs the literature has not yet annotated.

The output is always a hypothesis, never a claim. SLAM-style mining proposes "these three entities co-cluster in a meaningful fraction of relevant studies"; the rest of the pipeline tests whether that co-clustering is causally informative.

Layer 3: PRISM-9 reduction

Most well-posed research problems share a useful property: they're dominated by a small number of entities, often ≤9. The rest of the graph contributes noise.

PRISM-9 forces that reduction. It picks the 9 entities whose joint variance explains the most outcome variance for the current question, then prunes everything else. This is the step that makes the next layer tractable.

Layer 4: Cube-27 expansion

Each of the 9 dominant entities expands fractally into 3 sub-factors, giving a 3×3×3 cube of 27 leaf variables. The geometry preserves causal locality, so adjacent cells correspond to causally adjacent sub-factors.

Layer 5: QUBO solving

With 27 binary variables (active / inactive sub-factor), path planning becomes a Quadratic Unconstrained Binary Optimization. That is a problem class quantum annealers (such as D-Wave Advantage at 5,000+ qubits) are designed to address.

We get a near-optimal causal path in seconds on quantum hardware when available; classical simulated annealing handles it in minutes as fallback. Either way, the user sees ranked experiments quickly.

Why not just bigger LLMs?

Because LLMs hallucinate causal arcs, and grounding them in your specific domain manual requires retrieval plumbing that often costs more than the model. The graph layer eliminates a class of failure modes that pre-training does not fix.

LLMs remain in the loop — as our natural-language interface and our summarizer. They do not drive the science.