The Clinical Refinery

Drug development always forces the same tradeoff:

• Be rigorous enough to win (biology, safety, regulation)
• Move fast enough to survive (time and cost)
• Do not make strategic development decisions that cap the upside (uptake and peak year sales)

Both small biotechs and large pharma have to navigate that triangle. The failure modes are just different. Smaller teams do not have 20 to 30 FTEs and six months to throw at every development question. They are quickly running out of runway. Big companies technically have the people, but enough organizational friction that it can take near CEO level coordination to get true cross functional alignment, and that is not feasible for every decision. On top of that, for both big and small, the incentives are often misaligned. Stakeholders get rewarded for advancing their own ideas, not for retiring them when the evidence is weak.

The result is that a small number of high stakes decisions are made under time pressure, with incomplete evidence and fragmented context. Fibonacci has solved this bottleneck.

Our blog, "The Ontological Advantage: Building Knowledge Graphs That Reason," laid out a different approach. Instead of drowning in documents and one-off analyses, you:

• Build a shared ontology of the drug world: drugs, targets, indications, dosing regimens, safety, efficacy, sales.
• Use it to populate knowledge graphs from high-entropy public data (labels, SEC filings, ClinicalTrials.gov, press releases, IP, conference abstracts, global registries).
• Let LLM agents run Retrieval-Augmented Reasoning (RAR) over those graphs, so they don't just retrieve facts but actually reason over structured context.

We call that end-to-end system the clinical refinery inside our Medicine Engine. Fibonacci Bio is a drug development company that uses AI to make better decisions faster across the entire development lifecycle, from target selection through commercial strategy.

This piece is about what that looks like in practice. Two refinery runs, two very real problems drug developers wrestle with:

Both are questions that normally lead to many months of internal and/or consulting work, KOL boards, governance meetings, and internal debate. Here's what happens when they go through an AI-native engine instead.

The Clinical Refinery in Three Steps

Step 1: Ingest the noise

We start from high-entropy public data, plus a growing internal warehouse: regulatory labels and reviews, global trial registries, SEC filings and earnings transcripts, corporate decks, press releases, publications and case series, and IP filings. It is the information that matters, and it is scattered everywhere.

Step 2: Give it structure

We map that chaos into the ontology, then materialize it as a knowledge graph. Once the data obeys the same grammar, it becomes composable. You stop rebuilding the same spreadsheet for every question.

Step 3: Reason over it

On top of the graph, RAR agents do not just retrieve facts. They traverse relationships, assemble cohorts, compare like with like, and explain what is causal versus what is correlation or class mix. The output often looks like a cross-functional diligence memo or a strategy consulting deliverable. Under the hood, it is a reusable substrate you can query again tomorrow.

With that foundation, here are two refinery runs.

Refinery Run #1: Oncology mAbs

What dosing frequency really buys you

The question is easy to ask and difficult, nuanced, and expensive to answer.

In oncology, this anxiety is rational. Buyers and partners do care about differentiation. Convenience can matter. But oncology is also unforgiving. Efficacy dominates decision making, and teams are forced to commit early while the data is still noisy.

What we did

We built a high-resolution knowledge graph of FDA-approved oncology monoclonal antibodies. For each drug, the refinery ingested and normalized dosing regimens, infusion logistics, step-ups and premeds, pivotal efficacy readouts, safety profiles, and peak sales estimates, grounded in labels, trial publications, SEC and investor materials, and public disclosures.

The code that actually runs this? It compiles down to a simple declarative query:

kg = KnowledgeGraph(session=db)

results = (
    kg.assets()
    .source(source_data)  # Just the names of the antibodies we care about.
    .enrich(DosingCadence)
    .enrich(PeakSales)
    .enrich(PrimaryIndication)
    .enrich(DrugProfile)
    .parallel(max_concurrency=3)
    .collect_sync()
)

Then RAR agents ran the question in multiple cuts:

• By dosing cadence (weekly or more often, every 2 weeks, every 3 weeks, monthly or longer)
• By convenience burden (frequency plus infusion time plus premeds plus step-ups)
• By indication mix and mechanism class
• By commercial outcomes

What we found

Dosing frequency alone is not a reliable predictor of peak-year sales, safety, or efficacy across approved oncology mAbs. Once exposure and efficacy are adequate, cadence within common ranges is rarely the primary driver of value. But that is not the whole story.

The chart below shows each approved oncology mAb plotted by dosing cadence and peak-year sales. Hover over any dot to see the drug. Notice how the highest-revenue drugs are scattered across cadences, and the apparent Q3W cluster is driven by checkpoint inhibitors and myeloma blockbusters, not by the dosing interval itself.

Convenience matters most when efficacy is similar and competition is tight. But a drug with a clear efficacy and outcome advantage can overcome inconvenient dosing. Drug development and strategy teams often optimize convenience later, for example extended intervals or subcutaneous formulations post-market approval, once efficacy is proven.

What this changes for teams

You can make early dosing decisions based on PK and PD, tolerability, and operational feasibility, without treating cadence as an existential commercial choice. You can move faster, with facts, rather than with fear.

Refinery Run #2: CD3 Bispecifics

From precision immunology to outpatient reality

CD3 bispecifics are among the most potent precision immunology tools in the armory. In the right settings, they can eliminate disease-driving cell populations and reset immune function, not just modulate it. They can also engage T cells against targets with very low antigen copy number, expanding the target universe for immunology.

If you can make that safe in outpatient immune and inflammatory care, you are not building a niche product. You are building a platform with Humira-scale value (>$200 billion USD in cumulative lifetime sales).

But immunology is an outpatient business. CD3 today is not.

The core question

We know the biology can work. The question is what it takes to make it usable.

What we did

We used public trial registries, including US, EU, and China equivalents, plus public disclosures to build a structured dataset and knowledge graph of CD3 bispecific programs across immunology and oncology. The CD3 ontology made the design space explicit:

• CD3 and target affinity, where disclosed
• Antigen density and, critically, target cell abundance (total kill volume)
• Route (IV versus SC)
• Multi-step titration schemes and premeds
• Adverse events including cytokine release syndrome (CRS) and neurotoxicity (ICANS) incidence and timing, including delayed events
• Comparative preclinical NHP packages, where available
• Human safety and efficacy signals

Then RAR agents queried the graph to surface design rules and a target product profile filter for what outpatient-ready CD3 would require.

What we found

First, the uncomfortable truth: no current CD3 profile is truly outpatient-ready in publicly disclosed datasets and labels. CD3 bispecifics still carry CRS risk that drives inpatient initiation, step-ups, and long observation windows.

Second, the graph makes the levers legible, and not all levers are equal. Outpatient feasibility is driven by the combination of:

1. CD3 engagement (affinity and effective potency)
2. Antigen density
3. Target cell abundance (how many cells you are actually killing)
4. Specificity and off-target binding (a silent amplifier of CD3 engagement)
5. Temporal control (dose titration, micro step-ups, and PK shaping)
6. Mitigation burden (premeds that are acceptable only on early doses)
7. Comparative NHP benchmarks that show clean cytokine profiles at high dose versus the status quo

The parallel coordinates below let you trace individual CD3 programs across these dimensions. Drag along any axis to filter. You can see how programs cluster by route and CRS profile, and how few threads make it through to the lower-right corner where outpatient feasibility lives.

Target cell abundance turns out to matter more than most teams expect. Oncology often means debulking trillions of target-positive cells. Precision immunology is, ideally, deleting millions of disease-driving cells. That difference changes the cytokine math, but only if binding specificity and CD3 engagement are tightly controlled.

The refinery-derived TPP filter (what we would want to buy or build)

This analysis becomes a licensing and build criteria filter for drug development and strategy teams. For CD3 immunology assets we would seriously consider, especially de-risked development candidate and later-stage programs, we look for:

• Targets enriched on disease-driving subsets with low total cell burden and minimal expression on critical healthy tissues
• Tuned CD3 engagement, often in the hundreds of nM to micromolar effective range
• Highly specific target binding with minimal off-target engagement
• Slow ramping of killing activity via multi-step titration over weeks
• SC as the long-term route where feasible
• Premeds and heavier monitoring acceptable only on the first dose or two
• Strong comparative NHP packages (dosing tens of mg per kg without major cytokine spikes) benchmarked against known CD3 programs, e.g., CD20×CD3
• Early human datasets that explicitly rule out delayed CRS, with cytokines and AE timing through the early doses and observation windows that shrink over time (<1 hour)

Why the refinery matters

CD3 in immunology is a world problem. Whoever unlocks safe outpatient CD3 bispecifics in immune and inflammatory indications will cure disease and create hundreds of billions of dollars in value. The clinical refinery is how you attack that problem in a way that is nimble, rigorous, and unbiased, for both small biotechs and large pharma.

The bottom line

The traditional way to answer questions like these takes 6 to 9 months, 20 to 30 FTEs across functions, and $500k or more in external consulting on top of millions in internal overhead. The clinical refinery compresses that to about a week, one analyst supported by AI agents, and thousands of dollars in marginal compute. The substrate is reusable, so the next question is cheaper than the first.

But the real value is not in the cost savings. It is in the decisions you can now afford to make well. The tradeoff we opened with, rigor versus speed versus protecting your upside, stops being a tradeoff when the cost of rigor drops by two orders of magnitude. You can be thorough and fast. You can pressure-test a dosing strategy before locking it in. You can map the design space for a new modality before committing R&D dollars.

One good decision, made with evidence instead of intuition, can be worth more than the entire cost of the system that produced it.