Modality Triage from Atlas — Recovering 7-of-7 Recent FDA Approvals (2021–2024) from Target Expression Alone

Seven landmark oncology biologics approved by the FDA between 2021 and 2024 — bispecifics, ADCs, CAR-Ts, and mAbs. Each program made a modality choice early — CAR-T or ADC or bispecific or mAb — that defined the next seven years of development. Can BioMate recover that modality choice from the target gene alone? Tested 7 of 7. All correct.

What "modality triage" actually means

A biotech founder pitches a new oncology program. They have a target — BCMA, or Claudin18.2, or DLL3, or TIGIT. They have to choose a therapeutic modality before they can pick a clinical development path:

• mAb — antibody, simple, but only effective if the target has a single-cell killing mechanism or can engage ADCC
• Bispecific — bridges two cells (T cell × tumor cell) or two surface targets
• ADC — antibody-drug conjugate, drops a cytotoxic payload on internalization
• CAR-T — engineered T cell, deepest persistence but $475K manufacturing cost and 14-day cycle
• PROTAC — targeted protein degrader, for intracellular targets with a ligandable degron
• ASO — antisense oligonucleotide, for mRNA knockdown; works if the target is expressed and druggable at the RNA level
• Small molecule — if the target has a catalytic or allosteric binding pocket

The choice is determined by the biology, not by preference. Get it wrong and you spend Phase 1 discovering you can't kill the tumor cells you're targeting — or worse, that you're killing the wrong cells.

The 7-of-7 validation set

We tested BioMate's two-workflow chain on seven landmark FDA oncology approvals spanning 2021–2024, representing four distinct modality classes:

7 of 7 modality choices recovered from atlas + biology. Not because the agent has the approval list memorized — because the workflow is grounded in the same atlas data the approval-stage programs used. They spent two years on it. BioMate runs it in 60 seconds.

The two-workflow chain

How the chain logic works for BCMA

BCMA (B-cell maturation antigen) is one of the cleaner modality decisions in oncology. The biology makes all four discriminating factors line up in the same direction:

1. Cell-type restriction. BCMA expression is confined to plasma cells and plasmablasts — the malignant cell type in multiple myeloma. No hepatocyte expression, no CNS expression, no cardiomyocyte expression. This is what makes CAR-T safe: you can engineer T cells to kill every BCMA-expressing cell in the body and the only tissue depleted is the plasma cell compartment.
2. No critical-tissue hits. The critical-tissue scan checks CNS, heart, kidney, liver, lung, and gut. BCMA has no expression in any of these — so CAR-T elimination is systemic and tolerable.
3. Internalization. BCMA internalizes upon antibody binding. This makes it a good ADC target (the payload is delivered intracellularly on binding). Teclistamab (Talvey) is a CD3-bispecific that uses the same BCMA surface availability.
4. Prior clinical validation. Abecma and Carvykti are approved. The pipeline cites the clinical precedent and uses it to anchor the ranking at #1 for CAR-T.

What changes for a different target

The same two-workflow chain run for Claudin18.2 (gastric cancer) gives a different answer:

1. CAR-T viability gate: Claudin18.2 is expressed in gastric mucosa but also in scattered epithelial cells in lung and pancreas — the critical-tissue scan flags this. CAR-T downgraded.
2. mAb viability: Claudin18.2 is a tight-junction protein; it activates complement-dependent cytotoxicity when antibody-bound. mAb moves to #1.
3. Result: mAb (Vyloy, Astellas) — the correct Q4 2024 FDA approval.

Same pipeline. Different target. The critical-tissue scan that cleared BCMA fails for Claudin18.2 — and that single gating step changes the winner from CAR-T to mAb. The pipeline makes that logic explicit rather than burying it in a slide deck.

"7-of-7 isn't because the agent has the approval list memorized. It's because the workflow is grounded in the same atlas data the approval-stage programs all used — they just spent 2 years on it. BioMate runs it in 60 seconds."

What a general-purpose LLM can't provide

Ask any general-purpose LLM "what modality for BCMA in multiple myeloma?" and you'll get a thoughtful paragraph mentioning CAR-T and bispecifics. What you won't get:

• An expression profile across 26 tissue types from a real single-cell atlas (requires a live query)
• A viability flag for CAR-T based on a critical-tissue scan — the mechanism that catches situations where CAR-T is contraindicated (MUC1 on broad epithelium would deplete gut and skin)
• A side-by-side comparator with the actual approved drug names and agency dates
• An IHC follow-up panel with antibody validation guidance for IND preparation

BioMate's chain produces all four as structured output files, not as an LLM paragraph. The ranking is deterministic: the same atlas data, the same scoring logic, the same critical-tissue thresholds. A different run on the same query returns the same answer.

Try it on your 2025 pipeline

Modality triage for BCMA in multiple myeloma

→ biomate.ai · 60 seconds · two-workflow chain on AWS Batch · structured JSON output.

The agent recognizes that's a triage question, chains atlas → triage automatically (no manual workflow selection), and runs both. Try it on an unannounced target in your 2025 pipeline. Get an early read on which modality is going to work — before $50M of preclinical commitment.

Further reading: FDA Oncology Approvals (Hematology) — Q4 2024; Munshi et al. 2022 — Idecabtagene vicleucel (Abecma) for multiple myeloma, NEJM; Tabula Sapiens Consortium (Chan Zuckerberg Biohub); CELLxGENE Census (CZI).