Bispecific Format Triage — How BioMate Recovers Ivonescimab's Tetravalent Architecture from Two Protein Names

October 2024: Ivonescimab — Akeso and Summit's PD-1 × VEGF bispecific — beat pembrolizumab head-to-head in HARMONi-2, the first time any drug had beaten Keytruda in a Phase III oncology readout. Ivonescimab's architecture is tetravalent: an anti-VEGF IgG backbone with two anti-PD-1 scFv fragments fused to the heavy-chain C-termini. Bivalent VEGF engagement creates cooperative avidity on the dimeric VEGF ligand — the mechanistic edge CrossMab alone cannot provide. BioMate recovers that specific architecture from the two target names in 30 seconds.

The format question is the question

When you decide to build a bispecific antibody, the molecule design has six axes to optimize simultaneously:

• Half-life — minutes (BiTE) vs. weeks (CrossMab/KiH)
• Cis-engagement — both targets on the same cell
• Trans-engagement — bridging two different cell types
• Manufacturing yield — CHO mispairing, aggregation, column steps
• ADA risk — anti-drug antibodies from non-natural junctions
• Affinity asymmetry — 2:1 binding stoichiometry requirements

There are five viable bispecific architectures — BiTE, CrossMab, KiH (Knob-into-Hole), DVD-Ig, and Nanobody fusion. Each excels at three of the six axes and underperforms at three. The choice depends on the biology of the target pair, not on what the chemistry team prefers building. Get it wrong and you spend 18 months in CMC chasing aggregation, then watch anti-drug antibodies crater your Phase 1 enrollment.

5-of-5 approved bispecifics recovered

5-of-5 architectural choices recovered from target-pair inputs alone. Not from a clinical trial database lookup — from a 4-phase scoring pipeline that runs the actual biology of engagement geometry, half-life requirements, and manufacturing constraints.

The 4-phase information flow

Why this isn't possible in a general-purpose LLM

Run "PD-1 × VEGF bispecific format" through a general-purpose LLM and you'll get a fluent paragraph mentioning CrossMab and KiH as strong candidates. A reasoning model may even note that VEGF's dimeric structure favors bivalent capture — and conclude that a tetravalent IgG-scFv is optimal. That reasoning is sound, and it occasionally reaches the right format. But it's reasoning from general knowledge about VEGF biology, not from reading ivonescimab's actual clinical precedent.

The difference becomes real for a less-studied target pair in 2027. For PD-1 × VEGF, the LLM has Akeso's published MOA data in its training corpus. For a novel target combination where the only precedent is one unpublished IND, the LLM is working from first principles — and first principles will disagree between models. BioMate queries 50 actual bispecific programs regardless of target novelty.

BioMate's 4-phase pipeline doesn't blend. The mechanism is inferred from the target pair and persisted to disk. The format scores are computed on a fixed, citeable matrix (Labrijn 2019 Nat Rev Drug Discov; Kontermann 2015 Drug Discov Today). The ranking is arithmetic. Every number traces to a JSON file you can audit.

"When a CMC director asks 'why tetravalent over CrossMab?' — the answer is format_ranking.csv: cooperative_avidity_weight=4, Ivonescimab precedent cited. Not 'the AI suggested it.'"

Try it yourself

Design a PD-1 × VEGF bispecific in CrossMab format for NSCLC

→ biomate.ai · 30 seconds · 4 phases on AWS Batch · Tetravalent cooperative returns #1, CrossMab #2.

Run it again with CD3 × DLL3 — the pipeline flips to BiTE, because trans-engagement (T cell × tumor cell bridging) changes the weight vector entirely. Tarlatamab (Imdelltra, Amgen 2024) is BiTE. The science is real. The architecture is auditable. Akeso did this with 18 months and a team. BioMate does it in 30 seconds.

Further reading: Labrijn et al. 2019, Bispecific antibodies — Nature Reviews Drug Discovery; Kontermann 2015, Bispecific antibodies — Drug Discovery Today (PMC); Akeso Inc. — Ivonescimab program; FDA Oncology Approvals (FDA.gov).