We've been wanting to answer one question for a while: can AI actually accelerate preclinical drug discovery, and meaningfully cut real experimental workload? That's not a question you answer with theory — you have to actually run it. So we sent three small teams after three targets using our full Benchora + ChemOrchestra stack: KRAS G12D, BRAF (V600E), and EGFR (C797S).
KRAS G12D: the binding site was hard enough to locate that the team wrapped up after SAR analysis alone.
EGFR (C797S): slow start, strong finish — now our fastest-moving project, with 200+ candidate structures mapped across a Markush scaffold.
BRAF (V600E): went well. Two team members independently found 3 promising seed molecules each via scaffold hopping, and grew those into two separate Markush structures.
Today I'll walk through the BRAF V600E discovery strategy. EGFR (C797S) is a separate writeup, once we get sign-off to share it.
Phase 1: Problem Framing
BRAF inhibitor development aims to extend survival for melanoma patients with brain metastases. Two things make this hard: blood-brain barrier penetration (often needs intrathecal dosing, rough on patient compliance), and BRAF's dimer paradox (an inhibitor can inadvertently induce wild-type BRAF dimerization, activating the very MAPK pathway it's meant to shut down). BRAF is competitive, but nowhere near as saturated as KRAS or HER2 — there's still room to move.
~116K new invasive melanoma cases/year in US+Canada; ~5,500 new stage IV cases; 35.6% five-year survival once metastatic; ~85% of BRAF/MEK-inhibitor patients eventually develop resistance.
Phase 2: Benchora-Guided Scaffold Hopping
Benchora was still early at this point — its ability to design around existing patented Markush structures was limited, and its generative model was only okay. Even so, it handled basic scaffold modification and structure-activity analysis well enough. Across several dozen search rounds, it generated two molecule trees, and the two team members manually picked 20-50 candidates each.
Benchora scaffold-hopping tree
Phase 3: Boltz-2 Potency Validation
Boltz-2 was genuinely strong for initial pose and potency triage here. Calibrating against Pfizer's 2025 BRAF V600E patent (US 12,303,509 B2), two disclosed compounds — #33 and #71 — came back at predicted 10 nM and 144.6 nM against experimental 1 nM and 303 nM. Not exact, but right order of magnitude and correctly ranked.
Phase 4: ChemOrchestra Multi-Parameter Screening
Boltz-2-predicted potency ran alongside hERG risk and blood-brain-barrier permeability in one pass. Heavy attrition here — most candidates got filtered out, which we read as the model doing real discriminative work, not a bad sign.
ChemOrchestra BIND result + pharmacophore analysis
Phase 5: Selectivity and FEP
Wild-type vs. V600E selectivity was the hard wall — Boltz-2 alone tended to misjudge wild-type activity (really above 20 μM) as hundreds of nanomolar, not precise enough for real selectivity screening. Free-energy perturbation was needed to lock in genuinely V600E-selective structures before looping back.
Our lead compound QB-1021 aligned against the Zelboraf–BRAF V600E crystal structure — still meaningful room for optimization relative to the approved drug
A month in, we brought an agent in to lock down two Markush structures, corresponding to two patents.
Summary
We're actively shopping both structures for academic collaboration or commercial co-development — both leads lean commercial, though we're not ruling out a joint paper. If you or someone you know is interested in the BRAF target, let's talk.
QB-1021/QB-1022: 10-20 nM. QB-1023: 20-50 nM
AI can look deceptively strong on activity prediction alone once it has enough training data — but real selectivity work still runs into FEP as a hard requirement. Boltz-2 (OpenFold too) still meaningfully speeds up the front end with a strong initial pose and potency guess. It's just not the whole story.
Feel free to schedule a meeting with me if you have any specific questions/requests: https://calendly.com/yizhouma/chemorchestra-user-session
Join our community to learn more: https://discord.gg/u3Mb6C2aHv