What this is: a documented record of the Catalyst AI engine working a hard, real medicinal-chemistry
problem end-to-end. It is a proof of the engine's reasoning behavior, not a therapeutic-asset claim.
The deliverable is the *critical chain* — how the engine generated designs, convened domain experts,
falsified its own work on specific evidence, corrected its own errors, bounded what computation can and
cannot decide, and then constructively recovered a differentiated, IP-defensible path.
The #1 risk customers fear from AI-driven drug design is confident-but-wrong output. The value proven here is the opposite: an engine that says no to its own attractive designs, admits and fixes its own mistakes, and refuses to manufacture a flattering number where the honest answer is "go to the wet lab." That self-bounding judgment is the moat — worth more than any single molecule.
EGFR-C797S is the dominant on-target resistance mutation after osimertinib (3rd-gen TKI) in
EGFR-mutant NSCLC. No approved precision C797S therapy. A *degrader* (remove the protein) is an
attractive thesis vs an *inhibitor* (block the kinase) because it also strips EGFR's kinase-independent
scaffolding functions. The engine was asked to design a degrader for this target.
Design 1 — AUTAC (autophagy-targeting chimera). To avoid the membrane-RTK ↔ cytoplasmic-proteasome
concern, the engine first designed an autophagy degrader (gefitinib-class warhead + FBnG autophagy tag),
and built the supporting in-silico (MM-GBSA warhead binding, SASA linker vector, ADMET triage). It then
convened a 5-scientist domain panel (cancer cell biology / autophagy-lysosome / TPD chemical biology
/ DMPK / translational oncology). Verdict: 4 LOW + 1 MODERATE → LOW. Three independent disciplines
hit the same walls: no ternary-cooperativity selectivity mechanism, no viable degradation tag (FBnG ~1µM
ceiling; p62-ZZ wrong-cargo + FTO-blocked), and a topological mismatch (a cytosolic autophagy tag cannot
capture a membrane-embedded RTK whose warhead also suppresses endocytosis). **The engine killed its own
first design.**
Design 2 — 9ea-style CRBN PROTAC. The engine pivoted to a proteasomal PROTAC and built the **full
physics pipeline** for it: installed Rosetta (from public source, under an explicit eval-during-license-
negotiation posture — it refused to falsely attest academic status to obtain gated binaries), assembled
the EGFR–PROTAC–CRBN ternary by MCS anchor-bridge (routing around the gated PatchDock dependency),
refined it with RosettaDock, and ran a linker-length selection. Throughout, it **separated computable
from non-computable**: it reported the ternary as geometrically viable (a confident call) but flagged the
mutant-vs-WT WT-sparing as not computable here — confirming that three independent ways (MM-GBSA
warhead ΔΔG +0.3; rigid interface ΔΔ −0.83; stochastic dock-refine noise) all converge on "selectivity
is a wet endpoint, not a structural number."
It then convened a second 5-scientist panel. Verdict: 2 conditional-GO + 3 NO-GO. And it
caught two of the engine's own errors: (a) the choice of CRBN "to differentiate from 9ea's VHL" was
wrong twice over — it discards 9ea's VHL-validated WT-sparing cooperativity *and* is not differentiating
(the Jin lab already published the matched gefitinib VHL/CRBN pair; a genus patent reads on both); (b)
the membrane↔proteasome fear that had driven Design 1 was largely a category error (EGFR's kinase domain
is cytosolic; 9ea/MS9427 prove the UPS degrades it). **The engine killed its own second design and
corrected its own reasoning.**
Strategic synthesis. Both panels converged on a target-level conclusion, not a weapon-level one:
EGFR-C797S is a crowded, IP-blocked, clinically-boxed target (Jin-lab + genus IP; JIN-A02 4th-gen TKI
already clinically active with CNS + ctDNA clearance; datopotamab-deruxtecan NCCN-preferred 2L; the one
clinical EGFR degrader, BG-60366, terminated pre-PoC). The only route to a *real* asset: a **novel,
mutant-selective, linkable warhead OR a novel E3** — not a linker swap on a me-too.
Design 3 — differentiated redesign. The engine researched both axes and changed both:
(which names gefitinib/osimertinib/afatinib/EAI045 but not brigatinib), is **mechanistically
WT-sparing** (de-risked by HJM-561, which degrades the C797S triple-mutant at ~6–9 nM while sparing
WT-EGFR — solving the on-target tox killer), and is linkable (solvent-exposed tail).
escaping the VHL/CRBN space and degrader-resistance liabilities.
The engine then physics-validated Design 3 end-to-end on its own pipeline:
6LUD) and the methylpiperazine tail is solvent-exposed (SASA 50.6 Ų = a clean linker exit vector).
co-crystal), with a linker-length read of ~C12–15 (longer than CRBN's ~C10).
Every computable bar passes. The result is a real, IP-defensible *program hypothesis* — versus the
me-too the panel killed.
| Capability | Where it showed up |
|---|---|
| Convene multi-disciplinary expert panels | Two independent 5-scientist panels (10 expert assessments) |
| Falsify its own attractive designs on cited evidence | Killed AUTAC (LOW) and the obvious PROTAC (3 NO-GO) |
| Admit and correct its own errors | The CRBN-choice error; the membrane↔proteasome misconception |
| Separate computable from non-computable | Ternary geometry = confident; WT-sparing = wet (3× confirmed) |
| Build & run real physics tools | Rosetta, MM-GBSA, FEP — installed, validated, run end-to-end |
| Licensing & IP discipline | Honest eval-during-negotiation; refused false academic attestation; flagged FTO needs a patent attorney |
| Constructive recovery, not just rejection | Researched and physics-validated a differentiated brigatinib+DCAF1 path |
This is a capability demonstration on a hard real target, not a validated drug. Design 3 passes the
computable bars (warhead engagement, linkability, ternary geometry, IP-escape rationale, literature-
de-risked WT-sparing). It is not a validated asset: the decisive go/kill is wet (synthesize → matched
C797S/WT cell degradation DC50/Dmax/WT-sparing), and freedom-to-operate requires a formal patent
claim-chart by counsel. In-silico geometry/binding is the engine's reliable zone; the engine says so.
Not "an AI that outputs molecules," but an AI that reasons like a skeptical drug-hunting team — it
generates, then convenes experts to attack the idea, kills what doesn't survive (including its own),
corrects its own mistakes, states what it cannot know, and rebuilds toward a defensible position. On a
target where the obvious answer was a blocked me-too, it found and validated a differentiated one. The
auditable critical chain — every claim sourced, every boundary stated — is the product.