What this document is. A worked, end-to-end demonstration of what the Catalyst AI drug-discovery engine produces — and, just as importantly, what it *honestly refuses to claim*. It pairs a negative result the engine caught on itself (DLL3) with a positive result it delivered and fully characterised (PD-L1). The credibility claim is not "our AI designs drugs" — it is "our AI produces traceable, independently-validated, honestly-bounded candidates, and tells you exactly where the in-silico evidence stops and the wet lab must begin." Everything here is in-silico. i_pTM is a predicted-interface confidence, not an affinity (Kd). No wet data exists yet. The engine's output is a *de-risked, prioritised hypothesis ready for a cheap, bounded wet test* — not a drug.
### ⬛ Positioning (read this first) This is a proof of the engine's capability, NOT a PD-L1 therapeutic asset. Honest verdict: - As an AI-engine proof-of-capability: strong. - As a PD-L1 therapeutic asset: weak-to-moderate — PD-L1 is a crowded, mature target (atezolizumab/ durvalumab/avelumab are approved), and *this binder is non-PD-1-competitive, so it is explicitly not a checkpoint inhibitor.* Do not pitch it as "a new-generation PD-L1 checkpoint inhibitor" — that claim does not hold and a pharma expert would see through it instantly. - As a seed-deck / technical-validation case: very useful. The value created here is platform credibility, not the molecule. With only target dossiers, the story is "we analyse targets." With this case added, it becomes "we design binders, validate them with independent models + developability + specificity + FTO, and know when not to trust ourselves" — which is the technical closed loop a seed investor needs to see.
The engine's terminal deliverable is a computationally de-risked candidate package:
| Layer | What the engine produces | For the PD-L1 lead (§3) |
|---|---|---|
| Molecule | A synthesisable binder sequence (+ predicted complex structure) | 74-aa de-novo mini-protein |
| Binding | Two independent structure-model predictions (consensus bar) | AF2 0.85 and Boltz-2 0.82 |
| Energetics | Interface ΔG, shape complementarity, buried area, H-bonds | ΔG −43, SC 0.74, 1558 Ų, 5 H-bonds |
| Developability | Solubility, net charge, aggregation, stability, charge patches | soluble 0.98, net −3, no aggregation/instability |
| Specificity | Cross-reactivity vs the closest paralog(s) | vs PD-L2 = 0.27 → selective (margin 0.55) |
| Functional epitope | *Which* surface it binds + whether that is the functional site | binds PD-L1, non-PD-1-competitive epitope (honest) |
| FTO | Claim-type + epitope-overlap pre-screen (not legal advice) | low risk (de-novo, non-antibody) |
| Honest boundary | Exactly what is in-silico vs what needs wet | §4 |
| Provenance | Every number reproducible from inputs + tool + positive control | §6 |
It is not a drug. It is an open-ended discovery problem compressed into a **small, bounded, ~weeks /
~$10–20k wet test** with a defined kill criterion. That compression is the value.
The same pipeline was run on two targets. The discipline that matters — a two-INDEPENDENT-model
binding bar — is what separates a real signal from a design-time mirage.
| DLL3 (de-novo, EGF1) | PD-L1 (de-novo) | |
|---|---|---|
| AF2 trajectory (design-time co-fold) | 0.73 *(looked good)* | 0.85 |
| AF2 independent re-prediction | 0.06 ✗ (no binding) | 0.85 ✅ |
| Boltz-2 (independent 2nd model) | 0.83 | 0.82 ✅ |
| BindCraft full internal gate | 0 / many designs passed | passed (final design) ✅ |
| Developability | unstable, +7 charge (kidney risk) | clean (net −3) |
| Engine verdict | rejected — not validated | delivered (target engagement) |
DLL3's "AF2 0.73" was a trajectory co-fold — AF2 scoring its own design objective, optimistic by
construction. Asked to re-predict independently, AF2 returned ~0.06. Only one model ever liked it, so it
was never a real consensus. The engine now structurally cannot count a co-fold score as validation
(hardened in binder_optimize: a two-model consensus requires two *independent* predictions; positive
control barnase:barstar = 0.73 confirms the AF2 method is sound). On DLL3 the engine therefore **says no
to its own earlier candidate** — which is precisely the behaviour that makes a *yes* trustworthy.
The DLL3 failure is target-physics (a small, acidic, cysteine-dense EGF surface where binding is
electrostatically captive), not an engine defect — proven by switching to an appropriate surface (PD-L1)
and immediately getting independently-validated binders.
Lead: l74_mpnn2 (74 aa de-novo mini-protein)
SPLKGLKPEELPDEEELEKLLKSLNAPEVDKKRALSFLRFYKWLKEMYEETGQEHFKKNAEWAYEQIKELVEKH
real PD-1:PD-L1 crystal complex (PDB 4ZQK) — not guessed.
Boltz-2 independent 0.82. Both clear the 0.6 consensus bar. Three distinct backbones (l103/l72/l74)
independently reached two-model consensus → robust, not a single lucky design.
1558 Ų, 14–15 interface residues, 5 interface H-bonds — a substantial, well-packed interface.
no aggregation-prone regions, no predicted instability; one minor local cationic patch. (Across
backbones the engine *de-selected* an aggregation-prone, unstable backbone (l103) in favour of clean
ones — multi-criteria selection, not just top i_pTM.)
far below the on-target 0.82 (selectivity margin 0.55) → predicted selective for PD-L1.
l74_mpnn2 contactsPD-L1 residues ~32–52 / 96–106, overlapping the PD-1 binding face by only ~8%. So it is a
validated PD-L1 binder, but it binds a NON-PD-1-competitive epitope — it is not predicted to be
a checkpoint blocker. All three validated backbones behave the same way. A PD-1-blocking molecule would
require a hotspot-constrained re-design steered tightly onto the PD-1 contact residues. *The engine
surfaces this rather than quietly implying a functional MOA — that is the rigor, not a footnote.*
on the CDR-sequence claims of the approved anti-PD-L1 antibodies (atezolizumab/durvalumab class), and
its epitope only partially overlaps the claimed PD-1-competitive face. Composition is novel. Residual:
method-of-use claims + a formal opinion from counsel.
environment) — flagged as an open in-silico step, not silently skipped.
In-silico (done, here): target engagement (two-model), interface energetics, developability,
paralog selectivity, epitope/MOA characterisation, FTO pre-screen. This **de-risks "does the engine's
output have a plausible, selective, developable binding mode"** — and it does.
Wet (the truth, not yet done) — cannot be answered by any AI:
Recommended next spend: a pooled expression + BLI/SPR test of the top designs (~$10–20k,
~4–6 weeks). Hard, pre-registered decision criteria (turns this from a pretty story into an executable
experiment package):
GO (advance):
KILL (stop):
Do not spend on anything downstream until GO is met. If wet data passes, this case **upgrades from
"AI proof" to "real binder proof."**
A pharma/BD audience does not license a 74-mer; they assess whether a process is trustworthy and
repeatable. This case demonstrates exactly that:
count design-time optimism. Every number is reproducible (§6).
non-functional epitope instead of cherry-picking the highest i_pTM.
→ FTO, end-to-end, on demand, sovereign/local.
The honest one-liner: **"the engine produced a selective, developable, FTO-clean, independently
two-model-validated PD-L1 binder — and told us, unprompted, that it binds a non-checkpoint epitope and
still needs a $15k wet test to be real."** That is a process worth trusting.
It is not a checkpoint inhibitor. That does not make it useless — but the realistic uses are narrow
and must be stated honestly:
structural epitope → synthesisable mini-protein → independent-model validation → developability →
specificity → FTO pre-screen. This is the seed-deck value.
PD-1, a non-blocking binder can suit PET/SPECT imaging, IHC, or flow reagents (detect PD-L1 rather
than block it) — *contingent* on wet cell-binding, affinity, background and (for in-vivo) PK.
conceivable, but PD-L1 is expressed on immune cells and some normal tissue, so the therapeutic window
for a *cytotoxic/radioligand* payload is uncertain. There is industry interest (e.g. anti-PD-L1 ADC
programs in clinical development), which means both opportunity and real competition/toxicity work.
capability suite (target selection → design → validation → specificity → developability → FTO → wet
recommendation).
| Item | Honest positioning |
|---|---|
| DLL3 disease/RLT memo | target + modality investment intelligence |
| DLL3 l55/l87 | early de-novo binders — have NOT reached PD-L1's two-model independent consensus → need optimisation + wet |
| PD-L1 l74_mpnn2 | engine binder proof case (this doc) |
| DLL3 EGF1 (failed) | engine negative control (self-rejected) |
This layering — not packaging every result as a win — is itself a credibility signal.
Scorecard (engine ≠ asset): engine-proof value 9/10; scientific honesty 9.5/10; PD-L1
*asset* value 5.5/10 (crowded target + non-blocking MOA); platform value 8.5/10; wet-test
readiness 8/10; BD-story value 7.5/10 (great for a seed deck, not for selling a standalone asset).
binder_optimize (two-independent-model consensus), binder_engineering(developability / charge / epitope), patent_fto (FTO). All unit-tested; full suite green.