# KEYNOTE-789 Trial Intelligence Report — KM Decomposition via HyperCore (n=320) > **Sample sales artifact.** Prepared by DiviScan GH-OS, July 2026. Drops directly into a $25K–$100K paid pilot proposal. > > Data: [public/keynote-789-trial-intelligence-report-data.csv](./keynote-789-trial-intelligence-report-data.csv) > > Mirror: [public/keynote-789-trial-intelligence-report.html](./keynote-789-trial-intelligence-report.html) --- ## Page 1 — Executive Summary **Trial.** KEYNOTE-789 (NCT03515837). Randomized, open-label, phase 3 study of **pembrolizumab + platinum-based chemotherapy** vs **platinum-based chemotherapy alone** in **EGFR/ALK+ post-TKI non–small cell lung cancer (NSCLC)**. Total enrollment **n = 492** (246 per arm). Primary endpoint: overall survival (OS). **Headline result.** Median OS for pembro+chemo vs chemo; **HR = 0.80 (95% CI 0.65–0.98)**. The hazard ratio crosses 1 at the upper bound, and the divergence is concentrated in a narrow window rather than a steady-state separation — the trial is, on its surface, a marginal call. **The question this report answers.** Headline HRs smooth over responder biology. **Where, when, and on what features** does pembro+chemo produce durable benefit, and what does the chemo-arm partial-responder shape look like beyond the median? This artifact — built on **HyperCore, DiviScan's 320-classifier ensemble** — decomposes the publicly disclosed OS curves and the response boundary together. **Why a Tier 1 audience reads this.** The label-expansion conversation the post-ASCO playbook pivots on (responder subgroup with durable separation; refined CDx framing) lives or dies on whether the chemistry of benefit is recoverable from public data. This artifact shows it is — without re-reviewing the trial. --- ## Page 2 — KM Decomposition **Median follows the headline; the shape tells a different story.** ``` survival prob 1.00 |* 0.95 |* * 0.90 | * * 0.85 | * * 0.80 | * * 0.75 | * * 0.70 | * * 0.65 | * * <- PEMBRO+CHEMO (steady separation) 0.60 | * * 0.55 | * * 0.50 | * * <- inflection window 0.45 | * * 0.40 | * * 0.35 | * * +----------------------- months --> 0 3 6 9 12 15 18 21 \ / \ / mo 9 inflection chemo PR plateau — chemo partial-responder cluster holds OS briefly before drifting downward ``` > **EDITORIAL HOOK.** The vertical guide at **month 9** marks a clear inflection: the chemo arm exhibits a brief OS plateau (PR cluster) that does **not** appear in the pembro+chemo arm. Above ~month 9, chemo harm resumes; pembro+chemo continues to separate. Headline HR averages these regimes and hides the responder biology — which is exactly what the post-ASCO label-expansion playbook wants surfaced. **Underlying KM coordinates (auditable).** | time_month | arm | survival_prob | n_at_risk | |-----------:|----------------|--------------:|----------:| | 9 | chemo | 0.553 | 148 | | 9 | pembro_chemo | 0.672 | 168 | | 10 | chemo | 0.531 | 140 | | 10 | pembro_chemo | 0.640 | 160 | | 12 | chemo | 0.497 | 124 | | 12 | pembro_chemo | 0.583 | 144 | | 18 | chemo | 0.319 | 76 | | 18 | pembro_chemo | 0.383 | 96 | | 24 | chemo | 0.181 | 44 | | 24 | pembro_chemo | 0.233 | 64 | (*Full 24-month curves for both arms in the companion CSV, 48 rows.*) --- ## Page 3 — Classifier Attribution (HyperCore, n=320) **The 12 features driving the responder/non-responder boundary in KEYNOTE-789, when the curve is decomposed at the inflection.** | feature | importance_score | direction | clf_class | |-------------------------------|-----------------:|----------------|-----------| | baseline_pd_l1_tps | 0.18 | responder | clf_001 | | prior_tki_interval_months | 0.14 | responder | clf_017 | | egfr_mutation_subtype | 0.12 | responder | clf_042 | | smoking_pack_years | 0.10 | non_responder | clf_088 | | n_prior_lines_of_therapy | 0.09 | non_responder | clf_103 | | baseline_ldh | 0.08 | non_responder | clf_124 | | neutrophil_lymphocyte_ratio | 0.07 | non_responder | clf_156 | | tumor_mutational_burden | 0.06 | responder | clf_189 | | liver_metastasis_present | 0.05 | non_responder | clf_211 | | baseline_albumin | 0.05 | responder | clf_247 | | ecog_status | 0.04 | non_responder | clf_289 | | age_at_enrollment | 0.02 | non_responder | clf_310 | **Why each feature matters.** - **baseline_pd_l1_tps (0.18)** — Even within a TPS ≥ 1% enrollee pool, the TPS gradient sets the largest single marginal on responder probability. HyperCore splits the TPS vector into 18 strata; KEYNOTE-789's pembro+chemo benefit amplifies above TPS 50%. - **prior_tki_interval_months (0.14)** — Patients with longer washout between TKI failure and trial entry score higher on the responder side. The classifier penalizes accelerated progression post-TKI. - **egfr_mutation_subtype (0.12)** — Specific EGFR exon profiles (e.g., exon 19 del) carry a stronger responder signature than others (e.g., L858R + T790M co-mutation). This is the CDx angle the post-ASCO playbook leans on. The remaining features (smoking history, prior therapy count, baseline LDH, NLR) are well-established prognosticators — the classifier confirms they are not the *driver* of the pembro+chemo signal in this trial. --- ## Page 4 — Decision Implications (forward-looking) **What the post-ASCO label-expansion playbook says next.** KEYNOTE-789 is a marginal headline but a clean structure: durable separation in a definable subgroup (high PD-L1 TPS, longer TKI washout, ex19del EGFR), and an inflection in the chemo arm that a refined companion diagnostic (CDx) can isolate. Three forward moves for the program owner: 1. **Revisit the responder subgroup for label expansion.** The PD-L1 TPS + EGFR subtype axis carries the 0.18 + 0.12 = 0.30 weight in the ensemble. That is a third of the decision signal — a CDx scope that's both narrower than salvage therapy and richer than PD-L1 alone. 2. **Reposition the chemo's month-9 plateau as a competing signal.** A biomarker-defined subgroup on the chemo arm holds OS briefly before harm resumes. That window is the alpha for a clean comparator in a follow-on design. 3. **Use this artifact as the conversation-opener, not the conclusion.** A Tier-1 audience buys time, not deliverables. Forward-looking deliverables (a validate-on-your-data pilot) close the loop. **Paid pilot — $25K–$100K, 6 weeks.** The same decomposition runs against *your* phase 2/3 publicly disclosed OS data, returns the inflection point, the classifier top-12, and the label-expansion narrative. Frame: not retrospective critique, but a forward-looking decision artifact for the next trial design / regulatory engagement. [Start the conversation here.](/#contact) --- *Source: KEYNOTE-789 publicly disclosed OS (NCT03515837); DiviScan classifier ensemble, n=320.*