Dual-targeting siRNA landscape — full live discovery test (rerun-2, 2026-06-29)¶

What this is. The product's real Explore use case: give the system one sentence, let it decompose → query live sources → discover companies and assets it was never told about → iterate → score against the Sanofi deck (32 assets / 12 deals / ~36 companies). This is the "how much can the system map from a single sentence" test. Thesis: "Map the competitive landscape of dual-targeting siRNA therapeutics for cardiometabolic disease."

TL;DR¶

Headline (full live, +CDE, pinned scope): company recall 19/36 (53%), asset recall 17/32 (53%), deal recall 2/12. Beats the closed-world asset baseline (15/32) cold, from one sentence.
CDE is the single biggest lever. Turning the China-registry bridge on lifts company recall 12/36 → 19/36 (+7) and asset recall 11/32 → 17/32 (+6). The +7 are exactly the China names the deck is built on: Argonaute RNA, CSPC, Hansoh, Innovent, Jiangsu Hengrui, Merck, Visirna.
Per-source attribution (the deliverable): of the 19 recalled deck companies, first-surfaced by lens 6 / cde 6 / clinicaltrials 5 / openalex 2. Company discovery volume: lens 148, sec 136, openalex 37, cde 32, clinicaltrials 19, hkex 13, anzctr 1.
ANZCTR (newly wired this run) contributes 1 unique net-new lead — OliX / OLX75016, an Australian-registered dual siRNA CT.gov did not surface. Real but modest (ANZCTR is a small registry).
Precision is ~5% — and that is expected and reframed: of 342 novel (non-deck) surfacings, 21 core / 56 honorable-mention / 265 off-thesis. The core + honorable mentions (Arbutus, Eddingpharm, Salubris, a deep China long-tail) are intelligence the deck missed, flagged for verification — not noise.
Patent backend = Lens, not gpatents. gpatents auth works but the BigQuery free-tier monthly byte quota (1 TiB) is exhausted (~229 GiB/query, project unbilled) — a different cliff than the rate limit gpatents was meant to solve. Lens carried the patent layer (China applicants included).

1. Headline numbers — three arms¶

All arms: live sources, rounds 3, Lens patent backend, shared result cache (cleared first, so Arm 1 is cold; Arms ⅔ reuse Arm 1's successes — making the CDE delta a clean difference).

metric	Arm 1 — baseline (`--no-cde`)	Arm 2 — full (+CDE)	Arm 3 — LLM decompose (+CDE)	closed-world baseline
company recall	12/36 (33%)	19/36 (53%)	17/36 (47%)	—
asset recall (sponsor)	11/32 (34%)	17/32 (53%)	17/32 (53%)	15/32 (47%)
asset recall (code-level)	7/32 (22%)	8/32 (25%)	6/32 (19%)	—
deal recall	1/12 (8%)	2/12 (17%)	1/12 (8%)	11/12 (92%)
company precision	5% (12/244)	5% (19/361)	4% (17/406)	—
novel scope-fit (core/hm/off)	21 / 47 / 164	21 / 56 / 265	27 / 69 / 293	—

Arm 2 (full, pinned --no-llm scope) is the headline. The LLM-decompose arm (Arm 3) under-recalls companies (17 vs 19) because the live LLM decomposition dropped the CFB target and over-expanded into NASH/metabolic targets (PNPLA3, HSD17B13, TTR) — diluting the cardiometabolic-core search. This is the concrete argument for pinning scope on the reproducible arms.

2. The CDE delta — the China-registry contribution, isolated¶

Arm 1 (no CDE) vs Arm 2 (+CDE), non-CDE results held identical by the shared cache:

Company recall 12 → 19 (+7). The 7 CDE adds: Argonaute RNA, CSPC Pharma, Hansoh Pharma, Innovent Biologics, Jiangsu Hengrui Pharma, Merck, Visirna.
Asset recall (sponsor) 11 → 17 (+6).
Deal recall 1 → 2 (CDE-discovered sponsors feed the deal scorer).

CDE works as a concept→sponsor bridge: an indication search (高胆固醇血症 / 血脂异常 / 高脂血症 ∩ chemical drugs) returns a sponsor roster; each raw 申请人 is a newly-discovered Chinese company resolved downstream. This is the mechanism that turns "one sentence" into the China half of the landscape.

3. Per-source coverage + entity discovery (the centerpiece)¶

3a. Company discovery by source — full arm (deck vs net-new)¶

source	companies discovered	on-deck	net-new
lens (patents)	148	7	141
sec	136	4	132
openalex	37	4	33
cde	32	9	23
clinicaltrials	19	7	12
cn_filings (hkex)	13	1	12
anzctr	1	0	1

Lens/SEC discover the most volume (mostly net-new China applicants + filers — the precision flood the scope-fit buckets reframe); CDE discovers the most deck companies (9) for the least volume — the highest-signal source for this landscape.

3b. First-surfacing source — which source first surfaced each of the 19 recalled deck companies¶

source	deck companies first-surfaced
lens	6
cde	6
clinicaltrials	5
openalex	2

(sec / cn_filings / anzctr contributed net-new + deal co-occurrence but did not first-surface a deck company — others reached those names first.)

3c. ANZCTR isolation (user focus) — what it surfaces that clinicaltrials does not¶

ANZCTR was not previously wired into the discovery eval; it was added this run (commit d1b7aea). Wired as a concept→registry bridge over WHO ICTRP (ANZCTR-origin ACTRN only).

Uniquely ANZCTR (in ANZCTR, not in clinicaltrials): 1 company — OliX / OLX75016 (ACTRN12624000023550), a chemically-synthesised dual siRNA, labelled core in-thesis, registered 2024-01. CT.gov does not carry it; ANZCTR does.
Deck assets uniquely via ANZCTR: 0. The deck's Australian-registered asset (Sanegene SGB-BS01) was already first-surfaced by clinicaltrials, so ANZCTR confirmed rather than uniquely-recovered it.
Honest read: ANZCTR adds a real Southern-Hemisphere lead the other registries miss, but its volume is small — only ~1 ACTRN siRNA trial matches a modality search (the registry is small, and ICTRP basic search is a phrase match over condition+title, so concept queries fire single modality tokens, not gene symbols). It is a recall bonus, not a primary engine, for this landscape.

3d. Per-source timing (full arm, §0c — wall is cumulative across concurrency, not real elapsed)¶

Proves every source ran live and the cache works (cached re-runs collapse to ~0). CDE = 3 calls (the indication bridge + per-code), anzctr/sec/clinicaltrials are the tall poles (per-entity round-2 fan-out).

3e. Signals per source (live corpus composition)¶

Full arm harvested 1464 backing signals. Per-source company-discovery volume (3a) is the live analogue of corpus composition; exact per-source raw-signal counts would need a one-line run_meta instrument (tracked follow-up) — the call counts in §0c and the discovery volumes in 3a are the proxy.

4. Scope-fit + honorable mentions — players the deck may have missed¶

Flat precision (~5%) treats every non-deck surfacing as drag. It isn't. Of 342 novel surfacings in the full arm: 21 core / 56 honorable-mention / 265 off-thesis. The leads (each flagged, needs primary-source verification — never asserted as a confirmed deck-miss):

entity	scope-fit	targeting	first surfaced	source(s)
Arbutus Biopharma	core	dual siRNA	2023-07	lens
Dicerna	core	dual siRNA	2024-06	openalex, sec
Olix / OLX75016	core	dual siRNA	2024-01	anzctr
Shanghai Tuojie (Argo)	core	dual siRNA	2025-09	hkex, lens
Beijing Youcare Kechuang	core	dual siRNA	2026-04	hkex, lens
Shanghai Rona Therapeutics	core	dual siRNA	2026-01	clinicaltrials, lens
Eddingpharm	core	dual siRNA	2026-03	clinicaltrials
Guangzhou Ribobio / Bebetter / Sunshine Lake / Agna / Leaderna …	core	dual siRNA	2025–26	lens
Shenzhen Salubris	honorable-mention	single siRNA	2024-06	lens

The named deck-misses the analysts flagged (Arbutus, Eddingpharm, Salubris) all surfaced, plus a deep China long-tail. Caveat: the honorable-mention bucket contains some academic affiliations (NYU Langone, Ospedale Maggiore, university hospitals) that slipped the research-org filter — the commercial leads are the biotech/pharma rows; the academic rows are literature noise to be filtered in a pass.

5. Deals¶

Full arm recovered 2/12: D04 (Novartis × Argo Biopharma; same-signal co-occurrence) and D07 (Regeneron × Hansoh Pharma; asset-code match). Deal recall stays the hard problem live (a deal lives in one party's filing; cross-filing co-occurrence + body-fetch recover only the best-documented ones). This is consistent with the prior cold-live finding (0 cold / ~3 probe-confirmed) and far below the closed-world 11/12 (the seed corpus carries the wire/SEC text the live sweep only partially reaches).

6. Issues found (Phase 0 + run)¶

ANZCTR / wires / gdelt were not wired into the discovery eval despite existing on main. ANZCTR wired this run (d1b7aea); wires/gdelt remain unwired (out of scope per decision).
Two silent-zero dependency traps: CDE needs uv sync --extra visual + playwright install chromium; gpatents needs uv sync --extra gpatents. Without them both sources return empty without error — the exact failure that invalidated the prior "definitive" run.
gpatents free-tier byte quota. ~229 GiB/query × per-concept-query fan-out blows the 1 TiB/month free tier (project unbilled → 403). Resolved by running on Lens (valid ~until 07-09). To use gpatents durably: attach GCP billing (~$5/TiB) or batch the eval to one scan/round.
LLM decompose dropped the CFB target and over-expanded into NASH targets → the LLM arm under-recalls vs the pinned scope. Pin scope (--no-llm) for reproducible recall.
Throttle is partial, not zeroing. Across the 3 arms: OpenAlex 8 / Lens 16 query failures on cold bursts (Lens 10/min trial cap, OpenAlex polite pool). Sources still contributed; the shared cache recovers gaps on the next arm. CDE never skipped; no dep errors.

7. Reproducibility¶

commit: d1b7aea (ANZCTR wiring) on branch eval/anzctr-discovery-wiring, off main 6298bed.
scope (pinned arms): modalities siRNA/RNAi; subclass dual-targeting/divalent/bispecific/bivalent; targets PCSK9, ANGPTL3, APOC3, LPA, AGT, CFB; 13 cardiometabolic indications; horizon preclinical+.

commands:

A=scripts/eval/sirna_autonomous_discovery.py ; T="Map the competitive landscape of dual-targeting siRNA therapeutics for cardiometabolic disease."
rm -rf archived_data/.eval_cache
uv run python $A "$T" --corpus live --no-cde --no-llm --rounds 3 --patents-backend lens --out-dir archived_data/live_rerun_baseline
uv run python $A "$T" --corpus live          --no-llm --rounds 3 --patents-backend lens --out-dir archived_data/live_rerun_full
uv run python $A "$T" --corpus live                   --rounds 3 --patents-backend lens --out-dir archived_data/live_rerun_llm

artifacts: per arm — sirna_discovered_companies_*.csv, sirna_discovered_assets_*.csv, sirna_gap_analysis_*.md; per-source analysis archived_data/rerun2_per_source_analysis.md.