Signal Scanner

Signal Scanner — Reading the Second Derivative of Scientific Mindshare

By the time a target or modality is consensus — covered in Nature Reviews, multiple Phase 2 readouts, generalist-press coverage — the asymmetric return is gone. Edge comes from reading the leading indicators, and within each indicator, from reading the acceleration rather than the level. A field with 200 papers/year growing 60% YoY is a far stronger signal than a field with 2,000 papers/year growing 5% YoY. The second derivative of publication and preprint volume leads aggregate counts by 1-3 years and leads first-in-human by roughly 4-7 years for a novel target — it is the single most actionable literature signal.

This skill is the velocity primitive of the frontier-intelligence suite. It takes a target/modality query and emits, per entity, a velocity + acceleration + breadth profile across a layered set of sources ordered by lead time: deepest and slowest (NIH grants, primary literature), through the middle (preprints, citation bursts, patents), to the translation-event ground truth (first IND). It does not score attention or hype — that is mindshare-tracker's job. It quantifies data-generation velocity and de-noises it.

The discipline that separates signal from noise is three filters applied before any number is trusted: strip review/perspective publication types (reviews are a trailing maturity marker, not a leading one); compute an author-affiliation Herfindahl concentration index to subtract single-lab artifacts (over 50% of preclinical findings are irreproducible, $28B/yr in the US, and single-lab standardized studies systematically fail to predict effect size); and normalize against all-PubMed baseline growth (4%/yr) so that not every field looks like it is growing.

How to Run

Input

Steps

Step 1 — Build the Query Set

Define the entity precisely before counting anything. A sloppy query set is the dominant source of false velocity.

• Assemble the target/modality string plus synonyms and MeSH terms (e.g., for degraders: PROTAC OR "targeted protein degradation" OR "molecular glue" OR "degrader").
• Watch for renaming/MeSH drift — the same biology rebranded inflates apparent novelty. When using the PubMed tool, capture the expanded query_translation so the MeSH mapping is auditable.
• For narrow/novel targets (rare-disease, single-gene) note that absolute denominators are small, so acceleration will be noisy and a single lab or grant can dominate — flag for higher breadth-filter sensitivity downstream.

Step 2 — Publication Velocity and Acceleration (PubMed)

Primary literature is the workhorse signal. Source: PubMed E-utilities (the connected search_articles MCP tool returns an exact total_count per query + date window — the atomic operation of the whole method; e.g., a 2024-restricted PROTAC query returns 658).

LITERATURE VELOCITY — PubMed

For each rolling 12-month (or quarterly) window over the last ~8 years:
  N(t) = total_count for [query set] in window t
         WITH "Review"[Publication Type], Editorial, Comment EXCLUDED

  velocity(t)      = ΔN / Δt                 (papers per year)
  acceleration(t)  = Δvelocity / Δt          (the LEAD signal — 2nd derivative)
  RGI(t)           = field CAGR / all-PubMed CAGR   (~4%/yr baseline)

Flag inflection when: RGI > ~3 for 3 consecutive years
                      AND acceleration is positive and rising
                      AND primary-research fraction is rising (not review-driven)

Track the primary-research fraction separately: a healthy emerging field shows rising original-article velocity; a hype field shows reviews outrunning primary data.

Step 3 — Preprint Surge (bioRxiv / medRxiv)

The earliest formal-literature signal — preprints lead peer-reviewed publication by 6-18 months. Source: bioRxiv/medRxiv API (search_preprints, search_published_preprints, funder_search, get_statistics).

• Run the same rolling-window velocity/acceleration on preprint counts.
• Compute the preprint-to-publication conversion rate (search_published_preprints): fraction reaching a journal within 12-18 months. Rising conversion = the field is maturing and being validated; falling conversion alongside rising preprint volume = low-quality flooding (a warning).
• Track preprint citation velocity — preprinted papers receive their first citation ~1 month earlier and ~0.5x more citations on average; time-to-first-citation is a within-field momentum sub-metric.
• Preprints are unreviewed: the breadth filter (Step 6) is essential here. Verify subject category with get_categories.

Step 4 — Citation Burst Detection on Landmark Papers

A single landmark mechanism paper (Nature/Cell/Science) that triggers a citation burst is a higher-quality signal than diffuse incremental growth. Source: OpenAlex (free citation graph), Dimensions, Semantic Scholar; Web of Science/Scopus where licensed.

• Run Kleinberg burst detection (the CiteSpace-style algorithm) over the field's citation graph: a burst is a sudden, statistically significant jump in citation rate for a paper or cluster. Direct citation detects young, fast-growing clusters earlier than co-citation or bibliographic coupling.
• Identify landmark papers — the small core that the citing literature co-cites (the formal definition of a "research front").
• Use field-normalized citation acceleration (control for baseline field citation rates), not raw counts.
• Read polarity: discount citation-farming/self-citation rings; a paper cited because it is being refuted is a negative signal masquerading as positive.

Step 5 — Patent Velocity (PCT/WO) — Mid-Stage Confirmation

Forward-looking but with a structural 18-month visibility penalty, so this is a mid-stage confirmation signal, not an earliest-mover one. Source: Lens.org (free, strong biotech coverage), PatentsView (USPTO), Espacenet (EPO), Google Patents, WIPO PATENTSCOPE for PCT/WO.

• Count PCT (WO) applications and priority filings per quarter mentioning the target/modality (composition-of-matter and method-of-use claims).
• Critical caveat: applications publish ~18 months after the priority filing date. Velocity on publication date lags reality by 18 months; the true leading metric is velocity on priority date (visible only retrospectively). Always back-date to priority. Patents become public 18-36 months before any product — they front-run commercialization but trail the earliest literature.
• Discount defensive/blocking filings with no development intent, and a single incumbent's broad portfolio spray (breadth filter again — many independent filers > one big-pharma sprayer).

Step 6 — Author/Institution Breadth (Herfindahl De-Noising Filter)

The most important de-noiser. N papers from 2 labs ≠ N papers from 40 independent groups.

BREADTH FILTER — author-affiliation Herfindahl

For the primary-research + preprint corpus of the field:
  s_i = share of papers from institution/group i
  HHI = Σ (s_i)^2          (0 → maximally broad; 1 → single lab)

If HHI is in the top decile (single-lab dominated):
  → CAP the literature-derived sub-scores (Steps 2, 3, 4) regardless of raw acceleration
  → genuine emergence is broad-based; concentration = discount the signal

ORCID coverage is incomplete in older literature, so affiliation disambiguation is imperfect — when breadth is uncertain, treat it as a flag, not a hard gate.

Step 7 — NIH / Funder Grant Velocity — Deepest, Most-Lagging

The deepest leading indicator (~5-7 years to clinic) but the lowest velocity and most lagging to genuinely novel biology, because study sections are conservative and fund incrementally. Source: NIH RePORTER API (project counts, $ awarded, by RCDC category and free-text); plus CZI, Wellcome, ERC, ARPA-H portfolios.

• Compute new-grant velocity (new awards/year, excluding renewals — renewals lag) and total $ committed per year in the theme.
• Read this as confirmation that academic mindshare has institutionalized, not as a frontier signal.
• Check whether a spike is RFA/program-announcement-driven (an artifact) versus organic.

Step 8 — Translation-Event Cross-Check (ClinicalTrials.gov v2)

The ground truth that closes the lead-time measurement. Source: ClinicalTrials.gov v2 API (search_trials, search_by_sponsor).

• Pull the first IND / Phase 1 entry date for the target/modality. This is the moment mindshare converts to clinical reality.
• The gap between literature-velocity inflection and first-in-human is the realized lead time the whole framework exists to capture (typically 3-7 years for a novel target).
• Interpretation: a strong acceleration signal with no clinical entries = maximal asymmetry (you are early); a strong signal with multiple Phase 2 readouts = you are at consensus (the literature edge is gone).

Output

SIGNAL SCAN — [Target / Modality]
Query set: [string + synonyms + MeSH]    |    Window: [years]    |    Date: [scan date]

LITERATURE (PubMed, reviews stripped)
  Latest annual count (primary):  [N]
  Velocity (papers/yr):           [v]
  Acceleration (2nd derivative):  [a]   ← LEAD SIGNAL
  Relative Growth Index (RGI):    [x]   (>3 for 3yr = inflection)
  Primary-research fraction:      [%]   (rising / falling)

PREPRINTS (bioRxiv/medRxiv)
  Velocity / acceleration:        [v] / [a]
  Preprint→publication conversion:[%]   (rising = validating; falling = flooding)
  Time-to-first-citation:         [months]

CITATION BURST
  Landmark paper(s):              [citation + burst window]
  Field-normalized acceleration:  [x]
  Polarity check:                 [supportive / refutational / mixed]

PATENTS (PCT/WO, priority-date adjusted)
  Priority-date filing velocity:  [v]   (mid-stage confirmation; −18mo visibility)
  Filer breadth:                  [many independent / single incumbent]

GRANTS (NIH RePORTER)
  New-grant velocity:             [awards/yr]   (deepest, ~5-7yr lead, lagging-to-novelty)
  RFA-driven?                     [yes/no]

DE-NOISING
  Author-affiliation HHI:         [0-1]   (top decile → literature sub-scores CAPPED)

TRANSLATION GROUND TRUTH (ClinicalTrials.gov v2)
  First IND / Phase 1 date:       [date or NONE]
  Realized lead time (lit-inflection → FIH): [years]

READ
  Inflection?                     [Yes / Ambiguous / No]
  How early are you?              [pre-clinical-entry / first INDs filed / at consensus]
  Confidence flags:               [single-lab artifact / review-inflated / RFA artifact / narrow-denominator noise]

Error Handling

Cross-Domain Connections

This skill is deliberately dual-use, serving two readers from the same scan:

• For the clinical scientist learning a field: the scan is a map of where the data is being generated fastest and most broadly — which preprints to read first, which landmark paper ignited the front, whether the evidence base is broad or one-lab-deep, and how far the biology is from the clinic. It tells you where the frontier of a field actually is, before the review articles catch up.

• For the investor screening early opportunities: the scan is a timing instrument — the realized lead time (literature inflection → first IND) is the asymmetry window, and the de-noising filters (review-stripping, Herfindahl breadth, polarity) are the anti-hype discipline that separates a genuine emerging field from capital-manufactured froth.

• frontier-intelligence/mindshare-tracker (sibling): consumes signal-scanner's velocity/acceleration/breadth and adds the attention layer — conference share-of-voice (AACR/ASCO/ASH/ESMO/JPM), Altmetric as a high-noise tripwire, and the 0-100 reflexivity-aware momentum score with a Consensus evidence-maturity gate. Signal-scanner measures data generation; mindshare-tracker measures attention and reflexive froth.

• frontier-intelligence/emerging-target-radar (sibling): the integrator that fuses signal-scanner velocity with mindshare attention and data-engine convergence into a ranked watchlist of target × modality candidates.

• competitive-intelligence/pipeline-mapper (depends_on): the first-IND/Phase-1 translation event surfaced in Step 8 is the seed for a full pipeline map — once a target inflects and enters the clinic, pipeline-mapper enumerates the competitive set, stages, and timelines.

• research/spelunker (depends_on): supplies the deep-source-spelunking and confidence-vocabulary discipline behind the de-noising layer (polarity reads, breadth assessment, distinguishing primary data from echo) — signal-scanner is its quantitative, target/modality-scoped specialization.

• probability-of-success/pos-base-rates: a target that clears the inflection + breadth bar and reaches first-in-human feeds the PoS pipeline; the realized lead time and biomarker/genetics context inform which base-rate row applies.