COMPASS
Quick actions
Impact Factor Distribution (JCR)
Recent reports
🔬 COMPASS Keywords
Click to select · Type above
| # | Title ↕ | PMID | Journal | Year ↕ | IF ↕ | Status |
|---|
Papers PharosDB
| # | Title | PMID | Journal | IF | Cit. | Topics | Type | Pub Date | Source | Score | OA |
|---|
BioMind 生物智研
Research Trends
📈 Publication Volume by Year
📊 Avg Impact Factor by Year
🔬 Publications by Topic
⭐ IF Distribution Trend
🏆 Top High-IF Journals (2024-2026)
| # | Journal | Papers | Avg IF |
|---|
🔍 Cross-Topic Intersections
🌱 Emerging Keywords (2025-2026)
📈 Topic Growth & Gap Analysis
Author & Institution Network
👤 Top Authors
🏛 Top Institutions
🔗 Co-Authorship Network (Top Authors)
Research Teams & Rising Stars
Knowledge Explorer
Cross-dimensional literature mining. Filter by one dimension, see relationships to another. Click any result to drill down to papers.
Disease × Pipeline Stage
Reports
Subscribe & custom search
Email subscription
Subscriber overview
Check / Unsubscribe
Settings
API services powering this platform — and how to bring your own keys.
Service Status
Use Your Own API Key
By default, this platform uses server-configured keys. Keys are stored per-session, never persisted.
Apply & Test will verify your key. If it fails, the platform falls back to the server default.
Usage
| Service | When Used |
|---|---|
| LLM API | Report generation, Q&A AI |
| NCBI E-utilities | Paper search & metadata |
| Semantic Scholar | Citation tracking |
| Embedding Model | Knowledge base indexing (local) |
| SMTP Email | Report delivery (server-managed) |
SMTP email uses the platform's server-configured account — users cannot override this.
Help & Guide
What is this platform?
COMPASS (COMPutational biology Agent for Pharmaceutical Sciences) — an automated pipeline that continuously monitors the intersection of computational biology and pharmaceutical R&D. It searches PubMed daily, downloads papers, builds a searchable knowledge base, generates bilingual review reports, and delivers them to subscribers via email.
Built on PharosDB (structured DB) and AetherBase (vector KB). Full statistics report →
Page Guide
Dashboard
Global overview: paper trends, hot topics, recent reports, and an AI-powered Ask KB Q&A panel.
Papers
Browse and search all 69,000+ papers in PharosDB. Filter by year, impact factor, topic, or free text. Click any paper for full details.
Knowledge Explorer
Cross-dimensional mining: pick two dimensions (e.g. Disease × Method) to see distribution, sparklines, and heatmaps. Click any cell to drill into papers.
Network
Interactive co-authorship graph showing top researchers and their collaborations. Filter by topic, click nodes to explore connections.
Teams
Top 30 research teams globally, detected via co-authorship community analysis. Rising stars, research series, and team-level collaboration networks.
Reports
AI-generated bilingual review reports. Filter by type (weekly/monthly/trends), preview in-browser, download as .md/.html/.pdf.
Subscribe
Sign up for weekly or monthly email reports. Check your subscription status or unsubscribe anytime. No email addresses are publicly visible.
Settings
View platform service status. Bring your own LLM or NCBI API key — apply & test, and the key will only be used for your session.
📊 Ranking & Scoring (Network)
Top Authors and Top Institutions are ranked by a composite score that weighs the author's role in each paper, together with the paper's impact factor and its citation count.
Contribution = Weight × (1 + 0.1 × IF) × (1 + 0.5 × log₁₀(1 + Citations))
Co-author papers (weight 0.3, no citation boost):
Contribution = Weight × (1 + 0.1 × IF)
Role Weight:
• First author / corresponding author — 1.0 (lead)
• Co-first author — 0.8 (lead)
• Other co-author — 0.3 (no citation boost)
IF = journal impact factor of the paper.
Citations = total citations, log-scaled (log₁₀) to dampen outliers. The citation boost applies only to lead papers — citations reflect the contribution of the leading authors, so a 0-lead co-author cannot outrank a first/corresponding author via high citations alone.
Author disambiguation & quality control. Authors are merged across name variants and ORCIDs; institutions are normalized to parent entities (e.g. Harvard-affiliated hospitals → Harvard University; Max Planck institutes → Max Planck Society) so scores are not diluted by fragmentation. Papers classified as News / Interview / Editorial are excluded from ranking, so journal staff writers do not appear as researchers. An author's primary institution follows the same platform-wide rule as Teams (see below): ORCID-first, then the mode of the first-listed affiliation across research papers from the last two calendar years.
Topic view. When a domain (topic) filter is active, the Score is re-computed from the author's/institution's papers in that domain only (same formula above, over that topic's papers), so the same author shows a different score in each topic, and the list is ranked by that domain score (descending). Authors with no lead contribution in the domain get score 0 and sink. The all-topics view shows the overall composite score.
🏛 Team Scoring (Research Teams)
Teams are detected from the co-authorship graph with a two-level split: Louvain communities first, then large communities (≥50 members, institution-level collaboration circles) are re-split into PI-centric research groups. Each team is ranked by its composite score, which rewards papers where the team holds lead authorship and weights in the journal impact factor.
Paper Weight:
• Paper has a team lead (first / corresponding author among team members) — 1.0
• Team members appear only as regular co-authors — 0.3
IF = journal impact factor of the paper (0 when unknown).
Lead papers = papers where at least one team member holds first or corresponding authorship.
Total papers = all papers involving any team member.
Team discovery (two-level split). ① Level 1 — run Louvain community detection on the co-authorship graph (edges = ≥2 shared papers), keep communities with ≥3 members and ≥5 papers. ② Level 2 — for each large community (≥50 members), split it into PI-centric groups: PI candidates are members with ≥5 lead papers (first / corresponding); each other member joins the PI with whom they share the most co-authored papers, requiring ≥3 shared papers; members without a ≥3-paper link merge into the nearest PI (highest co-authorship, fallback = PI with most lead papers); groups with fewer than 3 members after splitting are dropped; communities with no clear PI core are kept whole. Teams are named "<Lead Author> Group". All qualified groups are persisted to the research_teams table in papers.db (full 1,900+ groups), while this page displays the top 30 by Team Score.
Teams are displayed in descending Team Score order. Under a domain (topic) filter, the Team Score is re-computed over that domain's papers only (same formula), paper / lead counts are rewritten to domain-level numbers, and teams are ranked by the domain score. Teams with no papers in the domain are hidden. Note: unlike the author score, the team score does not yet include a citation term.
Author institution labeling (作者单位标注规则). ① Authors with an ORCID — institution is resolved from the papers under that ORCID identity (the platform stores no separate ORCID registry, so the ORCID's own research papers are the authoritative source, merging across name variants); ② Authors without affiliation info — institution comes from their research-paper affiliations; ③ When multiple institutions exist — take the mode (众数) of the first-listed affiliation across papers published in the last two calendar years, falling back to all papers when no papers fall in that window; ties are broken by the most recent year, then alphabetically. This prevents a single recent joint affiliation (e.g. a co-supervised paper) from overriding an author's primary institution.
Bring Your Own Key (BYOK)
On the Settings page, you can enter your own API keys for:
- LLM API — enter provider, model, and key. Click Apply & Test. If the key works, your LLM calls will use your account. If it fails, the platform falls back to the server default.
- NCBI API — increases PubMed search speed from 3 req/s to 10 req/s. Free from NCBI account settings.
Keys are stored per-browser-session and never persisted to disk. They are cleared when you close the browser or click Clear.
Automation Schedule
| Task | Schedule | What |
|---|---|---|
| Citation Tracking | Mon 6:00 AM | Update citation counts for tracked papers |
| Author Extraction + Fulltext KB | Mon 6:30 AM | Extract authors, rebuild knowledge base |
| Weekly Pipeline (Windows Task Scheduler) | Mon 7:00 AM | Search new papers + generate bilingual review report + send to weekly subscribers (search → report → email chained in one task) |
| Team Discovery & Rising Stars | Mon 7:45 AM | Recompute teams / rising stars / research series |
| Monthly Report Generate | 1st Mon 7:30 AM | 30-day search + trend analysis + monthly report |
| Monthly Email | paused | Monthly report delivery is paused — send manually if needed |
Rising Stars Detection
Rising Stars surface new researchers and emerging teams. Candidates are drawn from 14,000+ disambiguated authors using multi-dimensional growth metrics, with a deliberate emphasis on true first authorship and recent career start rather than senior corresponding-author status. Only original research papers are counted — Review / Editorial / News / Interview / Comment / Letter / Guideline / Case Report / Meta-Analysis / Retracted papers are excluded (matched on raw PubMed types, since PubMed often tags a review as both "Journal Article" and "Review"), so journal staff writers, editors and review-heavy authors do not appear as researchers. Top 50 are selected each run.
Authorship Credit (per paper)
| Role | Weight |
|---|---|
| True first author (position 1) | 1.0 — full credit |
| True last/senior corresponding (corresponding & last-listed) | 1.0 — full credit |
| Co-first author (shared first) | 0.4 — down-weighted |
| Co-corresponding (corresponding but not last-listed) | 0.4 — down-weighted |
Weighted lead score = 1.0×(first + last-corresponding) + 0.4×(co-first + co-corresponding). Co-first / co-corresponding roles no longer inflate a researcher's standing.
Admission Thresholds
| Criterion | Threshold |
|---|---|
| Minimum papers | ≥ 3 |
| Average IF | ≥ 2 |
| Active years | ≥ 2 distinct years |
| Recent activity | ≥ 1 paper in 2025-2026 |
| Recent IF | ≥ 5 |
| True lead authorship | ≥ 1 first-authored or last-corresponding paper (co-first/co-corresponding alone does not qualify) |
Growth Scoring
| Metric | Formula |
|---|---|
| Paper growth | recent (2024+) ≥ early (<2024) × 0.5 |
| IF trajectory | recent_avg_if ≥ early_avg_if × 0.8 |
| Growth rate | recent_papers / max(early_papers, 1) |
Ranking Priority (descending)
- True first-author papers (emphasis on hands-on new researchers)
- Growth rate (faster rising = higher)
- Career start recency (later first publication = newer)
- Weighted lead score (co-first / co-corresponding already down-weighted)
- Recent average IF (quality gate, not the primary driver)
Paper Relevance
Every paper is scored on how strongly it connects computational biology methods to drug discovery applications.
| Rule | Score |
|---|---|
| Keyword match in title | +2 each |
| Keyword match in abstract | +1 each |
| Cross-domain: method + industry keywords both present | +2 ~ +5 |
| Open Access paper | +1 |
Filter threshold: ≥2. Score 0 papers are excluded from the knowledge base.