The scoring system answers a specific question: which drug–disease hypotheses are both well-evidenced through a bridge mechanism and genuinely unexplored?

This is not the same as finding total evidence (most papers mentioning both concepts), nor is it purely about finding balanced evidence (equal support on both sides). The goal is closer to:

• –Both sides must be credible. A bridge with 200 B↔A papers and 0 C↔A papers is not a lead — one side of the triangle doesn't exist.
• –The B–C connection must not already exist in print. If a drug–disease pair is already well-studied, it isn't undiscovered knowledge.
• –Evidence is subject to diminishing returns. The jump from 0 to 1 paper is enormous. The jump from 100 to 101 is nearly noise.

These three constraints shape every formula choice below.

Where p is the number of PubMed papers that already study the disease–compound pair directly (B–C). Verified via live PubMed query.

Why this shape

The log applies diminishing returns to the penalty for existing literature. A hypothesis with 1 existing paper is substantially less novel than one with 0. But the difference between 100 and 200 papers is almost irrelevant — both are well-studied. Using raw p in the denominator would treat 100 papers as 100× worse than 1 paper, which overcorrects.

The output is bounded (0, 1]. It reaches exactly 1.0 only when no existing B–C literature exists.

Behavior

Where B↔A is the number of distinct publications supporting the disease–bridge link and C↔A is the number supporting the compound–bridge link.

Why log-compress each side first

In v1.0 the log was applied after the geometric mean: ln(1 + √(B×C)). This accidentally rescued lopsided triangles — a bridge with B↔A = 2, C↔A = 100 produced √200 ≈ 14.1, then the log compressed that to roughly the same as a balanced (10, 10) case.

v1.1 applies ln to each side independently first, then takes the geometric mean of the compressed values. A single well-studied side cannot compensate for a weak other side. Zero-collapse is preserved: ln(1+0) = 0 collapses S to 0.

For symmetric cases (B↔A = C↔A), v1.0 and v1.1 produce identical output — the formulas are mathematically equivalent when both sides are equal.

Behavior

The lopsided case (2 / 100) now scores ≈ 0.69, below the balanced 10/10 (≈ 0.71) as intended. In v1.0 it incorrectly scored ≈ 0.73 — higher than 10/10.

A simple equal-weighted average of Novelty and Strength. Both are bounded to [0, 1], so H is also bounded to [0, 1].

Why equal weights

The 0.5 / 0.5 split is a deliberate starting point, not a calibrated value. It encodes the view that being undiscovered matters as much as being well-evidenced. A hypothesis that scores perfectly on strength but has 500 existing papers is not novel discovery — it's a database confirmation. The weights have not yet been tuned against real-world validated discoveries.

What ranks highest

• –B–C pair with zero existing literature (N = 1.0) and strong bilateral bridge evidence
• –Bridges where both the disease side and the compound side have meaningful, independent publication support
• –Compounds not yet associated with the disease in any published study

1. v1.1June 5, 2026current

   Strength formula: log compression moved inside the geometric mean

   • –Strength: 1 − 1 / (1 + √(ln(1+B↔A) × ln(1+C↔A))) — log-compress each side first, then geometric mean
   • –Fix: lopsided triangles (e.g. B↔A=2, C↔A=100) now score correctly below a balanced equivalent
   • –Symmetric cases: output identical to v1.0 — only lopsided hypotheses are affected
   • –Novelty and Combined formulas unchanged
2. v1.0May 27, 2026

   Initial production scoring system

   • –Novelty: 1 / (1 + ln(1 + p)) — log-compressed inverse of B–C paper count
   • –Strength: 1 − 1 / (1 + ln(1 + √(B↔A × C↔A))) — log applied after geometric mean (known limitation)
   • –Combined: 0.5 × Novelty + 0.5 × Strength — equal-weighted linear combination