Drug Discovery

In one sentence: The same engine that models consciousness discovers drug candidates — molecular fragments become “memes,” cultural evolution replaces random search, and each generation builds on the accumulated chemical knowledge of the previous one.

Theory sources: BM (meme dynamics), EMT (cultural accumulation), SM (cumulative cultural evolution, ratchet effect), AGI_F (multi-agent stigmergy)

Implementation: bmc-engine/bmc-discovery/ (Rust core + Python chemistry bridge + RDKit)

The Core Idea: One Engine, Any Domain

BMC claims its architecture is domain-agnostic: change the environment, and the same engine solves a different problem. Drug discovery is the strongest test of this claim.

No domain-specific architecture changes. The same code. Just a different environment and scoring function.

How It Works

The Agent Loop

Each BMC agent works like a molecular chemist with access to a shared library:

Four Ways to Modify Molecules

The Ratchet: Each Generation Builds on the Last

The critical innovation: agents don’t start from scratch each cycle. A shared knowledge store accumulates everything that worked:

• Fragment registry: Tracks which fragments have been tried and how they performed
• Co-success binding: Fragments that appear together in high-scoring molecules get linked (like memes that reinforce each other)
• Cultural library: Top-performing fragments are shared across all agents
• Domain isolation: Different target types get separate libraries (kinase fragments for kinase targets, GPCR fragments for GPCR targets)

This is the Super-Ratchet Effect — each generation builds on accumulated knowledge, and no information is lost between cycles.

Results

Tested on 4 standard pharmaceutical benchmarks (Therapeutics Data Commons):

The Ratchet in Action

Watch how scores improve as the cultural library grows:

The improvement accelerates as the library grows. DRD2 is the most dramatic: from 0.206 (barely above random) to 0.997 (near-perfect) in 10 cycles.

What These Numbers Mean

• QED 0.947: The discovered molecule scores near the theoretical maximum for drug-likeness properties (Lipinski rules, solubility, molecular weight, etc.)
• DRD2 0.997: 99.7% predicted binding to the dopamine D2 receptor — the agent learned to construct molecular scaffolds with near-perfect complementarity
• GSK3-beta 0.960: High predicted activity against a validated Alzheimer’s target

BMC Mechanisms at Work

Curiosity = Exploration

The SEEKING drive controls the mutation rate:

• Early cycles: High mutation (explore broadly across chemical space)
• Later cycles: Lower mutation (exploit known good scaffolds)
• Gap detection: Missing patterns in fragment co-occurrence trigger targeted exploration

Filter = Chemical Validity

Two-level quality control:

1. Can this molecule physically exist? (valence rules, ring stability, synthetic accessibility)
2. Does it score above the threshold? (oracle score check)

Fragments that fail repeatedly get downweighted — the system learns to avoid unproductive chemical motifs.

Care = Knowledge Consolidation

High-scoring fragments get reinforced:

• Fragments that co-occur in good molecules become linked
• Top assemblies are decomposed and their fragments shared with everyone
• The knowledge store persists between cycles — nothing is lost

BLEND = Fragment Crossover

The crossover strategy is BLEND: take sub-structures from two different high-scoring fragments and combine them into something new. The molecular version of “What if we combine feature A from drug X with feature B from drug Y?”

How It Compares

BMC Discovery is competitive with state-of-the-art molecular optimization methods (REINVENT, GraphGA, GFlowNets) on standard TDC oracle benchmarks — without any pretraining or training data.

The key insight: neural methods and BMC reach similar score ranges, but BMC does it through explicit cultural accumulation (Super-Ratchet) rather than gradient-based pretraining. The knowledge is inspectable, transferable, and compounds across projects.

What Makes This Different from Existing Approaches

The critical difference: BMC makes accumulated knowledge explicit and inspectable. You can look inside the library, see which fragments are working, and understand why a molecule was designed the way it was.

The Domain-Agnostic Test

The same codebase handles all 4 targets with only configuration changes:

One engine, four targets, competitive results on all of them. This is the BMC domain-agnostic principle in action.

Testable Predictions

Formalization

For readers interested in the mathematical treatment:

Super-Ratchet Effect:

Each cycle adds knowledge; nothing is lost between generations.

Domain-agnostic mapping:

Full formal treatment: SM Parts IV–V, BM Part IV, AGI_F Part IV.

Back to: Solutions Overview | Related: Financial Markets (same engine, different domain) | Theory: Swarm Memetics (cultural ratchet)