|
|
--- |
|
|
title: Neuromorphic Molecular Constraint Solver |
|
|
emoji: 🧬 |
|
|
colorFrom: blue |
|
|
colorTo: green |
|
|
sdk: streamlit |
|
|
sdk_version: 1.27.2 |
|
|
app_file: app.py |
|
|
pinned: false |
|
|
--- |
|
|
|
|
|
## Neuromorphic Molecular Constraint Solver (Vibecoded by amateur, it is a mess) |
|
|
|
|
|
This is a demonstration of a novel approach to *de novo* molecular generation. Instead of using a traditional generative model (like a VAE or GAN), this system translates chemical rules into a large Boolean Satisfiability (3-SAT) problem and solves it using a custom solver inspired by neuromorphic computing principles. |
|
|
|
|
|
### How it Works |
|
|
|
|
|
The process involves two main stages: |
|
|
|
|
|
1. **Encoding**: User-defined chemical properties (molecular weight, number of aromatic rings, forbidden functional groups, minimum atom count) are compiled into a massive 3-SAT problem. This includes complex chemical intelligence like valence rules (e.g., Carbon must have 4 bonds) and graph connectivity, which are encoded using cardinality constraints. |
|
|
|
|
|
2. **Solving**: A memory-efficient, sparse solver inspired by P-KAS (Phase-Keyed Associative Storage) and Kuramoto oscillators finds a satisfying assignment for the tens of thousands of variables and clauses. This method finds a solution by relaxing into a stable state rather than through algorithmic search. |
|
|
|
|
|
The key advantage is **generation by construction**. The output is guaranteed to satisfy the hard constraints, leading to a very high validity rate. |
|
|
|
|
|
### How to Use the Demo |
|
|
|
|
|
1. Use the sidebar on the left to set your desired molecular properties. |
|
|
2. **Crucially, set a "Minimum atom count" greater than 0** to avoid trivial solutions like H₂O. A value of 10-15 is a good starting point. |
|
|
3. Click the "Generate Molecules" button. |
|
|
4. Be patient. The encoding and solving process for such a large constraint problem can take 10-30 seconds per molecule. |
|
|
|
|
|
### Limitations & Current Status |
|
|
|
|
|
This is a proof of concept and has several important limitations: |
|
|
|
|
|
* **Graph Generation, Not Full Chemistry**: The solver's primary output is a **structural graph** of atoms and their connections. It does not yet solve for bond orders (single, double, triple). |
|
|
* **Visualization**: The RDKit visualizer assumes all bonds are `SINGLE` for drawing purposes. This means that even if the solver finds a valid graph where an atom has the correct *number* of bonds, the drawing may appear chemically incorrect (e.g., a Carbon with four single bonds to two atoms). The atom labels (`ID:Element`) are provided to help you inspect the raw graph structure. |
|
|
* **Approximate Solver**: The neuromorphic solver is a heuristic method that aims for very high satisfaction (99%+). It is not a formal, complete SAT solver and may not find a perfect 100% solution for extremely difficult or unsatisfiable problems. |
