Madgen Mass Spec Attends Attends To De Novo Molecular Generation

MADGEN - MASS-SPEC ATTENDS TO DE NOVOMOLECULAR GENERATION

1 Introduction

Their contributions:

• two stage framework. Generates target molecule structure.
• Utilizes fragmentation patterns in the MS/MS spectra to help with prediction.
• The scaffold based approach improves interpretability, which allows better analysis of biological functions and chem properties.
• Broadly applies in metabolomics, drug discovery, and environmental analysis.

2 Related Work

3 Methods

3.1 Scaffold Retrieval

Similarity score: $h(z^n_{spec},z^m_{mol})$ $h(z^n_{spec}, z^m_{mol}) = exp\pmatrix{\frac{z^n_{spec}\cdot z^m_{mol}}{\vert{\vert{z^n_{spec}}}\vert\vert \vert\vert{z^m_{mol}}\vert\vert τ}}$ Where $z^n_{spec}$ and $z^m_{mol}$ are the embeddings of the spectrum scaffold. SPA: Scaffold prediction accuracy Oracle Retrieval: uses a lookup table that gets the correct scaffold based on the MS/MS spectrum and chemical formula.

3.2 Scaffold-Conditioned De Novo Molecule Generation with Spectra Guidance

3.2.1 Notations and Problems Formulation

A molecule is represented as $G$ a graph where $G = (V, E)$ . The scaffold is then $S = (V^S, E^S)$ which is a subgraph of $G$.

3.2.2 Scaffold-Conditioned Generation via Markov Bridge

$p(E\vert E^S, V^G) = \sum_{E_0:E_{T-1}} \prod_{t-0}^{T-1} p(E_{t+1} \vert E_t, E^S, V^G)$ where $E_0 = ∅$ is the case where no bonds are formed from isolated atoms to others, and $E_T = E$ is the sequence of random variables $E_{0:T}$ is the progressively connecting atoms to form the final molecules.

3.2.3 Classifier-Free Guidance From Mass Spectrum

CFG: Classifier Free Guidance

4 Experiments

4.1 Datasets

MCES: Maximum Common Edge Substructure

4.2 Experimental Setup and Evaluation Metrics

4.3 Results