AI-Guided Design of\nMALDI Matrices: Exploring the\nElectron Transfer Chemical Space for Mass Spectrometric Analysis of\nLow-Molecular-Weight Compounds

The development of matrices for Matrix-Assisted Laser\nDesorption/Ionization\nMass Spectrometry (MALDI MS) has traditionally relied on experimental\nefforts. Here, we propose a Goal-Directed artificial intelligence\ngenerative model, fueled by computational chemistry calculated data,\nto construct a chemical space optimized for Electron Transfer (ET)\nprocesses in MALDI analysis. We utilized a group of 30 reported ET\nmatrices, subjected to structural enumeration and molecular properties\nprediction using semiempirical and <i>ab initio</i> calculations,\nto establish a comprehensive database comprising diverse structural\nand property data. Subsequently, employing a protocol of structural\nenumeration with 68 canonical SMILES of Bemis-Murcko (BM) fragments,\nwe expanded the structural complexity of the initial library. This\nprocess generated 82753 compounds organized into 10 scaffold levels,\nwith a p50 index from the Cyclic System Retrieval (CSR) curve of scaffolds\nof 50%. From the resulting enumerated library, a diverse subset of\nstructures was selected by using the Jarvis-Patrick clustering method.\nThese structures, along with their associated properties measured\nfrom quantum mechanics and experimental data, were used to train a\nMachine Learning (ML) model to predict ionization energy (<i>E</i><sub><i>i</i></sub>) values. Subsequently, a\nScoring Neural Network (SNN), coupled with our Goal-Directed generative\nmodel using a Recurrent Neural Network (RNN) with Deep Learning (DL)\narchitectures, was trained. The generative model was guided using\na prior network within a Reinforcement/Transfer Learning environment.\nThe final AI-generative model learned that structures with high unsaturation,\nH/C ratios under 1, and molecular weights between 100 and 300 u are\nfavorable for ET MALDI matrices, as well as those with few aromatic\nrings and zero aliphatic rings. Other molecular features were also\nfavored. The resulting AI-generated library exhibits <i>E</i><sub><i>i</i></sub> values over 8.0 eV, akin to those of\nreported “good” ET MALDI matrices, indicating successful\ndesign with high synthesis accessibility scores. In conclusion, our\ngenerative model provided valuable insights into the molecular features\nideal for ET MALDI compounds while generating a wide range of structurally\ndiverse molecules within a similar molecular property space. The next\ncritical step in this process is to synthesize a selection of these\ngenerated compounds for the experimental validation and further characterization.