An interpretable molecular descriptor for machine learning predictions in atmospheric science

TL;DR

ATMOMACCS presents an interpretable molecular descriptor that merges MACCS fingerprints with SIMPOL-inspired motifs to better capture atmospheric organic chemistry. Using kernel ridge regression across four atmospheric datasets, the authors demonstrate consistent improvements in predicting $P_{sat}$, equilibrium partition coefficients, $\Delta H_{vap}$, and $T_g$, with the integer-count version (v5) delivering the strongest performance. SHAP analyses reveal that carbon number and oxygen-related motifs predominantly influence volatility and partitioning, while carbon-hydrogen topology and non-oxygen heteroatoms govern phase-transition properties, underpinning the descriptor’s interpretability. The approach offers a scalable, open-source, chemically informed framework that outperforms traditional fingerprints and even some non-linear alternatives, with clear pathways for extension to non-covalent atmospherically relevant systems.

Abstract

The study of aerosol formation and chemistry using machine learning is limited by the lack of molecular descriptors suited to atmospheric compounds. Interpretable models are particularly affected because they often rely on dictionary-based descriptors tied to specific molecular substructures, which currently fail to capture the full range of organic atmospheric compounds, including large, highly oxidized molecules common in the atmosphere. We introduce ATMOMACCS, an interpretable descriptor combining the 166 binary keys of the MACCS fingerprint with motifs inspired by the SIMPOL method for estimating saturation vapor pressures. We show that ATMOMACCS based models improve predictions of saturation vapor pressures (7-8 % error reduction), equilibrium partition coefficients (5 % and 9 % error reduction), glass transition temperatures (22 % error reduction), and enthalpy of vaporization (61 % error reduction) on four datasets with atmospheric compounds. Feature analysis shows that saturation vapor pressure and partition coefficients are governed by carbon number and oxygen-related features, whereas other phase-transition properties (e.g., enthalpy of vaporization, glass transition temperature) depend on carbon-hydrogen bond types and the presence of heteroatoms other than oxygen. This highlights the generalizability of ATMOMACCS across different datasets and properties as an interpretable molecular descriptor.

Figures (15)

• Figure 1: Molecular emissions from various sources enter the atmosphere, where they undergo reactions with molecules like ozone and hydroxyl radicals. These reactions produce oxidized compounds with diverse functional groups, which are key to understanding atmospheric processes leading to particle formation.
• Figure 2: Molecular substructures present in atmospheric compounds can be identified and incorporated into a binary molecular representation called MACCS. By extending MACCS to include additional features relevant to atmospheric chemistry, we create a new representation, ATMOMACCS.
• Figure 3: Functional fragment counts (SIMPOL) for (a) Wang,Wang2017 (b) GeckoQ,Besel2023 (c) Ferraz-Caetano,FerrazCaetano2024 and (d) LiLi2020 datasets. Carbon number, oxygen number and carbon types are excluded. The colors represent the number of occurrences within the same molecule. In the datasets, most fragments appear once or twice per molecule on average.
• Figure 4: Distribution of target values in the datasets: (a) saturation vapor pressure ($P_{sat}$) from the Wang dataset Wang2017, (b) water–gas equilibrium partition coefficient ($K_{W/G}$) from the Wang dataset, (c) water-insoluble organic matter–gas equilibrium partition coefficient ($K_{WIOM/G}$) from the Wang dataset, (d) saturation vapor pressure ($P_{sat}$) from the GeckoQ dataset Besel2023, (e) enthalpy of vaporization ($\Delta H_{vap}$) from the Ferraz-Caetano dataset, FerrazCaetano2024 and (f) glass transition temperature ($T_{g}$) from the dataset of Li et al. Li2020
• Figure 5: The construction of ATMOMACCS follows a four step process. First, the two dimensional molecular structure is read from the molecular SMILES (Simplified Molecular Input Line Entry System) string. Next, the appearance of ATMO features is counted based on the specified molecular structure. These counts are then converted into a binary representation, with the specific encoding scheme varying across different ATMOMACCS versions. Finally, the MACCS fingerprint is concatenated with the ATMO keys to produce the complete ATMOMACCS molecular descriptor.