COBRA: A Computational Brewing Application for Predicting the Molecular Composition of Organic Aerosols

Abstract

Atmospheric organic aerosols (OA) represent a significant fraction of airborne particulate matter and can impact climate, visibility, and human health. These mixtures are difficult to characterize experimentally due to their complex and dynamic chemical composition. We introduce a novel Computational Brewing Application (COBRA) and apply it to modeling oligomerization chemistry stemming from condensation and addition reactions in OA formed by photooxidation of isoprene. COBRA uses two lists as input: a list of chemical structures comprising the molecular starting pool, and a list of rules defining potential reactions between molecules. Reactions are performed iteratively, with products of all previous iterations serving as reactants for the next. The simulation generated thousands of structures in the mass range of 120–500 Da, and correctly predicted ~70% of the individual OA constituents observed by high-resolution mass spectrometry. Select predicted structures were confirmed with tandem mass spectrometry. Esterification was shown to play the most significant role in oligomer formation, with hemiacetal formation less important, and aldol condensation insignificant. COBRA is not limited to atmospheric aerosol chemistry; it should be applicable to the prediction of reaction products in other complex mixtures for which reasonable reaction mechanisms and seed molecules can be supplied by experimental or theoretical methods.

1. Introduction

Atmospheric organic aerosols (OA), airborne particles comprised primarily of organic material, are estimated to contribute up to 50% of the total particulate matter mass at continental mid-latitudes and up to 90% at forested areas (1). OA impact climate and visibility by interacting with solar radiation, atmospheric oxidants, and water vapor (2), and are associated with adverse effects on human health as discussed in Ref. (3), and references therein. The climate, visibility, and health effects of OA are dependent on their chemical composition (4–6). However, the diversity of OA formation and growth mechanisms makes detailed molecular chemical composition of OA difficult to predict by traditional modeling or characterize by experimental methods. During their residence time in the atmosphere, OA undergo chemical aging processes (7, 8) that further enhance molecular complexity through, for example, the formation of nitrogen-containing organic compounds (NOC) (9–11).

Recent advances in high-resolution mass spectrometry (HR-MS) have enabled simultaneous detection of hundreds of individual molecules in OA (9, 12–15). HR-MS tools are useful in providing the molecular formulas for OA constituents. Ion fragmentation patterns observed in tandem mass spectrometry (MSⁿ) experiments provide additional information about the structures. However, interpretation of MSⁿ data is complicated by the presence of multiple structural isomers and lack of sufficiently diagnostic fragmentation patterns. In previous work, reliable structural information could only be extracted for the low molecular weight (MW) species (16, 17) and homologous oligomers (18–20) present in OA, leaving the structures of the majority of complex oligomers uncharacterized.

HR-MS experiments suggest that in many cases, oligomeric compounds in atmospheric OA are produced from repetitive reactions between the OA constituents (17, 20–22). There is strong evidence that condensation reactions such as esterification and aldol condensation (18, 20, 23–28), and addition reactions such as hemiacetal formation (29–31) are quite common in both organic aerosols and in aqueous solutions of OA. Advanced computational approaches are needed for predicting the OA composition using a bottom-up approach, in which the starting low-molecular weight compounds are known from experiments, and reaction rules for combining these compounds into oligomers can be unambiguously defined.

This work describes the first application of a Computational Brewing Application (COBRA) to modeling oligomerization products observed in the detailed composition of OA. Figure 1 shows a diagrammatic representation of COBRA, which is a customizable simulation engine for chemical reactions. Prior work from our group related to predicting the mechanisms of organic chemical reactions has focused on a rule-based approach with broad knowledge of general organic chemistry (32), or a machine learning approach to reaction prediction based on ranking mechanistic steps by productivity (33). COBRA differs from these approaches in that it is optimized to simulate the evolution of a complex mixture by combinatorial computation of many thousands of chemical reactions between mixture constituents, based on a chosen pool of starting (seed) compounds and specific reaction mechanisms. For the case study presented in this work, the formation of a large set of experimentally-observed high-MW oligomeric products in OA derived from the photooxidation of isoprene (20, 22) is modeled by considering chemical transformations of a basic set of monomers through the following oligomerization reactions: esterification, aldol condensation, and hemiacetal formation. We demonstrate that COBRA succeeds at modeling relevant oligomerization chemistry, predicting and visualizing high-MW compound structures, and predicting unique reaction products in OA. More broadly, the COBRA approach is applicable to studying the evolution of a wide range of organic mixtures as long as the starting components and reaction mechanisms can be suggested.

Figure 1.

A schematic representation of COBRA. COBRA converts a set of seed molecules into a set of predicted products using predefined chemical reaction rules. This process is repeated for several iterations, with predicted products going back into the pool of reacting molecules after each step. The four reactions used in this work are shown in the center panel, and the full list of 27 seed molecules is shown in Table 1.

2. Methods

2.1 COBRA

Chemical structures are input into COBRA using the widely-used chemical string representation SMILES (34). Reaction transforms are defined using the reaction transform language SMIRKS (35). The SMIRKS language is a superset of SMILES, and can be used to define reaction transforms to an arbitrary degree of specificity. We leverage the programming language Python in conjunction with OpenEye’s OEChem library (36) to process the input SMILES and SMIRKS into predicted products. Standard valence rules are explicitly programmed into the reaction transforms in order to predict chemically meaningful structures. Specifically, R-groups in the reactions shown in Fig. 1 were allowed to be alkyl or acyl groups, but not hydrogen alone. Filters can be imposed to prevent compounds with certain properties from participating in reactions, or from being included in the pool of final products.

The simulation includes the following iterative steps. 1) For each molecule and each pair of molecules in the reactant pool, all unimolecular reaction transforms and all bimolecular transforms, respectively, are applied. Identical reactions already performed in the previous iterations are recognized and skipped. 2) The product molecules are filtered. For this simulation, we automatically filter out any molecule with greater than 40 heavy (C, O, N) atoms. This filter restricts the molecular weights of the products below approximately 500 Da, the limit at which molecular assignments made with 0.1 mDa accuracy can be uniquely determined for C_cH_hO_oN_nS_s species (37). This restriction significantly decreased the model run time while still capturing the chemistry of the most abundant compounds in isoprene high-NOx SOA (22). 3) Molecules that were not filtered out during the previous step are added back into the reactant pool, and the system returns to step one. The simulation is complete either after the requested number of iterations has passed, or after a specified number of unique reactions has been simulated (30,000 reactions for the full simulation and the exclusion simulations, in this study). Results can then be conveniently visualized and searched for specific chemical structures matching any specified structural criteria. Additionally, we can compute the sequence of reactions that led to a given product’s formation.

2.2 Experimental methods

Isoprene OA was photochemically generated in a 5 m³ Teflon chamber, as described previously (20). Samples were generated in dry air, under high-NOx (VOC:NOx <1) conditions, and in the absence of inorganic seed particles. H₂O₂ was used as an OH precursor. The initial mixing ratio of isoprene was 250 ppbv (parts per billion by volume). Blank samples were produced in an identical manner as OA samples, without UV photooxidation. Samples were collected on Teflon filters (0.2 μm pore size, Millipore), vacuum sealed, and frozen prior to analysis. Negative ion mode direct injection ESI (tip voltage 4kV) was used as an ionization method for solvent-extracted OA samples. The solvents used for analysis were water and acetonitrile (both HPLC grade, Fluka) at a 1:1 volume ratio. Mass analysis was done with a high-resolution linear ion trap (LTQ-) Orbitrap^™ (Thermo Corp.) at Pacific Northwest National Lab (PNNL) Environmental Molecular Science Laboratory facility (EMSL), with a mass resolution of 60,000 m/Δm at m/z 400. MSⁿ studies were performed in the LTQ, with mass selection in the 0.5 m/z range and collision-induced-dissociation energies of 20–40 energy units. Product ions of MSⁿ were analyzed in the Orbitrap.

3. Results and Discussion

Table 1 lists the seed molecules for the simulation. These low-MW species have been identified as important building blocks in the formation of isoprene photooxidation OA (22). The majority of these molecules have been detected in isoprene OA (Ref. (38), and references therein), and are primarily multifunctional carbonyl, alcohol, and carboxyl compounds derived from the oxidation of isoprene with the hydroxyl radical (OH) (18, 22, 39–46). Some of these molecules have sufficiently high vapor pressure to exist primarily in the gas phase. For example, glycolaldehyde has a room temperature vapor pressure of 0.028 Torr (47), and may participate in aerosol mass-growth reactions initiated in the gas- or heterogeneous phase. Note that glyoxal is not included among the seed molecules used in these simulations because it was not found among the monomer units inferred from mass spectrometry of isoprene oxidation products as described in Ref. (22). The 3-nitrate ester of 2-methylglyceric acid (2MGA) (18, 20, 27, 45) represents the sole NOC in the list of seed molecules. Available information about NOC monomers is limited due to the small number of studies of NOC composition in the literature (Ref. (22) and references therein).

Table 1.

Seed molecules used for the COBRA simulations described in this work. Exclusion percent is a measure of the molecule’s contribution to the pool of correctly predicted products.

Formula	Name	Exclusion percent	Structure
C2H4O2	glycolaldehyde	0
C2H4O2	acetic acid	0
C2H4O3	glycolic acid	−0.6
C3H4O2	methylglyoxal	−0.3
C3H4O2	acrylic acid	−0.9
C3H6O2	hydroxyacetone	0
C3H6O2	propanoic acid	2.2
C3H4O3	pyruvic acid	2.8
C3H6O3	lactic acid	−2.2
C3H4O4	3-hydroxypyruvic acid	−5.6
C3H4O4	2-hydroxy-3- oxopropanoic acid	−2.5
C4H6O2	methacrylic acid	1.2
C4H6O3	2-methyloxopropanoic acid	−1.9
C4H8O3	2-methylglyceraldehyde	−0.3
C4H8O4	2-methylglyceric acid	−1.5
C4H8O4	erythrose/threose	−0.9
C4H6O4	methylmalonic acid	−0.6
C4H6O4	2-hydroxy-2-methyl-3- oxopropanoic acid	−2.5
C4H6O5	hydroxy(methyl) propanedioic acid	−0.6
C4H7O6N	2-methylglyceric acid 3- nitrate	18.8
C4H8O5	2,3,3-trihydroxy-2- methylpropanoic acid	−2.2
C5H8O3	2-hydroxy-2- methylbutanedial	2.8
C5H8O4	methylsuccinic acid	3.4
C5H8O4	2-hydroxy-2-Me-4- oxobutanoic acid	−4.6
C5H8O5	2-hydroxy-2- methylbutanedioicacid	−1.9
C5H10O3	2,4-dihydroxy-2- methylbutanal	6.2
C5H10O5	2,3,4-trihydroxy-2- Mebutanoic acid	−0.6

COBRA was used to simulate the evolution of a virtual OA mixture by applying the four reaction transforms shown in Figure 1 to the set of 27 starting molecules in Table 1. The resulting product pool contained 135,107 predicted structures after performing 30,000 unique reactions, representing a total of 758 unique elemental formulas of the type C_cH_hO_oN_n with MW<500 Da. The two orders of magnitude difference between the number of structures and elemental formulas reflects a large number of predicted structural isomers. The experimental HR-MS data from Ref. (22) (Table S1 in the Supporting Information section) contains 464 neutral elemental formulas, and 323 of these formulas (70%) were predicted by COBRA. Figure 2 shows the experimental HR-MS spectrum separated into two panels. The first panel shows all HR-MS peaks in the experimental spectrum that were predicted by COBRA (Fig. 2A), and the second one shows peaks that did not appear in the COBRA output (Fig. 2B). As Figure 2 demonstrates, the 70% fraction of the experimentally observed peaks predicted by COBRA represents the most abundant compounds with MW<500 Da. The remaining 30% fraction has average peak intensities that are more than an order of magnitude smaller compared to the average peak intensities associated with the successfully predicted compounds, suggesting that COBRA captures the essential chemistry producing oligomers in isoprene high-NOx OA. The high degree of overlap between the predicted and observed formulas confirms that the oxygenated hydrocarbon seed molecules proposed in our previous work are the relevant oligomer building blocks in this type of OA (22). The fact that a relatively small set of modeled reactions was sufficient to account for roughly 70% of the HR-MS peaks suggests that the oligomerization chemistry in isoprene OA is well constrained.

Figure 2.

Comparison between the HR-MS experiment and COBRA predictions. Panel (a) shows the HR-MS peaks that are predicted by COBRA (70%). Panel (b) shows the remaining 30% of HR-MS peaks that are not predicted by our simulation. In both panels, NOC peaks are colored green. Panel (c) indicates the number of isomeric structures predicted at each observed molecular mass.

The fraction of the experimentally observed compounds that are correctly predicted by the simulation can be a misleading metric of success in some cases. For example, if the simulation were significantly underdefined, and generated every possible C_cH_hO_oN_n formula (~ 10⁵) allowed by the valence rules, this fraction would become 100%. Therefore the reverse comparison of the fraction of predicted peaks that show up in the experiment is just as important. In the present case, 43% of predicted formulas correspond to compounds detected by HR-MS, and the remaining 57% are not observed. This level of agreement can be viewed as good, considering that the simulation does not use any kinetics restrictions and predicts the formation of compounds that may be below the limit of detection of HR-MS.

It is remarkable that a total of 62 NOC molecular formulas were predicted stemming from 2MGA-3-nitrate alone, representing 41% of the total NOC observed in the isoprene OA sample in the m/z 120–500 range (Fig. 2A). Note that this particular simulation’s ability to predict NOC compounds is limited because only a single NOC seed molecule is included. Nonetheless, the data confirm that 2MGA-3-nitrate is a prolific oligomer building block in isoprene OA that produces a variety of products through oligomerization in the condensed phase (18, 20, 27, 45).

Since structural complexity increases with molecular size, the number of structural and stereo isomers that can be produced by oligomerization reactions also grows exponentially with increasing mass of the individual products (Fig. 2C). For the 323 experimentally observed molecular formulas, COBRA predicts 102,650 unique structures, with number of isomers ranging from 1 to nearly 3,000. The apparent decrease in the number of hits as MW approaches 500 Da is a result of filtering; we artificially limit structures in the product pool to those with less than 40 heavy atoms, and do not consider structures with molecular weights in excess of 500 Da. Despite the large number of predicted isomers, the simulation results help constrain possible structures, especially for lower-MW compounds. For example, over 1,000 unique molecular structures associated with a molecular mass of 142.063 Da, assigned to C₇H₁₀O₃, can be retrieved from internet-based chemical inventories (e.g., SciFinder). In contrast, COBRA predicts only two structures of C₇H₁₀O₃, thereby dramatically reducing the complexity (Table 2). This represents more than a one hundred-fold increase in the level of confidence in assigning molecular structures to a given formula obtained by HR-MS. Even at high masses, COBRA predictions remain useful because the structural information for high-MW compounds is not readily available from chemical databases, and this type of modeling is a practical step toward understanding the possible structures of large oligomers.

Table 2.

A small fraction of the full output from COBRA. For each observed product, several different isomers (SMILES structures) are predicted (number of hits). The full number of hits for each mass is shown in Figure 2C.

Product Mass (Da)	Product Formula	Hits (n)	SMILES # 1	SMILES # 2	… SMILES # n
128.047	C6H8O3N0	4	CC(=C)C(=O)OCC=O	CC(=C)C(=O)OC(=O)C	CC(=O)COC(=O)C=C
130.027	C5H6O4N0	4	CC(=O)C(=O)OCC=O	CC(=O)C(=O)OC(=O)C	C=CC(=O)OCC(=O)O
132.042	C5H8O4N0	9	CCC(=O)OC(=O)CO	CCC(=O)OCC(=O)O	CC(C(=O)O)OC(=O)C
136.037	C4H8O5N0	2	C(C(O)OC(=O)CO)O	C(C(O)OCC(=O)O)O
142.063	C7H10O3N0	2	CCC(=O)OC(=O)C(=C)C	CC(=C)C(=O)OCC(=O)C
146.022	C5H6O5N0	8	CC(=O)C(=O)OC(=O)CO	CC(=O)C(=O)OCC(=O)O	CC(=O)OC(C=O)C(=O)O

A few published studies used MSⁿ fragmentation data to determine structures of the oligomers in isoprene high-NOx SOA (18, 19, 22). All molecular structures described in the previous studies match structures predicted by COBRA. To further test the validity of the COBRA predictions, we made a comparison to MSⁿ studies previously performed by our group of NOC oligomers observed in SOA generated from isoprene photooxidation (22). Figure 3 shows the predicted structures of two NOC oligomers, C₈H₁₃NO₉ (267.059 Da) and C₁₄H₂₀NO₁₃ (411.101 Da), that are most consistent with MSⁿ experiments. The MSⁿ experiments were performed using the negative ion mode; therefore, the molecular masses of the detected compounds are those of deprotonated molecules. Oligomer fragmentation is often characteristic of the functional groups and monomer units present within the compound. For example, HNO₃ and CH₃NO₃ are characteristic MSⁿ losses from organic nitrates, and C₄H₆O₃ is a characteristic loss from the carbonyl ester unit of 2MGA (18, 22).

Figure 3.

Experimental MS² fragmentation spectra for precursor (I) ions (a) C₈H₁₂NO₉⁻ and (b) C₁₄H₂₀NO₁₃^-. Two structures predicted by COBRA that are consistent with the fragmentation patterns are overlaid on the spectra. Product ions are obtained from fragmentation at the dashed bonds (after losses of neutral molecules). Product ions for (a) are: II. m/z 203.056 (I – HNO₃); III. m/z 189.040 (I-CH₃NO₃); IV. m/z 164.020 (I-C₄H₆O₃); V. m/z 119.035 (I-C₄H₅NO₅). Product ions for (b) are: II. m/z 347.098 (I – HNO₃); III. m/z 333.082 (I-CH₃NO₃); IV. m/z 308.062 (I-C₄H₆O₃); V. m/z 307.082 (I-C₃H₅NO₃).

For a given molecular formula, the proposed structures based on MSⁿ experiments described in Ref. (22) match at least one of the structures predicted by COBRA. We note that structural isomers may not have distinct fragmentation patterns, and therefore it is possible that the experimental MSⁿ data represents several structural isomers predicted by COBRA. COBRA predicted four isomeric structures for C₈H₁₃NO₉, and at least one predicted structure is consistent with the fragmentation pattern produced by MSⁿ studies. For C₁₄H₂₁NO_13, at least one out of 44 predicted structures is consistent with the MSⁿ fragmentation pattern. Therefore, it is reasonable to hypothesize that other structures predicted by COBRA may be present in the OA, perhaps but not necessarily, in lower quantity than the most abundant isomer.

To elucidate which seed molecules were the most important as oligomer building blocks for isoprene OA, we performed 27 “exclusion” simulations, each with a different seed molecule omitted. Each exclusion simulation was limited to the same number of reactions (30,000). The total number of experimental molecular formulas recovered in each case was compared against a simulation in which no seed molecules were omitted. The “exclusion percent” column in Table 1, calculated as a percent reduction in the number of predicted formulas matching experimental formulas, can be viewed as a measure of a compound’s relative importance in generating structures with experimentally observed molecular formulas. The highest exclusion percent (19%) was for the 3-nitrate ester of 2-methylglyceric acid, consistent with its special role of being the sole NOC precursor. Specifically, the removal of 2MGA-3-nitrate eliminated all NOC oligomers from the product pool. Removal of other seed molecules resulted in smaller variation in the number of predicted formulas, ranging from −6 to 6%. Note that because of the fixed number of reactions in these simulations, an exclusion of a less significant contributor to chemistry can actually lead to a slight increase in the total number of predicted formulas, yielding a negative exclusion percent, as happened with 2-hydroxy-2-Me-4-oxobutanoic acid. None of the nitrogen-free compounds appeared to stand out from the rest, suggesting some redundancy in the pool of seed molecules. The redundancy in seed molecules lowered the percent overlap between simulation and experiment. These results highlight the need for future work in determining missing precursors and reaction mechanisms. In general, this type of exclusion analysis is useful for assessing the contribution of a single molecule to the total product pool, e.g. the importance of glyoxal in heterogeneous reaction with various OA constituents (48–51), or the contribution of first vs. second generation VOC oxidation products to producing condensable organic compounds.

We performed additional exclusion simulations, in which one of the four reactions shown in Fig. 1 was disabled. The results of the simulations based on the three remaining reactions were compared to the results of the full simulation. Table 3 lists the percent reduction in the number of experimentally observed molecular formulas recovered by the simulation, the percent reduction in the number of theoretically predicted molecular formulas, and the percent reduction in the number of theoretically predicted structures generated by the simulation. The esterification reaction (Reaction 1) played the most significant role in recovering experimentally observed formulas, with an exclusion percent of 40.9%, followed by the hemiacetal formation reaction (Reaction 4) with an exclusion percent of 10.2%. The esterification and hemiacetal formation reactions also made the largest contributions to the total number of theoretically predicted structures. Exclusion of esterification from the simulation reduced the number of structures in the COBRA output by 37.8%, while exclusion of hemiacetal formation resulted in a 46.7% decrease. Although hemiacetal formation is prolific at generating theoretical structures, the results demonstrate that esterification is the most important reaction in promoting oligomerization in isoprene OA, while hemiacetal formation is of secondary importance. These results are consistent with a previous study, wherein reducing aerosol liquid water content changed the composition of isoprene high-NOx SOA most drastically by hindering the esterification reaction (20). They are also consistent with the fast and efficient reactions of OA compounds containing carboxylic acid groups and carbonyls with alcohols observed in Ref. (52). In contrast, aldol condensation (Reactions 2 & 3) does not appear to play a significant role in forming the oligomers. Removal of either Reaction 2 or Reaction 3 resulted in a small reduction in the number of theoretically produced structures and recovered HR-MS formulas.

Table 3.

Results from an exclusion analysis, in which one of the four reactions shown in Fig. 1 is disabled. The values listed in the last three columns represent: 1) the percent reduction in the number of experimentally observed formulas recovered by the simulation, 2) the percent reduction in the total number of theoretical formulas predicted by the simulation, and 3) the percent reduction in the total number of theoretical structures predicted by the simulation.

Reaction Transform	Reaction Type	Percent reduction in experimentally observed formulas	Percent reduction in theoretically predicted formulas	Percent reduction in theoretically predicted structures
1	esterification	40.9	12.8	37.8
2	aldol condensation	0.3	0.7	5.6
3	elimination of water from aldol condensates	0	1.5	0.01
4	hemiacetal formation	10.2	−0.1	46.7

Once a pool of COBRA-generated simulation products is available, it can easily be mined for desired products or classes of products. For example, we are interested in the potential for isoprene photooxidation to generate α,β-unsaturated aldehydes, a class of genotoxic compounds found in cigarette smoke and air pollution (53–58). The most common atmospheric α,β-unsaturated aldehydes with adverse health effects are gas-phase species such as acrolein (C₃H₄O) (59). However, adverse health effects are also strongly linked to inhalation of particulate matter (3, 60–64), which contains more complex and poorly characterized organics. COBRA predicted 6,131 structures corresponding to α,β-unsaturated aldehydes in the isoprene photooxidation SOA data set. This large set of structures can be filtered further, by mass, O/C ratio, or any other property required to yield a focused set of predicted structures. Applying a computational tool like COBRA along with the molecular-level experimental characterization of products is a powerful first step to identifying condensed-phase atmospheric toxins.

One limitation of COBRA is that it does not currently include kinetics information, and therefore cannot model the relative abundances of individual products. This limitation may be important in cases where the product branching ratios from a particular reaction are dissimilar or dependent on atmospheric conditions, in which case the results from COBRA will artificially enhance the importance of non-competitive products. For example, in the absence of nitrogen oxides (NOx) the same OH-initiated oxidation chemistry produces more hydroperoxides and fewer carbonyl products (65). However, a scaling factor can be incorporated in the transformation rules if the branching ratios are known or can be estimated with ab initio methods. For the comparison with the ESI-based MS data, this is not a major limitation because there are no simple relationships between the detection efficiency in ESI and the concentration of the analyte species (66).

Applying COBRA to the simulation of oligomerization in isoprene OA is a good example of the utility of computational tools for understanding complex natural mixtures and verifying experimental data. An important strength of COBRA is its ability to handle a complex simulation and generate a large number of predicted compounds. While the combinatorial explosion of generated structures means computation time is currently on the order of several days, this time can be significantly reduced in future work by using parallel computing algorithms. If pertinent experimental data are available, direct comparison between the observed and predicted compounds can be made to better constrain chemistry and chemical structures. Furthermore, as COBRA reduces the number of possible structures available for a given chemical formula, the predicted structures may be useful for discriminating isobaric species in a mass spectrum. Applying a small set of transformation rules to predict an experimentally determined high-resolution mass spectrum may become a powerful tool for better understanding the chemistry of complex systems comprised of hundreds or thousands of individual compounds.