Evaluation of Physicochemical Property Data in the ECHA Database

Physicochemical properties such as the octanol–water partition coefficient (K_OW), the octanol–air partition coefficient (K_OA), vapor pressure (p_V), or the solubility in water (S_W) are important substance-specific properties that determine the environmental fate of substances. K_OW influences, for example, the ability of a substance to partition into membranes and can thus be used to determine the baseline toxicity of a substance.¹ The physicochemical properties of a substance are also important parameters in the planning and realization of experiments. Volatile substances, for example, have to be handled differently than non-volatile ones and substances with a low solubility in water need a different experimental setup than substances with a high water solubility.² Accurate experimental and non-experimental physicochemical property data are therefore important elements in the risk and hazard assessment of chemicals and various approaches have been taken in the past to evaluate the data quality of larger data sets.^3,4

One of the largest and most important databases with physicochemical properties is the European Chemicals Agency (ECHA) database that contains data submitted under the EU regulation on the registration, evaluation, authorization and restriction of chemicals (REACH). Starting in 2007, manufacturers and importers of chemicals had to register their chemicals under REACH if the chemical was manufactured in and/or imported to the European Economic Area (EEA) above 1 tonne/year.⁵ The submitted data are very valuable and of great importance for the evaluation of the chemicals. However, the amount of data also presents a major challenge, as not only the substances have to be evaluated, but also the quality of the data has to be assessed. ECHA and the EU Member States evaluate the information substance-by-substance (sometimes also as groups of substances), either during a compliance check or a substance evaluation. However, this is very time-consuming and resource-intensive and therefore not all dossiers and substances have been evaluated so far.⁶ 

The aim of the present study was therefore to cross-evaluate the physicochemical property data available in the ECHA database. The evaluation here was done for all data points of one endpoint (often several 1000s) and not substance-by-substance. This allowed for the identification of systematic deviations and outliers. For the evaluation of the data, calculations were run with COSMOtherm for all mono-constituent, neutral organic substances that are fully registered under REACH. To ensure that the COSMOtherm data themselves did not contain systematic errors, they were evaluated beforehand with data from the PHYSPROP database, a manually curated physical properties database.⁷ Based on the results of our study, suggestions for method improvements and future research are provided.

The physicochemical properties (often referred to as endpoints) evaluated in this study are those that are of high importance for the chemical risk and hazard assessments and include: The logarithmic octanol–water partition coefficient (log₁₀ K_OW), the solubility in water (S_W), the vapor pressure (p_V), the logarithmic air–water partition coefficient (log₁₀ K_AW), the logarithmic soil adsorption coefficient (log₁₀ K_OC), the normal boiling point (T_b), and the second-order rate constant for the degradation with OH radicals (k_OH). An exact definition of each endpoint is provided in Sec. S1 in the supplementary material 1.

It is important to note that the analysis was carried out with data from the ECHA database because these data are readily available for a large number of substances and the data are used for the regulation of the chemicals in the EEA. However, it may well be that other large data sets containing data on physicochemical properties would show similar percentages of incorrect entries and would lead to similar conclusions.

The non-confidential substance data submitted to ECHA under the REACH regulation were downloaded from the International Uniform ChemicaL Information Database (IUCLID) website⁸ in April 2022 and contained the data from the registration dossiers as of September 15th 2021. The data contained 26 081 registration dossiers for 23 184 substances.⁸ The isomeric simplified molecular-input line-entry system (SMILES) string, which describes the chemical structure in line notation, was not available in the IUCLID data. We therefore contacted ECHA and obtained a list of the registered substances with the available IUPAC names, European Community numbers (EC numbers), Chemical Abstract Service Registry Numbers® (CAS RN), molecular formulas, and SMILES notations in January 2021. Additional SMILES codes were obtained manually from the brief profiles of the registered substances in February and May 2022. For substances for which no SMILES code was publicly available in the ECHA database, the SMILES was obtained via the CAS RN from PubChem⁹ and/or the CompTox Chemical Dashboard¹⁰ and cross-checked with SciFinderⁿ.¹¹ Information on registration status, registration type, substance type, composition, and tonnage band were retrieved in April 2022 from the ECHA website.¹² 

The substances investigated in this study are those that are neutral, organic, and mono-constituent with a full registration under REACH [including previously notified substances (NONS)]. Mono-constituent means here that one constituent is present at a concentration of at least 80% (w/w).¹³ Intermediates were not included. Substances for which the CAS RN and SMILES notation did not correspond to each other were also not included.¹⁴ Substances that were registered as neutral but had either an acidic pK_a value below 7.4 or a basic pK_a value above 7.4 were also not included (1300 substances), but are listed in the supplementary material 2. The final set of investigated substances consisted of 5318 substances (Table S1 in the supplementary material 1).

For substances with more than one component, only the main component was analyzed. For example, in N,N′-diallyl-1,3-diaminopropane dihydrochloride, only N,N′-diallyl-1,3-diaminopropane would be analyzed. Some of the substances with more than one component consisted also of stereoisomers that corresponded to a CAS RN without stereoinformation. The properties of one of the isomers were then used in the analysis. Additional information on the subcomponents of 687 mixtures that did not have a main component but were also not labeled as multi-constituent or unknown or variable composition, complex reaction products or biological materials (UVCB) is provided in the supplementary material 2 as well. However, these data were not included in the analysis since precise information on the composition of the mixtures is lacking. Charged substances were also not analyzed in the current study; information on their physicochemical properties will be published later.

Values of the various properties (log₁₀ K_OW, S_W, p_V, log₁₀ K_OA, k_OH, T_b, log₁₀ K_OC) as well as information on the physical state were extracted with Python 3.10.4 from the xml documents provided in the IUCLID files. These IUCLID-derived data were linked afterwards with the information provided on the ECHA website¹² (registration status, registration type, origin, etc.) and the SMILES codes. This was done via the EC number and/or the substance name. These two identifiers were the only ones provided in the IUCLID data sets. IUCLID data for substances without EC numbers and names (e.g., for substances that are named “no public or meaningful name is available”) could not be assigned and were excluded.

For the analysis of the data, only those values were selected that were not marked as unreliable, were determined for the registered substance, and were in a certain pH and temperature range so that the data points were comparable to other experimental or calculated data points. The specific criteria for removing a study are shown in Table 1.

The “best available value” is shown in some of the graphics in the supplementary material 1 and is also given in the supplementary material 3. For T_b, log₁₀ K_OW, log₁₀ K_AW, and log₁₀ K_OC, the “best available value” is calculated as the arithmetic mean of all values from studies that were labelled as “key studies” or, if no key study existed, as the arithmetic mean of all other values. For S_W, p_V, and k_OH, the “best available value” is calculated as the geometric mean of all values from studies that were labelled as “key studies” or, if no key study existed, as the geometric mean of all other values. Data with qualifier <, >, ≤, or $≥$ are not included in the “best available values.”

The data from the PHYSPROP database were used to evaluate the calculations carried out with COSMOtherm but also to define beyond which limits (for each endpoint) a deviation should be considered an outlier. A comparison of the PHYSPROP data with the data in the ECHA database was also performed and is included in the supplementary material 1. However, relatively few of the substances registered under REACH have experimental values in PHYSPROP (see Table S2 in the supplementary material 1). A validation of the data in the ECHA database just with the data from the PHYSPROP database would therefore not have covered a sufficient number of chemicals.

pK_a values were calculated with MarvinSketch 22.18.¹⁷ Substances were treated as neutral if the most acidic pK_a value was above 7.4 and/or the most basic pK_a value below 7.4. “Acidic pK_a” means pK_a values for acidic groups (e.g., COOH), “basic pK_a” means pK_a values for the conjugated acids of basic groups (e.g., NH₂).

COSMOtherm and the related program COSMOconf are both commercial software applications that are now distributed by BIOVIA.¹⁸ COSMOconf is a tool that can generate the conformers of a molecule that are most relevant for interactions with other molecules. These conformers can later be used in COSMOtherm to estimate the physicochemical properties of a substance. Both programs are based on the Conductor-like Screening Model for Real Solvents (COSMO-RS) theory.^19,20 Compared to many other property estimation methods, the calculations with COSMOconf and COSMOtherm are not based on training sets of physicochemical property data; instead they are based on chemical potential differences of molecules immersed in solvents. The data that originate from calculations with COSMOconf and COSMOtherm are internally consistent and have been shown to be more accurate than the data from many other estimation programs.^21–23 A very recent publication showed that COSMOtherm was not only superior to other estimation programs but also that the calculated log₁₀ K_AW values for 21 per- and poly-fluoroalkyl substances were within one log₁₀ unit of the experimental data.²⁴ COSMOtherm is also able to calculate physicochemical properties for various temperatures, which is especially important for strongly temperature-dependent endpoints such as p_V or log₁₀ K_OA.²⁵ 

COSMOconf was run on the ETH “Euler Cluster” on around 200 central processing units (CPUs) in parallel over almost 3 years. Calculations in COSMOconf for small molecules (<12 atoms) are very fast and are done within minutes. However, the typical calculation time increases exponentially with the number of atoms in a molecule and can be in the range of weeks for molecules with 80 atoms or more. We ran COSMOconf therefore only for molecules that had fewer than 80 atoms. This covered around 87% of the substances. 5% of the calculations also failed and no results could be obtained. The parametrization used in COSMOconf was BP-TZVPD-FINE COSMO+GAS.

Two kinds of statistical analysis were conducted in this study. The root-mean-square error (RMSE), R-square (r²), and the mean absolute error (MAE) were used to measure the differences/similarity between two data sets. For the calculation of the values, the scikit-learn 1.2.2 package in Python 3.10.4 was used.²⁶ In addition, for each endpoint, the limits beyond which a deviation should be considered an outlier, i.e., instances where there is clearly no agreement between measurement and prediction, were determined. For a normal distribution, one could quantify this via the standard deviation. Outliers would be detected if the values are outside of three standard deviations. In this study, we adjusted this procedure slightly.

1. For each endpoint, the errors between the experimental data from the PHYSPROP database and COSMOtherm were calculated. The error in this case was the difference between a datapoint in the PHYSPROP database and the COSMOtherm value. For p_V, S_W, and k_OH, this was done for the log₁₀-transformed data.

2. Subsequently, it was checked if the distribution of the errors followed a normal distribution. This was done by running the Shapiro-Wilk Test in Python using the SciPi package²⁷ and the shapiro function.

3. For all endpoints where the error distribution did not follow a normal distribution, the error distribution was analysed visually and if it was symmetric and similar in shape to a normal distribution, the 2.3rd and 97.7th percentile of the errors were calculated. The values within these percentiles cover 95.4% of the errors which corresponds to two standard deviations in a normal distribution. We chose two standard deviations and not three (which would be the default) since we wanted to sort out not only outliers but also values with high deviation in general.

4. To set the final threshold for the detection of outliers and other values with high deviation, the arithmetic mean of the 2.3rd and 97.7th percentile of the errors was calculated.

For the substances investigated in this study, 823, 853, 852, 1050, 597, and 392 experimental data points were available in the PHYSPROP database for log₁₀ K_OW, S_W, p_V, T_b, H_W, and k_OH, respectively. The comparison between these data points and the ones generated with COSMOtherm is shown in Fig. 1; the corresponding statistics are presented in Table 2. The data points agree in general quite well for log₁₀ K_OW, p_V, log₁₀ K_AW, and T_b (all have R-square values around 0.9). Higher deviations are observed for S_W and k_OH with R-square values of 0.76 and 0.79, respectively. A detailed analysis of the data is provided in the supplementary material 1 Sec. S3.

From this comparison, we derived for each endpoint a validated range for the COSMOtherm data (Table 3). The validated range gives for each endpoint the lower and upper bound for which PHYSPROP data were available; not counting those values from the PHYSPROP database that were outside of the thresholds defined in Table 2. The data outside the validated range are not per se inaccurate, but the uncertainty in these data is larger. A discussion of these data per endpoint is provided in Sec. 4.2.

The available data in the ECHA database differ between full registrations and NONS (Fig. 2). NONS registrations are for substances that were already registered under the previous regulation (Directive 67/548/EEC). These substances were automatically transferred into REACH and are regarded as already registered. Companies were then able to claim the registrations as their own.

For substances with full registrations, the largest numbers of data points are available for log₁₀ K_OW, T_b, S_W, and p_V. These endpoints are required for all substances that are manufactured or imported in quantities of 1 tonne/year or more.²⁸ Most of the submitted data for these four endpoints were experimental data (Fig. 2 left). Data for log₁₀ K_AW and k_OH are not required for any tonnage band and were therefore only provided in some registration dossiers. The submitted data for these two endpoints are a mixture of experimental data, (Quantitative) Structure-Activity Relationships [(Q)SAR] results, and other calculations.

Most of the submitted data for the NONS only include the main result but no further information on the study type [experimental study, calculation, results from (Q)SAR, etc.] or other details (Fig. 2 right). This is because dossiers for NONS complying with the standard information requirements need to be submitted to ECHA only when the dossiers are updated to increase the tonnage band.^29,30 According to Chemsafe-Consulting (2021), this happened until 2021 in only 8% of the cases. Otherwise, the dossier does not need to include information that was not required under the previous legislation.³⁰ 

It is important to note that, for many of the endpoints shown in Fig. 2, more than half of the studies submitted under REACH (and shown in Fig. 2) are either labeled as not reliable, were not conducted for the registered substance, or were outside the selected pH or temperature range. The final number of investigated substances and studies (excluding those just mentioned) is provided in Table S2 in the supplementary material 1. The studies that are listed in Table S2 are also the ones that were analyzed in Sec. 3.3.

The calculated COSMOtherm values that were used to evaluate the data from the ECHA database are provided in the supplementary material 2. Moreover, the supplementary material 3 includes for all endpoints those substances for which the “best available value” deviates more than the endpoint-specific threshold from the COSMOtherm values. The evaluation of the log₁₀ K_OC values is provided in Sec. S5.7 in the supplementary material 1; all other endpoints are described in the following subsections.

Relevant experimental data for log₁₀ K_OW were submitted for 2425 substances and included a total of 3390 studies (Table S2, supplementary material 1). 2994 of these studies could be analyzed by comparison with the COSMOtherm data. Relevant non-experimental data for log₁₀ K_OW were submitted for 2702 substances and included a total of 4140 studies. Out of these, 3448 could be analyzed by comparison with the COSMOtherm data. Figure 3 shows the data points for the experimental as well as non-experimental log₁₀ K_OW values.

The log₁₀ K_OW values from the ECHA database with no qualifier or qualifier “ca.” agree with a R-square value of 0.70 (Table S5, supplementary material 1) with the calculated values from COSMOtherm (Fig. 3, left). 4.3% of the ECHA data are more than 1.2 units above the COSMOtherm values, 11% are more than 1.2 units below the COSMOtherm data points. Higher deviations are observed for the more hydrophobic substances. A few of the detected outliers and values with higher deviation are outside the range validated here with the data from the PHYSPROP database (log₁₀ K_OW > 10), but most of them are within the range. More information on the outliers and values with high deviation can be gained from a closer look at the K_OW measurement methods.

The guidance on information requirements and chemical safety assessment - Chapter R.7a: Endpoint specific guidance³¹ recommends to use certain methods for specific ranges of log₁₀ K_OW. Specifically, the shake flask method (EU A.8, OECD TG 107) for −2 < log₁₀ K_OW < 4; the high-performance liquid chromatography (HPLC) method (EU A.8, OECD TG 117) for 0 < log₁₀ K_OW < 6 (10 in exceptional cases); and the slow-stirring method (OECD TG 123) for log₁₀ K_OW values up to 8.3. The plots in Fig. 4 are therefore also divided into log₁₀ K_OW ranges (<4, 4–8, and >8).

All methods perform (judged by comparison to the COSMOtherm values) quite well for substances with log₁₀ K_OW < 4 (Fig. 4, top). There are only a couple of data points that deviate by more than 1.2 units from the COSMOtherm values and no method seems to be better or worse than the other. For substances with log₁₀ K_OW between 4 and 8, most measurements were performed with the HPLC method, which seems to be the best-suited method together with the slow stirring method (OECD TG 123). However, some data points were also obtained with the shake flask method (which is not recommended for this range) or without a method stated, and these partition coefficients deviate in more than one-third of the cases by more than 1.2 units from the COSMOtherm values. It is therefore highly recommended to use the HPLC or slow stirring method for future measurements in the log₁₀ K_OW range 4–8. No method performed very well for substances with log₁₀ K_OW > 8. Most of the results were again obtained with the HPLC methods, but 21 of the 33 partition coefficients obtained by HPLC deviate by more than 1.2 units from the COSMOtherm values.

The log₁₀ K_OW values from the ECHA database with qualifier > or ≥ and qualifier < and ≤ are shown in Fig. S5 in the supplementary material 1. The experimental log₁₀ K_OW values for substances with a calculated log₁₀ K_OW < 5 and qualifier > or ≥ are very similar to the COSMOtherm values. However, more than half of the experimental log₁₀ K_OW values with a (calculated) log₁₀ K_OW > 5 and a qualifier > or ≥ are lower by more than 1.2 units than the COSMOtherm values. For the experimental data points with qualifier < or ≤, the deviations are less severe.

The partition coefficients that are not from experimental studies but from (Q)SARS, calculation, read across, and unspecified methods are shown in Fig. 3 (right). The R-square value is 0.64 and thus lower than the value for the experimental partition coefficients. One reason for this is that the registration dossiers of five alkenes contained 75 or 150 non-experimental data points for log₁₀ K_OW that ranged for all five alkenes from 1.5 to 16.3 (vertical lines in Fig. 3 right). These log₁₀ K_OW values were obtained from (Q)SARs and read-across of structurally similar compounds that had different chain lengths. However, it is even stated in the summary of e.g., dodec-1-ene that “Partition coefficient increases with increasing carbon number across the category, with a very similar trend for both the alpha olefins and olefins.” The partition coefficients of substances with different chain lengths should therefore not have been used for the registered compounds.

If the partition coefficients for the five alkenes are disregarded and the remaining log₁₀ K_OW values are compared to the COSMOtherm values, 20% of the non-experimental data points in the ECHA database are still at least 1.2 units below the COSMOtherm values and 6.3% are at least 1.2 units above the COSMOtherm values. These values are still higher than the comparable values for the experimental data points and show that (Q)SARs and calculation methods need to be selected more carefully than currently done.

Relevant experimental data for the solubility in water were submitted for 2170 substances and included a total of 2851 studies (Table S2, supplementary material 1). 2358 of these studies could be analyzed by comparison with the COSMOtherm data. Relevant non-experimental data for the solubility in water were submitted for 1324 substances and included a total of 1656 studies. Out of these, 1321 studies could be analyzed by comparison with the COSMOtherm data. Figure 5 shows the data points for the experimental as well as non-experimental solubilities in water.

The solubilities in water from the ECHA database with no qualifier or qualifier “ca.” agree with a R-square value of 0.72 (Table S5, supplementary material 1) with the calculated values from COSMOtherm. 3.9% of the ECHA data are more than 2.1 orders of magnitude above the COSMOtherm values; 3.4% are more than 2.1 orders of magnitude below the COSMOtherm values. The highest deviations are observed for substances with a solubility in water below 10⁻³ mg/l. Some of these solubilities are outside the range validated here with the PHYSPROP data (S_W < 10⁻⁴ mg/l). A discussion of the accuracy of these data is provided in Sec. 4.2.

Figure 6 shows the method types (flask method, column elution method, other) underlying the experimental data in the ECHA database. The general test guidelines (OECD TG 105, EU method A.6, EPA OPPTS 830.7840) are also provided in IUCLID, however all three guidelines include both the flask and column elution method and are therefore not really useful for method selection. According to the guidance on information requirements and chemical safety assessment - Chapter R.7a,³¹ the column elution method and the flask method with slow stirring are appropriate for low-solubility test substances (S_W < 10 mg/l). The flask method with fast stirring is appropriate for higher-solubility test substances (S_W > 10 mg/l).³¹ No specific recommendations are given for substances with a very low solubility (S_W < 10⁻³ mg/l).

In the data from the ECHA database, no distinction is possible between the flask method with fast and slow stirring (Fig. 6). This makes it very difficult to judge the performance of the individual methods. However, we can say that both the flask method and the column elution method worked well (with the expected variation) for substances with solubilities above 10⁻³ mg/l. For substances for which COSMOtherm calculated a solubility below 10⁻³ mg/l, neither of the methods worked well. But again, it was not possible to distinguish between fast and slow stirring so it cannot be said per se that also the flask method with slow stirring would not work.

The solubilities in water from the ECHA database with qualifier > or ≥ and qualifier < and ≤ are shown in Fig. S7 in the supplementary material 1. Almost all experimental solubilities in water with qualifier > or ≥ deviate less than 2.1 orders of magnitude from the COSMOtherm values. The water solubilities in the ECHA database with qualifier < or ≤ show much larger deviations. Especially the values for substances with a calculated solubility of less than 10⁻⁴ mg/l are in almost all cases more than 2.1 orders of magnitude above the COSMOtherm values. This is not problematic per se because the qualifier says that the values are the upper limit, but this also means that they should not be seen as real values, but as upper bounds.

The non-experimental values for the solubility in water from the ECHA database are shown in Fig. 5 (right). The R-square value is 0.57. There are 58 data points that have a water solubility of 10⁶ mg/l, i.e., 1 kg per kg of water in the ECHA database, many of them far above the solubility calculated by COSMOtherm. It was not possible to determine what caused these many occurrences of the same value. 10⁶ mg/l seems to be the maximum solubility that WATERNT in EPI Suite predicts.³² However, this does not explain why many of the data points are much higher than the COSMOtherm values. The meaning of these values is entirely unclear, and we recommend that they should not be used.

Without these values, 5.1% of the non-experimental water solubilities in the ECHA database are still at least 2.1 orders of magnitude below the COSMOtherm values and 9.0% are at least 2.1 orders of magnitude above the COSMOtherm values. The deviations are significantly higher for substances with a very low solubility in water (S_W < 10⁻³ mg/l).

Relevant experimental data for the vapor pressure were submitted for 1924 substances and included a total of 2515 studies (Table S2, supplementary material 1). 2254 of these studies could be analyzed by comparison with the COSMOtherm data. Relevant non-experimental data for the vapor pressure were submitted for 775 substances and included a total of 1112 studies. Out of these, 962 could be analyzed by comparison with the COSMOtherm data. Figure 7 shows the data points from the ECHA database for the experimental as well as non-experimental vapor pressure compared to the data from COSMOtherm.

The experimental vapor pressures from the ECHA database with no qualifier or qualifier “ca.” show large deviations for some substances (for some, more than 10 orders of magnitude) from the calculated values from COSMOtherm (Fig. 7). The overall R-square value is 0.63. 10.5% of the ECHA data are more than 1.9 orders of magnitude above the COSMOtherm values; 4.3% are more than 1.9 orders of magnitude below the COSMOtherm data points. Higher deviations are observed for the less volatile substances. Many of these data points are outside the range validated here with the data from the PHYSPROP database for COSMOtherm (<10⁻⁸ Pa). However, high deviations are also seen within the validated range.

Figure 8 shows the method types used for the experimental data in the ECHA database. Commission Regulation No. 761/2009 Annex I (A.4 Vapor pressure) recommends – with some restrictions – for substances with a vapor pressure above 1 Pa the static, dynamic, isoteniscope, or the gas saturation method.³³ For substances with a vapor pressure between 0.001 and 1 Pa, one of the three effusion methods, the gas saturation methods, or the spinning rotor method are recommended. For substances with a vapor pressure below 0.001 Pa, the Knudson cell effusion method, the isothermal thermogravimetry effusion method, the gas saturation method, and, in part, the spinning rotor method are recommended.³³ 

In the ECHA database, substances with a vapor pressure (calculated with COSMOtherm) above 1 Pa were mainly analyzed with the static or dynamic method and the deviations are quite small compared to the values from COSMOtherm (Fig. 8). Although the dynamic and the static method are not recommended for substances with a vapor pressure below 1 Pa, they were also used in 29% of the cases for substances with a calculated vapor pressure between 0.001 and 1 Pa. Especially the results from the static method show large deviations from the COSMOtherm values and we recommend not using these values. For substances with a calculated vapor pressure below 0.001 Pa, no method performed well. Most data in the ECHA database for these substances were derived with the vapor pressure balance method, although this method is not recommended for substances with a vapor pressure below 0.001 Pa. However, none of the methods that were recommended in Commission Regulation No. 761/2009 for a vapor pressure below 0.001 Pa performed better (judged by comparison to the COSMOtherm results). The worst results were obtained with the static method. Only 1 out of 39 values shows a deviation from the COSMOtherm values of less than 1.9 orders of magnitude.

The non-experimental values for the vapor pressure from the ECHA database are shown in Fig. 7, right. The R-square value is 0.47. 6.3% of the vapor pressures in the ECHA database are at least 1.9 orders of magnitude below the COSMOtherm values and 22.3% are at least 1.9 orders of magnitude above the COSMOtherm values. The deviations are significantly higher for substances with a low vapor pressure (<10⁻⁶ Pa).

Relevant experimental data for the Henry’s law constant in water were submitted for 54 substances and included a total of 62 studies (Table S2, supplementary material 1). All of these studies could be analyzed by comparison with the COSMOtherm data. Relevant non-experimental data for the Henry’s law constant in water were submitted for 227 substances and included a total of 322 studies. Out of these, 314 could be analyzed by comparison with the COSMOtherm data. The experimental and non-experimental data agreed with a R-square value of 0.74 and 0.70, respectively (Fig. S11).

Relevant experimental data for the normal boiling point were submitted for 2263 substances and included a total of 4295 studies (Table S2, supplementary material 1). 4035 of these studies could be analyzed by comparison with the COSMOtherm data. Relevant non-experimental data for the normal boiling point were submitted for 1432 substances and included a total of 2431 studies. Out of these, 2152 could be analyzed by comparison with the COSMOtherm data. Figure 9 shows the experimental as well as non-experimental normal boiling points.

The experimental normal boiling points from the ECHA database with no qualifier or qualifier “ca.” show large deviations from the COSMOtherm values for some substances (Fig. 9 left). The overall R-square value is 0.67. 0.4% of the ECHA data are more than 50 °C above the COSMOtherm values; 12% (476 of 3851 data points) are more than 50 °C below the COSMOtherm data points. The highest deviations are visible for substances with (calculated) normal boiling points above 500 °C, which is also outside the range validated here with the data from the PHYSPROP database. A discussion of the accuracy of these points is provided in Sec. 4.2. Some of the calculated normal boiling points might be above the decomposition temperature of the substance. However, we were not able to check this in all cases because decomposition temperatures were not available for all substances. No differences in accuracy were observed regarding the applied methods (Fig. S14). This is also in line with the guidance on information requirements and chemical safety assessment - Chapter R.7a, where it is stated that any determination method may be used within the scope and applicability specifications.³¹ 

The non-experimental values for the normal boiling point from the ECHA database are shown in Fig. 9 right. The R-square value is 0.56. Similar to the octanol–water partition coefficient, the registration dossiers of several alkenes (this time eight alkenes) contained 46 or 92 non-experimental data points for the normal boiling point that ranged for all eight alkenes from 60 to 367 °C (vertical lines in Fig. 9 right). The normal boiling points were obtained from QSARs and read-across of structurally similar compounds that had different chain lengths and were actually not suitable for read-across.

Without these substances, 26% of the normal boiling points in the ECHA database are at least 50 °C below the COSMOtherm values and 8.0% are at least 50 °C above the COSMOtherm values. Again, the deviations are significantly higher for substances with a normal boiling point above 500 °C.

For the photodegradation in air with OH radicals, the results are often stored in fields like “Remarks on Results” that are not exported to IUCLID. It was therefore only possible to extract experimental rate constants from 34 studies although 195 were labeled as experimental studies, which indicates that there might be more information on the ECHA website. We therefore manually extracted additional results and added these to the IUCLID data. The finally analyzed data set contains experimental data for 110 substances and includes in total 141 studies (Table S2). 136 of these studies could be analyzed by comparison with the COSMOtherm data. At least 20 studies were also labeled as experimental although they originated from (Q)SARs or calculations.

Non-experimental data for the photodegradation in air with OH radicals were available in IUCLID for 167 substances and included a total of 183 studies. Out of these, 164 could be analyzed by comparison with the COSMOtherm data. No additional data points were extracted manually for the non-experimental data. The available data points agree for some substances quite well with the calculated values from COSMOtherm [Figs. S11(c) and S11(d)]. However, some studies report very unrealistic values like −12 cm³ molecule⁻¹ s⁻¹ (CAS RN 507-20-0) or 1.24 cm³ molecule⁻¹ s⁻¹ (CAS RN 420-46-2), which also causes the very low R-square value of 0.05.

The approach that was chosen here – validation of the COSMOtherm data with data from the manually curated PHYSPROP database and a subsequent validation of the data in the ECHA database with the COSMOtherm data – has the advantage that far more data can be validated in the ECHA database than if only the PHYSPROP data had been used (Table S2). COSMOtherm is therefore an important element of our approach because the calculations in COSMOtherm can be used to evaluate data where no other reliable measurements or (Q)SARs are available. This is for example visible for the 665 substances whose “best available” log₁₀ K_OW values deviate more than expected (based on the variation normally observed for log₁₀ K_OW values from the COSMOtherm values). For these 665 substances, only 60 have experimental data in the PHYSPROP database and could have been validated with these values (supplementary material 3).

Additionally, it is important to note that some of the values from the PHYSPROP database may also be contained in the ECHA database. Including these values directly in the validation of the data from the ECHA database leads to a bias, suggesting an overall accuracy of the data in the ECHA database that might not be real. This means that even if we see high R-square values for the comparison, this does not imply that all other values in the ECHA database would also be very accurate. However, for completeness, the comparison between the experimental ECHA data and those in the PHYSPROP database is shown in Sec. S4 of the supplementary material 1.

The data in the PHYSPROP database did not cover for all endpoints the entire range of values that were calculated by COSMOtherm. The following sections therefore discuss the accuracy of the values outside of this validated range. The air–water partition coefficients and the second-order rate constants for the photodegradation in air with OH radicals are not discussed, as the data points in the PHYSPROP database cover the same range as the COSMOtherm values for these endpoints.

The log₁₀ K_OW values in the PHYSPROP database cover the range from −3 to 10. For a few substances, COSMOtherm calculated log₁₀ K_OW values that are two log₁₀ units below or above this range. We assume that also the values outside this validated range are reasonable, because the values from COSMOtherm agree over the entire log₁₀ K_OW range very well with the data from the PHYSPROP database.

COSMOtherm calculated solubilities in water that are up to two log₁₀ units below the validated range (10⁻⁴ to 10⁶ mg/l), for a few substances even five orders of magnitude below. Especially the very low calculated solubilities in water (10⁻⁶ to 10⁻⁹ mg/l) are all below the corresponding experimental values, sometimes up to 5 or 6 orders of magnitude. Without a confirmation with experimental values, it is very difficult to judge how reliable the low solubilities in water calculated by COSMOtherm are. On the other hand, solubilities in water in the ng/l range are also not easy to measure and there is per se no reason why the COSMOtherm values should be incorrect as they are all based on the same calculation principle. We would therefore still recommend using the very low solubilities from COSMOtherm, but stating that they are associated with a higher uncertainty.

The vapor pressures from COSMOtherm in the range 10⁻¹⁸ to 10⁻⁸ Pa, which are outside the validated range, are all significantly below the experimental values from the ECHA database. However, COSMOtherm also calculated values below those from the ECHA database for 22% of the substances that are in the validated range of 10⁻³ to 10⁻⁸ Pa and for which the other data from COSMOtherm agree well with the experimental data from the PHYSPROP database. The vapor pressures below the validated range (<10⁻⁸ Pa) from COSMOtherm are therefore not necessarily incorrect, but show that there is an urgent need for accurate measurements in this low-volatility range. If those measurements are not feasible, then it would at least be good to indicate that the measured values are upper bounds and not exact values.

The calculated normal boiling points from COSMOtherm are for 31% of the substances considered here outside the range that could be validated with the data from the PHYSPROP database. It is therefore very difficult to judge how reliable the calculated normal boiling points above 350 °C from COSMOtherm are that deviate from the measured ones. The values from COSMOtherm outside the validated range should therefore be regarded as quite uncertain.

Our analysis shows that there are some endpoints, and ranges within the endpoints, where the COSMOtherm values deviate considerably from the experimental values in the ECHA database. A summary of the findings is given in Table 4. These include, e.g., log₁₀ K_OW values above 8, solubilities in water below 10⁻³ mg/l, vapor pressures below 10⁻⁶ Pa and normal boiling points above 500 °C. These are also the ranges for which various problems have been reported with measurements.^31,34–36 Specific examples included e.g., decabromodiphenyl ethane for which a measured log₁₀ K_OW of 3.55 has been reported,³⁷ but which is actually much higher (11.1–13.6) (Refs. 38 and 39). Also, for trisiloxane, 1,1,1,3,5,5,5-heptamethyl-3-tetradecyl-, a measured solubility in water of 19.1 mg/l has been reported,⁴⁰ but the solubility in water is expected to be much lower based on the structure of the compound. WSKOW v.1.4.1 in EPI Suite calculates for trisiloxane, 1,1,1,3,5,5,5-heptamethyl-3-tetradecyl-, a solubility in water of 6.5 × 10⁻⁹ mg/l. In the case of log₁₀ K_OW, S_W, and p_V, we presume that the large differences were mainly caused by inaccuracies and uncertainties in the measurements. We recommend therefore that newly developed measurement techniques should be applied for these ranges to generate more reliable data. Examples for those techniques for log₁₀ K_OW are the slow-stirring dual-flask/solid-phase microextraction (SPME) method³⁶ or correlating the octanol–water partition coefficient to measured n-butanol–water partition coefficients.³⁵ In terms of solubility, the slow-stir water solubility method⁴¹ has been developed for substances with solubilities below 0.1 mg/l and although the method is recommended in the ECHA Guidance³¹ for low-solubility test substances, it is unclear how often it has been really used for the substances registered under REACH; this is mainly because there is no differentiation between fast and slow-stirring methods in the IUCLID data.

Methods for vapor pressures as low as 10⁻¹⁰ Pa are available,³³ however none of the recommended methods performed very well in the in the range <10⁻⁸ Pa compared to the COSMOtherm data. The static method (which is not recommended for this low-vapor-pressure range) performed particularly poorly. We recommend therefore that at least the results of the static method for low-volatile substances should be repeated with other, more suitable methods.

The average deviations between the non-experimental data and the COSMOtherm values are for all endpoints greater than the average deviations between the experimental data and the COSMOtherm values. This shows that more guidance is needed that details which estimation methods or (Q)SARs should or can be used for which endpoint. It would probably also be useful to check/calculate if a substance is within the applicability domain of a model and rate those data that are outside the applicability domain as less certain.⁴² Read-across from substances with deviating carbon-chain lengths should be avoided, because the values for most of the endpoints increase or decrease with changing carbon-chain length.

The ECHA database is one of the largest and most important databases with physicochemical properties. However, the quality of the data is very variable, and we also identified certain ranges that show values (compared to the COSMOtherm values, which we assessed as reliable) that are systematically too low or too high. This can be problematic as, for example, a high log₁₀ K_OW value is an indicator for a bioaccumulative substance. An underestimation of log₁₀ K_OW, therefore, likely also leads to an incorrect hazard assessment. With the publication of the COSMOtherm data for more than 4400 substances, this work contributes therefore also to more accurate and trustworthy non-experimental physicochemical property data than currently available that can be used in the future for the hazard and risk assessment of these chemicals.

The supplementary material 1 is a pdf and contains additional data and graphics related to the methods and results in this article. Supplementary material 2 is an Excel workbook and contains the physicochemical property data calculated with COSMOtherm. Supplementary material 3 is also an Excel workbook and contains for each endpoint those datapoints from the ECHA database that were identified as outliers or values with high deviation.

J.G. acknowledges funding from the Swiss Federal Office for the Environment. We thank Oleksandr Yushchenko, Rachel London, Katharina Sodnikar, Maya Amacha, Narain Ashta, and Sarah Partanen (all ETH Zürich or previously ETH Zürich) for their help with the COSMOconf calculations and Elvira Rudin for interesting discussions.

The authors have no conflicts to disclose.

Data curation, formal analysis, methodology, visualization, and writing of the original draft was done by J.G. Conceptualization, funding acquisition, methodology, supervision, and reviewing and editing of the draft was done by M.S.