Enzymes

The input for an enzyme must be a string containing the enzyme's amino acid sequence. The model currently does not handle missing information, such as the '*' character, and including it will prevent the model from proceeding.

Metabolites

There are three valid input types for the metabolite: SMILES string, KEGG Compound ID, and InChI string

InChI string

InChI strings are textual representations of chemical structures. Every InChI string is a unique identifier and contains detailed information about the structure of a small molecule. For more details on InChI, see this page from IUPAC.

KEGG Compound ID

The KEGG Compound database contains identifiers for many small molecules and drugs. A KEGG Compound ID starts with a "C" or "D" followed by a five-digit number. For more information see the KEGG homepage.

SMILES

Simplified Molecular Input Line Entry Specification (SMILES) allows to represent the structure of a molecule using ASCII strings. You can get the SMILES for a molecule e.g. by searching for the molecules name in PubChem. Since SMILES representations are not unique for all molecules, we recommend to use InChI string or KEGG Compound IDs instead, if possible.

Case study: ESP (Enzyme-Substrate Pair Prediction) step by step

If you are interested in predicting the interaction between a single enzyme and a single metabolite, you can use the single input option. You can enter the enzyme sequence and the metabolite in the respective fields and click on the "Predict" button. The metabolite can be entered as a SMILES string, KEGG Compound ID, or InChI string. To see an example, you can click on "Fill Fields with Example".

Once you have submitted, the job will be processed, and you will see the job status. If your position in the queue is 0, your job will be the next one to be processed. It can take up to a minute to get your result.

In the Main Results tab, the model will return a prediction score for the interaction between the enzyme and the metabolite. In the example above, the value is 0.96, which means the model predicts that the metabolite is likely a substrate for the enzyme. The prediction score is a value between 0 and 1. Scores close to 1 mean that the model predicts the metabolite is likely a substrate for the given enzyme, whereas scores close to 0 mean the model predicts that the metabolite is unlikely to be a substrate for the enzyme.
Prediction scores close to 0.5 (i.e., scores in the range of 0.3 to 0.7) should be considered with caution. The prediction model is unsure which class to predict in these cases. The model also returns how often the metabolite was present in the training set. In this case, the metabolite was present in the training set 8 times. We have shown that the prediction performance of our model is low when applied to metabolites that were not present in our training set. Therefore, we check whether every uploaded metabolite was part of our training set.

If you go to the "Metabolite Information" tab, you will see all proteins experimentally reported to interact with the metabolite. We provide a copy button for the amino acid sequence of the protein and a link to its UniProt page.

In the same way, if you go to the "Enzyme Information" tab, you will see all molecules experimentally reported to interact with your input protein.

Finally, the "Kinetic Parameters" tab informs you about the kinetic parameters of the enzyme-metabolite pair. These predictions are meaningful only if the ESP prediction score is above 0.5. To estimate k_cat, the model used the amino acid sequence information only, Without any information about the chemical reaction, this may lead to potentially underperforming predictions. The page will also display how these predicted kinetic values compare to all enzyme values from our datasets, through histograms.


TurNuP (k_cat prediction)

Predicting the turnover number k_cat of an enzyme-reaction pair works similarly to the case study above. The main difference is that, to optimize the prediction, the model requires information about the catalyzed reaction. This is done by providing several substrates and products, using semicolons as separators.

On average, predictions for new enzyme-reaction pairs deviate from the true k_cat value by a factor of 4.9 (see our manuscript for more details). On an independant test set the model achieve a coeficient of determination of R² of 0.44If similar enzymes or similar reactions were part of our training set, the model's accuracy increases. As an estimate of model performance, the single input option outputs an enzyme sequence identity and a reaction similarity score compared to the training data. We also provide additional information and predictions, such as other sequences catalyzing the same reaction, other reactions catalyzed by the same enzyme, and, finally, an estimate of the Michaelis-Menten constant for the substrate of the reaction. You can use the "Example" button to fill the fields with an example and examine the different outputs.

K_M prediction

Predicting the Michaelis constant K_M of an enzyme-substrate pair works similarly to the case study above. On average, predictions for new enzyme-substrate deviate from the true K_M value by a factor of 4.1 (see our manuscript for more details). On an independent test set, the model achieves a coefficient of determination R2R2 of 0.53. If similar enzymes or similar molecules were part of our training set, the model's accuracy increases. We also provide additional information and predictions, such as other sequences catalyzing the same substrate, other molecules catalyzed by the same enzyme, and, finally, an estimate of the turnover number k_catt​ for the enzyme-substrate pair. Here, k_cat is estimated solely based on the amino acid sequence and might therefore be less accurate than when reaction information is provided. You can use the "Example" button to fill the fields with an example and examine the different outputs.

SPOT (Transporter - Substrate Pair prediction)

Predicting the interaction between a transporter and a substrate works similarly to the case study above. The prediction score is a value between 0 and 1. Scores close to 1 mean that the model predicts that the molecule is likely a substrate for the given transporter, whereas scores close to 0 mean that the model predicts that the molecule is likely not a substrate for the transporter. Prediction scores close to 0.5 (i.e. scores in the range of 0.4 to 0.6) should be considered with caution. The prediction model is unsure which class it should predict in these cases. We have shown that the prediction performance of our model is better when substrates are present in our training set. Therefore, we check for every uploaded molecule if it was part of our training set. We also provide additional information, such as other transporter of the same substrate and other molecules that are substrates of the same transporter. You can use the "Example" button to fill the fields with an example and examine the different outputs.

Switch from CLS to XLSX format

Starting from December 9, 2024, we have switched from using CLS files to XLSX files specifically for k_cat predictions. XLSX files can be created using spreadsheet programs like Microsoft Excel or Google Sheets, or through pandas library in python. For more details on how to create an XLSX file, you can refer to Microsoft Excel or Google Sheets.

How should your file look like?

Your file format depends on the model you are using. Attention: InChI strings can contain commas (","), so be sure to properly structure your data in the required format.
You can download a sample file for each model below.



Example of multiple inputs with a file. The enzyme-substrate pairs and metabolites displayed here are not real.

Enzyme-Substrate Pair Prediction:

Your file must be in XLSX format and contain exactly two columns, one called "Protein" and one called "Metabolite". Each row should contain one enzyme and one metabolite in the format described above. The upper limit of accepted enzyme-metabolite pairs is 500. You can download a sample file here. We have shown that the prediction performance of our model is low when it is applied to metabolites which were not present in our training set. Therefore, we check for every uploaded metabolite if it was part of our training set. We return this information in the column "metabolite".

k_cat prediction:

Your file must be in XLSX format and contain three columns, titled "Enzyme", "Substrates", and "Products". Each row should contain one enzyme-reaction pair in the format described above. For both columns Substrates and Products, metabolites should be separated by a semicolon ";". The upper limit of accepted enzyme-reaction pairs is 500. You can download a sample XLSX file here.

K_M prediction:

Your file must be in XLSX format and contain exactly two columns, one called "Enzyme" and one called "Substrate". Each row should contain one enzyme and one metabolite in the described format. The upper limit of accepted enzyme-metabolite pairs is 500. You can download a sample file here.

SPOT:

Your file must be in XLSX format and contain exactly two columns, one called "Protein" and one called "Metabolite". Each row should contain one transporter and one molecule in the format described above. The upper limit of accepted transporter-molecule pairs is 1000. You can download a sample file here.