How to prepare my own training set and known amp sequences

[aslangabriel99]

Hi， I want to use your tools to design new peptides against Klebsiella pneumoniae, please give me some tutorials on how to prepare my own training set and known amp sequences? Could I use your tools to make it work? should I change configure files? Thanks a lot.

[mark0428]

Hi @aslangabriel99 ,

Thank you for reaching out with your questions. We have recently updated AMPd-Up to version 1.0.1, which now allows users to train the models using their own data.

To do this, simply specify the path to your training data directory using the -tr argument (the file should be in fasta format).

Hope this helps!

Best,
Chenkai

[aslangabriel99]

Thank you for your help. I still have another problem how to prepare my own training data? I asked chatgpt, it indicated that I should follow the processes such as "

1. Preprocessing Sequences
   Quality Filtering: Remove sequences with ambiguous amino acids or sequences below a given length threshold.

Redundancy Removal: Use tools like CD-HIT to cluster and remove redundant sequences; this helps to avoid training bias.

Normalization: Ensure that the sequences are in a standard format and consistent in terms of case and header conventions.

2. Annotation and Feature Extraction
   Physicochemical Features: Calculate properties such as charge, hydrophobicity, isoelectric point, and secondary structure tendencies. Libraries like Biopython or packages in R (e.g., ‘Peptides’ package) can be used.

Sequence Encoding: Depending on AMPd-Up’s requirements, encode the sequences using methods such as amino acid composition, dipeptide frequencies, or even more advanced embeddings if supported.

3. Data Partitioning
   Training vs. Testing: If you plan to use machine learning components in AMPd-Up, separate your dataset into training, validation, and testing sets, ensuring a balanced representation of active versus inactive peptides.

Cross-validation Setup: Prepare scripts or workflows to perform cross-validation if required to tune model hyperparameters.”. Would you mind giving me some advice? I'll appreciate it a lot.

[mark0428]

Hi @aslangabriel99 , please find my suggestions below, per bullet point:

1. Preprocessing Sequences:
   Retain only sequences composed of standard amino acids, as AMPd-Up does not support non-standard residues. Remove redundant sequences that are exactly identical. Keeping similar (but non-identical) sequences is acceptable—this follows the original AMPd-Up approach—if your final model is capable of generating a diverse set of sequences. Otherwise, consider removing highly identical sequences during preprocessing.

   I assume your training set consists of lab-validated AMPs that are active against Klebsiella pneumoniae? I'm not sure how large this set is, but ideally, it should contain at least several hundred samples to get AMPd-Up well-trained. If the set is too small, AMPd-Up may struggle to learn effectively.

2. Annotation and Feature Extraction:
   Not necessary. AMPd-Up only takes sequence information as input and is expected to learn relevant properties on its own.

3. Data Partitioning:
   If your training data consists entirely of AMPs, then this is not necessary. However, if you wish to perform more precise hyperparameter tuning, you can adapt the AMPd-Up code to perform cross-validation—though this is optional.

[aslangabriel99]

Thank you very much for your patient and comprehensive answer to my previous inquiry about AMPd‑Up. I greatly appreciate the time and expertise you have devoted to helping me understand your platform.

As I prepare my own training set, I have a question regarding the treatment of chirality during peptide‐sequence tokenization. In the cAMPr4 database, entries such as CAMPSQ16980 (a cyclic peptide with sequence “dYdF,” where the lowercase “d” denotes D‑amino acids) are annotated as “l = L form of amino acid.” When you tokenize peptide sequences for model input, how do you encode or distinguish L‑ versus D‑residues? In other words, what conventions or symbols do you use to preserve chirality information in the token stream, and do you recommend any best practices for representing D‑amino acids when constructing a machine‐learning–ready dataset?

Thank you again for your invaluable guidance.

[mark0428]

AMPd-Up only accepts or generates sequences composed of standard L-amino acids. Therefore, sequences containing D-amino acids should be removed from the training set. Alternatively, if your training set is too small, D-amino acids may be treated as their corresponding L-forms as a compromise.

[aslangabriel]

Thank you for your patience. I have successfully applied your tool to the design of common linear antimicrobial peptides (AMPs). However, I am now interested in extending its use to the design of cyclic peptides and am uncertain about the feasibility of this approach. Could you kindly provide some guidance or suggestions on this matter?

[mark0428]

Thank you for your message, and I'm glad to hear you have successfully applied AMPd-Up to the design of common linear AMPs. While the training set may include a few cyclic peptides, the current version of AMPd-Up is not designed to generate peptides with explicit cyclic structures.

If you are interested in cyclic peptide design, a feasible approach would be to curate a dataset of cyclic AMPs and train a custom model tailored to that structure type—assuming you have access to a sufficient number of well-characterized examples. I'd be happy to offer further guidance if you decide to pursue this direction.