henhaobio
Sign in

bioRxivBioinformaticsPreprint

Deep learning regression model for antimicrobial peptide design

Jacob J. WitteniD, Zack Witten

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Antimicrobial peptides (AMPs) are naturally occurring or synthetic peptides that show promise for treating antibiotic-resistant pathogens. Machine learning techniques are increasingly used to identify naturally occurring AMPs, but there is a dearth of purely computational methods to design novel effective AMPs, which would speed AMP development. We collected a large database, Giant Repository of AMP Activities (GRAMPA), containing AMP sequences and associated MICs. We designed a convolutional neural network to perform combined classification and regression on peptide sequences to quantitatively predict AMP activity against Escherichia coli. Our predictions outperformed the state of the art at AMP classification and were also effective at regression, for which there were no publicly available comparisons. We then used our model to design novel AMPs and experimentally demonstrated activity of these AMPs against the pathogens E. coli, Pseudomonas aeruginosa, and Staphylococcus aureus. Data, code, and neural network architecture and parameters are available at https://github.com/zswitten/Antimicrobial-Peptides.

In-depth summary

Claude Sonnet 4.6 · max effort · 8 Nov 2026 · from the full text of version 1 (12 Jul 2019) · CC BY 4.0

A CNN trained on 4050 E. coli peptide MIC records predicts activity, outperformed other AMP classifiers with random negatives, and guided two designed peptides active against three pathogens.

Earlier AMP machine learning mostly classified peptides, optimized around one known scaffold, or generated sequences with no activity prediction. This work pairs a large MIC database with a global regression model, so designed sequences carry a quantitative predicted MIC.

Key results

  1. ShownThe GRAMPA database merges MIC records from five public AMP databases. E. coli has the most data. Log MIC correlates more strongly within a gram class, and C. albicans correlates poorly with bacteria.

    6760 unique sequences and 51345 MIC measurements; at least 700 measurements for 10 microbes; 4559 E. coli peptides after preprocessing; each correlation used at least 50 measurements. · GRAMPA; Preprocessing; 3.1 Dataset characterization; Figure legends

  2. SuggestedA convolutional network predicted log MIC better than ridge regression and k-NN baselines, and it was insensitive to hyperparameters. It also beat the recurrent variants that were tried.

    Compared with two ridge models (amino acid counts; one-hot, 874 features) and k-NN (edit distance chosen). The text says the CNN was the clear winner but gives no numeric Table 1 values. · 3.2 Design of machine learning architecture; Regularized linear model; k-NN analysis and sequence alignment; Neural network model

  3. SuggestedUsed as a classifier with random-peptide negatives, the CNN ensemble outperformed other AMP classifiers. With UniProt negatives it ranked second after AMP Scanner v2. The comparison is confounded by training-set size and negative-data choice.

    Training sizes: 1778 (AMP Scanner v2), 2578 (CAMP), 3417 (iAMPpred), 4050 (this model). Second best by MCC and AUC with UniProt negatives; 10:1 negative ratio exceeded AMP Scanner v2; k-NN (k=7) only slightly worse than the CNN. · 3.3 Classification performance; k-NN analysis and sequence alignment; Tables

  4. ShownSimulated annealing with the 30-network ensemble produced peptides with low predicted MIC and elevated alpha-helical hydrophobic moment. The 140° turn control showed no such elevation, so the model appears to have learned helical amphipathicity.

    Each peptide compared with 1,000 shuffles; HM percentiles 99.0 (CNN-SA1) and 97.8 (CNN-SA2); predicted log MIC −0.2 and −0.11; 100,000 steps per run; ensemble 6×5=30 networks. · Ensemble model; 2.3 Simulated annealing for sequence design; 2.4 Hydrophobic moment analysis; 3.5 Peptide design

  5. SuggestedTwo designed, synthesized peptides were active against E. coli, P. aeruginosa and S. aureus. Their E. coli MICs were better than most active AMPs in the dataset.

    E. coli MICs lower than 79% of active AMPs in the dataset; activity also seen against S. aureus UAMS-1 and P. aeruginosa PAO1; n=2 peptides, three technical replicates; per-peptide MICs are in Table 3, not in text. · 2.5 Minimum inhibitory concentration…; 3.5 Peptide design; 3.6 Experimental validation; Tables

Methods, data and reproducibility

A 1D CNN on one-hot encoded peptides regresses log MIC against E. coli. It is trained with synthetic inactive negatives (log MIC 4) in a 30-network ensemble, then used as the objective for simulated-annealing peptide design.

  • GRAMPA: MIC records scraped from APD, DADP, DBAASP, DRAMP and YADAMP (Spring 2018) · multiple species (bacteria; C. albicans) · n = 6760 unique sequences; 51345 MIC measurements
  • Preprocessed E. coli log MIC peptides (geometric mean per pair); 3404 without cysteine, 1155 with cysteine · E. coli · n = 4559 peptides
  • Held-out test set (no-cysteine AMPs, length 46 or less) · E. coli · n = 499 AMPs
  • Training and validation split of remaining no-cysteine AMPs; recombined for final training · E. coli · n = 2316 train, 579 validation; 2895 or 4050 final
  • Negative data from UniProt cytoplasmic proteins, CD-HIT filtered at 40% identity; length-matched substrings
  • Experimental MIC validation of two designed peptides (CNN-SA1, CNN-SA2) by broth microdilution · E. coli BL21; P. aeruginosa PAO1; S. aureus UAMS-1 · n = 2 peptides; three technical replicates
Methods & tools
Convolutional neural network (Adam, mean squared error), Ensemble voting of 30 networks, Simulated annealing, sklearn RidgeCV, Biopython pairwise2 (ident and PAM30 matrices), k-NN (Levenshtein distance), CD-HIT, HeliQuest, AMP Scanner v2, iAMPpred, CAMP database classifiers, Broth microdilution MIC assay
Software & environment
Python sklearn RidgeCV; Biopython pairwise2 with matlist.ident and matlist.pam30; strains E. coli BL21, P. aeruginosa PAO1, S. aureus UAMS-1; BBL Mueller-Hinton II broth (BD Falcon). No package versions stated.
Experimental conditions
No computational hardware or runtime stated. Assay: 96-well broth microdilution, 37°C for 24 hours, 100 μM to 100 nM, inoculum approximately 3×10^4 CFU/mL, three technical replicates.
Reproducibility gaps
No hardware, runtime or package versions reported. Numeric results for Table 1 and Table 3 are not in the text. DADP is no longer maintained, and the designed peptide sequences are not printed in the text.
Limitations (6)
  • Negative training data reduced regression accuracy slightly, and some active peptides were predicted inactive (log MIC around 4), so part of AMP sequence space is inaccessible to the design algorithm. · stated by the authors
  • Classifier comparisons differ in training set size, labels (log MIC vs binary) and negative data (random peptides vs UniProt), so they do not isolate algorithm quality. · stated by the authors
  • Many designed peptides closely resemble known AMPs, with frequent LAK repeats from the training data; sequence diversity must still be balanced against predicted activity. · stated by the authors
  • The recurrent models may simply have been slower to train, and with more data architectures capturing long-range dependencies might show advantages. · stated by the authors
  • The designed peptides have not undergone experimental local sequence optimization, and the model does not predict toxicity, which the authors call a key future design criterion. · stated by the authors
  • Only two peptides, picked by manual inspection of predictions below log MIC 0, were tested, with technical replicates only. This gives no estimate of the design hit rate. · noted in this summary

Read this if you want a deep-learning MIC regression model trained on aggregated AMP data, with a simulated-annealing design loop and benchmarks against AMP classifiers.

Citations
100

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDong '25Majumder '18Bai '26Pandey '23Feng '24Hu '25
78% semantic matchBioinformatics
Ruihan Dong, Qiushi Cao, Chen Song·Advanced Science·Sep 12, 2025· 11 citations