henhaobio
Sign in

Communications BiologyBioinformaticsPeer-reviewed journal article

Prediction of Hemolytic Peptides and their Hemolytic Concentration (HC50)

Anand Singh RathoreiD, Nishant KumariD, Shubham ChoudhuryiD, Naman Kumar MehtaiD, Gajendra P. S. RaghavaiD

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

Abstract

Several peptide-based drugs fail in clinical trials due to their toxicity or hemolytic activity against red blood cells (RBCs). Existing methods predict hemolytic peptides but not the concentration (HC50) required to lyse 50% of RBCs. In this study, we developed a classification model and regression model to identify and quantify the hemolytic activity of peptides. Our models were trained and validated on 1924 peptides with experimentally determined HC50 against mammalian RBCs. Analysis indicates that hydrophobic and positively charged residues were associated with higher hemolytic activity. Our classification models achieved a maximum AUC of 0.909 using a hybrid model of ESM-2 and a motif-based approach. Regression models using compositional features achieved R of 0.739 with R{superscript 2} of 0.543. Our models outperform existing methods and are implemented in the web-based platform HemoPI2 and standalone software for designing hemolytic peptides with desired HC50 values (http://webs.iiitd.edu.in/raghava/hemopi2/). HighlightsO_LIDeveloped classification and regression models to predict hemolytic activity and HC50 values of peptides. C_LIO_LIA hybrid model combining machine learning and motif prediction excels in accuracy. C_LIO_LIBenchmarking of the existing classification methods on independent datasets. C_LIO_LIWeb server, standalone software, and pip package for hemolytic activity prediction of peptides/proteins. C_LI

In-depth summary

Claude Sonnet 4.6 · max effort · 8 Nov 2026 · from the full text of version 1 (24 Jul 2024) · CC BY 4.0

HemoPI2.0 predicts peptide hemolysis and pHC50 from sequence; a hybrid of MERCI motif scoring and ESM2-t6 reaches AUC 0.909 on held-out peptides, beating the earlier tools tested.

Earlier hemolysis predictors only classify peptides and rely on training sets generalized across vertebrates, and none estimates HC50, the concentration that lyses half of red blood cells. This work trains on experimentally measured HC50 values to give both a hemolytic label and a regression estimate.

Key results

  1. ShownSequence-feature and language-model classifiers separate strongly hemolytic (HC50 ≤ 100 μM) from weakly hemolytic peptides with similar AUCs on the held-out 20% set. Random forest (RF) on Pfeature descriptors was best among feature-based models.

    Independent-set AUC: RF on ALLCOMP excluding SOC (1090 descriptors) 0.888; AAC+DPC+PCP (450 descriptors) 0.881; AAC+DPC (420 descriptors) 0.875; ProtBERT PLM 0.875; MLPC on ProtBERT embeddings 0.882. · 3.2.1 Machine Learning Models; 3.2.2 Protein Language Models; 3.2.3 ML models based on word embeddings

  2. ShownAdding MERCI motif scores (+0.5 hemolytic, -0.5 non-hemolytic, 0 no match) to ML and PLM predictions raised AUC on the independent set. The ESM2-t6 hybrid was best.

    Hybrid AUC on independent dataset: ESM2-t6 + MERCI 0.909; RF (ALLCOMP excluding SOC) + MERCI 0.906, versus 0.888 for RF alone. · 3.3 Hybrid model for classification; 3.5 Model Finalization

  3. ShownOn the same independent dataset, HemoPI2.0 clearly outperformed the existing classifiers that could be run. Four other tools could not be evaluated because of service or code problems.

    Independent-set AUC: HAPPENN 0.736, PeptideBERT 0.739, Plisson et al. 0.740 versus HemoPI2.0 0.909. · 3.6 Benchmarking

  4. ShownRandom forest regression on Pfeature descriptors predicts pHC50 best. ESM2-t33 embeddings with Extra Trees were somewhat weaker, and no prior tool predicts HC50 for comparison.

    Independent set: RFR on ALLCOMP excluding SPC (1167 descriptors) R 0.739 and second metric 0.543; ETR on ESM2-t33 embeddings R 0.711 and second metric 0.495 (both second values are printed as R in the text). · 3.4.1 ML Regressors using Pfeature; 3.4.2 ML models based on word embeddings; 3.5 Model Finalization; 3.6 Benchmarking

  5. SuggestedHemolytic peptides are enriched in hydrophobic residues (Cys, Phe, Gly, Leu) and specific motifs, while positively charged residue composition correlates with higher HC50 (lower potency). This is descriptive and no correlation coefficients are quoted in the text.

    Top hemolytic-only motifs include CGET and TLLKKVLKA; non-hemolytic-only motifs include AKD and SKIK; AAC_K, AAC_R and DPC1_KK correlate positively with HC50, and AAC_G and DPC1_IG negatively. · 3.1.1 Amino acid composition analysis; 3.1.3 Motif analysis; 3.1.4 Correlation analysis; 4. Discussion

Methods, data and reproducibility

Peptides with experimental HC50 were curated from two databases, encoded with Pfeature descriptors (1092 features) or PLM embeddings, and modeled with ML classifiers (HC50 ≤ 100 μM) and pHC50 regressors. A MERCI-motif hybrid was added, and all models were tested on a held-out 20% set.

  • Peptides from DBAASP v3 with available HC50 values · n = 3147 peptides
  • Peptides from the Hemolytik database with available HC50 values · n = 560 peptides
  • Final curated HemoPI2 set (natural residues, at least six residues, mean HC50 per sequence): 891 hemolytic, 1035 non-hemolytic · n = 1926 unique peptides
  • Random split: 80% for five-fold cross-validation and training, 20% held-out independent evaluation set · n = 80% / 20% of the dataset
Methods & tools
Pfeature, MERCI motif tool, Two Sample Logo, ESM-2 (t36, t33, t30, t6 checkpoints), ProtBERT, BioBERT, ProtBERT + BiLSTM, Random Forest / Extra Trees / XGBoost / SVM / MLP, DBAASP v3, Hemolytik, HuggingFace model checkpoints
Software & environment
Source databases DBAASP v3 and Hemolytik; ESM-2 checkpoints esm2_t36_3B_UR50D, esm2_t33_650M_UR50D, esm2_t30_150M_UR50D, esm2_t6_8M_UR50D (HuggingFace); pip package and standalone software offered. No library versions stated.
Reproducibility gaps
Hyperparameters, software versions and hardware are not given in the text. Curated data are on the web server, with the full table in a supplementary file. Several competitor tools could not be re-run.
Limitations (4)
  • HLPpred-FUSE, HemoNet, haemolytic-Pred and the Capecchi et al. RNN could not be benchmarked because of broken web services or GitHub code missing input files, so the comparison is incomplete. · stated by the authors
  • The web server accepts only peptides built from the 20 natural amino acids; non-natural residues are not supported. · stated by the authors
  • The text gives 1926 peptides in the dataset section but 1924 in the classification and discussion sections, and does not explain the difference. · noted in this summary
  • Evaluation rests on a single random 80/20 split without confidence intervals, with HC50 averaged across experimental conditions and a fixed 100 μM cutoff, so the effect of these choices on the reported AUC and R is unclear. · noted in this summary

Read this if you need a sequence-only predictor of peptide hemolysis or HC50, with benchmark AUCs against HAPPENN, PeptideBERT and other tools and public code.

Citations
78
FWCI
23.02
Field percentile
top 1%

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
THISSingh '26Rathore '24Saraswat '24Lavertu '26Kaur '24Naorem '23
60% semantic matchBioinformatics
Ayushi Singh, Kavin Raj SA, Anand Singh Rathore+1 more·Chemical Research in Toxicology·Jan 12, 2026· 11 citations