AIPAMPDS

AIPAMPDS is a platform for designing antimicrobial peptides that couples a generative sequence model to a screening cascade for potency and toxicity. Its core is a compact GPT-style transformer decoder pretrained on a corpus of natural antimicrobial peptides and then fine-tuned on minimum inhibitory concentration data for specific pathogens, so that generation can be aimed at Escherichia coli, Staphylococcus aureus, or both. It was developed at China Pharmaceutical University with collaborators at Nanjing Agricultural University and posted to bioRxiv in March 2025.

Antimicrobial peptides are attractive against multidrug-resistant infections because they attack bacterial membranes rather than a single enzymatic target, which slows resistance. Their clinical translation stalls for two reasons the platform addresses together: the generative models most often applied to peptide design, GANs and VAEs, suffer from mode collapse and constrained diversity, and design pipelines typically optimize antimicrobial activity while treating hemolytic toxicity as a downstream afterthought. AIPAMPDS replaces the generator with an autoregressive decoder and makes hemolysis a first-class filter, applied to every candidate alongside activity.

Scope is deliberately narrow. The generator is small and was pretrained on roughly twenty thousand peptides rather than a protein-scale corpus, and target conditioning covers only the two organisms with sufficient public MIC data. Availability is split down the middle: the companion repository ships prediction scripts and model definitions for all four screening models — with the HemoRisk-Estimator weights committed in-repo and the remaining checkpoints served over Dropbox — while the GPT generator is reachable only through the hosted web server, with neither code nor weights released. The repository carries no license file, so the published scripts come with no explicit reuse terms.

#Key Features

• Target-conditioned generation: Three fine-tuned variants of the same pretrained decoder produce broad-spectrum, E. coli-directed, or S. aureus-directed peptides, selected by a radio button at generation time.
• Four-stage screening cascade: Candidates pass sequentially through an antimicrobial-activity classifier, two organism-specific MIC classifiers, and a hemolysis predictor, so activity and safety are evaluated in the same pass.
• Specificity-first thresholds: The activity classifier and the MIC classifiers both use a decision threshold of 0.998; for the MIC models that holds the false-positive rate below 0.02, at the cost of recall.
• Protein language model embeddings: The activity classifier builds on ProtTrans' ProtBERT, retaining the full token sequence rather than global average pooling before a convolutional and LSTM head.
• Experimental validation: Generated and natural candidates were both carried through synthesis and assay, giving a direct comparison of AI-designed peptides against metagenome-mined ones.

#Technical Details

The generator is a transformer decoder of eight stacked blocks; each block applies masked self-attention producing a 256-dimensional vector, expands it to 1,024 dimensions through a feed-forward network with GELU activation, and projects back to 256. Pretraining used 22,122 antimicrobial peptides for 30 epochs, followed by 30 epochs of fine-tuning with AdamW on MIC records covering 3,106 E. coli sequences, 5,266 S. aureus sequences, or 6,239 combined. The screening models are separate architectures: AMP-Identifier pairs ProtBERT embeddings with a convolutional layer, max pooling, and an LSTM, trained on 28,508 antimicrobial and 28,513 non-antimicrobial sequences; EcoliMIC-Classifier and StaphMIC-Classifier are TextCNN models with kernel sizes 2 and 4 over a 192-dimensional embedding of 50-residue padded sequences; and HemoRisk-Estimator combines convolution, max pooling, an LSTM, and an attention layer over 3,977 peptides annotated for hemolysis. On a held-out split, AMP-Identifier reached 99.18% accuracy and 0.99 ROC AUC against amPEP, Macrel, AMPlify, and AMPscanner, and HemoRisk-Estimator reached 84.67% accuracy against 78.64% for an embedding plus Bi-LSTM baseline and 76.63% for PeptideBERT. In validation, 100,000 sequences from the E. coli generator were reduced to 82 by screening, then to five by requiring at least 75% alpha-helix content under DSSP and a compact predicted fold from AlphaFold 2.

#Applications

The platform is aimed at microbiologists and medicinal chemists who want candidate peptides without building a generative pipeline themselves. The browser interface takes no registration, lets a user set the number and maximum length of sequences to generate, and returns each candidate annotated with length, molecular weight, net charge, and instability index, exportable as CSV. The screening arm accepts user-supplied FASTA, so peptides from any source — including metagenome mining — can be triaged for activity and hemolytic risk before committing to synthesis.

#Impact

The wet-lab results are the substantive contribution. Five generated and five metagenome-derived candidates were synthesized and tested; every natural candidate was active against both Gram-positive and Gram-negative strains, every generated candidate was active against Gram-negative strains as its conditioning intended, and nine of the ten showed negligible hemolysis of mouse erythrocytes at 128 µg/mL. One generated peptide, GAMP_4, inhibited all six ESKAPE pathogens at MICs of 8-64 µg/mL while causing 10.76% hemolysis at that concentration. The evidence base remains small — ten peptides, two target organisms, in vitro assays only — and the preprint has not been peer reviewed, so the platform is best read as a usable design tool with encouraging early validation rather than a settled benchmark.