Trained checkpoints for a conditional VAE that generates antimicrobial peptide (AMP)
sequences targeting a user-specified potency (MIC, minimum inhibitory concentration)
against E. coli, plus an independently trained CNN ("the Judge") that predicts MIC
from sequence and is used to evaluate generated candidates.
Generator (CVAE): a bidirectional, 3-layer GRU encoder (256 hidden units) maps a peptide
sequence + a shared learned embedding of the normalized target log10(MIC) to a 32-dimensional
diagonal-Gaussian latent. A 3-layer unidirectional GRU decoder (256 hidden units) is
re-conditioned on the latent sample and the score embedding at every timestep, generating
logits over a 22-symbol vocabulary (20 amino acids + PAD + EOS) autoregressively. Trained with
a β-rescaled, free-bits ELBO objective (free bits = 0.1, β annealed over 50 epochs) and 30%
word dropout to prevent posterior collapse.
Judge: parallel 1-D convolutions (kernel sizes 3, 5, 7) extract motifs at different
receptive fields, concatenated to 128 channels, expanded to 256, passed through a residual
block (with a 1×1-conv shortcut) back down to 128 channels, global-max-pooled, and regressed to
a scalar (normalized log10 MIC) through a small MLP head. Trained independently of the CVAE,
purely as a post-hoc evaluator — it never sees the conditioning score.
Usage
Requires the model definitions from the GitHub repo
(src/models/cvae.py, src/models/judge.py) and its data/dataset.py for the
vocabulary/encoding utilities.
python
1import torch
2from huggingface_hub import hf_hub_download
3from models.cvae import CVAE
4from models.judge import Judge
5from data.dataset import decode_sequence, normalize_score, denormalize_score
67cvae_path = hf_hub_download("Sloudis/controllable-amp-design","cvae_best.pt")8judge_path = hf_hub_download("Sloudis/controllable-amp-design","judge_best.pt")910cvae = CVAE()11cvae.load_state_dict(torch.load(cvae_path, map_location="cpu"))12cvae.eval()1314judge = Judge()15judge.load_state_dict(torch.load(judge_path, map_location="cpu"))16judge.eval()1718# See src/evaluation/generate.py in the GitHub repo for a full generation CLI,19# including score normalization against the training set's log_mic mean/std.
Evaluation
Measured on a held-out test set (1,000 sequences):
Judge: Spearman ρ = 0.70, Pearson r = 0.73 (MIC prediction from sequence alone)
Generator: 99.8% valid, 100% novel (not present in training data) sequences;
amino-acid composition matches natural AMPs
Conditioning accuracy: strongest in the densely-sampled mid-potency range
(~65% hit rate within ±0.5 log10 MIC units), degrading toward the extremes of the
potency range where training data is scarce
Limitations
Conditioning-accuracy numbers are only as good as the Judge itself (ρ=0.70, not a
ground-truth oracle) — no generated peptide was synthesized or tested against live bacteria.
Trained and conditioned on E. coli MIC only; says nothing about Gram-positive activity,
selectivity, hemolysis/cytotoxicity, or synthesizability.
Conditioning reliability degrades toward the extremes of the potency range (see the report,
Sec. 6.3).
Citation
Stavros Loudis. "Controllable Antimicrobial Peptide Design via Conditional Variational
Autoencoders." Technical University of Crete, 2026.
