Fung-AI generates candidate antifungal peptides with experimental hits

Per the bioRxiv preprint, Fung-AI is an AI/ML-driven pipeline that generates candidate antifungal peptides using a modified GAN framework and subjects candidates to downstream computational filtering before synthesis and experimental testing (bioRxiv). The authors report synthesizing 13 peptide candidates, of which five inhibited fungal growth in vitro and two showed low cytotoxicity to human hepatic cells, according to the preprint (bioRxiv). The PREreview archived on Sciety and an accompanying ResearchGate figure indicate the downselection pipeline included multiple antifungal classifiers and a hemolytic classifier to reduce predicted toxicity and prioritize candidates (Sciety; researchgate.net). The PREreview also notes that the tested peptides had higher minimum inhibitory concentration (MIC) values relative to existing antifungal peptides (Sciety).

Per the preprint, the core generative component is a modified GAN trained on available antifungal peptide data, followed by classifier-based rescoring and toxicity filtering prior to choosing sequences for synthesis (bioRxiv). Groups building peptide discovery pipelines commonly combine generative models with orthogonal computational filters to improve experimental hit rates; using separate activity and toxicity classifiers is a standard approach to trade off potency against safety during in silico prioritization. This hybrid strategy reduces the search space but depends heavily on the size, diversity, and label quality of training datasets, which the PREreview identifies as a constraint for antifungal peptides specifically (Sciety).

The PREreview frames this work against a backdrop of rising resistant fungal infections and slower therapeutic discovery compared with antibacterial agents; the authors present Fung-AI as a proof-of-concept to accelerate peptide discovery using AI-guided design and wetlab validation (Sciety). For ML practitioners focused on biologics, the study is notable because it moves beyond purely in silico benchmarks to include synthesis and biological assays, providing an end-to-end example of model-to-experiment workflow. However, the PREreview flags that the observed activity may reflect convergence on canonical antimicrobial peptide (AMP) features imposed by downstream filters rather than novel antifungal-specific motifs, which reduces confidence that the model alone uncovered new mechanistic features (Sciety).

Follow whether the authors or independent groups scale training data and diversify assay panels to lower MIC values and improve potency-to-toxicity balance. Critical indicators of progress will include expanded training sets with better negative controls, orthogonal assays across fungal species and human cell types, and replication of hits by independent labs.

The PREreview emphasizes limited dataset size and higher MIC values of validated peptides compared to existing antifungal peptides as the major constraints on immediate therapeutic relevance (Sciety). The review also questions whether downstream filtering biases selections toward well-known AMP properties rather than new antifungal mechanisms (Sciety). Overall, the sources document a methodologically careful proof-of-concept that combines generative modeling with experimental validation, while reviewers highlight data limitations and potency gaps that keep this work in an exploratory stage (bioRxiv; Sciety).