PREreview of Fung-AI: An AI/ML-driven pipeline for antifungal peptide discovery

Summary

While resistant fungal infections have been on the rise, the discovery of alternative treatment options lags severely behind. Due to limited databases and fungi’s proximity to eukaryotic cells, many current antifungal drug discovery models remain limited in a therapeutic scope. In this paper, the authors aim to bridge the gap between predicting potential antifungal peptides and generating synthesizable, experimentally testable peptide candidates for drug discovery. Fung-AI uses a modified GAN framework to generate potential antifungal peptides, then undergoes a series of computational filtering steps to narrow down peptides for testing. The authors show the proof-of-concept of their pipeline synthesizing 13 peptides, of which five inhibited fungal growth, with two also showing low cytotoxicity to human hepatic cells. However, a major limitation is that the tested peptides have higher MIC values compared to current antifungal peptides. The Fung-AI pipeline approach shows strong promise as a proof-of-concept framework for AI-guided antifungal peptide discovery, but it remains unclear whether the observed hit rate reflects the model uncovering broader antifungal peptide characteristics, or if instead it reflects convergence towards canonical AMP properties driven by the downstream filtering.

Major Points

1. Selection of peptides for experimental validation: From our reading of the text, the authors selected one peptide per cluster that best fit a qualitative representative secondary structure. We are concerned that this limited selection for each cluster may act as a confounding variable in the downstream activity comparisons, since that singular peptide may not fully represent the potential antifungal properties of the entire cluster. This is particularly relevant to the decision later to expand further experimental validation only on Cluster 17 based on the superior activity of Peptide 40 in the initial screen. Testing a more diverse set of peptides per cluster could clarify whether the observed effects are representative of the cluster as a whole, or if they are more specific to the chosen peptides.

2. Potential bias for canonical AMP structures: The enrichment for alpha-helices and random coil peptides during the experimental validation makes us wonder whether the GAN model may be inadvertently biased towards generating canonical AMP-like structural motifs. If so, this would mean that the pipeline is not generating the full structural diversity of possible novel antifungal peptides. Reporting the predicted secondary structure distribution across the generated peptide pool, and not just the synthesized peptides, would help clarify whether this reflects a limitation of the GAN model or an artifact of the downstream selection process.

3. Antifungal specificity: The authors acknowledge that the mechanism of action of the generated peptides remains unknown, which limits our ability to assess whether the Fung-AI pipeline is enriching for antifungal-specific features or generating broad spectrum antimicrobial peptides. We would suggest testing the selected peptides against non-fungal microbial populations to address this concern. If the generated peptides show comparable activity against bacterial targets, it would suggest that the pipeline is generating peptides with general membrane disrupting properties rather than antifungal specific ones. This would have important implications for how the training data, classifiers, and down-selecting criteria should be refined in future iterations of the pipeline. Alternatively, they could examine how some known anti-bacterial peptides would score/perform in their models/pipeline.

Minor Points

1. Clarity and justification of in silico pipeline decisions: We appreciate that the authors made the Github repository for Fung-AI readily available, however we had difficulty locating the referenced Supporting Text containing additional details regarding the model architecture and training procedures. While we do not have extensive expertise in generative deep learning methods, we found several aspects of the computational workflow difficult to fully understand from the main text alone. We believe that making the Supporting Text more readily accessible through Biorxiv and expanding on decision rationale within the text would greatly improve a broader scientific audience’s understanding.

   • Additional details regarding the choice of a 0.5 probability threshold for prioritizing antifungal peptide candidates would be helpful. Based on the classifier score distributions shown in Figure 2 and as described in the text, many generated peptides appear concentrated towards the tails of the distribution. Could the use of a higher threshold further enrich potential candidates with high antifungal properties, despite reducing the pool of candidates downstream?

   • The rationale for choosing each of the three classifiers in the pipeline is only inferable from the legend of Figure 2 and the Methods section of the pre-print, making the main narrative difficult to follow. We would suggest briefly describing each classifier in the main text and including what distinguishes them from one another and why a multi-classifier approach was chosen.

   • We would appreciate it if the authors could further discuss their reasoning for the down-selection criteria used in the Fung-AI pipeline. As described, the restriction of peptide length to 10-35 amino acids and filtering for positively charged, hydrophobic residues appear to bias the outputs toward canonical AMP properties. It would be helpful to hear the authors’ thoughts on how these constraints may shape the diversity and novelty of generated candidates, and whether less stringent filtering criteria could allow the model to explore a broader antifungal peptide design space.

   • Clarification on why the down-selection criteria was not integrated into the criteria for the discriminator in the GAN framework. This would provide a learning signal to the generator to not generate what would be down-selected anyway enabling more viable candidates to be generated.

1. Unreported data from initial validation screen: The authors state that MICs were measured for the five initially down-selected peptides, but no data is presented of the four peptides that did not advance to the second phase of testing. Additionally, Peptide 48 is described in the text as the strongest antifungal candidate with limited cytotoxicity, yet it does not appear to be included in the representative well-plate images provided. Including these experimental results would improve transparency regarding the down-selection process and help create a more cohesive narrative connecting the pipeline’s novel peptide generation with the downstream experimental validation.

2. Potential challenges at high peptide concentrations: Several of the reported MIC values occur at relatively high peptide concentrations. We were curious whether the authors considered the possibility that some observed effects could partially reflect nonspecific perturbations to the media or a broader physiochemical stress effect rather than antifungal-specific biological activity alone. Additional discussion regarding how these possibilities were accounted for or interpreted would further help contextualize the reported MIC results.

3. Future directions rationale: The authors suggest that newer generative approaches such as diffusion models could be explored as alternatives to the GAN-based pipeline but provide no explanation as to why these methods would be expected to improve upon the current approach. Given that this is presented as one of the future directions, we suggest including a brief discussion of the specific limitations of GANs that diffusion models would address in this specific scenario.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.