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

This study presents Fung-AI, an AI/ML-driven pipeline that successfully generated and experimentally validated novel de novo antifungal peptides with activity against *Fusarium graminearum* and *Candida albicans* and low cytotoxicity, demonstrating a proof-of-concept for rapid antifungal discovery.

The Gist

Imagine a world where invisible invaders (fungi) are attacking our crops and making people sick. These invaders are getting smarter, stronger, and harder to kill with our current medicines. Scientists at Johns Hopkins decided to fight back, not with a bigger hammer, but with a super-smart robot designer. They called their project Fung-AI.

Here is the story of how they built it, using simple analogies:

1. The Problem: The Fungal "Lock"

Fungi are tricky enemies. They are like living cells that are very similar to our own human cells. This makes it hard to design a weapon that kills the fungus without accidentally hurting us. Plus, the fungus is evolving to ignore our old drugs. We need new weapons, but finding them by hand is like looking for a needle in a haystack, one needle at a time.

2. The Solution: The "Imagination Machine" (The GAN)

Instead of guessing, the scientists built an AI "Imagination Machine" (technically called a Generative Adversarial Network, or GAN).

Think of this AI as a master chef who has read every recipe book in the world for "antifungal soups."

• The Chef (Generator): This part of the AI tries to cook up brand new, never-before-seen recipes (peptide sequences) that might kill fungi.
• The Critic (Discriminator): This part of the AI tastes the new recipes and says, "Nope, that looks fake," or "Yes, that looks like a real antifungal soup."

They play a game of "Cat and Mouse" over and over. The Chef gets better at cooking, and the Critic gets better at spotting fakes. Eventually, the Chef becomes so good at cooking that it can create thousands of brand-new, unique recipes that don't exist in nature.

3. The Filter: The "Security Checkpoint"

The AI cooked up about 10,000 new recipes. But we can't test 10,000 things in a lab; it would take forever and cost a fortune. So, they built a Security Checkpoint (a series of computer filters).

• Filter 1 (The Fungus Detector): Three different computer programs checked the recipes to see, "Does this look like it will kill a fungus?" They kept the 3,500 best candidates.
• Filter 2 (The Safety Guard): Since fungi are like humans, a weapon that kills fungus might also kill us. This filter checked, "Will this recipe hurt human cells?" They threw out the dangerous ones.
• Filter 3 (The Originality Check): They ran the remaining recipes through a giant library (BLAST) to make sure they weren't just copying existing recipes. They wanted new ideas.

After all this digital filtering, they were left with a tiny, elite group of just 13 candidates.

4. The Lab Test: The "Taste Test"

The scientists synthesized these 13 digital recipes into real, physical peptides (tiny protein chains) and put them in a lab to test against real fungi.

• The Targets: They tested them on a wheat-killing fungus (which threatens our food) and a human-killing fungus (which threatens our health).
• The Results:
  • 5 out of 13 actually worked! They stopped the fungi from growing.
  • 2 of those 5 were the stars of the show. They killed the fungi but were very gentle on human liver cells (meaning they are likely safe for people).
  • The "Miss": None of them worked against a specific, super-resistant fungus called C. auris. This taught the scientists that their AI needs to be trained specifically on different types of enemies in the future.

5. The Big Picture

This paper is a proof of concept. It shows that we don't need to wait for nature to give us new drugs. We can use AI to dream up new medicines from scratch.

The Analogy Summary:
Imagine you need to find a key that opens a specific lock (the fungus) but doesn't break the door (the human body).

• Old way: Try every key you can find in a giant pile.
• Fung-AI way: Build a robot that can instantly invent 10,000 new key shapes, use a computer to instantly test which ones fit the lock and won't break the door, and then hand you the top 13 best keys to try in real life.

The Takeaway:
While the new "keys" (peptides) aren't perfect yet (their "lock-picking" power is a bit weak, and they didn't work on the super-hard lock), the method works. It proves that AI can be a powerful partner in the race to save our food and our health from fungal threats.

Technical Summary

1. Problem Statement

Fungal pathogens pose a critical threat to global health and food security, causing an estimated 2.5 million deaths annually and significant crop losses. The current landscape of antifungal drug discovery is hindered by:

• Limited Drug Classes: A scarcity of known antifungal drug mechanisms.
• Resistance: The emergence of multi-drug and pan-drug-resistant clinical isolates (e.g., Candida auris).
• Biological Complexity: Fungi share eukaryotic cell biology with mammals, making it difficult to design drugs that are toxic to fungi but safe for humans.
• Data Scarcity: While antimicrobial peptide (AMP) data exists, high-quality, large-scale datasets specifically for antifungal peptides are limited, hindering the training of robust generative AI models.

2. Methodology: The Fung-AI Pipeline

The authors developed Fung-AI, a semi-automated, in silico pipeline designed to generate, filter, and prioritize novel de novo antifungal peptides. The workflow consists of four main stages:

A. Data Curation and Training

• Antifungal Dataset: Curated 7,335 peptides (3,423 antifungal, 3,912 non-antifungal) from literature, DRAMP 4.0, and CAMPR4. Sequences were filtered to 10–35 amino acids.
• Hemolytic Dataset: Used 3,804 peptides (1,576 hemolytic, 2,228 non-hemolytic) from Yaseen et al. to train toxicity filters.
• Generative Model (GAN): A custom Generative Adversarial Network was trained on the combined antifungal/non-antifungal dataset. To prevent mode collapse and handle small data sizes, the architecture included an encoder-decoder structure alongside the generator and discriminator.
• Discriminative Classifiers: Three binary classifiers were retrained to predict antifungal activity:
  1. TCN-based: Temporal Convolutional Network (Singh et al. architecture).
  2. One-hot Encoding: 1D CNN-BiLSTM.
  3. BLOSUM Encoding: 1D CNN-BiLSTM with pre-seeded BLOSUM62 embeddings.
• Hemolytic Classifier: A 1D CNN-BiLSTM model trained to predict hemolytic (toxic) activity.

B. In Silico Generation and Down-Selection

1. Generation: The GAN generated ~10,000 unique peptide sequences.
2. Antifungal Filtering: Peptides were passed through the three antifungal classifiers. Only sequences predicted as antifungal by all three models (probability > 0.5) were retained (~3,578 peptides).
3. Toxicity Filtering: The remaining peptides were screened by the hemolytic classifier. Non-hemolytic candidates were retained (~2,433 peptides).
4. Clustering and Novelty Check:
   • UMAP & HDBSCAN: Embeddings were reduced using UMAP and clustered using HDBSCAN to identify groups of similar, high-quality candidates.
   • BLAST: Candidates were queried against the NCBI Non-Redundant (nr) Protein Database. Sequences with high similarity (>90% query cover, >80% identity) to known proteins were discarded to ensure novelty.
5. Biophysical Filtering: Final candidates were selected based on:
   • Net charge: Mildly cationic (0 < charge < +6).
   • Hydrophobicity: 30–60% hydrophobic amino acids.
   • Structure: Predicted via PEP-FOLD4 to ensure diverse secondary structures (e.g., alpha-helical).

C. Experimental Validation

• Synthesis: 13 peptides were synthesized (initially crude, then >95% purity for validation).
• Antifungal Assays: Minimum Inhibitory Concentration (MIC) was tested against:
  • Fusarium graminearum (Wheat pathogen)
  • Saccharomyces cerevisiae (Model organism)
  • Candida albicans (Human pathogen)
  • Candida auris (Emerging multidrug-resistant pathogen)
• Cytotoxicity Assays: LC50 values were determined using HepG2 human liver carcinoma cells.

3. Key Results

• Generation Success: The pipeline successfully generated ~10,000 novel sequences, of which 78 mapped to specific "antifungal-like" clusters distinct from known proteins.
• Hit Rate: From an initial pool of 13 synthesized peptides (across two phases), 5 peptides (38%) demonstrated confirmed antifungal activity.
• Antifungal Activity:
  • Peptide 12 & 40: Showed the strongest activity with MICs of 250 µg/mL against F. graminearum and 500 µg/mL against C. albicans and S. cerevisiae.
  • Peptide 48 & 65: Showed broad but weaker activity (MIC 500 µg/mL).
  • Peptide 17: Active only against F. graminearum (MIC 500 µg/mL).
  • C. auris: None of the peptides showed significant activity against C. auris at concentrations up to 500 µg/mL.
• Cytotoxicity & Therapeutic Window:
  • Peptide 12: High cytotoxicity (LC50 = 66.06 µg/mL), limiting its therapeutic potential.
  • Peptide 48: No observed toxicity (LC50 > tested range).
  • Peptide 65: Low cytotoxicity (LC50 = 704.2 µg/mL).
  • Peptide 40 & 17: Intermediate cytotoxicity (LC50 ~315 and 85 µg/mL, respectively).
• Structural Prediction: Active peptides were predicted to form cationic alpha-helical structures, consistent with known membrane-disrupting AMP mechanisms.

4. Key Contributions

1. Proof-of-Concept Pipeline: Demonstrated that a generative AI pipeline (GAN) combined with discriminative classifiers can successfully design de novo antifungal peptides with minimal experimental screening.
2. Novel Sequence Discovery: Validated that the AI-generated sequences are truly novel (low BLAST similarity to known proteins) yet functionally active.
3. Dual-Target Efficacy: Identified peptides active against both agricultural (F. graminearum) and human (C. albicans) pathogens, highlighting the pipeline's versatility.
4. Safety Screening: Successfully integrated hemolytic prediction and cytotoxicity testing, identifying candidates (Peptides 48 and 65) with favorable safety profiles for future optimization.

5. Significance and Limitations

• Significance: The study addresses the urgent need for new antifungal agents by leveraging AI to accelerate the discovery timeline. It proves that even with limited antifungal-specific data, a generative approach can yield viable therapeutic scaffolds.
• Limitations & Future Directions:
  • Data Scarcity: The limited size of antifungal datasets resulted in peptides with relatively high MICs compared to natural peptides (e.g., VLL-28). Future work should utilize pathogen-specific fine-tuning.
  • Mechanism of Action (MOA): The study did not experimentally determine the MOA; future iterations should incorporate MOA prediction.
  • Pathogen Specificity: The pipeline failed to inhibit C. auris, likely due to the scarcity of training data for this specific emerging pathogen. The authors suggest that targeted data collection and diffusion models could improve performance against specific resistant strains.

In conclusion, Fung-AI represents a significant step toward automating the discovery of antifungal therapeutics, offering a scalable framework to combat rising fungal resistance in both clinical and agricultural sectors.