Modification-aware AI enables terminal chemical modifications for peptide design and discovers potent antimicrobials

This study introduces Termini, a modification-aware generative AI framework that designs potent antimicrobial peptides with optimized N- and C-terminal modifications, achieving a 92.5% in vitro activity rate and demonstrating robust therapeutic efficacy against a broad spectrum of pathogenic bacteria.

The Gist

Imagine the world of bacteria is like a massive, fortified castle. For decades, we've been trying to break in with our weapons (antibiotics), but the bacteria have learned to lock the gates, change the locks, and even turn our weapons against us. This is Antimicrobial Resistance (AMR), and it's becoming a global emergency.

The scientists in this paper decided to stop trying to fix the old weapons and instead built a super-smart robot architect to design brand-new ones. They call this robot "Termini."

Here is the story of how they built it and what it achieved, explained simply:

1. The Problem: Too Many Keys, Too Few Locks

Antimicrobial peptides (AMPs) are tiny protein chains that act like "molecular keys" to unlock and destroy bacteria. But there are so many possible combinations of these keys (more than the number of grains of sand on Earth) that humans can't test them all one by one. It's like trying to find the one key that opens a specific lock by testing every key in a giant, dusty junk drawer.

2. The Solution: The "Modification-Aware" Robot

The team built an AI system (Termini) that doesn't just guess random keys. It learned from millions of existing keys found in nature (like those made by frogs, plants, and humans) to understand what makes a key work.

The Secret Sauce:
Most AI systems design the "body" of the key but ignore the "cap" on the end. In chemistry, the ends of these peptide chains can be capped in different ways (like putting a little hat on the top or bottom).

• The Analogy: Imagine designing a screw. Most designers just focus on the threads. This AI realized that putting a rubber cap on the head or a metal tip on the bottom changes how well the screw fits into the wood.
• The Innovation: Termini was trained to explicitly design these "caps" (N-terminal and C-terminal modifications) right from the start. It treats the cap not as an afterthought, but as a crucial part of the design.

3. The Factory: Generating and Filtering

The robot went to work:

1. Generation: It created 26,000 new, unique peptide designs.
2. Simulation: Before making them in a lab, the AI ran them through a virtual simulation against 15 different types of bacteria (including the super-bugs that kill people in hospitals). It also checked if they would hurt human cells (toxicity).
3. Selection: It picked the top candidates that looked like they would kill bacteria but leave humans safe.

4. The Real-World Test: The "Hit Rate" Miracle

The team synthesized (built) 120 of these AI-designed peptides in the lab and tested them against 11 dangerous bacteria.

• The Result: 92.5% of them actually worked!
• Why this is huge: In the past, when scientists used AI to design drugs, usually only 1 out of 10 or even 1 out of 20 would work. Getting nearly 1 in 1 to work is like throwing darts and hitting the bullseye almost every single time.

5. The "Caps" Make a Difference

The study proved that the "caps" (terminal modifications) were a game-changer.

• The Analogy: It's like taking a good pair of running shoes and adding a special grip to the sole. The shoe is the same, but the grip makes it run faster on wet pavement.
• The AI predicted that adding a specific cap would make the peptide stronger against bacteria, and the lab tests confirmed it. In some cases, the capped version was significantly more powerful than the uncapped version.

6. How Do They Kill? (The Mechanism)

The researchers wanted to know how these new keys worked. They found that the AI didn't just create one type of "sledgehammer."

• Some peptides acted like drill bits, poking holes in the bacterial outer wall.
• Some acted like magnets, pulling the bacteria's internal energy apart.
• Some were shape-shifters, changing their structure to fit the bacteria perfectly.
• Key Finding: The AI discovered that you don't need a perfect, rigid shape to kill bacteria. Sometimes, a flexible, "floppy" peptide works just as well, or even better, than a rigid one.

7. The Final Boss: Testing in Living Animals

To prove these weren't just lab tricks, they tested the best candidates on mice with skin infections caused by a super-bug (Acinetobacter baumannii).

• The Outcome: The AI-designed peptides cleared the infection just as well as the strongest existing antibiotic (Polymyxin B), but with a much better safety profile.
• The Takeaway: The "capped" versions of the peptides stayed effective longer in the body, proving that the AI's design choices translated to real-life healing.

Summary

This paper is a breakthrough because it shows that AI can design life-saving drugs if you teach it the right rules. By teaching the AI to pay attention to the tiny "caps" on the ends of the molecules, the researchers created a pipeline that is:

1. Faster: It designs drugs in days, not years.
2. Smarter: It understands that small chemical changes matter.
3. More Successful: It has a hit rate that dwarfs previous methods.

It's like upgrading from a blindfolded archer to a sniper with a thermal scope, finally giving us a fighting chance against the rising tide of superbugs.

Technical Summary

1. Problem Statement

The global rise of antimicrobial resistance (AMR) poses an existential threat to public health, with drug-resistant pathogens projected to cause more deaths than cancer by 2050. While Antimicrobial Peptides (AMPs) offer a promising alternative due to their multifaceted mechanisms of action, their discovery is hindered by:

• Immense Sequence Space: The combinatorial possibilities of peptide sequences (even limited to 20 canonical amino acids) make exhaustive experimental screening infeasible.
• Limitations of Existing AI Pipelines: Previous generative models for AMPs often fail to explicitly account for terminal chemical modifications (N-terminal acetylation and C-terminal amidation). These modifications are critical for peptide stability and bioactivity but are typically treated as post-hoc optimizations rather than integral design variables.
• Narrow Validation: Many studies validate candidates against limited microbial panels or achieve low success rates (hit rates) in experimental synthesis.

2. Methodology: The "Termini" Framework

The authors developed Termini, an integrative, modification-aware machine learning framework designed to systematically discover potent AMPs. The pipeline consists of four main stages:

A. Data Curation and Generative Modeling

• Dataset: A comprehensive dataset of 32,095 peptides (5–30 residues) was compiled from major AMP databases (e.g., APD, dbAMP, DRAMP) and training sets of 24 existing AMP classifiers.
• Generative Model: A diffusion-based generative model (adapted from ESM-2 protein language models) was trained in the latent embedding space.
  • It generates candidate sequences conditioned on physicochemical properties (net charge, hydrophobicity).
  • Crucially, the model generates sequences that are subsequently expanded into four variants for each backbone: Free/Free, Acetyl/Free, Free/Amide, and Acetyl/Amide.

B. Predictive Screening (The "Funnel")

Generated candidates undergo a multi-tier computational filtering process using Chemprop (a graph neural network framework based on SMILES representations):

1. Activity Classification: Binary classifiers predict antimicrobial activity (Active vs. Inactive) for 15 bacterial/fungal species (including ESKAPEE pathogens).
2. Potency Regression: Regression models predict quantitative Minimum Inhibitory Concentration (MIC) values (log(MIC+1)).
3. Toxicity Filtering: Dedicated classifiers predict cytotoxicity (CC50) and hemolysis (HC50) to filter out toxic candidates.

• Key Innovation: The predictive models were explicitly trained on data including terminal modifications, allowing them to predict how capping affects activity and toxicity before synthesis.

C. Experimental Validation

• Synthesis: 120 peptide entries (representing 60 unique backbone sequences with different terminal states) were synthesized.
• In Vitro Testing:
  • Antimicrobial Activity: Tested against 11 pathogenic species (Gram-negative and Gram-positive).
  • Toxicity: Hemolytic (RBC) and cytotoxic (HEK293T) assays.
  • Mechanism: Circular Dichroism (CD) for secondary structure; NPN uptake and DiSC3-5 assays for membrane permeabilization and depolarization.
• In Vivo Testing: A murine skin scarification model infected with Acinetobacter baumannii was used to evaluate lead candidates.

3. Key Contributions

• Explicit Terminal Modification Modeling: Unlike previous works, Termini treats N-terminal acetylation and C-terminal amidation as first-class design variables, integrating them directly into the generative and predictive loops.
• Broad-Spectrum Validation: The study validated candidates against the broadest panel of pathogens reported to date in a single AMP discovery study (11 species, including MDR strains).
• High Hit Rate: The framework achieved a 92.5% hit rate (111/120 peptides) for antimicrobial activity in vitro, significantly outperforming many previous generative approaches.
• Mechanistic Diversity: The study demonstrates that effective AMPs do not require a single canonical structure (e.g., $\alpha$-helix) but can function via diverse structural ensembles (mixed $\alpha$/$\beta$, disordered) and membrane interaction mechanisms.

4. Key Results

A. Performance of Predictive Models

• Classification: Models achieved AUCs >0.80 for 9/15 species in cross-validation and >0.75 for all 15 species on held-out test sets.
• Regression: Strong correlations (Pearson 0.50–0.72) were observed between predicted and experimental log(MIC) values.
• Modification Prediction: The models successfully predicted the direction of activity changes caused by terminal modifications (68% match rate), allowing for the prioritization of specific capped variants.

B. Experimental Outcomes

• Activity: 111 of 120 synthesized peptides (92.5%) showed antimicrobial activity.
  • 89/120 had MIC $\le$ 16 $\mu$M against at least one organism.
  • 63/120 exhibited broad-spectrum activity (active against both Gram-negative and Gram-positive species).
• Terminal Effects: C-terminal amidation frequently enhanced potency (lower MICs) for specific backbones, while N-terminal acetylation could improve safety profiles (higher CC50) without sacrificing efficacy.
• Toxicity: The toxicity filter was highly effective; 97.5% of peptides selected by the model were non-hemolytic, and 71.6% were non-cytotoxic.
• Structure-Function: CD spectroscopy revealed that active peptides spanned a wide conformational landscape (highly helical, $\beta$-enriched, and disordered). Effective membrane disruption did not strictly depend on stable $\alpha$-helical formation.
• In Vivo Efficacy: Three lead candidates significantly reduced bacterial burden in a murine A. baumannii skin infection model at Day 2, with efficacy comparable to Polymyxin B.

C. Mechanistic Insights

• Membrane Interaction: Peptides exhibited diverse mechanisms: some strongly permeabilized and depolarized, others decoupled these processes, and some showed weak membrane signals yet retained activity.
• Sequence Dependence: Terminal modifications acted as "tuners" of membrane interaction kinetics and selectivity rather than universal switches. For example, amidation could improve potency for one sequence while increasing toxicity for another, highlighting the need for sequence-specific modeling.

5. Significance and Conclusion

This study establishes a scalable, modification-aware paradigm for AMP discovery. By integrating terminal chemistry directly into the AI workflow, the authors overcame a major bottleneck in peptide design: the inability to predict how chemical capping alters biological function.

• Translational Impact: The framework successfully translated computational designs into potent, low-toxicity therapeutics with proven in vivo efficacy against multidrug-resistant pathogens.
• Future Directions: The authors note that while the current model covers acetylation and amidation, future iterations could incorporate lipidation, glycosylation, and non-natural amino acids to further expand the chemical space of peptide therapeutics.

In summary, Termini provides a robust, data-driven strategy for accelerating the development of next-generation antimicrobials, moving beyond simple sequence generation to the precise engineering of chemically modified peptide therapeutics.