Machine learning that designs new peptides.

My MEng thesis at NTUA, 2025–2026. I built the whole pipeline: generating new peptides, filtering them and training the models that ranked them. A lab then made the top 15 and tested them.

NTUA Circuits & Systems Group
Houston Methodist Research Institute

5MPeptides screened
90%Precision in the top 10%
13/15Worked in the lab

Poster (PDF) Presented at Greeks in AI 2026, Jul 15–17

LLAALIKKLLIARARKAVAILVcationichydrophobicillustrative sequence

Machine Learning–Driven De Novo Antimicrobial Peptide Design Targeting ESKAPE Pathogens

From five million designs to fifteen peptides.

ESKAPE is the six bacteria behind most drug-resistant hospital infections. A lab can make only a handful of new peptides, so I judged the models on the few at the top of their ranking, the only ones that ever get made.

~5,000,000sequences designedPeptides that had never existed, generated inside the chemical space of charged, helix-forming AMPs.
1,450,996passed the filtersA seven-criterion physicochemical filter, 29.6% of the designs.
Rankedby two modelsA Random Forest per Gram class scored every survivor by predicted activity.
148synthesis-ready candidatesPicked for physicochemical diversity, in three design groups.
15made in a labThe top five of each group, assayed against all six ESKAPE species.

13 of those 15 killed at least one ESKAPE species.

One model for each Gram class.

Gram-positive and Gram-negative bacteria are killed in different ways, so each got its own model. Neither was given any rules about how peptides work, and SHAP shows each relying on the features the literature points to: hydrophobicity for Gram-positive, charge and amphipathicity for Gram-negative.

Gram-positive

hydrophobicity-driven

90.2%of its top-ranked 10% were really activeAUROC 0.767

What drove it

Hydrophobic fraction0.034
logP0.028
Morgan fingerprint bit0.022
GRAVY0.020
β-sheet fraction0.012

Gram-negative

charge & amphipathicity

90.6%of its top-ranked 10% were really activeAUROC 0.827

What drove it

Net charge0.038
H-bond donors0.020
Helical moment0.019
Isoelectric point0.016
Bertz complexity0.015

Models selected on precision in the top 10% of the ranking. Features ranked by mean |SHAP| on the held-out test set.

What the lab found.

Each cell is the lowest dose that stopped the bacteria growing, in µg/mL; the brighter the rose, the less it took. Four candidates, GN-3, GN-5, BR-3 and BR-5, cleared four or more species without harming red blood cells.

≤ 4
8
16
32
> 32, no effect
more potent ←

Measured MIC in µg/mL for each candidate against each ESKAPE species
Candidate	Ec	Sa	Kp	Ab	Pa	Ef	HL50
Gram-positive-biased
GP-4	>32	4	>32	16	>32	8	16†
GP-5	>32	8	>32	32	>32	8	32
Gram-negative-biased
GN-1	16	32	>32	8	>32	32	>32
GN-3	16	8	8	8	>32	8	>32
GN-4	32	>32	>32	8	>32	>32	>32
GN-5	32	16	16	4	>32	16	>32
Broad-spectrum
BR-2	32	16	16	32	>32	8	>32
BR-3	16	8	16	16	>32	16	>32
BR-5	16	32	16	16	>32	16	>32

the four best. † toxic at its lowest active dose. No candidate stopped P. aeruginosa, whose outer membrane keeps peptides out; that is the next problem to work on.

What I learned

90% of each model’s top 10% were active
Precision at the top of the list
A lab makes only the first few candidates, so that is the part of the ranking that has to be right.
each model's top driver, by SHAP
Models you can explain
Someone has to decide which peptides are worth making, and they need to see why each one ranked high.
5M → 1.45M → 148 → 15 → 13
Reproducible from scratch
Every number in the thesis comes from the pipeline’s saved state, and the whole run can be repeated from the start.

From five million designs to fifteen peptides.

One model for each Gram class.

What the lab found.

What I learned

Precision at the top of the list

Models you can explain

Reproducible from scratch