Paving the way for new antimicrobial peptides through molecular de-extinction

Abstract

Molecular de-extinction has emerged as a novel strategy for studying biological molecules throughout evolutionary history. Among the myriad possibilities offered by ancient genomes and proteomes, antimicrobial peptides (AMPs) stand out as particularly promising alternatives to traditional antibiotics. Various strategies, including software tools and advanced deep learning models, have been used to mine these host defense peptides. For example, computational analysis of disulfide bond patterns has led to the identification of six previously uncharacterized β-defensins in extinct and critically endangered species. Additionally, artificial intelligence and machine learning have been utilized to uncover ancient antibiotics, revealing numerous candidates, including mammuthusin, and elephasin, which display inhibitory effects toward pathogens in vitro and in vivo. These innovations promise to discover novel antibiotics and deepen our insight into evolutionary processes.

Abbreviations

AEP - archaic encrypted peptide,

AMP - antimicrobial peptides,

EP - encrypted peptide,

MEP - modern encrypted peptide,

MIC - minimal inhibitory concentration,

ML - machine learning,

NCBI - National Center for Biotechnology Information.

INTRODUCTION

Molecular de-extinction represents a promising area of scientific research that enables the identification, synthesis, and understanding of molecules’ biological functions throughout evolution 1. However, it also raises interesting bioethical and philosophical debates within the scientific community 2. Advances in ancient DNA sequencing methods 3 have increasingly allowed us to access biological data from the past. Ancient DNA sheds light on historical protein-coding sequences that may not exist in our current time or have been changed throughout evolution 4, 5. Furthermore, advancements in computational biology and artificial intelligence 6, 7, 8, 9 have shifted the discovery of promising molecules from a chance-based approach to a more intentional and data-driven methodology.

Among the extensive range of molecules that can be found through proteomics or genomics, antimicrobial peptides (AMPs) stand out. AMPs have played crucial roles in the defense mechanisms of animals, evolving over millions of years to protect hosts against various pathogens, thereby ensuring survival in ancient environments 10. These molecules continue to function in the innate immune systems of various organisms today, fighting multiple microorganisms. They can modulate the immune system, disrupt cell membranes, target intracellular processes, and inhibit biofilm formation 11, 12, 13, 14. Despite considerable variability in AMP structures 15, identifying specific features is critical for mining these potential antibiotics, particularly when leveraging computational tools.

However, it is important to highlight the gap between the amount of discovered antimicrobial peptides and those that successfully advance the clinical trials and, finally, reach the market 16. Issues related to stability and toxicity frequently hinder the development cycle, which, in the best-case scenarios, still averages 13 years to be launched 17. The advancement of artificial intelligence and molecular de-extinction offers a valuable opportunity not only to discover new antimicrobials but also to provide accurate in silico predictions, thereby shortening the path to addressing the global antibiotic resistance crisis.

Figure 1 . Workflow for identifying ancient antimicrobial peptides. Genome and proteome data serve as sources for mining ancient AMPs. This mining can be performed using (i) machine learning (ML) methods to generate encrypted peptides (EPs), or (ii) computational tools that identify defensins by analyzing disulfide bond patterns. The activity of these ancient AMPs can be predicted using deep learning (DL) models, and their structures can be elucidated through molecular dynamics simulations or circular dichroism. The predicted AMP activity can be validated through in vitro and in vivo assays. Figure created with BioRender.com.

BIOINFORMATICS TOOLS

Molecular de-extinction employs various methodologies to mine compounds such as AMPs and other peptide antibiotics from extinct organisms (Figure 1). These approaches utilize diverse bioinformatics tools to discover novel AMPs from the genomes and proteomes 1, 18 of extinct organisms. Digital repositories like the National Center for Biotechnology Information (NCBI) and the Protein Data Bank (PDB) provide access to genomic and proteomic data 19, 20. One current methodology involves identifying AMPs with specific characteristics of β-defensins from the genomes of these organisms 21. Programs such as AUGUSTUS can locate protein-coding genes within genomic sequences obtained from NCBI 21, 22. To refine the identification process, tools such as HMMER 23 and InterPro 24 determine which protein sequences within the selected genomic data belong to the β-defensin protein family. Subsequent structural and physicochemical analyses are conducted to evaluate the potential antimicrobial properties of these proteins. Advanced tools such as AlphaFold 2 25 or AlphaFold 3 26 can accurately predict protein structures, while features such as cationicity and amphipathicity can be analyzed using ExPASy to complement the AMP characterization 21.

A different strategy for molecular de-extinction employs machine learning (ML) models to identify and classify encrypted peptides (EPs) — protein fragments with antimicrobial properties — from the proteomes of extant and extinct organisms 1, 18, 27. The panCleave ML model, for example, applies a pan-protease cleavage site classifier to conduct computational proteolysis, identifying potential EPs within protein sequences 1. This open-source model achieved over 80% accuracy for proteases, with at least 100 observations in the test set. Specifically, for cysteine catalytic types, the average accuracy was 81.3%, based on 1,858 correct predictions out of 2,286 observations, whereas for threonine catalytic types, the accuracy was 34.6%, with 18 out of 52 observations predicted correctly 1. Other genome mining tools, such as ThioFinder 28, RODEO 29, and RiPPMiner-Genome 30, have also been widely used to discover new AMPs in extant organisms 31.

Progress in this field is further demonstrated by the development of the deep learning model APEX (Figure 2), which has been used to mine all extinct organisms as sources of antibiotics 18. This state-of-the-art model consists of a peptide sequence encoder coupled with neural networks for predicting antimicrobial activity, enabling the extraction and classification of peptide antibiotics based on their potential Minimal Inhibitory Concentration (MIC) 18. APEX achieved a significant Pearson correlation (>0.3) for predicting species-specific antimicrobial activity, showing correlations between predicted and experimentally validated activities for several strains, including Escherichia coli strains AIC221, AIC222, ATCC 11775; Acinetobacter baumannii ATCC 19606; Pseudomonas aeruginosa strains PAO1 and PA14; and Enterococcus faecium ATCC 700221. Predicting antibiotic activity with advanced artificial intelligence like APEX (Figure 2) brings us closer to mining effective candidates for novel antibiotic alternatives. The antimicrobial activity of these ancient compounds can potentially also be predicted using various deep learning models, such as AMP-Bert 32 or AMPlify 33.

Figure 2 . APEX Model Architecture. The APEX model combines recurrent and attention neural networks to analyze peptide sequences for antimicrobial prediction. The model first extracts physicochemical features of peptide sequences using the AAindex library. These features are processed through a two-layer attention neural network (a₁ and a₂), enhancing global feature interactions and compressing the representation into a lower-dimensional format. The output of this attention network is then processed by two separate recurrent neural networks (RNNs) (h₁ and h₂): one predicts species-specific antimicrobial activity (o₁), and the other classifies the peptide as antimicrobial (AMP) or non-antimicrobial (non-AMP) (o₂). Figure created with BioRender.com.

UNVEILING ANTIMICROBIAL PEPTIDES FROM EXTINCT ORGANISMS

Natural AMPs and other peptide antibiotics can originate from four main processes: (i) genome-encoded peptides 21, 34, (ii) cleavage via proteolysis 27, 35 , (iii) synthesis by non-ribosomal means 36, and (iv) small open reading frames (smORFs) 37, 38. Due to the vast nature of genomic and proteomic databases, molecular de-extinction of AMPs has thus far been driven by strategic exploration to unveil antimicrobials encoded in the genome 21 or encrypted within proteins 1, 18. For instance, six β-defensins were predicted through computational analysis of genomes from extinct and critically endangered species. This prediction was based on intrinsic disulfide bonding patterns (Cys1–Cys5, Cys2–Cys4, and Cys3–Cys6), as well as the characteristic structural features of β-defensins 21, including three antiparallel β-strands and a right-handed α-helix. Two defensins, Ad-AvBD5 and Ad-AvBD10, were identified from Anomalopteryx didiformis, the New Zealand moa that became extinct approximately 600 years ago 39. Three β-defensins (Cs-AvBD1, Cs-AvBD9, and Cs-AvBD10) were derived from Cyanopsitta spixii, Spix’s macaw, which is endemic to Brazil. Additionally, one β-defensin was identified from Diceros bicornis minor, a critically endangered subspecies of the black rhinoceros. Ad-AvBD5, Cs-AvBD1, and Cs-AvBD10 were noted for their high stability in molecular dynamics analyses, displaying cationic charges of +3, +7, and +2, respectively 21.

Using ML to predict patterns of protein cleavage into peptide fragments, we have detected encrypted peptides resulting from proteolysis in Homo sapiens neanderthalensis (Neanderthals) and Homo sapiens subsp. Denisova (Denisovans) 1. Among the 69 archaic protein fragments identified, six showed in vitro antimicrobial activity, four from Neanderthals and two from Denisovans 1. The molecule PDB6I34D-ALQ29, a fragment from Chain D of Neanderthal glycine decarboxylase, displayed broad-spectrum antimicrobial activity against both P. aeruginosa and E. coli strains, with MIC values ranging from 32 to 128 mmol.L^-1. This archaic encrypted peptide (AEP) possesses a net charge of +5 and an amphiphilicity index of 0.99. Conversely, compound A0A343EQH4-LAM11 (also known as neanderthalin), an AEP with a net charge of 0 and an amphiphilicity index of 0.63, demonstrated significant efficacy in preclinical animal models by reducing bacterial loads by several orders of magnitude against A. baumannii. It was observed that AEPs have lower net charge and normalized hydrophobicity, more basic residues, and fewer acidic residues and polar residues compared to modern encrypted peptides (MEP) 1. The differences in amino acid composition led to distinct physicochemical traits in AEPs, including lower amphiphilicity, a greater tendency toward disordered conformations, and reduced aggregation.

Moreover, the deep learning model, APEX, has facilitated the identification of several ancient encrypted peptides within the proteomes of extinct animals and plants 18. Species such as the New Zealand moa (Anomalopteryx didiformis), the South American giant sloth (Mylodon darwinii), the giant elk (Megaloceros sp.), Grant’s zebra (Equus quagga boehmi), the woolly mammoth (Mammuthus primigenius), the straight-tusked elephant (Elephas antiquus), the ancient sea cow (Hydrodamalis gigas), as well as extinct plant species like Magnolia latahensis and Hesperelaea palmeri, revealed peptide molecules with excellent antimicrobial traits. These peptides exhibited low MIC values against some ESKAPE pathogens: Enterococcus spp., Staphylococcus aureus, Klebsiella pneumoniae, A. baumannii, P. aeruginosa, and E. coli 18. Testing these encrypted peptides in two different preclinical mouse models indicated that mammuthusin-2 (MEP), elephasin-2 (AEP), and mylodonin-2 (AEP) possess high potential for antibiotic and anti-infective efficacy. Additionally, mammuthusin-2 exhibited slower degradation kinetics. Notably, although the AEPs and MEPs identified through APEX showed an atypical prevalence of low amphiphilicity and uncharged polar residues, they were primarily characterized by helical structures, resulting in more effective membrane disruption 18. These findings highlight the potential of computational and artificial intelligence tools in uncovering extinct peptides with antimicrobial properties (Table 1).

Table 1. Extinct peptides identified through molecularde-extinction.

Extinction	Species	Name	Peptides sequence	Amino acids	Net charge	Ref.
~600 years ago	Anomalopteryx didiformis	Ad-AvBD5	TRQDCESRGGFCSRGSCPLGITRIGICSLQDFCCRRKMGE	40	3
		Ad-AvBD10	VSFADTEECRSQGNFCRPVSCPPVFSVSGSCYGGAMKCCKKEYGQ	45	1
Extinct in the wild in 2000	Cyanopsitta spixii	Cs-AvBD1	NKAQCHREKGFCALLKCPFPYVISGRCTKFTFCCKKGA	38	7	20 a
		Cs-AvBD10	DPLFPDTTECKNQGNFCRAGTCPPTFAISGSCHGGLLRCCSKKISS	46	2
		Cs-AvBD9	PAYSQVDADTAACRQNRGSCSFVECSSPMVNIGTCRSGKLKCCKXYV	47	3
~40,000 years ago	Homo sapiens neanderthalensis	PDB6I34D- ALQ29	ALQLCYRHNKRRKFFVDPRCHPQTIAVVQ	29	5
		A0A384E0N 4-DLI09	DLIERIQAD	9	-2
		A0A343EQH 4-LAM11	LAMVIPLWAGA	11	0	10 b
		A0A343EQH 0-NVK38	NVKMKWQFEHTKPTPFLPTLITLTTLLLPISPFMLMIL	38	2
~50,000 years ago	Homo sapiens subsp. 'Denisova'	A0A343AZS 4-FMA25	FMAEYTNIIMMNTLTTTIFLGTTYN	25	-1
		A0A0S2IB02 -AYT38	AYTTWNILSSAGSFISLTAVMLMIFMIWEAFASKRKVL	38	2
1889	Ara tricolor	AWH62785.1-RLA27	RLATLQLWTINKITKQLMIPLNKPGHK	27	5
~781,000–30,000 years ago	Elephas antiquus	AQU14158.1-LHL12	LHLKILKIIRLL	12	3
		AQU14158.1-IFL14	IFLHLKILKIIRLL	14	3
1883	Equus quagga boehmi	ABN79624.1-CVL25	CVLLFSQLPAVKARGTKHRIKWNRK	25	7
		ADN88909.1-RAY26	RAYICRKKFLSLRKASIKLQSLVRMK	26	9
1875	Hesperelaea palmeri	CED79820.1-KLL26	KLLRKVLKETKKWVIKSVVFFKKIRK	26	10
1768	Hydrodamalis gigas	AKN52354.1-LYC24	LYCRIYSLVRARGRRLTFRKNISK	24	8	7 c
~103,000 to 42,000 years ago	Lophiomys imhausi maremortum	QYC36821.1-HWI16	HWITINTIKLSISLKI	16	2
~6050–5050 BCE	Mammut americanum	ABQ86189.1-WMT15	WMTIHALKLSLSFKL	15	2
~1.8 million to 12,000 years ago	Mylodon darwinii	AWK29290.1-WFH14	WFHFNSKILLLTGL	14	1
		SMQ11516.1-KRK18	KRKRGLKLATALSLNNKF	18	6
		SMQ11516.1-KIY25	KIYKKLSTPPFTLNIRTLPKVKFPK	25	7
1952	Pinguinus impennis	ASB29243.1-KFI13	KFILNFKIPISFK	13	3

^a List of extinct b-defensinsidentified by software but not validated its activity experimentally.

^b List of active AEPs identified by panCleaveand validated experimentally.

^c List of active AEPs identified by APEX andvalidated experimentally.

CONCLUSION

Just as the evolutionary loss of AMPs is evident 40, microbial resistance genes also impose significant fitness costs on organisms due to their energy burdens, leading to the eventual loss of some of these genes over time 41. Molecular de-extinction thus emerges as an innovative concept in drug discovery, offering the possibility of uncovering ancient molecules that may exhibit unique mechanisms of action or target sites not addressed by contemporary antibiotics. Reintroducing these ancient antimicrobials paves the way for exploring alternative therapeutic approaches to combat contemporary drug-resistant pathogens. Moreover, AI-driven molecular de-extinction has the potential to enrich our understanding of evolution, ecology, and biodiversity.

CONFLICT OF INTEREST

Cesar de la Fuente-Nunez provides consulting services to Invaio Sciences and is a member of the Scientific Advisory Boards of Nowture S.L., Peptidus, European Biotech Venture Builder, and Phare Bio. He is also a member of the Advisory Board for the Pep-tide Drug Hunting Consortium (PDHC). All other authors have no conflicts of interest to declare.