ApexGO uses generative AI to optimize antibiotics, achieving 85% success against superbugs

A generative artificial intelligence approach for peptide antibiotic optimization - Nature Machine Intelligence

In particular, the optimized peptides demonstrated significant improvements in antimicrobial potency, including activity against multidrug-resistant strains. In preclinical mouse models of A. baumannii infection, several optimized molecules -- specifically those derived from mammuthusin-3 and mylodonin-2 -- exhibited potent anti-infective activity comparable with or exceeding that of polymyxin B, a last-resort antibiotic. These findings highlight the effectiveness of ApexGO as a generative AI approach for antibiotic optimization, offering a potential path forward in tackling AMR. We recently developed APEX, a deep learning approach to predict antibiotic function from amino acid sequence. APEX efficiently predicts the minimal inhibitory concentrations (MICs) of peptides against a variety of Gram-negative and Gram-positive bacterial pathogens (see the 'APEX 1.1' section). Although APEX successfully discovered antibiotic molecules from extinct organisms, its discriminative learning nature limits its capacity for de novo design or property optimization. Moreover, the vastness of the peptide sequence space renders exhaustive virtual screening infeasible. To address these challenges, we developed ApexGO, which integrates APEX with black-box BO and generative modelling methods to perform constrained template-based optimization of antibiotic activity in peptide sequences (Fig. 1a). ApexGO uses a transformer-based VAE to map the peptide sequences into a continuous latent space, effectively transforming the discrete optimization problem into a tractable continuous one. This approach enables the efficient exploration of the sequence space to identify molecules with improved antimicrobial properties. In ApexGO, the generation process involves sampling latent space points and decoding them into peptide sequences using the VAE decoder. The evaluation process uses APEX to predict the antimicrobial potential of the generated peptides. A surrogate model, implemented as a parametric Gaussian process regressor, models the correlation between the latent space points and the APEX-predicted MICs. The BO algorithm utilizes this surrogate model to propose latent space points probably to decode into peptides with enhanced antimicrobial activity. This iterative process continues until peptides with optimized properties are identified. Key features of ApexGO include the following. In this work, we focused on template-based AMP optimization, using ten peptides mined from the proteomes of extinct organisms as templates. These templates were selected based on their varied activity profiles, ranging from selective to broad-spectrum effects. Importantly, none of the template de-extinct peptides were highly active (MIC ≤ 16 μmol l), ensuring room for 4- to 32-fold potency gains (Supplementary Data 1). We intentionally focused on a starting set of peptides with mid-micromolar MICs to give ApexGO headroom to demonstrate meaningful improvements, and better simulate the typical setting of lead development. Starting from nanomolar-active peptides would (1) collapse this dynamic (at such low concentrations, assay noise can be comparable with or exceed true improvements) and (2) require the APEX oracle to extrapolate well beyond the range of peptides it was trained on, rendering its predictions less accurate. Additionally, for most membrane-related or non-specific mechanisms, peptides must reach a threshold concentration to exhibit activity unless they are receptor mediated, which were not included in APEX's training and were, therefore, excluded from our optimization. ApexGO performed constrained peptide sequence optimization, ensuring that each optimized sequence maintained at least 75% sequence similarity to its template (Methods). The optimization aimed to improve antimicrobial activities against either seven Gram-negative pathogen strains or all 11 strains (including Gram-positive bacteria) that APEX predicts (see the 'APEX 1.1' section). Starting from the selected template sequences with relatively high average MICs against Gram-negative bacteria, ApexGO generated sequences with progressively lower predicted MICs as optimization progressed (Fig. 1b). ApexGO was tasked to produce ten optimized peptides from each of the ten starting templates, for a total of 100 product peptides, of which we will go on to experimentally validate all of them to avoid and characterize any potential selection bias. Compared with the MIC distributions of the APEX training peptides and the template peptides, the optimized peptides showed a substantial shift towards lower experimental MICs, indicating enhanced predicted antimicrobial activities (Fig. 1c). Statistical analysis confirmed that the prediction improvements were significant (P values of one-sided Mann-Whitney U-test were 1.19 × 10 and 1.12 × 10 when comparing optimized peptides with in-house and template peptides, respectively). To assess the contribution of each amino acid change proposed by ApexGO, we generated intermediate variants by systematically reverting individual mutations in the optimized sequences back to the template amino acids (Fig. 1d). By evaluating these variants with APEX and ranking them based on the predicted MICs, we found that the optimized peptides were almost always ranked first (mean ranking of 1.23; standard deviation of 0.42), suggesting that each substitution contributed incrementally to the predicted antimicrobial improvement according to APEX; however, these insights are based on computational evaluation and were not experimentally confirmed. To benchmark ApexGO against other deep generative AMP frameworks, we directly compared it with HydrAMP and the PepDiffusion latent diffusion model on the same template-constrained optimization task. Using the official HydrAMP implementation and generator/decomposer weights, we generated analogues for each of the ten extinct peptide templates under the 'discovery' criterion at decoder temperatures of T = 3, 5 and 10, applied the same ≥75% sequence-similarity constraint used in our main experiments, and scored all feasible candidates with the APEX Gram-negative objective (Supplementary Fig. 1). Across all seeds and temperatures, the best feasible HydrAMP analogues consistently exhibited higher (less-active) APEX-predicted MIC values than the corresponding ApexGO designs, and only 22.2%, 10.8% and 0.2% of HydrAMP proposals at T = 3, 5 and 10, respectively, satisfied the similarity threshold, with no feasible analogue obtained for one seed at T = 10 even after exhausting the full proposal budget. By contrast, when we used PepDiffusion with the study's best-performing checkpoints and default sampling hyperparameters to generate 10,000 AMP-conditioned sequences (9,962 unique), none of the candidates met the ≥75% similarity threshold to any of the ten templates, preventing a meaningful objective comparison. Together, these benchmarks indicate that although HydrAMP and PepDiffusion are well suited for unconstrained, diversity-oriented AMP discovery, they either produce few or no viable candidates when repurposed for per-template, similarity-constrained optimization, whereas ApexGO reliably identifies high-scoring derivatives under these design constraints. To validate the predictive accuracy of ApexGO, we synthesized and tested two sets of optimized peptides per de-extinct template: (1) five peptides predicted to be active against Gram-negative bacteria and (2) five peptides predicted to be broad spectrum. ApexGO produced a total of 100 optimized peptides, which were synthesized and experimentally tested for antimicrobial activity against 11 clinically relevant bacterial pathogens, including strains resistant to conventional antibiotics (Fig. 2 and Supplementary Fig. 2). Although ApexGO produces a much larger number of intermediate peptides during optimization and running to produce the final optimized peptides, the experimental performance of these intermediate peptides is ultimately related to the rank correlation of the APEX model itself, not to the ApexGO procedure introduced here. Of the 100 synthesized peptides, 86 exhibited detectable antimicrobial activity (MIC ≤ 64 μmol l) against at least one bacterial strain, resulting in an 86% hit rate. Pearson and Spearman correlation coefficients between experimentally determined MICs and APEX predictions for the 100 optimized sequences were 0.463 and 0.462, respectively, underscoring APEX's predictive power and the effectiveness of using APEX as the oracle for BO. Comparing the mean experimentally determined MICs between optimized peptides and their corresponding templates, we observed that 68% of the optimized peptides showed improved antimicrobial activity after optimization with ApexGO (Fig. 2). When considering the Gram-negative strains only, the improvement rate reached 72%. Template-wise and bacterial-strain-wise analyses (that is, in terms of average mean MIC improvement rate and average MIC log-fold change) indicated that ApexGO was particularly effective at optimizing peptides against Gram-negative pathogens (85% hit rate), consistent with APEX's stronger predictive performance for these bacteria. That being said, peptides optimized for broad-spectrum activity had significantly lower MIC distributions than those optimized specifically for Gram-negative bacteria (P value of one-sided Mann-Whitney U-test, 0.006), highlighting the value of broad-spectrum optimization in ApexGO. Specific observations include modifications to equusin-4 derivatives, such as replacing histidine with isoleucine at position 18 (H18I), which improved activity against Gram-negative strains, particularly A. baumannii, Escherichia coli and P. aeruginosa. Arctoterin-1 optimized peptides showed increased activity against Gram-negative bacteria, with substitutions such as histidine to glycine at position 6 (H6G) enhancing activity against S. aureus. Positively charged modifications in lophiosin-1 at the N terminus maintained antimicrobial activity, whereas negatively charged substitutions failed to improve the broad-spectrum efficacy. In mylodonin-2, substitutions like threonine to arginine or tryptophan at position 10 (T10R or T10W) enhanced activity against K. pneumoniae, especially when combined with amphiphilic N-terminal modifications. For mammuthusins 2 and 3 peptides, increases in normalized hydrophobic moment generally correlated with improved activity, except in specific cases in which crucial residues were altered, such as isoleucine at position 6 in mammuthusin-3-5. Hydrodamin-2 peptides showed increased or maintained activity with various modifications, except for hydrodamin-2-10, which had a unique tyrosine to tryptophan substitution at position 9 (Y9W). Finally, hesperelin-3 peptides lost activity with the introduction of a proline residue in place of glycine at position 6 (G6P), whereas the most effective broad-spectrum peptides preserved the arginine at position 1 (R1). Many of the most potent ApexGO-optimized peptides featured lysine substitutions or insertions (Supplementary Table 1), a modification previously shown to promote Gram-negative accumulation. This observation aligns with findings from small-molecule studies, where the addition of primary amines notably improved their uptake by E. coli. Although our work focuses on peptides rather than small molecules, this convergence -- through increased cationic character -- suggests a potentially generalizable strategy for enhancing Gram-negative permeability across diverse compound classes. Since the optimized peptides sometimes differed substantially from the templates in terms of physicochemical descriptors and antimicrobial activity profiles, we investigated whether the optimization led to changes in their secondary structure compared with the templates. To assess the secondary structure of the peptides, we exposed them to three different media (Fig. 3a and Supplementary Figs. 3-5): water, helix-inducing medium (trifluoroethanol (TFE) in water, 3:2, v:v) and membrane-mimicking environment (sodium dodecyl sulfate (SDS) at 10 mmol l). Equusin-4 derivatives exhibiting higher helicity (Supplementary Figs. 3a-c and 5), particularly in the helix-inducing medium, were more active, especially against K. pneumoniae (Fig. 2). By contrast, arctoterin-1 derivatives (Supplementary Figs. 3d-f and 5) showed no clear correlation between helicity and activity; the optimized peptides that were more active against Gram-negative bacteria were less helical. Lophiosin-1 (Supplementary Figs. 3g-i and 5) and mylodonin-2 (Fig. 3a and Supplementary Figs. 3m-o and 5) derivatives were mostly β-structured or unstructured. In these cases, lower helicity correlated with higher activity, whereas more α-helical derivatives were inactive. Although mylodonin-3 derivatives were primarily unstructured, peptides with slightly higher β-sheet content were more active (Supplementary Figs. 3j-l and 5). Mammuthusin-2 (Supplementary Figs. 4d-f and 5) and mammuthusin-3 (Fig. 3a and Supplementary Figs. 4a-c and 5) derivatives exhibited low helicity with a slight increase in the α-helical structure in lipid bilayers; however, no clear correlation with antimicrobial activity was observed. Similarly, hydrodamin-2 (Supplementary Figs. 4g-i and 5) and hydrodamin-3 (Supplementary Figs. 4j-l and 5) derivatives showed a slight increase in α-helicity in membrane-mimicking environments, but their antimicrobial activity did not correlate with the secondary structure. Hesperelin-3 derivatives were mostly unstructured (Supplementary Figs. 4m-o and 5). An exception was one derivative that showed an increase in α-helicity is the presence of lipid bilayers; however, this did not strongly correlate with activity. Overall, no consistent trend was observed between the secondary structure and the antimicrobial activity across the peptides studied. These findings illustrate that ApexGO-generated peptides can adopt a wide range of structural conformations -- α-helical, β-like or disordered -- highlighting that antimicrobial activity can emerge from diverse structural backgrounds, and that potency improvements are not strictly dependent on converging towards a specific secondary structure. The bacterial membrane is a common target for most AMPs, where they engage in non-specific interactions with the lipid bilayer. The antimicrobial activity of AMPs is influenced by their amino acid composition, distribution and various physicochemical characteristics such as amphiphilicity and hydrophobicity. To investigate the underlying mechanisms by which the AI-optimized peptides predicted by ApexGO kill bacteria, we tested whether differences in the composition of the derivatives -- resulting from substitutions and/or insertions made to the original sequences -- would affect their mechanism of action. For these assays, we selected A. baumannii ATCC 19606, which was the most sensitive strain in our MIC assays. First, we tested whether the peptides permeabilized the bacterial outer membrane using 1-(N-phenylamino)naphthalene (NPN) assays (Fig. 3b and Supplementary Fig. 6). NPN, a lipophilic dye, faintly fluoresces in aqueous solutions but fluoresces more when it encounters lipidic environments such as bacterial membranes. NPN can penetrate the bacterial outer membrane only if it is disrupted or compromised. Among the mylodonin-3 derivatives, mylodonin-3-7 was the only analogue that exhibited superior permeabilization compared with both control antibiotics (Supplementary Fig. 6g,h). Mammuthusin-3 derivatives mammuthusin-3-3, mammuthusin-3-8 and mammuthusin-3-9 also showed slightly better permeabilization than the controls (Fig. 3b and Supplementary Fig. 6k,l), although not as effective as mylodonin-3-7. Within the mammuthusin-2 family, mammuthusin-2-7 had the highest permeabilization activity, with derivatives mammuthusin-2-2, mammuthusin-2-8, mammuthusin-2-9 and mammuthusin-2-10 being slightly more effective than the antibiotics (Supplementary Fig. 6m,n). Arctoterin-1-9 was the most effective permeabilizer among its family, whereas derivatives arctoterin-1-3 and arctoterin-1-7 demonstrated slight permeabilization improvements over the controls (Supplementary Fig. 6c,d) Equusin-4-7 was the strongest permeabilizer among its derivatives, with equusin-4-10 showing marginally better depolarization than the antibiotics (Supplementary Fig. 6a,b). By contrast, the only lophiosin-3 derivative active against A. baumannii was not an effective permeabilizer (Supplementary Fig. 6e,f). Additionally, none of the mylodonin-2 (Fig. 3b and Supplementary Fig. 6i,j), hesperelin-3 (Supplementary Fig. 6s,t), hydrodamin-3 (Supplementary Fig. 6q,r) or hydrodamin-2 (Supplementary Fig. 6o,p) derivatives demonstrated notable permeabilization activity. Next, we tested whether the optimized peptides depolarized the cytoplasmic membrane of A. baumannii (Fig. 3c and Supplementary Fig. 7). We used the potentiometric fluorophore 3,3'-dipropylthiadicarbocyanine iodide (DiSC-5), whose fluorescence is suppressed by its accumulation and aggregation within the cytoplasmic membrane. On disturbances in the transmembrane potential of the cytoplasmic membrane, this fluorophore migrates to the outer environment and emits fluorescence. Polymyxin B was used as a positive control in these experiments, as it is a depolarizer that also permeabilizes and damages bacterial membranes. All the tested peptides demonstrated stronger depolarization activity compared with the antibiotics polymyxin B and levofloxacin. In this work, polymyxin B and levofloxacin are used solely as positive controls to contextualize effect against bacterial pathogens; mechanistic class, macrocyclization and pharmacokinetic differences preclude a direct comparison of these antibiotics with our linear peptides. Among the mylodonin-3 derivatives, mylodonin-3-7 was the most effective depolarizer of A. baumannii, followed closely by mylodonin-3-3, mylodonin-3-4, mylodonin-3-6, mylodonin-3-8 and mylodonin-3-10 (Supplementary Fig. 7g,h). Mylodonin-2-9 was the top depolarizer among the mylodonin-2 analogues (Fig. 3c and Supplementary Fig. 7i,j). Mammuthusin-3 derivatives mammuthusin-3-1, mammuthusin-3-3, mammuthusin-3-6 and mammuthusin-3-7 were slightly better depolarizers than the others (Fig. 3c and Supplementary Fig. 7k,l), whereas mammuthusin-2-7 exhibited the highest depolarization activity within its family, followed by derivatives mammuthusin-2-2, mammuthusin-2-3, mammuthusin-2-8, mammuthusin-2-9 and mammuthusin-2-10 (Supplementary Fig. 7m,n). Lophiosin-1-3 (Supplementary Fig. 7e,f) and several hydrodamin-3 analogues (hydrodamin-3-6, hydrodamin-3-7, hydrodamin-3-9 and hydrodamin-3-10) were also effective depolarizers (Supplementary Fig. 7q,r). Among the hydrodamin-2 derivatives, hydrodamin-2-5, hydrodamin-2-6 and hydrodamin-2-10 showed the best activity (Supplementary Fig. 7o,p), although slightly less than the top performers from other groups. Hepserelin-3 derivatives, namely, hepserelin-3-4 and hepserelin-3-6, and arctoterin-1-5 were the strongest depolarizers in their respective families (Supplementary Fig. 7s,t). Equusin-4 derivatives, particularly equusin-4-2 and equusin-4-7, exhibited uniformly high depolarization activity (Supplementary Fig. 7a,b). Overall, the best depolarizers from each group displayed comparable efficacy, with hydrodamin-2 derivatives being slightly less potent than the rest. These results indicate that although many optimized peptides displayed measurable membrane permeabilization or depolarization, these effects did not consistently correlate with MIC values, suggesting that improvements in antimicrobial potency may arise from a combination of membrane and non-membrane-related mechanisms. All derivatives were tested for cytotoxic activity against human embryonic kidney (HEK293T) cells and compared with their templates (Fig. 3d). This assay is widely used to assess the toxicity of antimicrobials because the results are highly reproducible. Of the peptides tested, 84 displayed no detectable cytotoxicity at the concentration range tested (4-64 μmol l). Among the equusin-4 derivatives, only 5, 6, 7 and 10 showed mild to moderate toxicity, whereas the most active derivatives were non-toxic. Arctoterin-1-1, the most active and broad-spectrum derivative, was the only cytotoxic peptide in its family. Interestingly, although the lophiosin-1 template was cytotoxic, all its derivatives were non-toxic. Most mylodonin-2 derivatives were non-toxic except for mylodonin-2-4 at high concentrations. Mylodonin-3 derivatives optimized for broad-spectrum activity exhibited higher toxicity compared with those optimized for Gram negatives, with mylodonin-3-9 and mylodonin-3-10 being the most toxic, though still at concentrations 10 to 20 times higher than their MICs. Among the mammuthusin-2 derivatives, only mammuthusin-2-5, which had lower antimicrobial activity, was toxic. Mammuthusin-3 and hydrodamin-3 derivatives were non-toxic within the tested concentration ranges. Hydrodamin-2 derivatives with double or triple tryptophan substitutions, optimized for broad-spectrum activity, showed higher toxicity than those targeting Gram negatives. Hesperelin-3 derivatives were generally non-toxic, except for hesperelin-3-9, which had a MIC five to ten times lower than its cytotoxic concentration. Overall, the cytotoxicity varied across peptide families and was often related to the specific optimizations made for antimicrobial activity. Proteolysis assays in human serum are resource-intensive and consume relatively large quantities of peptide. Before conducting in vivo mouse experiments, we assessed the stability of the parent peptides mammuthusin-3 and mylodonin-2, as well as their most active and non-toxic derivatives, mammuthusin-3-6 and mylodonin-2-3, in the presence of serum proteases. Interestingly, after 2 h of incubation, approximately 40% of mylodonin-2-3 remained intact. We attribute this enhanced stability to the replacement of the original arginine residues in mylodonin-2 with aliphatic isoleucine residues, which are less susceptible to proteolysis. By contrast, mammuthusin-3-6 was degraded within the first 30 min, most probably because an arginine residue was introduced relative to the parent peptide (Supplementary Fig. 8). To evaluate the in vivo anti-infective efficacy of the most active optimized derivatives from two highly active template peptides, we used two preclinical mouse models: skin abscess and intramuscular thigh infection. Two de-extinct compounds active against A. baumannii and their respective most active, non-toxic AI-optimized derivatives were tested with a single dose at their MIC concentration after the infection was established. The peptides had a wide range of MIC values (16-64 μmol l) when tested in vitro against A. baumannii: mammuthusin-3 and mammuthusin-3-6 (MIC values of 32 and 16 μmol l, respectively) from the woolly mammoth Mammuthus primigenius and hydrodamin-3 and hydrodamin-3-9 (MIC values of 16 and 8 μmol l, respectively) from the extinct manatee Hydrodamalis gigas; and mylodonin-2 and mylodonin-2-3 (MIC values of 64 and 16 μmol l, respectively) from the extinct giant sloth Mylodon darwinii. In the skin abscess infection model, mice were infected with a bacterial load of 5 × 10 cells of A. baumannii (Fig. 4a). Each peptide or antibiotic (positive control) was administered as a single dose at their MIC over the infected area. After 2 days, the bacterial counts revealed that all the tested peptides significantly reduced the bacterial load by up to four orders of magnitude. In particular, mylodonin-2-3 demonstrated faster bacterial clearance in the skin abscess model compared with all controls. By day 2 post-treatment, colony-forming units (CFU) levels in the mylodonin-2-3 group were markedly reduced relative to polymyxin B, levofloxacin and the parent peptide, indicating a strong therapeutic response at an early time point. These findings underscore the potent anti-infective properties of these molecules against A. baumannii infections and the efficacy of our ApexGO optimization model in generating high-activity peptides. After 4 days, the optimized derivative mylodonin-2-3 exhibited slightly higher activity than their parent compound, showing antibacterial efficacy similar to that of widely used antibiotics polymyxin B and levofloxacin (Fig. 4b). No variations in weight, damage to the skin tissue, or other harmful consequences induced by the molecules or their derivatives were observed in the mice throughout our experiments (Supplementary Fig. 9a). In the murine deep thigh infection model, the efficacy of mammuthusin-3-6 and its template was assessed following the establishment of an intramuscular infection (Fig. 4c). This well-established preclinical model is suited for evaluating the translatability of potential antibiotics. Mice were rendered neutropenic with cyclophosphamide before intramuscular injection of 9 × 10 A. baumannii cells (Fig. 4d). A single dose of each compound at its MIC was administered intraperitoneally. Two days post-treatment, mammuthusin-3-6 demonstrated activity comparable with the positive control antibiotics, polymyxin B and levofloxacin, reducing the bacterial counts by three orders of magnitude. Four-days post-treatment, the peptides showed bacteriostatic activity reducing the bacterial load by two orders of magnitude compared with the untreated control group (Fig. 4d). No toxicity was observed in treated mice, as indicated by stable weight monitoring (Supplementary Fig. 9b). These robust in vivo results demonstrate that the optimized derivatives possess potent anti-infective efficacy under physiologically relevant conditions. The optimized peptides not only improved on their template counterparts but also performed comparably with widely used antibiotics, highlighting their potential as effective antimicrobial agents.

Penn researchers develop ApexGO to enhance promising antibiotic candidates

University of Pennsylvania School of Engineering and Applied ScienceMay 13 2026 Researchers at the University of Pennsylvania have developed ApexGO, a novel, AI-powered method for turning promising but imperfect antibiotic candidates into more potent ones. Unlike many existing AI approaches to antibiotic discovery, which screen large databases for molecules that might work, ApexGO starts with a small number of imperfect candidates and improves them step by step, using a predictive algorithm to evaluate each modification and guide the next. "Antibiotic discovery is fundamentally a search problem across an enormous molecular space. ApexGO gives us a way to navigate that space with far more direction," says César de la Fuente, Presidential Associate Professor in Bioengineering and in Chemical and Biomolecular Engineering in the School of Engineering and Applied Science, in Psychiatry and Microbiology in the Perelman School of Medicine and in Chemistry in the School of Arts & Sciences, and co-senior author of a new paper describing the method in Nature Machine Intelligence. "ApexGO begins with a promising but imperfect peptide," explains de la Fuente, referring to a short string of amino acids, "proposes precise edits, predicts whether those changes are likely to enhance antimicrobial activity, and then keeps moving toward versions that are more likely to work when we make and test them." Laboratory tests against disease-causing bacteria supported ApexGO's predictions: 85% of the AI-generated molecules halted bacterial growth, while 72% outperformed the peptides from which they were derived. In mice, two antimicrobial peptides created by ApexGO reduced bacterial counts at levels comparable to polymyxin B, an FDA-approved antibiotic used as a last-resort treatment for some drug-resistant infections. What is striking is that ApexGO's predictions held up in the real world. ApexGO was optimizing against another computer model, so one concern was that it might find molecules that looked good to the model but failed in the lab. Instead, the majority of the molecules it designed actually worked." Jacob R. Gardner, Assistant Professor in Computer and Information Science (CIS) and the paper's other senior co-author From screening molecules to making new ones For years, the de la Fuente lab has looked for antibiotic candidates in unlikely places, from frog secretions to ancient microbes. Two years ago, the group released APEX, an AI model that predicts whether or not a given peptide is likely to have antimicrobial properties. "APEX helped us find promising antibiotic candidates in enormous biological datasets," says Marcelo Torres, Research Assistant Professor of Psychiatry in the Perelman School of Medicine and co-first author of the paper, referring to work that revealed antibiotic candidates everywhere from woolly mammoths to giant sloths. "ApexGO takes the next step: once we have a promising molecule, it helps us ask how to make it better." That's where Gardner's lab comes in. The group specializes in methods like Bayesian optimization, which helps AI systems explore large numbers of possible solutions efficiently. "It would be impossible to test every possible peptide," says Yimeng Zeng, a doctoral student in CIS and co-first author of the paper. "Bayesian optimization helps the model make informed choices about what to try next, balancing candidates that look promising with candidates that could teach the model something new." Essentially, one part of ApexGO - short for APEX Generative Optimization - suggests molecular tweaks, while the previously published APEX model predicts whether those changes are likely to increase antimicrobial activity. ApexGO then uses those predictions to guide the next round of proposed edits. "If a region of the search space looks promising, the model can spend more effort exploring nearby variants," says Zeng. "But it can also move into less certain regions, where there may still be hidden improvements." Searching more systematically Until now, the researchers point out, antibiotics have largely been found by accident. The most famous example is also the first: penicillin, which Alexander Fleming discovered after noticing that mold in a petri dish was restricting the growth of bacteria. "In a sense, we've been incredibly lucky," says de la Fuente. "ApexGO points to a more systematic way forward." The space of all possible antimicrobial peptides is huge: Like searching a vast forest for something small or rare, finding an antibiotic peptide is normally prohibitively time-consuming. Even a short peptide can be modified in an enormous number of ways, making it impossible for researchers to synthesize and test every possible version by hand. That ApexGO could identify antibiotic candidates with laboratory activity against disease-causing bacteria, simply by searching this space computationally, points to a different approach. "We ran ApexGO for a few months and found hundreds of candidates," notes Gardner. "If we ran that process for a year, how many thousands of these could we find?" "This result points toward a future in which we can optimize molecules for a desired function in a fraction of the time," adds de la Fuente, "using machines to guide discovery through chemical spaces too vast for humans to explore by trial and error." Future directions While some of the molecules proposed by ApexGO showed promising antibiotic activity, the researchers emphasize that even the best-performing peptides are still early-stage candidates. Before any could be used to treat infections in humans, they would need to be further optimized for safety, stability and how long they remain active in the body. Still, the study suggests that AI can help researchers decide which molecules are worth making and testing in the first place. Instead of synthesizing one candidate after another by trial and error, tools like ApexGO could help narrow the search to molecules more likely to work. For de la Fuente, that approach could eventually extend beyond antibiotics. "In this case, we wanted to optimize peptides for antimicrobial activity," he says. "But you could imagine applying the same idea to peptides with other biological functions, like modulating the immune system or targeting tumors." Gardner's lab is already exploring related approaches using AI agents, which may be able to draw on scientific knowledge and reason through design choices. "The larger idea is that AI can help scientists search spaces that are too large to explore by hand," says Gardner. "ApexGO is one example of that. The next generation of tools may be able to explore these spaces in even more flexible ways." "ApexGO shows that AI can do more than predict which molecules might work: it can help us improve them," adds de la Fuente. "At a time when antibiotic resistance is rising worldwide, we need technologies that help us move faster from an idea to a real therapeutic candidate. ApexGO is an important step toward that future." University of Pennsylvania School of Engineering and Applied Science Journal reference: Torres, M. D. T., et al. (2026). A generative artificial intelligence approach for peptide antibiotic optimization. Nature Machine Intelligence. DOI: 10.1038/s42256-026-01237-5. https://www.nature.com/articles/s42256-026-01237-5