TACTIC: A transfer learning framework to predict drug interactions in emerging pathogens

Summary

Machine learning (ML) is necessary to efficiently identify potent drug combinations within a large candidate space to combat drug resistance. However, existing ML approaches cannot be applied to emerging and understudied pathogens with limited training data. Hence, we developed a computational framework (TACTIC [transfer-learning and cross-species training to infer combination therapies]) to train ML models on data from multiple model bacteria. TACTIC was built using 2,965 drug interactions from 12 bacterial strains and outperformed traditional ML models in predicting drug interaction outcomes for species with limited training data. Upon analyzing ∼600,000 predicted drug interactions across 18 pathogenic and commensal strains, we identified interactions that are selectively synergistic against Gram-negative and mycobacterial pathogens. We experimentally validated synergistic drug combinations containing clarithromycin, ampicillin, and mecillinam, against clinical strains of M. abscessus, an emerging drug-resistant pathogen. We also confirmed 11 synergistic combinations against Pseudomonas aeruginosa and Staphylococcus aureus strains that cause endophthalmitis.

Graphical abstract

Highlights

• •TACTIC computationally identifies antibiotic combinations for emerging pathogens
• •Orthology mapping between reference and target strains enables transfer learning
• •Phylogenetic and metabolic patterns are correlated with strain-specific outcomes
• •Synergistic combinations were validated in vitro for M. abscessus and P. aeruginosa

Motivation

Machine learning (ML) tools can accelerate the search for combination therapies effective against drug-resistant pathogens. However, the scarcity of drug-response and drug-interaction data for emerging pathogens such as Mycobacterium abscessus impedes the development of ML-based approaches. We bridge this gap by curating data across multiple species and using orthology mapping to perform transfer learning, an ML strategy in which a model trained on one task is applied to a related task, thereby reducing data requirements. The resulting TACTIC framework uses model-organism data to build ML models that predict drug-interaction outcomes for emerging pathogens with available genome sequences.

Chung et al. develop TACTIC, a framework that leverages data from model organisms to identify promising antibiotic combinations against emerging pathogens. TACTIC uses transfer learning to train from drug response and interaction data sourced across 12 bacterial strains and accurately prioritizes promising synergistic combinations against Gram-negative and Mycobacterium pathogen strains.

Introduction

Antibiotic resistance is becoming a global health crisis as growing resistance outpaces current antibiotic drug development. In the United States alone, more than 2.8 million antibiotic-resistant infections occur each year, with more than 35,000 cases resulting in death.¹ These infections are caused by a diverse set of bacterial pathogens that pose major health threats, especially if left unchecked. Some of the most concerning species include carbapenem-resistant Acinetobacter baumannii, multidrug-resistant Pseudomonas aeruginosa, and drug-resistant Mycobacterium tuberculosis (henceforth M. tb).^1,2 At the same time, a few new classes of antibiotics have been brought to the market for decades.³

One possible solution for overcoming antibiotic resistance is to design synergistic drug combination therapies.⁴ Such combinations could engage multiple cellular targets to suppress growing resistance, which is difficult to achieve with a single active compound. A major drawback, however, is the challenge in exploring a vast combinatorial space that exponentially increases when considering new drugs or dosage levels. Due to the empirical nature of designing drug combination therapies, the discovery of new combined regimens has been slow. There is, therefore, a dire need for alternative approaches that can efficiently screen and prioritize promising drug combinations.

Machine learning (ML) has recently been leveraged to design synergistic drug combinations for various complex diseases.^{5,6,7,8,9,10,11,12,13} In the context of antibiotic resistance, an approach called INDIGO was developed to find combination therapies effective against Escherichia coli and M. tb.^14,15 Specifically, omics data for E. coli and M. tb treated with individual drugs, along with corresponding drug combination assay data, were used to construct two separate models predictive of drug interaction outcomes (e.g., synergy) for each organism. The E. coli INDIGO model was further extended to predict drug interaction outcomes for Staphylococcus aureus by leveraging omics data for genes that are orthologous between the two species. Although INDIGO was shown to yield predictions that correlate with experimental data, the INDIGO models are not highly predictive for species with limited data compared to E. coli (e.g., mycobacteria).

Within this study, we leverage transfer learning and literature mining to construct predictive models of drug interactions for emerging and clinically relevant pathogens, for which little or no data exist for training traditional ML models. Transfer learning is an ML concept where a model trained on one task is re-applied to solve a different but similar task, typically after some fine-tuning.¹⁶ In this case, we hypothesized that an ML model built using abundant data from model organisms can be applied to pathogen species with limited training data using transfer learning. We integrate this idea onto a modeling foundation inspired by INDIGO to develop a new approach called TACTIC: transfer learning and cross-species training to infer combination therapies. We show that TACTIC generates more accurate predictions than INDIGO for Gram-negative and mycobacteria strains. We then interpret the TACTIC model to uncover genetic and metabolic patterns correlated with cross-species and species-specific drug interaction outcomes. Finally, we apply TACTIC to predict drug combinations that show narrow-spectrum synergy against clinically relevant pathogens and additivity or antagonism against common commensal microbes.

Results

Drug interaction data collection

To build and validate our framework, we compiled drug interaction data from 17 publications^{14,15,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31} (Figure S1A; Table 1; Data S1A). These studies were selected because numerical Bliss and/or Loewe scores were reported for at least one bacterial strain, and we filtered out any drug combination for which drug response data were not available (see STAR Methods). In total, we curated 2,965 drug interactions involving 88 conditions (86 drugs and 2 media) measured across 12 strains representative of six species: A. baumannii, E. coli, M. tb, P. aeruginosa, S. aureus, and Salmonella enterica serovar Typhimurium (henceforth S. Typhimurium) (Figure S1B; Data S1B and S1C). Drug interactions were quantified based on the Loewe additivity³² or Bliss independence³³ model. Based on the drug interaction classification scheme defined for each study, our pooled data consist of 913 synergistic (31%), 826 neutral (28%), and 1,226 antagonistic (41%) interactions (Figure S1C). Our pooled data are representative of 2- to 10-way drug combinations, although more than 85% (n = 2,522) consists of pairwise interactions.

Table 1.

TACTIC data collection

Organism	Strain	Ntotal	Nhigh	Model	Interaction outcome						Reference
–	–	–	–	–	Synergy		Neutral		Antagonism		–
A. baumannii	ATCC 17978	45	30	Loewe	23	(0.51)	10	(0.22)	12	(0.27)	Cokol et al.28
E. coli	BW25113	473	–	Both	117	(0.25)	155	(0.33)	201	(0.42)	Brochado et al.25 and Ocampo et al.31
E. coli	MC4100	98	26	Both	12	(0.12)	24	(0.25)	62	(0.63)	Russ et al.19
E. coli	MG1655	623	185	Both	123	(0.20)	218	(0.35)	282	(0.45)	Chandrasekaran et al.,14 Yeh et al.,21 Cokol et al.,28 and Katzir et al.30
E. coli	iAi1	316	–	Bliss	116	(0.37)	66	(0.21)	134	(0.42)	Brochado et al.25
M. tuberculosis	ATCC 35801	154	134	Bliss	85	(0.55)	41	(0.27)	28	(0.18)	Silva et al.20
M. tuberculosis	H37Rv	389	68	Loewe	181	(0.47)	56	(0.14)	152	(0.39)	Ma et al.,15 Rey-Jurado et al.,17 Rey-Juradoet al.,18 Yilancioglu et al.,22 Zhao et al.,23 Bhusal et al.,24 Chandramohan et al.,26 Coelho et al.,27 and Gonzalo et al.29
P. aeruginosa	PAO1	163	–	Bliss	44	(0.27)	47	(0.29)	72	(0.44)	Brochado et al.25
P. aeruginosa	PA14	163	–	Bliss	51	(0.31)	54	(0.33)	58	(0.36)	Brochado et al.25
S. aureus	ATCC 29213	45	–	Loewe	5	(0.11)	23	(0.51)	17	(0.38)	Chandrasekaran et al.14
S. Typhimurium	14028s	248	–	Bliss	77	(0.31)	68	(0.27)	103	(0.42)	Brochado et al.25
S. Typhimurium	LT2	248	–	Bliss	79	(0.32)	64	(0.26)	105	(0.42)	Brochado et al.25

Drug interaction outcome measurements gathered from 17 independent studies.N_total, total number of interactions; N_high, number of higher-order interactions.See also Figure S1 and Data S1.

Of the 1,370 unique drug combinations represented in our data collection, 757 (∼55%) were measured more than once (Figure S1D; Data S1D). For drug combinations repeatedly measured in different strains (n = 664), the vast majority (>86%) either agreed on a single outcome class (n = 215) or disagreed between neighboring class labels (synergy vs. neutral or neutral vs. antagonism, n = 357) (Figure S1E). For example, polymyxin B (PMB) paired with benzalkonium, rifampicin (RIF) (PMB-RIF), or verapamil was found to be synergistic across six strains, suggesting broad-spectrum activity that may be a strategic therapeutic option when the causative agent is unknown. In contrast, only ∼14% (n = 92) of drug interactions measured in more than one organism were found to have diverging outcomes across different strains. For example, A22 combined with mitomycin C was antagonistic or additive across most strains except E. coli BW25113 (Figure S1F). Such drug interactions may represent those with narrow-spectrum activity, which could be leveraged to treat bacterial infections where the causative agent is known while minimizing unintended damage to other bacteria (e.g., the host microbiome).

The TACTIC approach

TACTIC is an extension of INDIGO,¹⁴ a computational approach for predicting drug interaction outcomes between antibacterial agents. The inputs for both TACTIC and INDIGO are known drug interaction scores for model training and drug response features (gene-drug interactions) from transcriptomics or chemogenomics. Unlike INDIGO models, which are trained on data for a single organism (i.e., E. coli or M. tb), a TACTIC model is trained on omic and drug interaction data measured in multiple organisms. TACTIC specifically expands on the orthology aspect, which enables information transfer from a reference dataset measured for one organism (e.g., E. coli) to another (e.g., A. baumannii). Each organism in TACTIC is represented by its unique list of genes and orthologs, which serve as mapping information for transfer learning.

The TACTIC model is trained on features and labels based on single-drug response omics data and drug interaction data, respectively (Figure 1A). ML features are defined by a sigma-delta vector, which mathematically represents the combined and drug-unique effect for a unique drug combination. The sigma-delta vector is then filtered based on orthology mapping between a reference strain (e.g., E. coli MG1655) and a strain of interest (e.g., A. baumannii), where feature information is retained for conserved genes (Figure 1B). This data processing step enables transfer learning by fine-tuning the model for a given strain based on the identity of orthologous genes. Drug interaction scores, which are directly provided from drug interaction data, are used as ground-truth labels to train a regression-based TACTIC model (Figure 1C).

Figure 1.

TACTIC approach schematic

The transfer-learning and cross-species training to infer combination therapies (TACTIC) approach uses (A) drug response and drug interaction data from multiple species as input, (B) which both undergo data processing via orthology mapping and concatenation to (C) define machine learning (ML) features and labels to train a base model that (D) predicts drug interaction outcomes for new drug combinations and/or organisms. To apply this base ML model to a new species using transfer learning, orthology mapping is performed, which allows the transfer of drug response data measured in one organism to another by conserving data between orthologous genes. Orthology-mapped drug response data are subsequently processed to define ML features for a given organism and a set of drug combinations. This is analogous to modifying a base ML model built for identifying domestic cats using abundant public data to identify tigers using transfer learning. In this study, a base ML model built using abundant data from diverse model organisms like E. coli is then modified based on orthology and applied to pathogen species with limited or no training data.

The ability to incorporate multi-species drug interaction and omics data collected from multiple sources enables the TACTIC model to learn from a larger and more varied set of data compared to when training on data from a single organism. This allows TACTIC to use the base ML model trained on multi-species data and infer drug interaction outcomes for an unseen strain of interest based on gene orthology information alone (Figure 1D). Given that each organism is represented by a unique list of genes and orthologs, the drug response features in TACTIC and subsequent drug-interaction predictions are different for each strain. Further details on the exact steps involved in developing a TACTIC model (e.g., defining ML model features, the ML algorithm used) are described in the STAR Methods section.

Benchmarking zero-shot TACTIC predictions in unseen test strains

We benchmarked TACTIC against INDIGO by evaluating their accuracy in predicting drug interaction outcomes for an unseen test strain, wherein each of the 12 bacterial strains for which data were collected above were hidden from the models iteratively. This represents a zero-shot prediction, wherein no additional data from a species of interest is used to fine-tune the TACTIC model. For the TACTIC approach, we trained a separate model iteratively for each test strain, where drug interactions measured for the strain of interest (N) were set aside for testing, while the remaining drug interaction data (2,965−N) were used for training (Figures S2A and S2B). INDIGO models were trained using the drug interaction data that were previously used to define the E. coli¹⁴ and M. tb¹⁵ models (Figure S2C). Model accuracy was measured based on the Spearman rank correlation (Spearman R) and the mean squared error (MSE) between model predictions and experimentally measured drug interaction scores.

We found that TACTIC models were considerably better at predicting drug interaction outcomes for A. baumannii, P. aeruginosa, S. Typhimurium, and most E. coli strains compared to INDIGO models and randomized TACTIC predictions (baseline comparison) (Figures 2A and 2B). Unsurprisingly, the INDIGO models built using E. coli and M. tb data yielded slightly more accurate predictions than TACTIC for E. coli MC4100 and M. tb, respectively. Interestingly, all models were less accurate in predicting drug interaction outcomes for S. aureus (Spearman R < 0.4 and MSE >1 for all models). This is likely because S. aureus (Gram-positive) is phylogenetically distant from all other bacteria used for constructing TACTIC and INDIGO (Gram-negative or Mycobacterium strains). Hence, the level of transfer learning from data measured in E. coli and M. tb to generate predictions in S. aureus is limited due to the small amount of overlap in orthologous genes. It is also worth noting that ∼80% of drug interaction data measured outcomes in Gram-negative species (Figure S1B), which may bias TACTIC models to be more accurate for these strains (e.g., A. baumannii).

Figure 2.

Benchmarking TACTIC against INDIGO using leave-one-strain-out validation

(A and B) To assess how well TACTIC generalizes to unseen microbial species, each microbial strain in the TACTIC compendium was iteratively left out of the training process, and its interactions were predicted using the other species data. Strain-specific drug interaction outcome prediction accuracy was compared between TACTIC models and INDIGO models developed from E. coli or M. tb data.^14,15 Random TACTIC predictions were defined by random shuffling of strain-specific TACTIC model predictions for each strain and served as baseline comparison. Model performance was quantified based on (A) the Spearman rank correlation (Spearman R) and (B) the mean squared error (MSE) between model predictions and experimentally measured drug interaction scores.

(C) Strain-specific TACTIC model confidence (defined as the fraction of top predictive features explaining 95% of the variance in model predictions conserved by orthologous genes between E. coli MG1655, M. tb H37Rv, and the given strain) is correlated with model performance as measured by the Spearman R and MSE.

See also Figures S2, S7, and Data S1.

To allow users to determine beforehand the potential effectiveness of transfer learning, we quantified the conservation of top predictive genes in the TACTIC model to create a confidence score. As expected, this confidence score shows a strong linear relationship with model performance in benchmarking test sets across all 12 strains (Figure 2C). We also confirmed that INDIGO and TACTIC model performance is low for other Gram-positive strains based on an independent drug interaction dataset measured in Bacillus subtilis 168, S. aureus Newman, and Streptococcus pneumoniae D39V³⁴ (Data S1E and S1F).

Genetic predictors of cross-species and species-specific drug interaction outcomes

Having confirmed that TACTIC provides an improvement over INDIGO in predicting drug interaction outcomes cross-species, we next trained a TACTIC model with the entire data collection on the 2,965 drug interactions measured across 12 bacterial strains. Using this fully trained TACTIC model, we sought to uncover which features the model most relied on for generating predictions and how changes in these feature values correlated with synergistic or antagonistic outcomes. For this task, we first determined the set of ML features, ranked by decreasing importance score, that explained 95% of the variance in model predictions (Data S2A, see STAR Methods). We then conducted a gene set enrichment analysis of all genes associated with the top ML features (N ∼1,400) against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database,³⁵ an extensively curated resource that maps gene function information onto all known cellular processes for a given organism. We determined four pathways that are significantly enriched (Figure S3A, hypergeometric test, adjusted p value <0.05); interestingly, all pathways relate to bacterial metabolism, with the biosynthesis of amino acids enriched by genes specific to both E. coli MG1655 and M. tb H37Rv. We found that over 50% of genes mapped to KEGG were highly conserved across the 12 strains represented in the training dataset (Figure S3B), which may explain why metabolic pathways, which are highly conserved, were uniquely found to be enriched.

To gain a more fine-grained insight, we evaluated how the values for ML features associated with genes belonging to these pathways varied between synergistic and antagonistic interactions (Data S2B, see STAR Methods). We determined 11 genes in central metabolic pathways for which feature values significantly differed between synergistic and antagonistic interactions (Table 2, Data S2C). The trends for all 11 genes indicate that drug interactions involving at least one drug that perturbs central metabolic pathways (i.e., glycolysis, tricarboxylic acid [TCA] cycle, and oxidative phosphorylation) likely result in antagonistic outcomes. This aligns with the hypothesis that active bacterial metabolism potentiates drug efficacy,^36,37 although these observations are correlative and require experimental validation.

Table 2.

E. coli MG1655 gene knockouts associated with antagonistic outcomes

Locus	Gene	Description	Primary pathway	KO fitness
b0115	aceF	Pyruvate dehydrogenase, E2 subunit	Glycolysis/gluconeogenesis	Positive
b0118	acnB	Hypothetical protein	Citrate cycle (TCA cycle)	Positive
b1136	icd	Isocitrate dehydrogenase	Citrate cycle (TCA cycle)	Positive
b3737	atpE	ATP synthase Fo complex subunit c	Oxidative phosphorylation	Positive
b4042	dgkA	Diacylglycerol kinase	Glycerolipid metabolism	Positive
b3041	ribB	3,4-Dihydroxy-2-butanone-4-phosphate synthase	Riboflavin metabolism	Positive
b0844	ybjI	5-Amino-6-(5-phospho-D-ribitylamino) uracil phosphatase	Riboflavin metabolism	Negative
b3620	waaF	ADP-heptose-LPS heptosyltransferase 2	Lipopolysaccharide biosynthesis	Positive
b3625	waaY	Lipopolysaccharide core heptose (II) kinase	Lipopolysaccharide biosynthesis	Negative
b3628	waaB	UDP-D-galactose: (glucosyl)lipopolysaccharide-1,6-D-galactosyltransferase	Lipopolysaccharide biosynthesis	Negative
b3632	waaQ	Lipopolysaccharide core heptosyltransferase 3	Lipopolysaccharide biosynthesis	Negative

KO, knockout.

See also Figure S3 and Data S2.

We next investigated the associations between ML features and species-specific drug interaction outcomes by analyzing 32 drug combinations measured in 6 different strains within the same study.²⁵ Within this subset, we found that the deviation in cross-species drug interaction outcomes aligns with the phylogenetic distance for combinations, especially for those involving bacteriostatic-bacteriostatic interactions (Data S2D). This may indicate that bacteriostatic drug effects are more dependent on the genetic state of a given organism, as opposed to bactericidal agents, which are believed to be more dependent on the phenotypic (e.g., metabolic) state of a bacterial cell.^36,38,39,40 This hypothesis aligns with our re-analysis of experimental measurements of the impact of various carbon sources on drug interactions.⁴¹ We observed a larger variation in drug interaction outcomes for cefoxitin (bactericidal) paired with tobramycin (TOB) (bactericidal) compared to cefoxitin paired with tetracycline (bacteriostatic) on E. coli growth across 95 carbon sources (Figure S3C). However, a larger investigation of combined drug efficacy by mode of action (bacteriostatic vs. bactericidal) across defined metabolic conditions is necessary to confirm this trend.

Interaction landscape of ∼3,600 drug pairs in 18 strains across 9 media conditions

We next applied the fully trained TACTIC model to generate drug interaction predictions across new strains and media conditions. Given that metabolic genes were top features in our model, we sought to understand the degree to which drug interaction outcomes differ across distinct organisms and growth environments. Recent studies have shown that the growth media condition may sometimes have a larger influence on drug interaction outcomes than strain-specific differences⁴²; hence, we applied TACTIC to generate strain- and media-specific drug interaction predictions. We included all 12 strains for which the TACTIC model was trained on as well as two non-tuberculous mycobacteria (NTM) species (Mycobacterium abscessus and Mycobacterium smegmatis) and four bacteria representative of gut microbiome commensals (Bacteroides vulgatus, Eubacterium eligens, Eubacterium rectale, and Lactobacillus rhamnosus). These new strains were selected based on growing concerns over NTM infections⁴³ and studies that have deduced the species-level identities of commensal gut microbes,^44,45 for which there is increasing appreciation over their symbiotic relationship with human hosts.^46,47,48,49 For the media conditions, we accounted for 9 nutrient sources that are present in the input omics data for TACTIC. These included seven carbon sources (acetate, glucosamine, glucose, glycerol, maltose, N-acetylglucosamine, and succinate), one nitrogen source (ammonium chloride), and one nutrient-rich condition (LB).

In total, we generated ∼600,000 outcome predictions spanning all possible drug pairs among 86 drugs (n = 3,655) across 18 bacterial strains and 9 media conditions (Figure S4A; Data S3A). Given the large scale of these data, we applied principal-component analysis (PCA) to inspect how each strain contextualized within each media condition cluster among one another (Figure S4B). Interestingly, we found that differences in predicted outcomes were primarily driven by the media context, where predictions made in certain media conditions (e.g., N-acetylglucosamine) clustered within a narrow range across the first principal component (PC1). The patterns seen across PC1 indicate that the growth environment may sometimes have a larger impact on the combined drug effect across a diverse range of bacteria, aligning with findings reported by Davis et al.⁴² Nonetheless, we observed that clustering patterns along the second principal component (PC2) were strongly driven by the strain identity, with some organisms clustering along both PC1 and PC2 (i.e., S. aureus, Mycobacterium spp.). Of note, we found that TACTIC predictions correlated with experimentally measured outcomes for M. tuberculosis grown in minimal media supplemented with glycerol⁵⁰ (R = 0.441, p = 0.03, Figure S4C).

To further understand strain-specific differences in predicted drug interaction outcomes, we decided to conduct strain-to-strain comparisons for predictions contextualized in LB, the media condition where strain-to-strain variance was the highest (Data S3B, see STAR Methods). Using TACTIC predictions on 3,655 drug interaction outcomes across 18 bacterial strains contextualized in LB, we first evaluated the similarity between strain-to-strain predictions. By defining similarity as the Pearson correlation between strain-specific predictions, we found that drug interaction outcome similarities are primarily based on the Gram stain of an organism (Figure S4D). Of note, outcome predictions for S. aureus were the most dissimilar from all other strain-specific predictions (lowest correlation <0.6). Unsurprisingly, strain-specific predictions for the same species were most similar to one another, and mycobacteria-specific predictions clustered among each other. Of note, strain-specific predictions seemed to cluster based on the strain-specific relationship with the host (i.e., pathogen vs. commensal) within Gram stain groups. This finding supported the potential of identifying drug combinations with selective synergy against pathogenic bacteria.

Identifying narrow-spectrum drug synergies against distinct pathogen groups

Prior to experimental evaluation of promising synergies, we first evaluated how TACTIC predictions aligned with existing evidence on drug interaction outcomes measured in organisms accounted for in our model.^{51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70} Of note, these outcomes were reported as qualitative observations; hence, they were not part of the training set and instead served as validation data for TACTIC model predictions. We found that TACTIC correctly predicted broad-spectrum synergy for most drug combinations determined to be synergistic across a wide range of bacteria, regardless of the combined mechanism of action (Table 3). We also found that TACTIC predictions generally aligned with literature evidence pointing to selective synergistic drug combinations. For example, TACTIC correctly predicts that clarithromycin (CLA) combined with RIF (CLA-RIF) is synergistic against mycobacteria.^62,63,64,65 Interestingly, TACTIC predicts that PMB-RIF is selectively synergistic against Gram-negative and mycobacteria species; although prior evidence of synergy against Gram-negative bacteria exists,⁶⁷ the effect of this drug combination has yet to be measured in mycobacteria. Another interesting drug combination is ampicillin (AMP) paired with another beta-lactam drug aztreonam (AZT) (AMP-AZT), which was previously shown to be synergistic against various Gram-negative pathogens^59,60; yet TACTIC predicts broad-spectrum synergy for this combination, which warrants its evaluation against other bacteria. Of note, we also found that TACTIC correctly predicted non-synergistic outcomes (additivity or antagonism) in P. aeruginosa, which has been reported across multiple independent studies (Table 3). Given that published data are heavily biased toward reporting synergy compared to non-synergistic outcomes, we were unable to find additional evidence for additive or antagonistic drug interactions in other bacteria.

Table 3.

TACTIC prediction agreement with literature-based evidence

Interaction	Mechanism	Organism measured	Literature consensus	TACTIC prediction
AMP-AMK	Beta-lactam and aminoglycoside	Gram-negative bacteria51,52,53; S. aureus51	Broad-spectrum synergy	Broad-spectrum synergy
AMK-AMX	Beta-lactam and aminoglycoside	E. coli54	Broad-spectrum synergy	Broad-spectrum synergy
AMK-AZT	Beta-lactam and aminoglycoside	P. aeruginosa54,55,56,66	Broad-spectrum synergy	Broad-spectrum synergy
AZT-TOB	Beta-lactam and aminoglycoside	P. aeruginosa57	Broad-spectrum synergy	Broad-spectrum synergy
AMP-GEN	Beta-lactam and aminoglycoside	Gram-negative bacteria58	Selective synergy in Gram-negative bacteria	Selective synergy in E. coli and S. Typhimurium
AMP-AZT	Beta-lactam pair	Gram-negative bacteria59,60	Selective synergy in Gram-negative bacteria	Broad-spectrum synergy
SMM-TMP	Double targeting of folic acid biosynthesis	Gram-positive and Gram-negative bacteria61	Broad-spectrum synergy	Broad-spectrum synergy
CLA-RIF	Target RNA and protein biosynthesis	NTM species62,63,64,65	Selective synergy in mycobacteria	Selective synergy in mycobacteria
PMB-RIF	Target RNA and lipopolysaccharide biosynthesis	Gram-negative bacteria67	Selective synergy in Gram-negative bacteria	Selective synergy in Gram-negative and mycobacteria
PMB-TMP	Target lipopolysaccharide and folic acid biosynthesis	Gram-positive and Gram-negative bacteria68,69,70	Broad-spectrum synergy	Selective synergy in mycobacteria and S. aureus
AMK-CIP	Target DNA and protein biosynthesis	P. aeruginosa66	Additivity or antagonism	Additivity or antagonism
AZT-CIP	Target DNA and protein biosynthesis	P. aeruginosa66	Additivity or antagonism	Additivity or antagonism
RIF-SM	Target RNA and protein biosynthesis	P. aeruginosa67	Additivity or antagonism	Additivity or antagonism
RIF-TET	Target DNA and protein biosynthesis	P. aeruginosa67	Additivity or antagonism	Additivity or antagonism

Comparison between 14 drug interaction outcomes predicted by TACTIC against consensus derived from prior studies.Refer to Data S1C for more information on abbreviated drug names.See also Figures S4 and S5 and Data S3.

To identify drug combinations with narrow-spectrum synergy, we compared TACTIC model predictions for commensals vs. three strain groups: all pathogens, Gram-negative pathogens, and NTM pathogens (Figure S5A). A visual inspection of the top 20 narrow-spectrum synergies revealed that broad synergistic outcomes are hard to achieve for all pathogens, most likely because this group includes both Gram-negative and Gram-positive species. For the second comparison with Gram-negative pathogens, 34 drug combinations were predicted to have comprehensive narrow-spectrum synergy (Figure S5B), with stronger synergies predicted against E. coli and S. Typhimurium strains compared to A. baumannii or P. aeruginosa strains. For the third comparison, 16 drug combinations were predicted to have strong comprehensive narrow-spectrum synergy against NTM pathogens (Figure 3A). Interestingly, most of these combinations (n = 14) include CLA, which is known to have potent effect against NTM strains.⁷¹

Figure 3.

Validation of narrow-spectrum synergy predictions for M. abscessus

(A) Sixteen drug pairs were predicted to have strong comprehensive synergy (all predicted scores < −0.2) against non-tuberculous mycobacteria (NTM) pathogens.

(B) Comparison of TACTIC model predictions against an independent dataset that measured outcomes for 96 drug interactions in M. abscessus ATCC 19977 revealed that predictions for synergistic outcomes (log2FIC50 < 0, n = 52) are statistically lower than predictions for antagonistic (log2FIC50 ≥ 0, n = 44) interactions. ∗∗∗ p = 0.0007605 (one-sided Student’s t test).

(C) Bliss score quantification of outcomes for two drug interactions (AMP-CLA and AMP-CLA-MEC) measured in triplicate biological replicates across three M. abscessus strains (ATCC 19977, smooth morphotype, and rough morphotype). AMP, ampicillin; CLA, clarithromycin; and MEC, mecillinam.

(D) Dose-response checkerboards showing percent cell viability across three M. abscessus strains after combined treatment with AMP-CLA or AMP-CLA-MEC. Checkerboards highlighted in red indicate those showing single-drug effect (≥ 50% change in cell viability) across both therapeutic agents for a given combination.

See also Data S3 and S4.

While the focus of TACTIC was to uncover species-level differences, we also examined whether any drug interactions were predicted to exhibit selective synergy between strains within the same species. This is a challenging task due to the lack of extensive training data documenting strain-level differences. For this analysis, we compared predicted outcomes for three strain pairs: E. coli BW25113 vs. iAi1, P. aeruginosa PAO1 vs. PA14, and S. Typhimurium LT2 vs. 14028s. These strains were specifically selected, given that drug interaction outcomes had been previously measured and compared for each pair.²⁵ Overall, we found that very few drug interactions (<5%) had diverging outcomes (synergy vs. additivity or antagonism) between strains of the same species (Figure S5C). Focusing on interactions with strong narrow-spectrum synergy (Figure S5D), we found four interactions that were not previously measured. Interestingly, all four are predicted to be synergistic against environmental or clinical strains (E. coli iAi1 and S. Typhimurium 14028s).

Experimental validation of TACTIC predictions for M. abscessus

Recent trends in antimicrobial surveillance indicate that infections caused by NTM pathogens are on the rise,⁷² with some species (e.g., M. abscessus) already developing concerning levels of antibiotic resistance.⁷³ Given the lack of comprehensive drug response and drug interaction datasets for NTM, it is challenging to build ML models tailored for these pathogens. To verify whether TACTIC can address this gap, we compared TACTIC predictions against experimental data from a recent study by Van et al. that measured 153 pairwise outcomes between 18 antibiotics in M. abscessus ATCC 19977.⁷⁴ Based on the drugs accounted for in TACTIC as well as drug interaction measurements that passed quality control in the Van et al. study, we were able to compare TACTIC predictions against experimental outcomes for 96 drug pairs (Data S4A). We classified interactions based on the authors’ (Van et al.) reported convention for defining synergy (log₂FIC₅₀ < 0) and antagonism (log₂FIC₅₀ ≥ 0). We found that TACTIC predictions generally agreed with experimental outcomes, where predictions for synergistic interactions were statistically lower than those for antagonistic interactions (Figure 3B, one-sided Student’s t test, p value <0.001). Of note, TACTIC correctly predicted novel synergies determined by Van et al. (e.g., bedaquiline combined with amikacin or azithromycin) as well as known outcomes in M. abscessus from prior studies (e.g., synergy for linezolid combined with amikacin and antagonism for linezolid combined with moxifloxacin).⁷⁵

In addition to validating TACTIC predictions using an independent dataset, we sought to experimentally verify outcomes for the top two narrow-spectrum synergy predictions against NTM pathogens (Figure 3C). For this task, we measured the combined effect of AMP-CLA and CLA-MEC in M. abscessus ATCC 19977 and two clinical isolates from a patient with cystic fibrosis (Data S4B, see STAR Methods). Of note, these two clinical isolates exhibited differences in morphotypes (smooth vs. rough), which is known to influence pathogenesis and antibiotic resistance in M. abscessus.⁷⁶ Given the fact that we could not measure the combined effect for CLA-MEC due to an indeterminate minimal inhibitory concentration for mecillinam (MEC), we measured the combined effect of all three antibiotics (AMP-CLA-MEC), where CLA and MEC were pre-combined in a 1:1 solution. Of note, TACTIC predicted that this three-drug combination also exhibits narrow-spectrum synergy against NTM pathogens (Data S3A).

We quantified drug interactions using a manual calculation based on the Bliss independence model³³ as well as SynergyFinder,⁷⁷ a web-based tool for analyzing muti-drug response data (see STAR Methods). Manually determined Bliss scores indicated that both AMP-CLA and AMP-CLA-MEC were synergistic against all three M. abscessus strains, while SynergyFinder Bliss scores also indicated synergistic outcomes except for AMP-CLA in two M. abscessus strains (ATCC 19977 and Rough morphotype, Figures 3C and Data S4C). Visual inspection of the dose-response checkerboards for AMP-CLA in these two strains showed that tendency toward synergy or antagonism varied based on the concentration of CLA (Figure 3D). Specifically, a lower dosage of CLA (≤0.25 μg/mL) appeared to potentiate higher concentrations of AMP (4–64 μg/mL) in both M. abscessus strains. On the other hand, higher concentrations of CLA (≥0.5 μg/mL) combined with lower dosages for AMP (≤2 μg/mL) appeared to yield more antagonistic outcomes.

Prior studies have shown that combinations of CLA with beta-lactam drugs (e.g., AMP and MEC) were antagonistic or neutral in E. coli, S. aureus, and Streptococcus pneumoniae,^14,78 supporting TACTIC’s prediction of narrow-spectrum synergy in NTM. These combinations are predicted to be strongly synergistic in NTM strains and weakly synergistic or antagonistic in all other strains (Data S3C).

Predicting narrow-spectrum synergies against bacterial endophthalmitis

Motivated by the capacity to pinpoint synergistic drug combinations for the emerging pathogen M. abscessus, we next applied TACTIC to predict effective combination therapies against bacterial cases of endophthalmitis, a serious eye infection that can lead to blindness if improperly treated.⁷⁹ In addition to P. aeruginosa, the following species are considered key pathogens: Pseudomonas mendocina, Pseudomonas putida, Pseudomonas stutzeri, Neisseria spp., and Gemella haemolysans. To treat endophthalmitis, 15 antibiotics are typically used to clear bacterial infections (Data S5A), with the combination of gatifloxacin and moxifloxacin (GAT-MOX) used as the first-line treatment. According to literature evidence on the ocular microbiome,^80,81 species belonging to the Corynebacterium, Propionibacterium, and Staphylococcus genera are predominantly found on the eye surface of healthy individuals.

Based on the combined information above, we applied TACTIC to predict drug interaction outcomes for all possible pairwise combinations between 73 drugs against 12 bacterial strains (Data S5B). The drug selection included 12 of the 15 clinically used antibiotics for treating endophthalmitis that are present in the TACTIC model. In total, we generated more than 30,000 predictions that account for 2,628 unique drug pairs across 12 strains (Data S5C). GAT-MOX, which is commonly used as a broad-spectrum regimen to treat bacterial cases of endophthalmitis,⁸² is predicted to be mildly synergistic against all P. aeruginosa and Neisseria strains, but antagonistic for the other Pseudomonas species (Figure S6A). For the commensal strains (Corynebacterium kroppenstedtii and Propionibacterium propionicum), this combination is predicted to be additive (i.e., neutral). These predictions indicate that GAT-MOX may be an effective first-line therapy for cases of endophthalmitis primarily caused by P. aeruginosa or Neisseria spp., but not for other cases. Hence, GAT-MOX may not be the optimal first-line treatment due to its incomplete and weak coverage against all identified pathogenic strains.

Prior to scanning the TACTIC prediction landscape for alternative treatment options, we compared predictions against experimental data from studies that measured combined antibiotic efficacy against clinical and drug-resistant strains of P. aeruginosa^83,84 (Figure S6A). Of 9 drug combinations involving PMB, the TACTIC model correctly predicts that PMB yields stronger synergies when paired with AZT, chloramphenicol, fosfomycin (FOS), minocycline, and RIF compared to other drugs. TACTIC model predictions also indicate that ceftazidime (CFZ) paired with TOB (CFZ-TOB) is relatively more synergistic than CFZ combined with ciprofloxacin (CIP) (CFZ-CIP) or CIP-TOB, aligning with literature-based trends.

After confirming that TACTIC predictions align with existing evidence measured in P. aeruginosa, we identified 48 of 2,628 possible drug combinations having selective synergy against pathogenic bacteria (Figure 4A; Data S5D). Inspection of the 10 most synergistic drug interactions against pathogenic strains revealed combinations that involve gentamicin and CFZ, which are commonly used to treat endophthalmitis and are more effective against Gram-negative pathogens.⁸⁵ Interestingly, the remaining top synergistic set includes combinations involving at least one antibiotic active against the bacterial cell wall (mechanism of CFZ activity), but these combinations have not yet been investigated in treating bacterial endophthalmitis.

Figure 4.

Validation of TACTIC predictions for P. aeruginosa and S. aureus in endophthalmitis

(A) Of 2,628 drug combinations, 48 were predicted to have narrow-spectrum synergy against Gram-negative agents of endophthalmitis (predicted score < 0 against pathogenic strains, predicted score ≥ 0 for commensal strains). Of 48, 20 were predicted to be synergistic in S. aureus, the causative agent in Gram-positive infection cases. Blue, commensal strains; yellow, Neisseria spp.; red, P. aeruginosa strains; purple, other Pseudomonas spp.; pink, S. aureus ATCC 29213.

(B) Dose-response checkerboards showing percent cell viability in a clinical P. aeruginosa strain (PS) after treatment with six drug combinations that showed single-drug effect (≥ 50% change in cell viability) across both therapeutic agents.

(C) Dose-response checkerboards showing percent cell viability in a clinical S. aureus strain (SA) after treatment with five drug combinations that showed single-drug effect (≥ 50% change in cell viability) across both therapeutic agents.

See also Figure S6 and Data S3, S4, and S5.

To verify whether TACTIC predictions of selective synergy for bacterial endophthalmitis were meaningful, we measured up to 41 of the 48 drug combinations shown in Figure 4A against four P. aeruginosa strains isolated from endophthalmitis patients (Data S5E, see STAR Methods). Of the 41 combinations tested, 33 were confirmed to be synergistic based on manual Bliss score calculation (Data S4C). Of these, six showed single-drug inhibitory effect, i.e., ≥ 50% change in cell viability, for both therapeutic agents within the clinically relevant dosage range (0–32 μg/mL, Figure 4B). Bliss scores based on both manual calculation and SynergyFinder confirmed that all six combinations yield synergistic outcomes in drug-susceptible P. aeruginosa strains (Data S5F). Previous studies have shown that synergy between FOS and ofloxacin (OFL) is dependent on OFL susceptibility in P. aeruginosa.^86,87 Our experimental data agree with these outcomes, where FOS-OFL was found to be synergistic in PS (a drug-susceptible strain), but not in PS_MDR (a drug-resistant strain, Data S4C). Another study found that PMB enhances P. aeruginosa killing by FOS in a dose-dependent manner,⁸⁸ which is what we observed in both the PS and PS_MDR strains (Figures 4B; Data S4C). To the best of our knowledge, the combined effect of cefotaxime or cefsulodin with GAT or MOX has not been directly measured in P. aeruginosa.

For all other drug combinations measured across P. aeruginosa strains, we found that neither drug in the combination showed strong independent activity (n = 64) or only one drug showed independent activity (n = 22) within the clinically relevant dosage range (0–32 μg/mL) (Data S5G). Due to the nature of these checkerboards, interaction outcomes quantified using the manual calculation and SynergyFinder disagreed for the majority of cases (n = 52, Figures S6B and S6C; Data S4C). The manual Bliss approach does not require each drug to have antimicrobial activity and hence can generalize well to potentiators and adjuvants predicted by TACTIC. We have highlighted six combinations that show synergy across all methods of quantification in Data S5F. To validate TACTIC model predictions against all experimental outcomes, we elected to compare our prediction data against manually calculated Bliss scores considering that the accuracy of SynergyFinder quantification of drug interactions is sensitive to whether dose-response fitting assumptions are met within a given checkerboard assay.⁸⁹ We found that TACTIC predictions for all drug interactions measured in P. aeruginosa (n = 93) significantly agreed with experimental outcomes compared to random TACTIC predictions based on both the MSE (one-sample t test, p value < 1E−24) and percent overlap between model predictions and manual Bliss scores (one-sample t test, p value < 1E−75) (Figure S6D).

Although endophthalmitis is primarily caused by Gram-negative pathogens such as Pseudomonas and Neisseria spp., S. aureus can sometimes drive Gram-positive infection cases.⁹⁰ To further evaluate the clinical utility of the narrow-spectrum synergies we determined above, we experimentally measured outcomes for 45 of the 48 drug combinations in a clinical S. aureus isolate (Data S5E, see STAR Methods). TACTIC predicted 20 of the 48 drug interactions to be synergistic in S. aureus (Figure 4A), of which we experimentally confirmed 17 to be synergistic based on manual Bliss score calculation (Data S4C). Notably, five combinations were synergistic within the clinically relevant dosage range (0–32 μg/mL, Figure 4C), and to the best of our knowledge, all five have not previously been measured in S. aureus. We also found that TACTIC predictions for all 45 drug combinations significantly agreed with experimental outcomes compared to random TACTIC predictions based on both the MSE (Figure S6E, one-sample t test, p value < 1e−34) and percent overlap between model predictions and manual Bliss scores (one-sample t test, p value < 1e−23). However, we note that TACTIC predictions for S. aureus were not as accurate as those for P. aeruginosa (Figure S6D).

Discussion

Here, we demonstrate how data from model organisms can be leveraged for drug discovery in emerging pathogens. We introduce TACTIC, a new computational approach that leverages transfer learning and cross-species data training via orthology mapping to learn drug-response patterns from multiple model organisms and predict strain-specific drug interaction outcomes in unseen strains (Figure 1). Using a data collection of ∼3,000 drug interactions measured across 12 bacterial strains (Table 1, Data S1B), we show that TACTIC models generate more accurate predictions for Gram-negative and mycobacteria strains compared to INDIGO (Figure 2). Notably, these predictions are zero shot, requiring no additional training on the species of interest. This simulates a worst-case scenario in which no drug combination training data are available or are difficult to generate for a species of interest. Additional training on a species of interest (few shot) further improves the model’s performance, as shown in Figure S7. The conservation of genes predictive of drug interactions across species enables accurate transfer learning with relatively less training data than a traditional single-species model. In contrast, existing drug discovery models are primarily built using chemical structures, but these data cannot differentiate between pathogen strains and models need to be rebuilt separately for each organism. TACTIC is also effective because it can observe more drug combinations and their interaction outcomes in various bacterial pathogens possessing differing levels of antimicrobial susceptibility.

Based on a fully trained TACTIC model, we show that species-specific drug interaction outcomes are statistically associated with the phenotypic (e.g., metabolic state), genetic (e.g., Gram stain), and pathogenic nature of a given organism (Table 2; Figure S3; and Data S2). TACTIC feature importance analysis revealed several genes in amino acid metabolism to be predictive of drug interactions across species. Drug interactions in a wide range of organisms are influenced by their metabolic state, and while central metabolism is highly conserved, bacteria show distinct species-specific patterns of amino acid consumption.⁹¹ Our results suggest that in addition to the causative pathogen strain, the in vivo metabolic environment should be considered while selecting antibiotic combination therapies. However, we emphasize that our findings are correlative and require follow-up experimental validation to establish causality.

We then use the TACTIC model to generate a comprehensive prediction landscape of interaction outcomes for ∼3,600 drug pairs across 18 bacterial strains and 9 media conditions (Data S3A). By comparing TACTIC predictions for commensal strains against different groups of pathogenic bacteria, we identified small sets of drug interactions that are predicted to have narrow-spectrum synergy against specific pathogens (Figure S5). Importantly, we validated predictions for M. abscessus, an emerging pathogen with limited data, using an independent dataset of 96 drug interactions measured in this species⁷⁴ (Figure 3B).

Additionally, we experimentally verified that top predictions for selective synergy against NTM pathogens (AMP-CLA and AMP-CLA-MEC) are synergistic against clinical M. abscessus strains (Figures 3C and 3D). Considering that both AMP and CLA can be administered as inhalable formulations, their combination may prove a more effective treatment for respiratory infections than systemically administered antibiotics, which often have poor penetration into the lung.^92,93 To the best of our knowledge, both drug combinations have not been previously investigated as treatment options for respiratory infections caused by M. abscessus.

The synergy involving CLA is especially notable given that M. abscessus ATCC 19977 is known to harbor the erm(41) gene, which is associated with macrolide (e.g., CLA) resistance.⁹⁴ However, this gene is only induced after prolonged macrolide exposure⁹⁵ and may not counter CLA potency when combined with other drugs. Additionally, CLA remains one of the antibiotics recommended for treating NTM infections involving M. abscessus.⁹⁶

Variation in the combined drug effect for AMP-CLA and AMP-CLA-MEC between M. abscessus strains may be explained by differences in biophysical factors like morphology, which was also observed in the Van et al. study.⁷⁴ Based on both manual and SynergyFinder Bliss scores (Figures 3C; Data S4C) and the dose-response matrices (Figure 3D) for both combinations, TACTIC predictions appear to best align with outcomes in M. abscessus strains with smooth morphotype, which includes ATCC 19977. Rough morphotype cells lack glycopeptidolipids, resulting in a more hydrophobic cell envelope that decreases susceptibility to hydrophilic antibiotics like AMP,⁷⁶ and are prone to clumping, which increases drug response variability. Distinct metabolic properties associated with each morphotype⁹⁷ may also be driving differences in drug interaction outcomes. Interestingly, Van et al. found that outcomes were comparatively more antagonistic for the isolate with the rough morphotype and slowest growth rate, which agrees with what we observe in the dose-response matrices for our experimental data.

We further applied the TACTIC model to determine promising combinations with narrow-spectrum synergy for treating bacterial cases of endophthalmitis (Figure 4A). Notably, we confirmed that 6 of the 48 narrow-spectrum synergy predictions yielded synergistic effects in P. aeruginosa within the clinically relevant dosage range (0–32 μg/mL, Figures 4B; Data S5F). We confirmed that two interactions (FOS-OFL and FOS-PMB) yield synergistic effects against P. aeruginosa based on prior independent studies.^86,87,88 We also found that experimental outcomes measured in P. aeruginosa significantly agreed with TACTIC predictions compared to random predictions (Figure S6D), where the confidence in the measured outcomes for drug interactions shown in Figure 4B is higher due to strong single-drug activity for both agents within the clinically relevant dosage range. Additionally, we confirmed that 5 of the 48 narrow-spectrum synergy predictions were synergistic at clinically relevant dosing against S. aureus (Figure S6E), a common agent in Gram-positive endophthalmitis cases.⁹⁰

The TACTIC approach enables the investigation of species-specific drug interaction outcomes, thus representing a unique computational framework that explores the existence of combination therapies with broad- or narrow-spectrum synergy. Broad-spectrum synergistic combinations are promising candidates for empirical antibiotic therapy when the causative pathogen is unknown. Although broad-spectrum antibiotic treatments are commonly used today and have high translational potential, they may increase antibiotic resistance and potentially damage the native microbiota, leading to a range of adverse effects.^98,99 There is hence a need for narrow-spectrum treatments, which are more effective once the causative pathogen is known.

Limitations of the study

A key limitation of the current TACTIC model is that the prediction accuracy decreases for species that are phylogenetically distant (e.g., S. aureus) from organisms that are accounted for during model training (e.g., E. coli and M. tb) (Data S1F). This was readily seen in the benchmarking results as well as in the S. aureus validation results in the context of endophthalmitis. To address this limitation, we introduced a confidence score that quantifies the extent of transfer learning between species based on the ratio of predictive features that are conserved by orthologous genes (Figure 2C). This metric could be used to quantify the predictive power for strain-specific drug interaction outcomes for a novel strain of interest and to track model performance as new data are introduced into TACTIC. In the future, omics data measuring the drug response in phylogenetically distinct (e.g., Gram-positive) organisms could be integrated into TACTIC models to improve prediction accuracy and extend application to a diverse range of bacterial species.

Additionally, transfer learning via orthology mapping is currently limited to exchanging drug-response information cross-species at the binary level and cannot yet factor the effect of specific genetic mutations onto drug interaction outcomes. Such information would be valuable in differentiating predicted outcomes for drug-sensitive vs. drug-resistant strains of the same organism (e.g., P. aeruginosa). This limitation could be addressed by incorporating strain-specific data from highly curated databases like the Pathosystems Resource Integration Center (PATRIC).¹⁰⁰ Orthology mapping is also currently restricted to bacterial strains with known orthology information, such as from OrtholugeDB,¹⁰¹ a curated database of gene orthology that was queried for all strains considered in this study. We note that the fully trained TACTIC model presented here is best suited for predicting drug interaction outcomes for Gram-negative and mycobacteria strains with orthology information; however, TACTIC can be extended to consider orthology information from other databases such as OrthoDB¹⁰² and PhylomeDB¹⁰³ as long as gene identifiers for organisms can be mapped against the gene annotation used in the drug response data.

We further note that within this study, we applied TACTIC to specifically identify synergistic interactions, one of the three selection inversion strategies proposed by Baym et al.¹⁰⁴ As a regression-based model, TACTIC can identify “suppressive” and “collateral” interactions, as described by Baym et al., by including concentration- and time-course information during model training as previously shown in a prior study.⁴¹ Beyond traditional ML methods like random forest, deep learning (DL) approaches are being explored for drug synergy prediction. While currently limited to large-scale cancer datasets,¹⁰⁵ new efforts are extending DL to open new avenues for antimicrobial discovery.^106,107,108

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Sriram Chandrasekaran (csriram@umich.edu).

Materials availability

This study did not generate new, unique reagents.

Data and code availability

• •All data curated or generated in this study are available in the supplemental information and listed under deposited data in the key resources table.
• •All original code is deposited in GitHub: https://github.com/sriram-lab/TACTIC and publicly available in Zenodo: https://doi.org/10.5281/zenodo.20385233.
• •Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request.

Acknowledgments

This work was supported by faculty start-up funds from the University of Michigan (UM); grants R01AI150826, R56AI150826, and R21AI144536 from the National Institute of Allergy and Infectious Diseases; grant R35GM137795 from the National Institute of General Medical Sciences; grants from UM Precision Health and UM Office of the Vice Provost of Research; UM Endowment for Basic Sciences Accelerator Award; and UM Research Scouts Award to S.C. This work was also supported by funds from the NIH/Fogarty International Center D43TW012027 to the University of Michigan.

Author contributions

Conceptualization, S.C.; methodology, C.H.C., D.C.C., K.S., and S.C.; investigation, C.H.C., D.C.C., N.M.R., M.R.S., M.A.O., P.L., G.R., R.K., and M.V.; visualization, C.H.C.; writing – original draft, C.H.C., D.C.C., and S.C.; writing – review and editing, C.H.C. and S.C.; funding acquisition, S.C., A.D.B., and K.S.; resources, K.S., A.D.B., and S.C.; and supervision, P.L., A.D.B., and S.C.

Declaration of interests

S.C. is an inventor on a pending patent application by UM related to drug combination discovery with machine learning.

Declaration of generative AI and AI-assisted technologies in the writing process

No generative AI or AI-assisted technologies were used during the preparation of this work.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Bacterial and virus strains
M. abscessus strain L948	ATCC, USA	19977
M. abscessus clinical isolate, rough morphotype	U-M Health System	Internal ID 0746A/B
M. abscessus clinical isolate, rough morphotype	U-M Health System	Internal ID 0746A/B
P. aeruginosa strain Boston 41501	ATCC, USA	27853
P. aeruginosa clinical isolate, PS	AEH Tamil Nadu, India	Internal ID 2281/2024
P. aeruginosa clinical isolate, PS_MDR	AEH Tamil Nadu, India	Internal ID 15308/2024
S. aureus clinical isolate, SA	AEH Tamil Nadu, India	Internal ID M6196/2020
Chemicals, peptides, and recombinant proteins
Ofloxacin	Aarti Drugs Ltd, India	Cat# 82419-36-1
Theophylline	Aarti Pharmalabs Ltd, India	N/A
Mueller Hinton II Broth Cation Adjusted	BD BBLTM, USA	Cat# B12322
Middlebrook 7H9 Broth	BD DifcoTM, USA	Cat# 271310
Middlebrook 7H10 Agar	BD DifcoTM, USA	Cat# DF027-17-4
Nutrient Agar	BD DifcoTM, USA	Cat# DF0001-17-0
Ampicillin (sodium salt)	Cayman Chemical Company	Cat# 14417
Clarithromycin	Cayman Chemical Company	Cat# 19455
Mecillinam	Cayman Chemical Company	Cat# 9002008
Azithromycin	Century Pharmaceuticals, India	N/A
Chloramphenicol	Century Pharmaceuticals, India	Cat# 56-75-7
Gentamicin	Fuan Pharmaceuticals, China	Cat# 1405-41-0
Amikacin	HiMedia Laboratories, India	Cat# CMS644
Ampicillin	HiMedia Laboratories, India	Cat# PCT1101
Bacitracin	HiMedia Laboratories, India	Cat# CMS2648
Cefotaxime	HiMedia Laboratories, India	Cat# MB134
Ceftazidime	HiMedia Laboratories, India	Cat# CMS1194
Ciprofloxacin	HiMedia Laboratories, India	Cat# TC447
Doxorubicin	HiMedia Laboratories, India	Cat# TC420
Doxycycline	HiMedia Laboratories, India	Cat# PCT1138
Epigallocatechin Gallate	HiMedia Laboratories, India	Cat# RM10179
Kanamycin	HiMedia Laboratories, India	Cat# TC136
Mitomycin C	HiMedia Laboratories, India	Cat# TC406
Mueller Hinton Broth (MHB)	HiMedia Laboratories, India	Cat# M391
Mueller Hinton Broth No. 2 Control Cations	HiMedia Laboratories, India	Cat# M1657
Oxacillin	HiMedia Laboratories, India	Cat# CMS5372
Phenazine Methosulfate	HiMedia Laboratories, India	Cat# TC484
Puromycin	HiMedia Laboratories, India	Cat# CMS8861
Sodium Chloride (NaCl)	HiMedia Laboratories, India	Cat# MB023
Trimethoprim	HiMedia Laboratories, India	Cat# CMS216
Dimethyl Sulfoxide EMPLURA® (DMSO)	Merck Millipore, India	Cat# 1.16743.0521
Moxifloxacin	MSN Laboratories, India	Cat# 186826-86-8
Benzalkonium	Novo Nordisk Pharmatech, Denmark	Cat# 63449-41-2
Gatifloxacin	Panchsheel Organics Ltd, India	Cat# 112811-59-3
A22 Hydrochloride	Sigma-Aldrich, India	Cat# SML0471
Amoxicillin	Sigma-Aldrich, India	Cat# A8523
Aztreonam	Sigma-Aldrich, India	Cat# PZ0038
Carbonyl cyanide m-chlorophenylhydrazone	Sigma-Aldrich, India	Cat# C2759
Cefoxitin	Sigma-Aldrich, India	Cat# C4786
Cefsulodin	Sigma-Aldrich, India	Cat# C8145
Cefuroxime	Sigma-Aldrich, India	Cat# C0695000
Cerulenin	Sigma-Aldrich, India	Cat# 219557
Cycloserine D	Sigma-Aldrich, India	Cat# C6880
Erythromycin	Sigma-Aldrich, India	Cat# E6376
Fosfomycin	Sigma-Aldrich, India	Cat# F0400000
Fusidic acid	Sigma-Aldrich, India	Cat# Y0001411
Nitrofurantoin	Sigma-Aldrich, India	Cat# N7878
Paraquat	Sigma-Aldrich, India	Cat# 36541
Phleomycin	Sigma-Aldrich, India	Cat# P9564
Polymyxin B	Sigma-Aldrich, India	Cat# PHR1595
Spiramycin	Sigma-Aldrich, India	Cat# 46745
Thioridazine	Sigma-Aldrich, India	Cat# T9025
Amoxicillin Trihydrate	Tokyo Chemical Industry, Japan	Cat# A20995G
Cefuroxime Sodium Salt, 95%	Thermo Scientific, USA	Cat# J66999.03
Reazurin Sodium Salt	Thermo Scientific, USA	Cat# B21187.03
Deposited data
Cross-species drug interaction data	This study	Data S1B
Experimental validation data	This study	Data S4B
Software and algorithms
INDIGOpy v1.0.0	Chandrasekaran et al.14	https://github.com/sriram-lab/INDIGOpy
Python v3.10.8	Python Software Foundation	https://www.python.org/
Scikit-learn v1.2.0	Pedregosa et al.109	https://scikit-learn.org/
TACTIC	This study	GitHub: https://github.com/sriram-lab/TACTIC; Zenodo: https://doi.org/10.5281/zenodo.20385233
Other
Filter Pipette Tips 200 μL	BiotixTM, USA	Cat# 12-111-126
96-well Clear Flat Bottom, with Lid, Sterile	Costar, USA	Cat# 07-200-656
Multi-channel Pipette	Eppendorf, India	Cat# 3125000052
Single-channel pipettes 10 μL	Eppendorf, India	Cat# 3123000020
Single-channel pipettes 200 μL	Eppendorf, India	Cat# 3123000055
Single-channel pipettes 1000 μL	Eppendorf, India	Cat# 3123000063
Autoclave	Equitron Medica Private Limited, India	Cat# 7441SLEFA
Petri Dish, Stackable Lid (100 mm × 15 mm)	Fisher Scientific, USA	Cat# FB0875712
Cuvette, Semi-micro, 1.5ML	Fisher Scientific, USA	Cat# 14955127
Multi-channel Pipette	Gilson, France	Cat# TA72306
SpectraMax® M3 Multi-mode Microplate Reader	Molecular Devices, China	Cat# MDEVM3
Incubator	Sub-Zero Lab Instruments, India	Cat# SZ-BOD
96 Well Micro Test Plate, Polystyrene, Type-U bottom	Tarsons, India	Cat# 941296
Cryochill TM vial 1.8 mL Capacity Self-standing, Sterile	Tarsons, India	Cat# 523182
Magnetic Stirrer	Tarsons, India	Cat# 6030
Microtips 10 μL	Tarsons, India	Cat# 521000
Microtips 200 μL	Tarsons, India	Cat# 521010
Microtips 1000 μL	Tarsons, India	Cat# 521020
Spinex Vortex Shaker	Tarsons, India	Cat# 3020
Spinwin Conical Bottom Tube Sterile 15 mL	Tarsons, India	Cat# 546021
Spinwin Conical Bottom Tube Sterile 50 mL	Tarsons, India	Cat# 546041
V-bottom Pipetting Reservoir, Sterile	Tarsons, India	Cat# 524092

Experimental model and study participant details

Laboratory strains for M. abscessus (ATCC 19977) and P. aeruginosa (ATCC 27853) were obtained from the American Type Culture Collection (ATCC). Clinical isolates of M. abscessus (Smooth and Rough morphotypes) and P. aeruginosa (PS4) were obtained from de-identified patient sample repositories in the University of Michigan Health System. Regulatory agency approval was not required for the use of these samples. Of note, the M. abscessus clinical isolates exhibited a baseline ampicillin MIC of >264 μg/mL and clarithromycin MIC of >8 μg/mL, as reported by the University of Michigan Health System. The remaining P. aeruginosa strains (PS and PS_MDR) and S. aureus strain (SA) were isolated from clinical patients at the Aravind Eye Hospital in Tamil Nadu, India. The Aravind Eye care system’s Internal Ethics committee and the IRB approved these studies (IRB number: RES2023084BAS). No human studies were conducted; all sample collections were taken with informed consent from patients.

Method details

Experimental design

TACTIC interaction predictions for drug combinations were confirmed using a microbroth dilution checkerboard assay,¹¹⁰ which was performed according to CLSI¹¹¹ and EUCAST¹¹² standards. Experiments were conducted at two independent institutions in the following manner.

• (1)The University of Michigan Health System: combinations involving ampicillin sodium (Cayman Chemical Company, USA), clarithromycin (Cayman Chemical Company, USA), and mecillinam (Cayman Chemical Company, USA) were tested against M. abscessus strains, while the combination of amoxicillin trihydrate (Tokyo Chemical Industry, UK) and cefuroxime sodium (Thermo Scientific, USA) was screened against P. aeruginosa PS4. Resazurin dye was used as an indicator of bacterial growth, where metabolism of the dye produces a color change from blue (absorbance maximum of 570 nm) to pink (absorbance maximum of 600 nm). M. abscessus was thawed from frozen stock, streaked onto Middlebrook 7H10 agar plate, and incubated for 72 h at 37°C. Antibiotics dissolved in Middlebrook 7H9 broth (in the case of M. abscessus) or Cation Adjusted Mueller Hinton II Broth (in the case of P. aeruginosa) were incubated in combination with bacteria inoculum (6 × 10⁴ CFU; determined using a standard curve) in a 96-well plate for 24 h at 37°C and 150 RPM, at which point the resazurin dye was added. The absorbance of each well was read at 570 nm and 600 nm, 24 h following the addition of the dye. Each microplate additionally included a sterile control (broth and resazurin only) and a growth control (inoculum, broth and resazurin only). The percentage inhibition for each antibiotic combination was calculated by comparing the absorbance of the treated wells to that of the control wells. At least three separate replicates were averaged to obtain the final data.
• (2)The Aravind Eye Hospital in Tamil Nadu: Mueller–Hinton Broth (MHB, 10 mL) and 2× Cation-Adjusted Mueller–Hinton Broth (CAMHB, 50–100 mL) were prepared and sterilized by autoclaving at 121°C for 15 min. Sterile saline (0.9% NaCl) was similarly prepared. Stock solutions of Drug A and Drug B were prepared according to CLSI/EUCAST guidelines or stability data. Single colonies of the three P. aeruginosa strains (PS_ATCC, PS, PS_MDR) and one S. aureus strain (SA) were inoculated into 10 mL MHB and incubated at 37°C for 18 ± 2 h at 200 rpm to reach ∼10⁹ CFU/mL. A 96-well microplate was filled with 100 μL of 2× CAMHB per well using a multichannel pipette. For Drug A dilution, 1.5 mL and 500 μL of Drug A stock (StA and 2×StA) were prepared in 2× CAMHB to achieve 4× and 8× the highest test concentration, respectively. 100 μL of StA was added to wells A1–A11; 2×StA to well A12. A 2-fold serial dilution was performed vertically from row A to G. For Drug B dilution, 1.0 mL of Drug B stock (StB) was prepared in 2× CAMHB to achieve 4× the highest test concentration. 100 μL of StB was added to wells in column 12. A 2-fold serial dilution was performed horizontally from column 12 to column 2. The final concentration of drug A was 32 μg/mL to 0.5 μg/mL and drug B was 32 μg/mL to 0.0312 μg/mL. The overnight culture was diluted in 0.9% NaCl to a final concentration of 10⁶ CFU/mL. After vortexing for 20 s, 100 μL of inoculum was added to each well, resulting in a final inoculum of 5 × 10⁵ CFU/mL per well. A no-growth control plate (without inoculum) was prepared, and well H1 served as the positive growth control. Plates were incubated statically at 35 ± 2°C for 18 ± 2 h, with optional sealing or humidification to reduce evaporation. After incubation, OD was measured at 600 nm. Percent growth was calculated relative to H1, and drug interaction was assessed across 10 combinations for PS_ATCC, 41 combinations for PS and PS_MDR, and 45 combinations for SA (Data S5E).

Experimental data for all drug interactions tested in M. abscessus, P. aeruginosa, and S. aureus strains are provided as a supplementary file (Data S4B). This data was quantified based on the Bliss Independence model³³ using both a manual calculation and SynergyFinder.⁷⁷ The Bliss framework quantifies drug synergy based on the combined drug effect on cell growth (i.e., percent cell viability) by comparing the observed growth value against the expected value, which is computed for a drug pair as follows:

$$G_{AB}=G_A+G_B-G_A{G}_B\text{,}$$

where G_AB is the expected growth achieved by the combination based on the growth observed for each drug independently (G_A and G_B). Manual Bliss scores reported in Figures 3C, Figure S6, Data S4C, and Data S5F were defined as the averaged difference between the observed and expected G_AB across the dose-response matrix, with negative scores indicating synergy.

Data acquisition and curation

Two distinct data types were used to construct TACTIC models: drug response and drug interaction data. Drug response information was comprised of E. coli chemogenomic data,¹¹³ where fitness of single-gene knockout strains was measured in response to individual stressors, and M. tb transcriptomic data,¹⁵ which measured differential gene expression in response to individual stressors. Each omic dataset was previously Z score normalized within its originating study; hence, the normalized omic data was directly used for defining TACTIC model features. The drug interaction data comprised of 2965 combinations collected from 17 literature sources covering 12 bacterial strains (Table 1, Data S1B). Drug combinations ranged between two-to ten-way interactions and were quantified based on the Loewe Additivity³² and/or the Bliss Independence³³ model based on the originating study. A total of 114 unique drug interactions were measured based on both mathematical models, and the Pearson correlation between Bliss- and Loewe-based scores indicated a strong agreement between metrics (R = 0.582, p-value < 1E−11, Data S6A). This suggested that Bliss and Loewe drug interaction scores are concordant and could be trained together. Of note, the drug interaction data collection used in this study represents the subset for which the full set of ML features could be determined based on the omics datasets from which drug response information was extracted. For exploratory analysis of the data (Figure S1), each drug interaction was labeled into one of the three outcome classes (i.e., synergy, neutral, antagonism) based on the classification criteria specified in its originating publication (Data S1A). Continuous interaction score values were used for regression-based ML model construction.

Defining ML features for drug combinations

Similar to INDIGO,¹⁴ the TACTIC model also defines drug combination features based on single-drug response omics data. Drug response chemogenomics or transcriptomics data is first binarized to indicate substantial gene-level changes (i.e., z > 2 and z < −2) in cell response to a given stressor. The binarized vectors for all drugs involved in a given combination are then combined to define sigma and delta features for each gene, which mathematically represent the combined drug and drug-unique effect, respectively. In this study, we used data measured in E. coli MG1655 and M. tb H37Rv. Specifically, the E. coli data measured the drug effect on gene knockout fitness across ∼4000 genes (chemogenomics), while the M. tb data measured gene expression changes in response to drug treatment across ∼4000 genes (transcriptomics). Of note, these two omics datasets had been previously used to construct two separate INDIGO models, one for E. coli¹⁴ and another for M. tb.¹⁵ To remove duplicated information captured by orthologous genes between the two organisms, a feature vector was defined by the sigma-delta transformation of all genes annotated for M. tb H37Rv and the sigma-delta transformation for genes unique to E. coli MG1655.

Orthology mapping for transfer learning

The E. coli and M. tb omics datasets were directly used to define ML features for drug combinations measured against E. coli str. K-12 substr. MG1655 (henceforth E. coli MG1655) and M. tb H37Rv. These datasets contained 3949 genes for E. coli MG1655 and 3968 genes for M. tb H37Rv. For all other organisms, ML features were defined by first determining the sigma-delta feature vector for a given drug combination then filtering this vector based on the orthology mapping between E. coli MG1655, M. tb H37Rv, and the given strain (Figure 1B). Gene orthology was determined by mapping the E. coli MG1655 and M. tb H37Rv genomes against the genomes for all other strains according to the genome annotation available in OrtholugeDB,¹⁰¹ a curated database that infers orthologs based on the phylogenetic distance or the reciprocal-best-BLAST hit (RBBH) rate between gene pairs.

For all strain-specific predictions, the strain selection was based on the genomes that were available in OrtholugeDB for both pathogenic and commensal bacteria. For the endophthalmitis predictions, the 12 strains represent the following eight species: P. aeruginosa, P. mendocina, P. putida, P. stutzeri, N. gonorrhoeae, N. meningitidis, C. kroppenstedtii, and P. propionicum (Data S5B). Of note, no genomes were available for species belonging to the Gemella genus, and the Staphylococcus genus was not included in the commensal set due to evidence of Staphylococci commonly being the causative agent for Gram-positive endophthalmitis cases.⁹⁰ Data S6B shows the total number of orthologs mapped against the E. coli MG1655 and M. tb H37Rv genomes for all strains considered in this study.

TACTIC model construction using Random Forests

All TACTIC models were constructed in Python v3.10.8 (Python Software Foundation) using the regression-based Random Forests (RF) algorithm supplied through the scikit-learn package v1.2.0¹⁰⁹. Of note, the RF algorithm was confirmed to yield the best performing TACTIC models compared to other regression-based ML algorithms, namely: linear regression, ridge regression, support vector regression, decision tree regression, and gradient boosting regression (Data S6C). Briefly, RF is an ensemble method comprised of multiple decision trees that independently learn to associate feature information (i.e., sigma-delta features for a given drug combination) to a target variable (i.e., the measured drug IS) during the training phase. Each decision tree is then tasked with estimating the target variable (i.e., drug interaction outcome) given feature information alone during the testing phase, and the average estimated value (i.e., the predicted drug IS) is returned by the RF algorithm. Randomization is factored into the model construction process via random sampling with replacement of the training dataset for each decision tree. This yields a diverse collection of decision trees that together reduce the variance between model predictions against true values for the target variable. Of note, all TACTIC models were constructed using the default parameter values for the RandomForestRegressor method in scikit-learn.

For TACTIC models reported in Figures 2A and 2B, predictions for strain-specific models were determined for the total number of drug interactions where data was collected for a given strain (Table 1). Predictions for random TACTIC models were determined based on random sampling of drug interaction scores in the training set for each strain-specific TACTIC model.

TACTIC model interpretation via feature importance

Feature importance for the full TACTIC model was determined by using the built-in computation for the scikit-learn RF algorithm, which defines importance as the mean decrease in variance for a regression task. Based on this definition, the importance score for each feature is calculated by first constructing a model without the given feature then comparing its prediction accuracy against that of the model trained on all features. Specifically, importance is quantified as the difference in the mean squared error between model predictions and true values for the target variable. TACTIC model interpretation was completed by using the set of features ranked by decreasing importance score with a cumulative sum of 0.95, or the set of the most important features that explain 95% of the mean decrease in variance. In total, ∼2500 features associated with 1258 E. coli MG1655 genes and 411 M. tb H37Rv genes (both sets including 251 orthologous genes between the two organisms) met this criterium (Data S2A).

Of note, confidence scores reported for strain-specific TACTIC models in Figure 3C were defined as the ratio of the ∼2500 features that are conserved by orthologous genes between E. coli MG1655, M. tb H37Rv, and the strain of interest. For example, the confidence score for the S. aureus TACTIC model was 0.248 given that this model can only rely on the predictive power of 620 features (out of ∼2500) that are associated with S. aureus genes that share orthologs with E. coli MG1655 and M. tb H37Rv.

Based on the KEGG pathways that were significantly enriched by the E. coli MG1655 and M. tb H37Rv genes identified above (Figure S3A), we evaluated how the values for TACTIC model features that are associated with genes belonging to these pathways varied between synergistic and antagonistic interactions. Specifically, we conducted a two-sample Student’s t test with equal variance and the chi-square test with Yate’s correction to compare sigma and delta feature value differences, respectively. Using a stringent significance level (adjusted p-value < 1E−6) based on the adjusted p-value distribution for all feature-specific tests, we determined 79 sigma or delta features that differed based on the drug interaction outcome type (Data S2B). From this set, we identified 11 genes for which both sigma and delta feature values significantly different between synergistic and antagonistic interactions along the same direction for gene-specific changes (Data S2C).

Finding associations between ML feature patterns and drug interaction outcomes

We assessed interaction scores for 32 drug combinations measured in all strains that were considered within the same study.²⁵ We defined strain-specific drug impact based on the equation below:

$${DI}_{interaction}=\frac{\sqrt{\sum _{i=1}^{N_{features}}{v}_{interaction}}}{N_{orthologous}}\text{,}$$

where DI_interaction is the drug impact for a given drug interaction, v_interaction is the vector of ML features determined for the given drug interaction, N_features is the total number of ML features, and N_orthologous is the number of E. coli MG1655 and M. tb H37Rv genes that have corresponding orthologs annotated within the genome of the strain for which the drug interaction pertains to. We then quantified the correlation between strain-specific drug impact and interaction scores for all 32 drug combinations, 3 of which were found to have a significant correlation (Data S2D).

PCA of strain- and media-specific prediction landscape

PCA was applied onto the strain- and media-specific drug interaction outcome prediction data (3655 interactions x 162 strain-media pairs). PCA results were then visualized onto the 2D space defined by the first two principal components (Figure S4B). The cumulative pairwise distance between strains within the PCA-defined space was calculated for each of the 9 media conditions to determine which growth environment yielded the largest strain-to-strain differences in outcome predictions (Data S3B).

Identifying narrow-spectrum pathogen-specific drug interactions

For these comparisons between pathogen strains and commensal strains, we imposed two criteria to define narrow-spectrum synergy. The first was to select drug combinations that are predicted to be additive or antagonistic (predicted interaction score (IS) ≥ 0) across all commensal strains. The second criterium selected pathogen-specific synergies based on one of four thresholds: strong comprehensive synergy (predicted IS < −0.2 for all strains), comprehensive synergy (predicted IS < 0 for all strains), strong average synergy (mean predicted IS < −0.2 for all strains), and average synergy (mean predicted IS < 0 for all strains). For the first comparison involving all pathogens, only the average synergy threshold returned narrow-spectrum synergies.

Mean squared error and percent overlap calculation for P. aeruginosa validation

The mean squared error (MSE) and percent overlap between experimentally measured outcomes and TACTIC model predictions were determined for 93 drug interactions measured across four P. aeruginosa strains and 45 drug interactions for one S. aureus strain (Figure S6D, Data S4C and S5E). For both metrics, manually calculated Bliss scores were compared against actual and random TACTIC predictions determined for P. aeruginosa strains UCBPP-PA14 (henceforth PA14) and PA7 and S. aureus strain ATCC 29213 (Data S5B). Specifically, predictions made for strain PA14 (one of the reference strains for P. aeruginosa)¹¹⁴ were matched with experimental outcomes for drug-susceptible strains (PS4, PS_ATCC, PS), predictions made for strain PA7 (known to be drug-resistant)¹¹⁵ were matched with experimental outcomes for the PS_MDR strain, and predictions made for strain ATCC 29213 were matched with experimental outcomes for the SA strain. Random TACTIC predictions were determined via random sampling from all predictions determined for PA14 and PA7 (N = 2,628 per strain). The MSE was defined as the average sum of squared errors between the manual Bliss score and TACTIC prediction across all 93 drug interactions. Percent overlap was defined as the accuracy in predicting synergistic outcomes (Bliss score <0) as synergistic (prediction <0) for all 93 interactions.

Quantification and statistical analysis

Three statistical tests were used for downstream model interpretation based on the set of sigma-delta features explaining 95% of model importance. First, the two-sample Student’s t test with equal variance was used to determine sigma features that statistically differ in mean value between synergistic or antagonistic drug interaction outcomes (Data S2B). For delta features, the chi-square test with Yates’ correction was applied to determine any associations between synergistic or antagonistic drug interaction outcomes and drug-unique effects on a single gene (Data S2B). Finally, a hypergeometric test was performed to determine KEGG pathways that were significantly represented by genes associated with the top model features (Figure S3A, those explaining 95% of the variance). This test was conducted for KEGG pathways annotated from both the E. coli MG1655 genome (organism code: eco) and the M. tb H37Rv genome (organism code: mtu). Of note, the number of pathways annotated for each genome served as the population size specified in the hypergeometric test. Importantly, p-values from all three statistical tests were adjusted using the Benjamini-Hochberg correction¹¹⁶ to minimize the false discovery rate. Additionally, a one-sided Student’s t test was used to report the findings in Figure 3C, a one-sample Student’s t test was used to report the findings in Figures S6D and S6E, and a two-sided Student’s t test was used to report the findings in Figure S3C.

Published: July 7, 2026

Supplemental information

Data S1. Cross-species drug interaction data, related to Table 1 and Figure 2

(A) Source metadata.

(B) TACTIC training drug interaction outcomes.

(C) Drug acronyms and descriptions.

(D) Unique drug interactions.

(E) Gram-positive drug interaction outcomes.

(F) Benchmarking results for Gram-positive strains.

Data S2. TACTIC model interpretation, related to Table 2

(A) Importance scores for all TACTIC features.

(B) Significance test results TACTIC features associated with genes in enriched pathways.

(C) Top TACTIC features associated with metabolic genes.

(D) Association between strain-specific drug interaction scores and drug impact.

Data S3. Cross-species and cross-media TACTIC predictions, related to Table 3 and Figure 3

(A) Prediction landscape across 18 strains and nine media conditions.

(B) Cumulative strain-to-strain distance across media conditions.

(C) Predictions for clarithromycin combined with beta-lactam drugs.

Data S4. Experimentally measured validation data, related to Figures 3 and 4

(A) Drug interaction scores from Van et al.⁷⁴ and corresponding TACTIC predictions.

(B) Experimentally measured response (% cell viability) for M. abscessus, P. aeruginosa, and S. aureus.

(C) Manual and SynergyFinder Bliss scores for all drug interactions measured in house.

Data S5. Endophthalmitis application data, related to Figure 4

(A) Antibiotics used to treat endophthalmitis.

(B) Bacterial strains relevant to ocular infections.

(C) Prediction landscape across 13 strains.

(D) Narrow-spectrum drug interaction predictions.

(E) Drug interactions measured in house.

(F) Bliss scores for validated synergies in P. aeruginosa.

(G) Effect outcomes for all checkerboards measured for P. aeruginosa.

Data S6. TACTIC methodology data, related to STAR Methods

(A) Agreement between Bliss and Loewe scores.

(B) Orthology coverage across all strains.

(C) Comparison between regression-based ML algorithms.