Exploratory AI-Assisted In Silico Evidence That Bacterial Signaling Molecules May Occupy a Drug-Like Pharmacokinetic Space: A Preliminary Comparative Cheminformatics Analysis and Candidate Identification

1. Introduction

1.1. The Antibiotic Discovery Crisis

Antimicrobial resistance (AMR) represents a growing global health threat while discovery pipelines remain optimized for a single endpoint: minimum inhibitory concentration (MIC) in standardized growth assays (Brown & Wright, 2016). This framing may systematically discard molecules whose primary bioactivity manifests at sub-inhibitory concentrations (Andersson & Hughes, 2014; Fajardo & Martínez, 2008).

1.2. The Signaling-First Hypothesis

A companion opinion article (Abegg, 2025, in preparation) proposes that molecules capable of modulating bacterial cell signaling at sub-inhibitory concentrations represent a systematically underexplored source for antibiotic discovery. This “signaling-first” triage framework rests on the empirical observation that many antibiotics act as intercellular signals at low concentrations (Linares et al., 2006; Davies & Davies, 2010; Romero et al., 2011), and that concentration-dependent transitions from signaling to growth inhibition have been documented for multiple compound classes (Beyersmann et al., 2017; Hoffman et al., 2005). Recent studies have strengthened this basis: Williams et al. (2024) showed that sub-inhibitory antibiotics elicit structured signaling responses in Pseudomonas aeruginosa, while Amer et al. (2025) demonstrated growth-phase-dependent modulation of quorum sensing (QS) by sub-MIC antibiotics.

1.3. An Untested Prediction

An implicit prediction of this hypothesis is that molecules with demonstrated bacterial signaling activity should exhibit favorable drug-likeness properties. The evolutionary logic is that the physicochemical constraints enabling effective intercellular signaling — small molecular size for passive diffusion, moderate lipophilicity for membrane transit — overlap with the properties that predict drug-likeness (Lipinski et al., 2001; Veber et al., 2002). Critically, this prediction applies specifically to small-molecule diffusible signals. Gram-positive bacteria frequently employ peptide-based quorum sensing systems, and intracellular second messengers operate under entirely different physicochemical logic. Including these non-drug-like signaling molecules in the analysis provides an essential internal control.

1.4. Objectives

This exploratory technical note addresses: Part 1: Do small-molecule bacterial signaling compounds exhibit systematically different drug-likeness profiles compared to both peptide/intracellular signals and randomly selected drug-like molecules? Part 2: Can specific signaling-active molecules with favorable profiles be identified that lack systematic antibiotic testing at elevated concentrations?

2. Methods

2.1. AI-assisted Workflow

This analysis was conducted with AI assistance using Claude (Anthropic; Claude Opus 4.6 Extended for all analysis, dataset construction, statistical verification, and manuscript revision, December 2024 – February 2025). The specific workflow was:

Step 1 — Dataset construction. The author provided Claude with the conceptual framework, inclusion/exclusion criteria, and iterative corrections over multiple sessions. Claude was prompted to identify bacterial signaling molecules with documented, reproducible signaling activity from the quorum sensing and bacterial communication literature, including: (a) all major classes of Gram-negative autoinducers (AHLs with varying acyl chain lengths and substitutions, quinolone signals, DSF family, AI-2); (b) Gram-positive peptide autoinducers (AIPs, CSPs, competence pheromones, lantibiotics); (c) intracellular second messengers (c-di-GMP, c-di-AMP, ppGpp, pppGpp, cAMP); (d) Streptomyces butyrolactone signals (A-factor, SCB1, VB-A); (e) interspecies signals (indole derivatives, CAI-1, photopyrones); (f) QS inhibitors and antivirulence compounds with documented signaling modulation; and (g) antibiotics with documented sub-MIC signaling effects. A comparison set of 71 drug-like molecules spanning diverse therapeutic areas was independently constructed. Selection was based on Claude’s training knowledge of published literature, not a systematic database query with reproducible search terms — this is a key methodological limitation.

Step 2 — Property compilation. For each molecule, Claude retrieved MW, LogP, HBA, HBD, TPSA, and rotatable bonds from PubChem compound records (preferred source) or DrugBank. LogP values correspond to XLogP3 as reported in PubChem unless otherwise noted. For Gram-positive peptide autoinducers (MW > 500 Da), properties were estimated from peptide sequence-based calculations and published structural data, as PubChem records are incomplete or absent for many of these molecules. Properties were compiled from database records, not computed de novo from canonical SMILES using a single cheminformatics toolkit (e.g., RDKit, CDK, OpenBabel). This means that (a) values for the same property may derive from different computational algorithms across molecules, (b) results cannot be precisely reproduced by re-running a computational pipeline, and (c) inter-source inconsistencies may exist. This is the most critical methodological weakness of this study. 

Step 3 — Filter application and statistical analysis. Claude applied seven drug-likeness filters to all datasets using the published threshold criteria for each filter (Section 2.3). Statistical tests were performed using Python (scipy.stats v1.11): Mann-Whitney U tests with rank-biserial correlation for continuous property comparisons, Fisher’s exact test for categorical pass/fail comparisons, and Enrichment Factors (EF) calculated as EF = (% passing in test set) ÷ (% passing in reference set). Bonferroni correction was applied separately for property comparisons (6 tests, adjusted α = 0.0083) and filter comparisons (7 tests, adjusted α = 0.0071). The 95% confidence interval (CI) for the Lipinski Odds Ratio (OR) was computed using the Woolf logit method. All scripts were generated by Claude and executed in the same session.

Step 4 — ADMET prediction and PAINS screening. Simplified ADMET predictions were derived from threshold-based classifications (Section 2.4). All 26 candidate molecules were screened for PAINS alerts via manual structural assessment guided by AI (Section 2.5).

Step 5 — Testing gap assessment. For each candidate molecule, Claude assessed whether published MIC or dose-response data at growth-inhibitory concentrations could be identified. This was based on Claude’s training knowledge of published literature (training data cutoff: approximately April 2024), not a formal systematic review with defined search protocols. Consequently, unpublished negative results (“file drawer problem”) and literature published after the training cutoff are not captured.

Step 6 — Figure generation and manuscript drafting. Python scripts (matplotlib v3.8, scipy v1.11) were generated by Claude and executed to produce all figures. The author critically reviewed all outputs, challenged assumptions, identified errors, and directed iterative refinement across multiple revision cycles. All Python scripts and raw data are provided as supplementary materials.

2.2. Dataset Construction

Signaling molecule dataset (n = 74 total; three subsets).

Small-molecule diffusible signals (n = 48, “Primary” subset): AHL autoinducers including unsubstituted (C4- through C14-HSL), 3-oxo-substituted (3-oxo-C6- through 3-oxo-C12-HSL), and 3-hydroxy-substituted (3-OH-C4-HSL, 3-OH-C14:1-HSL) variants (n = 12); quinolone signals (PQS, HHQ, HQNO, NHQ; n = 4); universal/interspecies signals including AI-2 (DPD) and indole derivatives (indole, 5-hydroxyindole, isatin, indole-3-carboxaldehyde, indole-3-acetic acid; n = 6); DSF family (DSF, BDSF, CDSF; n = 3); Streptomyces butyrolactone signals (A-factor, SCB1, virginiae butanolide A; n = 3); Vibrio CAI-1 (n = 1); Photorhabdus photopyrone A (n = 1); dual signal/antibiotic tropodithietic acid (TDA; n = 1); bacterial signal metabolites (pyocyanin, phenazine-1-carboxylic acid, 2-aminoacetophenone, tyramine, hordenine; n = 5); QS inhibitors with documented signaling modulation (furanone C-30, 5-fluorouracil, cinnamaldehyde, eugenol, ajoene, curcumin, baicalin; n = 7); and antivirulence signal modulators (savirin/M64, mBTL, virstatin, LED209, LsrK inhibitor 11e; n = 5).

Peptide autoinducers and intracellular messengers (n = 17, “Negative” subset): Gram-positive peptide autoinducers and lantibiotics (AIP-I, AIP-II, AIP-III, AIP-IV from S. aureus; CSP-1 and CSP-2 from S. pneumoniae; GBAP from E. faecalis; ComX from B. subtilis; SHP from Streptococcus; iAM373 and cCF10 from E. faecalis; nisin A; n = 12) and intracellular second messengers (cyclic-di-GMP, cyclic-di-AMP, cAMP, ppGpp, pppGpp; n = 5). These serve as an internal negative control — they are bona fide signaling molecules but are predicted to fail drug-likeness filters due to fundamentally different physicochemical constraints (large peptide size for receptor specificity; high charge density for intracellular enzyme recognition).

Antibiotics with sub-MIC signaling effects (n = 9, “Excluded” subset): Azithromycin, tobramycin, ciprofloxacin, tetracycline, erythromycin, trimethoprim, rifampicin, colistin, and daptomycin. These are excluded from the primary enrichment analysis to avoid circular reasoning (antibiotics are already known drugs, so including them would artificially inflate drug-likeness metrics).

The complete dataset with all 74 molecules and their physicochemical properties is in Supplementary Table S1.

Random drug-like molecule dataset (n = 71). Molecules selected to span diverse therapeutic areas: cardiovascular (amlodipine, atorvastatin, clopidogrel, digoxin, lisinopril, losartan, metoprolol, warfarin; n = 8), CNS/psychiatric (carbamazepine, diazepam, donepezil, fluoxetine, gabapentin, levetiracetam, risperidone, sertraline; n = 8), oncology (cyclophosphamide, doxorubicin, erlotinib, imatinib, methotrexate, paclitaxel, sorafenib, tamoxifen; n = 8), metabolic/endocrine (glipizide, levothyroxine, metformin, pioglitazone, sitagliptin; n = 5), anti-inflammatory (celecoxib, colchicine, ibuprofen, indomethacin, naproxen, prednisone; n = 6), antiviral (acyclovir, lopinavir, oseltamivir, remdesivir, ribavirin, tenofovir; n = 6), antifungal (fluconazole, griseofulvin, ketoconazole, terbinafine; n = 4), miscellaneous FDA-approved drugs (allopurinol, chloroquine, finasteride, hydroxychloroquine, montelukast, omeprazole, ranitidine, sildenafil; n = 8), failed/withdrawn (BMS-986165/deucravacitinib, lorcaserin, rofecoxib, torcetrapib; n = 4), large/complex (cyclosporine, ivermectin, sirolimus, vincristine; n = 4), and additional diverse drugs (apixaban, aspirin, caffeine, canagliflozin, cetirizine, diphenhydramine, empagliflozin, loratadine, pantoprazole, rosuvastatin; n = 10). Complete dataset in Supplementary Table S2.

Candidate molecule dataset (n = 26). Subset curated for Part 2: all molecules from the Primary subset with documented signaling activity and without established clinical antibiotic use. Complete with testing gap annotations, PAINS flags, and priority classifications in Supplementary Table S3.

2.3. Drug-Likeness Filters

Seven independent filters were applied using published threshold criteria:

1. Lipinski Rule-of-5 — strict (0 violations: MW ≤ 500, LogP ≤ 5, HBA ≤ 10, HBD ≤ 5) (Lipinski et al., 2001)

2. Lipinski Rule-of-5 — relaxed (≤1 violation allowed) (Lipinski et al., 2001)

3. Veber Criteria — Rotatable bonds ≤ 10 AND TPSA ≤ 140 Ų (Veber et al., 2002)

4. O’Shea & Moser Antibiotic-Specific — MW < 600, −2 < LogP < 6, HBD ≤ 5, HBA ≤ 12 (O’Shea & Moser, 2008)

5. Ghose Filter — 160 ≤ MW ≤ 480, −0.4 ≤ LogP ≤ 5.6, 40 ≤ atoms ≤ 130, 20 ≤ MR ≤ 130 (Ghose et al., 1999). Note: atom count and molar refractivity (MR) were not available in the compiled dataset; Ghose pass rates are based on MW and LogP criteria only and may therefore be overestimated.

6. Egan Filter — LogP ≤ 5.88, TPSA ≤ 131.6 Ų (Egan et al., 2000)

7. Muegge Filter — 200 ≤ MW ≤ 600, −2 ≤ LogP ≤ 5, TPSA ≤ 150, HBA ≤ 10, HBD ≤ 5, RotBonds ≤ 15, rings ≤ 7 (Muegge et al., 2001)

A composite drug-likeness score (0–6) was calculated for candidate molecules as the number of filters passed (excluding the Lipinski relaxed variant to avoid double-counting).

2.4. ADMET Predictions

Simplified ADMET predictions were derived from threshold-based classifications applied to the physicochemical properties in the dataset: membrane permeability (Good: TPSA < 75 Ų AND 0 < LogP < 5; Poor: TPSA > 140 Ų OR LogP < −2 OR LogP > 6; Moderate: all intermediate values) and BBB penetration (High: TPSA < 60 Ų AND LogP > 1; Low: TPSA > 120 Ų OR LogP < −1; Moderate: intermediate). These threshold-based classifications are substantially less reliable than validated computational tools (SwissADME BOILED-Egg model, pkCSM, ADMETlab 3.0) and should be treated as rough estimates only.

2.5. PAINS Screening

All 26 candidate molecules were evaluated for known PAINS alerts based on published criteria (Baell & Holloway, 2010). Molecules containing Michael acceptors, reactive electrophiles, quinone/redox-cycling scaffolds, or known promiscuous motifs (catechols, thioallyl groups) were flagged. This was a manual structural assessment guided by AI, not a computational PAINS filter (e.g., RDKit FilterCatalog).

2.6. Statistical Analysis

Mann-Whitney U tests for continuous property comparisons; Fisher’s exact test for categorical pass/fail comparisons; rank-biserial correlation (r) as effect size metric. Bonferroni correction applied separately for property comparisons (6 tests, adjusted α = 0.0083) and filter comparisons (7 tests, adjusted α = 0.0071). Both uncorrected and Bonferroni-corrected p-values are reported. Power: with n = 48 and n = 71, the Mann-Whitney U test has approximately 80% power to detect medium effects (r ≈ 0.27) at α = 0.05.

3. Results

3.1. Part 1: Three-Group Comparison Reveals Specificity of Drug-Likeness Enrichment

3.1.1. Property Distributions

Small-molecule signaling molecules showed significantly lower molecular weight, fewer hydrogen bond acceptors, and lower TPSA than random drug-like molecules — all surviving Bonferroni correction (Table 1, Figure 1). LogP, HBD, and rotatable bonds did not differ significantly after correction. Critically, peptide autoinducers and intracellular messengers showed property profiles dramatically different from both groups, confirming that drug-likeness is not a universal property of bacterial signaling molecules.

3.1.2. Drug-Likeness Filter Enrichment: Three-Group Comparison

The three-group comparison reveals a striking pattern: small-molecule signals showed high pass rates (56–98%), random drugs showed intermediate rates (63–83%), and peptide/intracellular signals showed near-zero rates on all filters (Table 2, Figure 2). After Bonferroni correction, three filters reach significance: Lipinski strict (adjusted p = 0.009), Egan (adjusted p = 0.004), and O’Shea antibiotic-specific (adjusted p = 0.016).

Key finding: Lipinski compliance. Small-molecule signaling: 46/48 (95.8%); Peptide/intracellular: 1/17 (5.9%); Random drugs: 52/71 (73.2%). Fisher’s exact (small-mol. vs. random): OR = 8.40, 95% CI: 1.86–38.1, p = 0.001; Bonferroni-adjusted p = 0.009.

3.1.3. Chemical Space Mapping

The chemical space map (Figure 3) shows the three distinct populations. Small-molecule signals cluster tightly within both the Lipinski zone (MW ≤ 500, LogP ≤ 5) and the antibiotic-relevant O’Shea zone, while peptide autoinducers (MW 680–3354 Da) and intracellular messengers (LogP −5 to −7) fall far outside drug-like chemical space. Random drugs show the expected broad distribution.

3.1.4. Predicted ADMET Properties

72.9% of small-molecule signals were classified as “Good” predicted membrane permeability vs. 32.4% of random drugs and 0% of peptide/intracellular signals (Table 3).

3.2. Part 2: Candidate Identification

Among 26 curated small-molecule signaling molecules (Supplementary Table S3), 10 were classified as HIGH priority based on composite drug-likeness score ≥ 5/6 and a documented testing gap. After PAINS assessment, three were flagged for structural alerts (Table 4). The six highest-confidence candidates — Savirin, LED209, LsrK inhibitor 11e, 3-oxo-C12-HSL, Hordenine, and Indole-3-acetic acid — are free of major PAINS alerts.

4. Discussion

4.1. Drug-Likeness Enrichment Is Specific to Small-Molecule Diffusible Signals

The three-group comparison provides stronger evidence than a simple two-group design. Small-molecule diffusible signals show ~96% Lipinski compliance (OR = 8.40, Bonferroni-adjusted p = 0.009), while peptide autoinducers and intracellular messengers show ~6% compliance. This demonstrates that drug-likeness enrichment is not a universal property of all bacterial signaling but is specific to the class of small molecules that function via passive diffusion and membrane transit. Three filters now survive Bonferroni correction (Lipinski, Egan, O’Shea), up from one in the smaller initial dataset.

This suggests a practical application: in silico filtering for signaling-like properties simultaneously enriches for drug-likeness (EF 1.12–1.31), membrane permeability (72.9% vs. 32.4% “Good”), and biological relevance to bacterial physiology. This convergence is further supported by findings that indole signaling modulates efflux pump expression (Salama et al., 2024) and that resistance machinery components can serve as signal transducers (Koberska et al., 2021).

4.2. The Testing Gap and Dose-Escalation Rationale

The antivirulence compounds (Savirin, LED209, Virstatin, mBTL) were explicitly designed NOT to kill bacteria — the question “what happens at 10–100× the anti-virulence EC₅₀?” has not been systematically published. 3-oxo-C12-HSL induces eukaryotic cell apoptosis at ~100 µM (Tateda et al., 2003), yet its effects on bacterial viability remain uncharacterized. Indole-3-acetic acid directly controls antibiotic production in bacteria (Matilla et al., 2018).

We acknowledge that unpublished negative data may exist, that most virulence-associated QS systems are not essential for growth under laboratory conditions, and that our testing gap assessment was AI-mediated rather than a systematic review.

4.3. Biological Considerations

Species-specificity. QS systems differ between Gram-negative (AHL-based) and Gram-positive (peptide-based) bacteria (Whiteley et al., 2017). Universal signals (AI-2, indole) may offer broader targets. Toxicity. 3-oxo-C12-HSL modulates NF-κB signaling (Kravchenko et al., 2008); Hordenine is sympathomimetic. Resistance. Efflux pumps export AHL and quinolone signals; AHL lactonases degrade AHL candidates; LasR loss-of-function mutants are common in chronic infections (Feltner et al., 2016). PAINS. Furanone C-30, curcumin (Nelson et al., 2017), and cinnamaldehyde contain reactive electrophilic motifs and were deprioritized (Baell & Holloway, 2010).

4.4. Limitations

(1) AI-assisted methodology. This work used Claude (Anthropic) as the primary computational tool. The author lacks formal cheminformatics training. AI analyses are stochastic and may differ across sessions. Results should be considered provisional until reproduced using deterministic workflows (RDKit, OpenBabel).

(2) Compiled rather than computed properties. Properties were sourced from PubChem/DrugBank, not computed from canonical SMILES with a single validated toolkit. Different algorithms may underlie values across molecules. This is the most critical methodological weakness.

(3) Peptide property estimation. Properties for Gram-positive peptide autoinducers (n = 12) were estimated from published data and sequence-based calculations. These values are sufficient to demonstrate peptides fall far outside drug-like space but are not experimentally determined.

(4) Sample sizes. n = 48 vs. n = 71 provides ~80% power for r ≈ 0.27. Larger validation against ChEMBL/ZINC (n > 10,000) is needed.

(5) Reference set bias. The random set consists primarily of FDA-approved drugs, already biased toward drug-likeness.

(6) Simplified ADMET and PAINS. Threshold-based ADMET (Section 2.4) and manual PAINS assessment are less reliable than validated tools (SwissADME, pkCSM, RDKit FilterCatalog).

(7) No experimental validation. Drug-likeness does not predict antibacterial activity.

4.5. Recommended Validation Pathway

Computational (immediate): Recompute all properties from canonical SMILES using RDKit; validated ADMET via SwissADME/pkCSM; RDKit PAINS filters; expand to n > 100 from ChEMBL; compare against 10,000+ unbiased compounds.

Literature: PRISMA-compliant systematic review for MIC data on all candidates.

Experimental: MIC at up to 256 µg/mL against ESKAPE pathogens; dose-response profiling; cytotoxicity (HepG2, HEK293); checkerboard synergy assays.

Definitive test: Prospective comparison of signaling-enriched vs. random library hit rates.

5. Conclusions

This AI-assisted exploratory analysis provides preliminary evidence that small-molecule bacterial signaling molecules exhibit drug-like physicochemical profiles, while peptide-based and intracellular signaling molecules do not. The strongest evidence is Lipinski compliance (OR = 8.40, Bonferroni-adjusted p = 0.009); three of seven drug-likeness filters now survive Bonferroni correction. Ten candidates with testing gaps were identified, six free of PAINS alerts.

The ultimate question is empirical: do signaling molecules yield antibiotic hits at rates higher than random libraries?