Can Flux Balance Analysis Predict Antibiotic Synergy? 107,296 Simulations Say No — and What to Do Instead

Corrected October 2026: an earlier version called iYS1720 an S. aureus USA300 model. In BiGG it is the Salmonella Typhimurium reconstruction (Seif et al. 2018, PMID 30218022), so two of the three models are not ESKAPE pathogens; the text now says so.

Companion to two preprints (Research Square, v1, not yet peer-reviewed): the negative result — Why LP-Based FBA Cannot Detect Synergy for Essential Gene Pairs (rs-9398278) — and the constructive method — Continuous-Valued Training Data from Genome-Scale Metabolic Models (rs-9374605). Code (open source, MIT): github.com/shoo99/ai-drug-target. This is computational, hypothesis-generating work — not a clinical or production tool.

TL;DR (Quick Answer)

A popular idea in computational drug discovery is to use flux balance analysis (FBA) on genome-scale metabolic models to predict which antibiotic combinations will act synergistically. We tested that idea exhaustively, and it doesn't hold:

• The negative result. Across 107,296 partial-inhibition simulations on three bacterial metabolic models (E. coli, Salmonella and the ESKAPE pathogen K. pneumoniae), standard LP-based FBA detected exactly zero synergistic essential-gene pairs. This isn't a tuning problem — all 39 essential genes show step-function dose-response with α* = 1.00, which is a mathematically predictable consequence of linear-programming sensitivity analysis. FBA-as-a-synergy-calculator is structurally incapable of the task for essential genes.
• The reframe that works. Use FBA as a feature generator, not a synergy calculator. Feeding FBA-derived metabolic features plus pathway identity into a classifier gave a preliminary signal (F1 = 0.847, AUROC = 0.804 under gene-level cross-validation) — above chance, but explicitly not clinical predictive power.
• A method to make that honest. Replace binary knockouts with partial gene inhibition (10–100% flux reduction) to generate 945 continuous-valued training targets, feed them to compact neural nets, and integrate a knowledge graph, local-LLM literature mining, and AlphaFold to curate 29 ESKAPE drug targets that lack approved therapeutics.
• Everything is open. Full pipeline, a Streamlit dashboard, a 40-test suite, and reproducible figures — MIT-licensed at shoo99/ai-drug-target.

The throughline is epistemic: a model that confirms what you already assume isn't evidence. Knowing exactly why FBA fails here is what tells you how to use it correctly.

Background: the AMR clock, and the appeal of FBA

Antimicrobial resistance has been projected to cause ~10 million deaths a year by 2050 (the O'Neill Review on Antimicrobial Resistance, 2016), and antibiotic combinations are one of the few levers against it. Screening every pair experimentally scales as O(n²), so a computational pre-filter is attractive — and FBA is the obvious candidate, because genome-scale metabolic models (GEMs) already predict single-gene essentiality well.

The leap people make is from "FBA predicts essentiality" to "FBA can predict whether knocking out gene A and gene B is synergistic." That leap is where things break.

The experiment: 107,296 simulations, three pathogens

We took three curated GEMs — iML1515 (E. coli K-12 MG1655; 2,712 reactions, 1,516 genes), iYS1720 (Salmonella enterica serovar Typhimurium; 3,357 reactions, 1,707 genes), and iYL1228 (K. pneumoniae MGH 78578; 1,228 genes) — and, for all 39 curated essential genes, swept pairwise inhibition from 0% to 100% in 10% steps, scoring each grid cell for Bliss-independence synergy. That's 107,296 partial-inhibition simulations.

The result was zero synergistic pairs, in every model. And not noisily zero — cleanly, structurally zero. Every essential gene produced a step-function response: growth is flat until inhibition crosses a threshold, then collapses, with a synergy parameter α* = 1.00 across the board.

Why FBA cannot find synergy here (it's the math, not the biology)

This is the part worth internalizing. FBA solves a linear program: maximize biomass flux subject to stoichiometric and capacity constraints. For an essential gene, partial inhibition does nothing to the optimum until the constraint actually binds, at which point growth drops — a single breakpoint. Combining two such genes under LP gives you the more binding of two breakpoints, not a multiplicative interaction. There is no mechanism in standard LP-FBA for two partial inhibitions to interact super-additively. So "FBA found synergy" in prior work is, for essential genes, largely re-confirming essentiality, not discovering an interaction.

The clean way to say it (and the papers' central distinction): FBA as a synergy calculator fails; FBA as a feature generator might help. Those are different jobs.

The honest ML proof-of-concept (and why the caveats are the point)

As a feature generator, FBA can supply metabolic context to a supervised model. On a small curated set of 45 antibiotic combinations (28 synergistic, 10 antagonistic, 7 additive; 17 target genes, 11 pathways; Bliss scores literature-derived, not measured here), a gradient-boosting classifier using pathway identity + FBA features reached F1 = 0.847, AUROC = 0.804 under gene-level GroupKFold cross-validation (so combinations sharing a gene can't leak between train and test). Permutation testing put it above chance (z = 2.58, p < 0.001).

What makes this credible is that the paper then argues against its own result:

• Pathway identity carried more weight than the FBA features (pathway-only F1 = 0.828 vs. FBA-only F1 = 0.807) — so FBA isn't doing most of the work.
• Remove ribosome-targeting combinations and AUROC falls to 0.627 (near random) — the signal is largely one pathway subgroup, not a general synergy rule.
• 71% of combinations had at least one gene unmapped in iML1515, and a "gene-mapping-status" artifact ranked as the second most important feature — i.e., the model partly exploited a data-coverage artifact.
• And the explicit caveat: above-chance signal in a small, biased n = 45 set "does not imply clinical predictive power."

This is what a useful negative-leaning result looks like — reported with the knife pointed at itself.

The constructive half: continuous training data from partial inhibition

The companion method paper turns the "FBA as feature generator" idea into a reusable pipeline that fixes three things people usually skip: knockout-only FBA gives binary phenotypes (useless for regression), there are no gene-level toxicity datasets, and pipelines rarely report negative validation.

• Partial-inhibition sampling. Applying 10–100% flux reduction to a mixed set of essential and non-essential genes (39 + 30 of iML1515's 1,516 genes) yields 945 continuous-valued FBA simulations to use as regression targets — explicitly framed as training targets, not drug-response predictions.
• Compact networks. A subsystem-structured ANN that mirrors metabolic organization cuts parameters by 61.5% versus a fully-connected baseline, and a dual-head ANN jointly regresses potency and toxicity.
• An honest toxicity heuristic. With no labeled toxicity data, a multi-evidence score (sequence homology 35%, pathway overlap 30%, conservation 20%, cross-reactivity 15%) — with the weights flagged as an initial proposal pending experimental calibration, not a validated model.
• Four-way integration. A Neo4j knowledge graph, local-LLM literature mining (46% effective precision — reported, not hidden), and AlphaFold structural analysis.

Applied across the three GEMs, the pipeline curates 29 targets that lack approved therapeutics. A sequence-homology audit found 11 of 21 assessed targets have no detectable human homolog (good for selectivity), while folA shows ~30% identity to human DHFR2. That is a reminder that sequence identity alone is a weak guide to selectivity: trimethoprim targets bacterial DHFR and is used clinically precisely because it binds the bacterial enzyme far more tightly than the human one.

Why this matters

• For modelers: before you read synergy out of an FBA combination screen, ask whether LP can even represent the interaction. For essential genes, it can't — you're measuring essentiality twice.
• For ML-on-biology: the value here isn't a leaderboard number; it's the discipline — gene-level splits to stop leakage, ablations that expose where the signal really comes from, and naming the data artifact your model latched onto.
• For the field's hygiene: publishing the negative result and the negative validation is the contribution. Continuous-valued partial-inhibition data and the compact ANNs are reusable parts, not a finished drug-discovery oracle.

Honest Limitations

1. Preprints, not peer-reviewed. Provisional.
2. Synergy labels are literature-derived, from heterogeneous conditions — not measured in this work.
3. The ML signal is small and biased (n = 45), pathway-driven, and partly artifact-driven; it is not evidence of clinical synergy prediction.
4. The toxicity heuristic is unvalidated — its weights await experimental calibration.
5. FBA-derived targets are hypotheses. Curation narrows a list; it does not confirm a drug.

FAQ

Q: Does this mean FBA is useless for drug discovery?

No. It means FBA-as-a-synergy-calculator fails for essential gene pairs (a structural LP limitation, shown across 107,296 simulations). FBA-as-a-feature-generator — supplying metabolic context to ML and to target curation — is still useful. The two roles are different.

Q: Why can't linear-programming FBA capture synergy?

For an essential gene, partial inhibition has no effect until the constraint binds, giving a step-function with a single breakpoint. Combining two such genes under LP yields the more-binding breakpoint, not a super-additive interaction — so there's no mechanism for synergy to emerge. Capturing it would need non-LP formulations (e.g., kinetic or regulatory models).

Q: Is the ML model ready to pick antibiotic combinations?

No. It shows above-chance signal on 45 literature-curated combinations, but ablations show the signal is largely one pathway subgroup plus a data artifact. Treat it as a proof-of-concept and a cautionary tale, not a predictor.

Q: Can I reproduce all of this?

Yes — the complete platform (pipelines, a 10-tab Streamlit dashboard, a 40-test suite, and figure-regeneration scripts) is MIT-licensed at github.com/shoo99/ai-drug-target. The metabolic models come from BiGG Models.

Resources

• Preprint 1 — the negative result (rs-9398278): https://doi.org/10.21203/rs.3.rs-9398278/v1
• Preprint 2 — the method (rs-9374605): https://doi.org/10.21203/rs.3.rs-9374605/v1
• Code (open source, MIT): github.com/shoo99/ai-drug-target
• Metabolic models: BiGG Models (iML1515, iYS1720, iYL1228)
• Related on this site: Building ARIA, an LLM-driven RNA-seq framework