Pathway-IC50: fitting dose-response curves to functional enrichment, then closing the loop with recombinant enzymes

Half-maximal inhibitory concentration ($\text{IC}_{50}$) — the inhibitor concentration that halves an enzyme's catalytic rate — is the workhorse metric of pharmacology, obtained by fitting a sigmoid curve to residual enzyme activity across a serial dilution of inhibitor. In parallel, functional enrichment analysis turns a list of genes whose expression changed in a treated cell into the biological pathways actually being perturbed. These two worlds rarely speak: transcriptomic studies enrich at a single, typically high dose, and enzyme inhibition studies run on purified protein with no cellular context. Meanwhile, the purified protein itself comes from recombinant production, whose failure modes (insoluble aggregates, misfolded active sites, solubilizing tags that alter the binding pocket) are never treated as a source of error in the $\text{IC}_{50}$ numbers that downstream conclusions rest on.

This proposal unifies the three directions into one closed loop. The core mechanism is dose-resolved functional enrichment: treat cells with a full inhibitor dilution series, score pathway activity at each dose, and fit the same sigmoid used in enzymology to obtain a pathway-$\text{IC}_{50}$ for every pathway. The recombinant side supplies the ground truth: the nominated target enzyme is produced heterologously, and its in vitro $\text{IC}_{50}$ is compared against the pathway-$\text{IC}_{50}$ of the pathway it governs. Disagreement between the two numbers is itself signal — it flags either a cellular off-target, a permeability failure, or an expression artifact (a tag or misfolded fraction shifting the in vitro curve). A third ingredient, expressibility priors from solubility predictors, is inserted into the target-ranking step so that enrichment-nominated targets are ones we can actually produce and assay.

Core Claim

No framework in the reviewed literature fits dose-response models across serial-dilution transcriptomics to compute pathway-level $\text{IC}_{50}$ values, and none cross-validates those against recombinant in vitro $\text{IC}_{50}$ measurements. The claim we will test is threefold. First, that the target's own pathway shows the lowest pathway-$\text{IC}_{50}$ among enriched pathways, and that this number correlates with the recombinant enzyme's $\text{IC}_{50}$ across a compound panel. Second, that weighting target-nomination rankings by predicted expressibility raises the fraction of nominated targets that survive to experimental $\text{IC}_{50}$ validation. Third, that systematic deviations between pathway- and enzyme-$\text{IC}_{50}$ can be attributed to specific causes — off-target engagement, poor cell entry, or expression artifacts — rather than remaining unexplained scatter.

Research Program

The closest existing work stops one step short on each side. The knowledge-graph deconvolution pipeline of A novel approach for target deconvolution from phenotype-based screening using knowledge graph shrinks 1,088 candidate proteins to 35 via a protein-interaction knowledge graph and validates the top hit (USP7) by surface plasmon resonance, but it uses a static graph rather than dose-resolved perturbation data, relies on binding affinity rather than inhibition curves, and never confronts the production bottleneck of its nominated target. Network-pharmacology studies such as Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection and Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation nominate multi-target mechanisms from enrichment but confirm none of them on purified recombinant enzyme. On the production side, Functional and Mechanistic Characterization of an Enzyme Family Combining Bioinformatics and High-Throughput Microfluidics mines 2,905 putative enzymes, expresses 45, and characterizes kinetics in droplet microfluidics — but with no transcriptomic layer and no inhibition profiling. Our program welds these halves together at the dose axis.

Concretely, three components will be built. The first is a pathway activity curve fitter: for each compound and cell line, cells are exposed to a 10-point, three-log dilution series bracketing the known in vitro $\text{IC}_{50}$; RNA-seq at each point yields a rank-based pathway activity score $A_p(d)$ for pathway $p$ at dose $d$, computed without the binary gene cutoffs that plague standard overrepresentation analysis (which discards genes with meaningful but sub-threshold changes). We then fit a Hill equation per pathway,

$$A_p(d) = A_{\min} + \frac{A_{\max} - A_{\min}}{1 + (d/\text{IC}_{50}^{p})^{h_p}}$$

where $h_p$ is a pathway-specific slope and $\text{IC}_{50}^{p}$ is the dose at which pathway $p$ loses half its activity. Curve fitting at the 50% point, rather than reading single concentrations near the saturated plateaus where noise dominates, follows the same logic that makes $\text{IC}_{50}$ the robust inflection metric in enzyme assays (Marani et al.). The second component is an expressibility-weighted ranker: candidate targets are scored by a combination of enrichment evidence (which pathways titrate earliest), network centrality over STRING associations (The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest), and a solubility prior from SoluProt (SoluProt: prediction of soluble protein expression in Escherichia coli) — a gradient-boosting model over 251 sequence features. The third component is a recombinant validation bench: top-ranked targets are expressed in E. coli with standard solubilization strategies from the host-engineering repertoire — low-temperature induction, chaperone co-expression, and solubilizing fusion tags (Characterization of Escherichia coli as a Recombinant Protein Expression Host and its Optimization); difficult eukaryotic targets fall back to the secreted Drosophila S2 route of FAS2FURIOUS: Moderate-Throughput Secreted Expression of Difficult Recombinant Proteins in Drosophila S2 Cells. Each purified enzyme gets a concentration-dependent $\text{IC}_{50}$ by an optical readout matched to the enzyme class — the Ellman colorimetric route for hydrolases, HTRF fluorescence for kinases as in the VEGFR-2 work of Wei et al., or cocktail substrate assays for cytochrome P450 panels as in Human liver microsomes study on the inhibitory effect of plantainoside D on the activity of cytochrome P450 activity. For each construct we produce both a tagged and a tag-cleaved (or tag-minimized) version, because solubilizing tags can sterically alter ligand access — the XXA-tagged TMPRSS2 study (Improving Soluble Expression of SARS-CoV-2 Spike Priming Protease TMPRSS2 with an Artificial Fusing Protein) improved soluble yield but never quantified whether the tag shifted kinetic or inhibition parameters, an open question we answer directly.

Baselines

The naive reference point is standard practice: a single high-dose RNA-seq experiment plus one overrepresentation analysis against Gene Ontology and KEGG, exactly as deployed in the static network-pharmacology studies of morusin in aortic dissection (Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection) and resveratrol in premature ovarian insufficiency (Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation). This baseline gets the same RNA-seq budget as our titration series (we spend its reads on deeper sequencing of one dose point), and it is beaten only if dose-resolved pathway-$\text{IC}_{50}$ ranking places the true target pathway first more often than single-dose overrepresentation does.

The strongest existing competitor for target deconvolution is the knowledge-graph pipeline of A novel approach for target deconvolution from phenotype-based screening using knowledge graph, which combined protein-interaction centrality, literature mining, and docking to cut 1,088 candidates to 35 and validated USP7 with a dissociation constant of $1.92 \times 10^{-8}$ M by surface plasmon resonance. We will re-implement its ranking stage (knowledge-graph centrality plus docking, without literature mining) and run it on the same compound panel with the same compute budget, so the only difference is whether dose-resolved enrichment evidence and expressibility weighting are available. It is beaten if our ranker recovers the true target within its top $k$ for more compounds, and, more importantly, if a larger fraction of its top nominations survive recombinant production to yield a measurable $\text{IC}_{50}$ at all — a success criterion the knowledge-graph paper never had to face because it leaned on pre-purchased protein.

For the expressibility prior, the strongest plug-in baseline is SoluProt (SoluProt: prediction of soluble protein expression in Escherichia coli), which on the independent NESG test set of 9,644 proteins reports accuracy of 58.5% and AUC 0.62; we use its scores unmodified as the prior, so any gain must come from the integration with enrichment ranking rather than from a better solubility model. The UniRep-based B. subtilis classifier (Deep protein representations enable recombinant protein expression prediction, best AUC 0.64, MCC 0.20) is retained as a secondary prior, and we note honestly that its host-specific numbers warn against assuming these priors transfer across hosts.

Two ablations isolate the key ingredients. Removing the dose axis (collapsing the titration to one high-dose point) isolates the value of pathway-$\text{IC}_{50}$ fitting itself. Removing the expressibility weight from the ranker isolates the contribution of the production prior. Both run on identical compound panels and sequencing budgets.

Why not just…?

Why not just run molecular docking on structure models and skip recombinant production entirely? The evidence in the file says docking does not yet substitute for physical measurement. On the Therapeutics Data Commons benchmarks (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development), drug-target interaction models drop from a Pearson correlation near 0.7 on in-distribution data to near 0.4 on temporal splits, and scoring functions are easily gamed. The knowledge-graph study (A novel approach for target deconvolution from phenotype-based screening using knowledge graph) found docking against the full 1,088-protein candidate set computationally prohibitive and still required a biophysical validation step at the end. Docking therefore remains a component of our baseline ranker, but the proposal's decisive numbers are physical $\text{IC}_{50}$ values, and the comparison against the docking-based baseline will settle empirically how much the enrichment and expressibility layers add.

Why not just use simple amino-acid composition instead of learned sequence representations for the expressibility prior? In the B. subtilis expressibility study (Deep protein representations enable recombinant protein expression prediction), models on raw amino-acid frequencies performed poorly, and the boost to AUC 0.64 came from the 1,900-dimensional UniRep representations. But the same file cautions the other way: SoluProt's gradient-boosted trees over hand-engineered features outperformed deep architectures on the NESG test set, and curated RDKit descriptors matched graph neural networks on TDC benchmarks. We therefore do not claim deep representations are the answer; we test both priors head-to-head inside the same ranker, and the primary baseline is the better-documented, independently benchmarked SoluProt (58.5% accuracy, AUC 0.62). The proposal survives either outcome because the contribution is the integration, not the prior.

Why not just do single-dose RNA-seq phenotyping? Because the sigmoid's information is in the middle, not the extremes. Near full inhibition or no inhibition the response curve flattens and noise dominates, which is precisely why enzymology reports the 50% inflection point (Marani et al.); a single high dose likewise pushes cells into generic stress and death pathways whose enrichment terms ("apoptotic process", "cell death") mask the primary high-affinity target. The single-dose baseline in our experiments is exactly this alternative run with matched sequencing budget, so if it recovers targets as well as the titration does, the core claim fails and we will report that.

Why not just measure the recombinant enzyme's $\text{IC}_{50}$ and skip the cells? For ranking potency of a single purified target, this suffices, and it is a real threat to the cellular component's value. But an in vitro $\text{IC}_{50}$ against a purified recombinant enzyme is an operational, assay-dependent number that fails to predict permeability, target engagement in cells, and metabolic turnover — the documented in vitro-to-in vivo disconnect. The decisive check is a divergence analysis: across the compound panel we test whether compounds whose pathway-$\text{IC}_{50}$ greatly exceeds their enzyme-$\text{IC}_{50}$ are systematically the poorly permeable or rapidly effluxed ones, and whether tag-variant constructs show shifted enzyme-$\text{IC}_{50}$ values that propagate (or do not) into the cellular readout. If pathway-$\text{IC}_{50}$ never adds information beyond enzyme-$\text{IC}_{50}$, we will say so and the proposal reduces to a tag-artifact study — itself still an unanswered question raised by the TMPRSS2 work.

Proposed Experiments

The pilot panel uses twelve compound-enzyme pairs with documented inhibition profiles and recombinantly accessible targets drawn from the file: butyrolactone I against JAK1 (Butyrolactone I blocks the transition of acute kidney injury to chronic kidney disease in mice by targeting JAK1), dibenzoxazepinone derivatives against glycogen phosphorylase (Discovery and Activity Evaluation of Novel Dibenzoxazepinone Derivatives as Glycogen Phosphorylase Inhibitors), azelnidipine against flaviviral RNA polymerase (Azelnidipine Exhibits In Vitro and In Vivo Antiviral Effects against Flavivirus Infections by Targeting the Viral RdRp), and plantainoside D against the cytochrome P450 isoforms it weakly inhibits (Human liver microsomes study on the inhibitory effect of plantainoside D on the activity of cytochrome P450 activity) — the CYP cases deliberately stress the expressibility prior, since membrane-bound human enzymes are exactly where naive E. coli expression fails. Known $\text{IC}_{50}$ values for these compounds are pulled from BindingDB (BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data, nearly 2.9 million measurements), whose documented factor-of-1,000 unit errors and target misidentifications mean every literature value we compare against is itself re-measured on our own recombinant material rather than taken on faith.

Experiment one tests target recovery. For each compound, the expressibility-weighted ranker must place the true target enzyme within its top five; the metric is top-$k$ recovery rate against the single-dose overrepresentation baseline and the docking-plus-centrality baseline. Experiment two tests the quantitative link: Spearman correlation between pathway-$\text{IC}_{50}$ of the target-governed pathway and re-measured recombinant enzyme-$\text{IC}_{50}$ across the panel; we pre-register a threshold of $\rho \ge 0.7$ as support for the core claim, chosen to match the in-distribution performance ceiling that computational drug-target models achieve before their documented degradation on harder splits (Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development). Experiment three tests the production side: the fraction of enrichment-nominated targets that yield soluble, catalytically active protein with and without the expressibility weight. The reference for what unaided selection achieves is the haloalkane dehalogenase campaign, where in silico solubility prediction reached only 66.7% accuracy and just 24 of 45 selected candidates proved both soluble and functional (Functional and Mechanistic Characterization of an Enzyme Family Combining Bioinformatics and High-Throughput Microfluidics); our weighting is beaten only if it does not improve on that roughly 53% realized success rate under a matched selection budget. Experiment four is the tag-artifact study: for each produced enzyme, the $\text{IC}_{50}$ of tagged versus tag-cleaved construct is compared, and any shift is checked for propagation into the pathway-$\text{IC}_{50}$. High-throughput droplet microfluidics at up to 20,000 reactions per day with 1,000-fold reduced protein consumption (Functional and Mechanistic Characterization of an Enzyme Family Combining Bioinformatics and High-Throughput Microfluidics) makes this panel affordable in purified protein.

The decisive check, which falsifies the whole program if it fails, is experiment two's correlation. If pathway-$\text{IC}_{50}$ values do not track recombinant enzyme-$\text{IC}_{50}$ values on known-target compounds, then dose-resolved enrichment is noise at the resolution needed for pharmacology, and the loop between cells and purified protein stays open.

References

  1. A novel approach for target deconvolution from phenotype-based screening using knowledge graph
  2. Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
  3. Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation
  4. Functional and Mechanistic Characterization of an Enzyme Family Combining Bioinformatics and High-Throughput Microfluidics
  5. A pyrazolopyran derivative preferentially inhibits the activity of human cytosolic serine hydroxymethyltransferase and induces cell death in lung cancer cells
  6. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest
  7. SoluProt: prediction of soluble protein expression in Escherichia coli
  8. Characterization of Escherichia coli as a Recombinant Protein Expression Host and its Optimization
  9. FAS2FURIOUS: Moderate-Throughput Secreted Expression of Difficult Recombinant Proteins in Drosophila S2 Cells
  10. Design, Synthesis, and Biological Evaluation of Axitinib Derivatives
  11. Human liver microsomes study on the inhibitory effect of plantainoside D on the activity of cytochrome P450 activity
  12. Improving Soluble Expression of SARS-CoV-2 Spike Priming Protease TMPRSS2 with an Artificial Fusing Protein
  13. Deep protein representations enable recombinant protein expression prediction
  14. Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development
  15. Butyrolactone I blocks the transition of acute kidney injury to chronic kidney disease in mice by targeting JAK1
  16. Discovery and Activity Evaluation of Novel Dibenzoxazepinone Derivatives as Glycogen Phosphorylase Inhibitors
  17. Azelnidipine Exhibits In Vitro and In Vivo Antiviral Effects against Flavivirus Infections by Targeting the Viral RdRp
  18. BindingDB in 2024: a FAIR knowledgebase of protein-small molecule binding data