When a drug binds its target reversibly, the effect fades as the drug washes out of the body. When a drug instead forms a permanent chemical bond with its target — an irreversible covalent inhibitor — the target stays inactivated until the cell manufactures a fresh copy of the protein. This single mechanistic fact, established by pharmacokinetic studies of covalent kinase inhibitors showing near-100% target occupancy even after blood drug levels fall below detection Target modulation and pharmacokinetics/pharmacodynamics translation of the BTK inhibitor poseltinib for model-informed phase-II dose selection and by work on irreversible cytochrome P450 inhibition, where functional recovery strictly requires new protein synthesis Inhibition of Human Liver Cytochrome P450 by Star Fruit Juice, is completely invisible to the standard tools biologists use to interpret drug effects.
Those tools are functional enrichment methods: given a list of genes or proteins a drug touches, they ask which annotated biological pathways are statistically overrepresented in that list, typically with a Fisher's exact test against ontologies like Gene Ontology or KEGG. The problem is that these methods treat every target as an identical binary hit. A permanently-inactivated, slowly-turned-over protein and a fleeting weak off-target contact contribute the same weight. For covalent inhibitors this is not a minor imprecision — it collapses the entire pharmacodynamic logic of the modality.
This proposal builds a covalent-aware, turnover-weighted enrichment statistic, and uses it to answer two questions nobody has asked: where in the human proteome are covalently druggable proteins concentrated at the pathway level, and can turnover weighting predict which pathways stay suppressed late after drug clearance when standard enrichment cannot.
Core Claim
The claim is that pathway-level interpretation of irreversible covalent inhibitors requires weighting each potential target by two quantities that current enrichment pipelines discard: the probability of covalent engagement, and the persistence of inactivation, which is governed by the target protein's turnover half-life. Concretely, for a protein $p$, let $\ell_p \in [0,1]$ be a covalent-ligandability score for its best-positioned nucleophilic residue (a cysteine reactivity prediction from an existing structure-based model), and let $t_{1/2}^{(p)}$ be the protein's turnover half-life. If the drug has largely cleared at time $T$, the expected fraction of $p$ still covalently inactivated is approximately
$$w_p \;=\; \ell_p \cdot 2^{-T/t_{1/2}^{(p)}}$$
because after drug clearance the covalent adduct can only disappear as the protein itself is degraded and resynthesized. Pathway scores are then computed by a weighted rank statistic over the proteome sorted by $w_p$, with significance assessed by permutation — rather than by counting binary hits against a hypergeometric urn model, which is what all standard tools do. When per-pair inactivation kinetics ($k_{\text{inact}}/K_i$, the second-order efficiency of the two-step covalent reaction) are available for a specific inhibitor, they replace $\ell_p$; ligandability serves as the engagement prior when they are not.
This is the piece missing from both sides of the intersection. On the covalent side, CovalentInDB 2.0 CovalentInDB 2.0: an updated comprehensive database for structure-based and ligand-based covalent inhibitor design and screening pairs the DeepCoSI graph neural network — which profiles the covalent ligandability of 144,864 human cysteines — with 8,303 inhibitors and 368 targets, yet treats every protein in isolation, with no pathway or network context of any kind. DeepCoSI itself Proteome-Wide Profiling of the Covalent-Druggable Cysteines with a Structure-Based Deep Graph Learning Network reports an AUROC of 0.92 and AUPRC of 0.82 on its cysteine-ligandability benchmark, so the per-residue engagement signal exists at proteome scale — it has simply never been pushed through a functional-interpretation layer. On the enrichment side, the most advanced existing method, DeepSyn Deep functional synthesis: a machine learning approach to gene functional enrichment, synthesizes functional context from a 3-million-edge knowledge graph and 16.7 million mined PubMed abstracts, but its edges are generic drug–gene associations drawn from STITCH, DrugBank, and text co-occurrence: it has no representation of covalent versus reversible binding, no warhead chemistry, no inactivation kinetics, and no protein turnover, so it cannot in principle condition its output on the modality. The network-pharmacology studies that come closest to combining target prediction with enrichment, such as the morusin and resveratrol analyses Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection 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, apply off-the-shelf overrepresentation analysis to non-covalent docking outputs, again with all targets weighted equally.
Research Program
Three components. First, the statistic itself: an implementation of the weighted rank enrichment test over GO Biological Process and KEGG annotations, with a permutation scheme that shuffles half-lives and ligandability scores across proteins matched for pathway membership structure, so that the test does not merely rediscover that long-lived proteins cluster in certain compartments. Significance will be reported with term-redundancy reduction using Enrichment Map clustering Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation, since flat lists of overlapping terms are a known interpretability failure — Merico et al. condensed 156 significant gene-sets in one breast cancer study into a handful of clusters precisely because raw lists obscure signal. Protein–protein association context from STRING The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest will be used to check that enriched terms form coherent modules rather than scattered annotations.
Second, a descriptive landscape analysis: run the graded-weight test on the full DeepCoSI ligandability profile of 144,864 cysteines from CovalentInDB 2.0 and determine, for the first time, whether covalently ligandable proteins are uniformly spread across the proteome or concentrated in specific pathways. This directly tests whether covalent screening libraries — CovalentInDB 2.0 alone indexes 2,030,192 commercially available candidate compounds — are being deployed with or without knowledge of which cellular systems carry systemic covalent vulnerability.
Third, the temporal validation, which is where the proposal's novelty lives: the turnover weighting is only justified if it predicts something the binary version cannot. The prediction is about time-course behavior. A covalent inhibitor hitting a fast-turnover target produces transient pathway suppression; the same warhead hitting a slow-turnover target produces suppression that persists after drug clearance, as the poseltinib occupancy data and the CYP resynthesis requirement demonstrate. We will assemble matched time-course RNA-seq datasets from covalent-inhibitor treatment experiments (BTK-family inhibitors are the best-documented class in the file, e.g. Irreversible inhibition of BTK kinase by a novel highly selective inhibitor CHMFL-BTK-11 suppresses inflammatory response in rheumatoid arthritis model), and test whether early-timepoint turnover-weighted enrichment predicts which pathways remain significantly suppressed at late timepoints. No benchmark for covalent-inhibitor time-course transcriptomics is indexed in the Lacuna research file; constructing this evaluation set from public deposits is itself a deliverable, and its absence is a limitation we flag openly.
A data-integration risk deserves honesty: the file indexes no dedicated protein-turnover database. Half-lives will be taken from large-scale proteomics turnover measurements and, where missing, imputed, with a sensitivity analysis under the turnover-ablated model to quantify how much imputation noise the method tolerates. If imputation destroys the signal, that is itself a decisive negative result about how much turnover resolution this kind of inference needs.
Baselines
The naive reference point is classical overrepresentation analysis: a Fisher's exact test of a binary target list against GO and KEGG, the urn-model approach that dominates the field and is packaged in tools like funkea funkea: Functional Enrichment Analysis in Python. It is the strongest competitor for the descriptive landscape question because it is exactly the right tool when inputs are genuinely binary and annotation is complete; we will run it on thresholded DeepCoSI hit lists under identical ontologies, permutation budgets, and false-discovery control, and beating it means recovering known covalent-target pathway families (kinase signaling, cysteine protease processing) at the same FDR with fewer redundant terms, and — critically — ranking truly annotated mechanism pathways higher in the retrospective inhibitor test.
The strongest existing method overall is DeepSyn, which replaces static database lookup with machine-learned functional synthesis over a knowledge graph and achieved 96% accuracy in ranking target genes for specific drugs, against 60% for traditional baselines, and outperformed standard methods by at least 18% when recovering Gene Ontology categories Deep functional synthesis: a machine learning approach to gene functional enrichment. We will run DeepSyn on the same input target lists with its published configuration and the same evaluation sets. It cannot accept graded weights, occupancy persistence, or turnover — its input vocabulary has no such fields — so the comparison is deliberately asymmetric in our favor only on the covalent-specific tasks: predicting late-timepoint sustained suppression, and ranking mechanism pathways for poorly curated targets where its literature-mined edges are thin. On literature-rich, well-curated drugs it will likely win the retrospective ranking task, and we will report that honestly. Beating DeepSyn means higher recall of mechanism-of-action terms at matched FDR on the covalent-inhibitor subset, and nonzero accuracy on the temporal prediction task where it has no applicable output at all.
The key ablation isolates turnover: the identical weighted pipeline with $w_p = \ell_p$, i.e. ligandability weighting only, half-life information removed. This ablation is the decisive test of the proposal's central ingredient. If it matches the full model on the temporal task, the turnover hypothesis is wrong and the contribution collapses to graded-weight enrichment — a real but much smaller idea. There is a second, cheaper ablation: binary inputs with DeepCoSI thresholded at 0.5, which reproduces the "just run ORA on the profiler output" recipe and bounds the value of using graded scores at all.
Why not just…?
Why not just run standard ORA on the covalent target list? This is the reviewer's first instinct and it is partially right — for the static, descriptive question it is a fair baseline and we include it as such. But it commits the equal-weighting fallacy the whole proposal attacks: a primary target held at near-100% irreversible occupancy after drug clearance Target modulation and pharmacokinetics/pharmacodynamics translation of the BTK inhibitor poseltinib for model-informed phase II dose selection counts the same as a marginal off-target. It also inherits the documented failure modes of static databases: DeepSyn showed that standard tools return nothing significant for uncurated mechanisms — the discovery paradox — and beat them by at least 18% on GO recovery Deep functional synthesis: a machine learning approach to gene functional enrichment. Most importantly, binary ORA has no output vocabulary for the temporal question, which is our decisive experiment.
Why not just use DeepSyn? For any literature-rich drug, this is a genuine threat, and we say so. But DeepSyn's knowledge graph is built from drug–gene association edges mined from STITCH, DrugBank, and text co-occurrence; it contains no concept of covalent versus reversible binding, no warhead class, no $k_{\text{inact}}$, and no protein half-life. It therefore cannot distinguish an acrylamide Michael acceptor from a reversible binder of the same target, and cannot represent occupancy persistence at all. Our experiments settle the comparison empirically: if DeepSyn matches our method on the temporal prediction task through indirect literature signal, the mechanistic weighting adds nothing and we will report that.
Why not just threshold DeepCoSI at 0.5 and enrich the top proteins? For the landscape question (where are ligandable cysteines concentrated), this is essentially our own stage one, and we concede it. The DeepCoSI paper itself shows the thresholding is well calibrated — 98% success identifying ligandable cysteines in apo structures at a 0.5 threshold — and that pocket-geometry-only modeling without the cysteine-interaction layer is insufficient, which is why we use DeepCoSI scores rather than naive pocket detection. The thresholded-ORA recipe, however, discards two things: the graded information in scores below and above the cut, and the entire turnover dimension. Both are precisely what the ablations measure. If the graded-weight ablation adds nothing over thresholded ORA on every task, the proposal's strong form fails and only the descriptive landscape result stands.
Why not just dock the covalent library proteome-wide with existing tools and then enrich? Non-covalent docking fails outright on covalent chemistry: on the CovDocker benchmark, AutoDock Vina placed only 6.1% of covalent poses within 2 Å RMSD and Smina 9.9%, versus 37.2% for the reaction-aware pipeline, and the traditional tools also placed reactive atoms in physically implausible orientations CovDocker: Benchmarking Covalent Drug Design with Tasks, Datasets, and Solutions. Speed is also disqualifying at proteome scale — over 167 seconds per complex for AutoDock4 — though we note GNINA 1.3's covalent mode reaches 36.2–66.6% success within 2 Å RMSD depending on prior conformational information GNINA 1.3: the next increment in molecular docking with deep learning, and its authors report that empirical Vina scoring still beats the CNN scorer on covalent complexes because the CNNs were trained on non-covalent data. Our design deliberately avoids per-pair docking: we consume precomputed ligandability profiles rather than re-docking, and reserve GNINA covalent docking as a spot-check on a small set of high-weight predicted off-targets.
Decisive Checks
The proposal lives or dies on three numbers. First, in the landscape analysis, whether any GO/KEGG categories show significant overrepresentation of high-ligandability proteins after graded-weight testing with matched permutation — reported as effect sizes and FDR-controlled term counts, with no precedent in the literature to anchor a specific expected value, so any structured, non-random concentration is a finding. Second, in the retrospective test on CovalentInDB's 368 annotated targets, recall of mechanism-of-action pathway terms at 5% FDR for the full model versus the $\ell_p$-only ablation, binary ORA, and DeepSyn. Third, and decisive for the core claim, late-timepoint prediction: Spearman correlation between turnover-weighted early-timepoint pathway scores and observed late-timepoint transcriptomic suppression, compared against the same correlation from every ablation. If the turnover term does not improve that correlation, the persistence mechanism adds nothing measurable and the honest conclusion is that graded ligandability weighting alone — a useful but modest contribution — is all the data supports.
References
- Target modulation and pharmacokinetics/pharmacodynamics translation of the BTK inhibitor poseltinib for model-informed phase II dose selection
- Inhibition of Human Liver Cytochrome P450 by Star Fruit Juice
- CovalentInDB 2.0: an updated comprehensive database for structure-based and ligand-based covalent inhibitor design and screening
- Proteome-Wide Profiling of the Covalent-Druggable Cysteines with a Structure-Based Deep Graph Learning Network
- Deep functional synthesis: a machine learning approach to gene functional enrichment
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- 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
- Enrichment Map: A Network-Based Method for Gene-Set Enrichment Visualization and Interpretation
- The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest
- Irreversible inhibition of BTK kinase by a novel highly selective inhibitor CHMFL-BTK-11 suppresses inflammatory response in rheumatoid arthritis model
- funkea: Functional Enrichment Analysis in Python
- CovDocker: Benchmarking Covalent Drug Design with Tasks, Datasets, and Solutions
- GNINA 1.3: the next increment in molecular docking with deep learning