Chemical Biology and Target Engagement
Chemical biology is where molecule generation becomes a biological test. A molecule that docks well is not necessarily a useful probe. A compound that changes a phenotype is not necessarily acting through the intended target. AI systems may help prioritize compounds, infer mechanisms, design perturbation experiments, and search degrader space, but target engagement and biological mechanism remain experimental questions.
Use this chapter to:
- Connect molecules to mechanisms through probes, target engagement, perturbation profiles, degraders, molecular glues, and assay design.
- A docking pose or phenotype is a hypothesis; target engagement, selectivity, cellular activity, and orthogonal assays decide whether chemistry explains biology.
What is this field trying to solve? Connect molecules to mechanisms through probes, target engagement, perturbation profiles, degraders, molecular glues, and assay design.
What is the core idea? A docking pose or phenotype is a hypothesis; target engagement, selectivity, cellular activity, and orthogonal assays decide whether chemistry explains biology.
What is the current state of the field? AI can help rank hypotheses and organize profiles, but direct mechanism assignment still depends on chemical and cellular evidence.
What do we know, and what remains open? Known reference points include Target 2035, Chemical Probes Portal, ChEMBL, BindingDB, PubChem, DUD-E, L1000, Cell Painting, CETSA, thermal proteome profiling, PROTAC studies, and molecular-glue profiling. What remains open is whether profile similarity or docking can become mechanism evidence without orthogonal target-engagement assays.
Why does this matter? Chemical biology links small molecules to target discovery, cell phenotyping, systems biology, degraders, biomarkers, and therapeutics.
Introduction
Small-molecule AI often stops at generation, docking, or ADMET. Chemical biology asks the next question: what does the molecule teach about the biological system? The answer requires target engagement, selectivity, cellular activity, concentration-response behavior, orthogonal assays, and a mechanism that survives perturbation.
Target 2035 frames the chemical-probe gap at field scale: the goal is potent and selective pharmacological modulators for human proteins, with high-quality protein-ligand data and iterative prediction-testing loops as enabling infrastructure (Edwards et al., 2025). That is the right evidence posture for AI in chemical biology: model output matters when it creates testable, selective, cell-active perturbations.
The Chemical Probes Portal and the broader probe-quality literature make the same point from another direction. Arrowsmith et al. warned that poorly selective or weak probes produce misleading biology and recommended potency, selectivity, and cellular activity as minimum requirements for credible probe use (Arrowsmith et al., 2015). That standard is more important when AI expands the number of plausible molecules. More candidates do not mean more biology unless the validation standard rises with the search space.
What Counts as Chemical Biology Evidence?
Chemical biology sits between computational chemistry and functional biology. It is not enough to ask whether a compound has a predicted binding pose. The useful questions are more concrete:
- Does the compound engage the intended target in the relevant cellular context?
- Is the response concentration-dependent and time-dependent in a biologically coherent way?
- Does an inactive analog lose the phenotype?
- Does an orthogonal chemical series reproduce the effect?
- Does genetic perturbation of the target change compound response?
- Does rescue or pathway readout support the proposed mechanism?
- Does selectivity profiling explain both intended and unintended phenotypes?
Each question reduces a different false-positive class. Docking controls pose plausibility. Biochemical assays control affinity. Cellular target-engagement assays control permeability and intracellular binding. Chemoproteomics controls target and off-target scope. Phenotypic assays control biological relevance. Genetics controls causality.
Current Models, Datasets, and Benchmarks
Chemical biology uses many of the same chemical data resources as small-molecule discovery, but the benchmark target differs. The goal is not only potency or affinity. The goal is interpretable perturbation.
| Resource or assay family | What it supports | Where it fails if overread |
|---|---|---|
| ChEMBL, BindingDB, PubChem | Bioactivity and compound metadata for training and lookup (Gilson et al., 2016; Kim et al., 2023) | Inconsistent assay context, publication bias, and inactive data sparsity |
| DUD-E and virtual-screening benchmarks | Retrospective docking or virtual-screening comparisons (Mysinger et al., 2012) | Decoy bias, analog leakage, and poor transfer to prospective campaigns |
| L1000 Connectivity Map | Gene-expression profile matching for small molecules and perturbagens (Subramanian et al., 2017) | Mechanism inference without direct target engagement |
| Cell Painting | Multiplexed morphological profiling for phenotypic screening (Bray et al., 2016; Cimini et al., 2023) | Image similarity mistaken for molecular mechanism |
| CETSA and thermal proteome profiling | Cellular target engagement and proteome-wide thermal-stability changes (Martinez Molina et al., 2013; Savitski et al., 2014) | Indirect thermal shifts mistaken for direct binding |
| PROTAC ternary-complex studies | Degrader mechanism, cooperativity, selectivity, and linker effects (Sakamoto et al., 2001; Gadd et al., 2017) | Binary affinity mistaken for degradation activity |
| Molecular-glue profiling | Discovery of compounds that induce target-ligase interactions (Mayor-Ruiz et al., 2020) | Overconfident rational design where interface rules remain incomplete |
The main benchmark lesson is that random splits and retrospective enrichment are weak evidence for biological mechanism. A model that ranks known actives above decoys may still fail when asked to find a new probe, in a new scaffold, with cellular engagement and selectivity.
Probe Quality and Assay Selection
Chemical biology programs should treat probe quality as a release criterion for biological interpretation. A compound is useful only if its measured behavior fits the question. A covalent electrophile may be valuable for target discovery and chemoproteomics, but unacceptable for a clean pathway probe if it reacts broadly. A weak reversible binder may be useful in a biochemical assay and uninformative in cells if exposure is poor. A degrader may show strong target loss and still create hard-to-interpret biology if degradation is slow, partial, or mediated by a highly context-specific E3 ligase.
The practical assay stack has layers:
| Evidence layer | Example assays | Interpretation |
|---|---|---|
| Chemical identity and purity | LC-MS, NMR, stereochemical assignment, solubility check | Confirms the material being tested |
| Biochemical binding or activity | SPR, ITC, enzyme assay, fluorescence polarization, AlphaScreen | Tests purified-system interaction or inhibition |
| Cellular engagement | CETSA, NanoBRET, occupancy assay, activity-based probe competition, target pull-down | Tests whether the compound reaches and engages the target in cells |
| Selectivity | Kinome panel, proteome-wide profiling, chemoproteomics, matched inactive analog | Tests whether interpretation survives off-target assessment |
| Phenotype | Viability, morphology, transcriptomics, secretion, reporter assay, functional cell assay | Tests whether engagement produces a biological response |
| Causality | Genetic knockout or knockdown, rescue, orthogonal chemotype, pathway epistasis | Tests whether the proposed target explains the phenotype |
The order is not fixed, but the logic is. A phenotypic screen may start before target identity is known. A degrader campaign may start with a known binder. A chemoproteomic screen may begin from a reactive fragment library. In every case, the claim should stop at the strongest measured layer. If the strongest evidence is profile similarity, the claim is MoA hypothesis. If the strongest evidence is cellular target engagement without downstream response, the claim is engagement. If the strongest evidence is phenotype without target engagement, the claim is activity, not mechanism.
Degraders and Molecular Glues
Degraders and molecular glues deserve separate treatment because their biology is not reducible to affinity. PROTACs bring a target protein and an E3 ligase into proximity. Molecular glues induce or stabilize interactions that were weak or absent without the compound. The productive event is therefore a cellular complex, not only a compound-target pair.
For PROTACs, binary target affinity and E3 affinity are incomplete evidence. Linker length, linker exit vector, cooperativity, target abundance, E3 abundance, subcellular localization, ubiquitination kinetics, proteasome dependence, and target re-accumulation rate all affect observed target loss. A strong binder may fail as a degrader if the ternary complex is geometrically poor. A weak binder may become effective if the ternary complex is cooperative. This is why structural ternary-complex evidence and degradation kinetics are more informative than docking alone (Gadd et al., 2017).
For molecular glues, target identity is often discovered through sensitivity patterns, resistance mutations, proteomics, or genetic screens. The model output should be treated as a proposal about induced proximity. The validation package should show target engagement, target or substrate depletion, E3 dependence where relevant, pathway coherence, and proteome-wide selectivity. Without those elements, a molecular-glue claim is usually only phenotypic activity with a proposed mechanism.
Invalid Docking Traps
Docking is useful when it is read as a triage layer. It becomes dangerous when the pose is converted into a mechanism claim. Three traps recur.
First, a predicted pose may be physically strained, chemically implausible, or incompatible with protein flexibility. Second, a benchmark may reward recognition of known ligand classes rather than transfer to new chemical matter. Third, a compound may bind in vitro and fail in cells because permeability, efflux, metabolism, protein binding, or target state differs from the model assumptions.
The antidote is not to discard docking. The antidote is to connect docking to prospective tests: make the compound, confirm identity and purity, measure binding, test cellular engagement, profile selectivity, and run a negative control. If the docking hypothesis survives those layers, it becomes useful chemistry.
What is demonstrated?
Demonstrated capability includes AI-supported hit finding and property prediction when the assay, target, chemical series, and validation data are specified.
Demonstrated capability also includes mechanism-of-action and target-engagement inference as prioritization when followed by orthogonal experiments such as competition assays, thermal-shift assays, proteomics, genetics, or chemical rescue.
Demonstrated capability includes profile-based hypothesis generation. L1000 showed that large-scale gene-expression profiles can connect small molecules, genes, and biological states (Subramanian et al., 2017). Cell Painting extends that logic to morphology (Bray et al., 2016). These assays are valuable because they create a measurable phenotype across many perturbations. They do not identify a direct molecular target without follow-up.
Demonstrated capability includes target-engagement measurement in cells. CETSA established that ligand binding can stabilize proteins in intact cells and tissues (Martinez Molina et al., 2013). Thermal proteome profiling expanded this concept across the proteome using mass spectrometry (Savitski et al., 2014). AI-supported analysis may help triage noisy proteomics and profile data, but the evidence claim comes from measured engagement.
Demonstrated capability includes degrader mechanism analysis when ternary-complex and cellular degradation evidence are present. PROTACs were introduced as chimeric molecules that direct target proteins to an E3 ligase for degradation (Sakamoto et al., 2001). Later structural work showed that cooperative ternary-complex formation can drive selectivity (Gadd et al., 2017). Molecular-glue discovery has similarly benefited from systematic profiling, but proof requires induced interaction, target loss, and pathway-consistent response (Mayor-Ruiz et al., 2020).
What is theoretical?
Theoretical capability includes end-to-end chemical-probe design that jointly optimizes binding, selectivity, cellular engagement, permeability, stability, and phenotype. Current systems handle parts of that objective but need experimental iteration.
Theoretical capability also includes degrader and molecular-glue design from structure and cellular context. The biology depends on ternary-complex formation, expression, localization, and degradation machinery, not only binary binding.
Theoretical capability includes mechanism-of-action inference from multi-modal perturbation profiles. A future workflow may combine docking, chemoproteomics, transcriptomics, morphology, proteomics, metabolomics, CRISPR perturbation, and dose-response data in one model. That would be useful because mechanism often lives in the relationship among modalities. The current evidence base supports hypothesis ranking, not automatic mechanism assignment.
Theoretical capability includes active-learning campaigns for probe discovery. The ideal loop would propose compounds, run target-engagement and phenotypic assays, update the model, and choose the next set. The chemistry and assay operations exist in pieces. The general-purpose probe engine remains unproven because the hard part is not only search. It is selecting evidence that distinguishes direct target biology from correlated cell stress.
What is beyond current capability?
Beyond current capabilities includes inferring mechanism of action from a single phenotype or docking pose. Phenotypic similarity and predicted binding are hypotheses until tested.
Beyond current capabilities includes calling a molecule a probe without selectivity and cellular-engagement evidence. A probe must be an instrument, not only a binder.
Beyond current capabilities includes treating a degrader docking model as proof of degradation. Degradation depends on target binder, E3 binder, linker geometry, ternary-complex stability, cooperativity, expression of target and ligase, subcellular localization, proteasomal routing, and pharmacology. Binary binding is one part of the system.
Beyond current capabilities includes assuming that a foundation model trained on chemical structures has learned biological mechanism. Chemical similarity is useful. Biological mechanism requires assays.
What would make this more promising?
A chemical biology result becomes more promising when it moves a molecule from plausible binder to interpretable biological perturbation. The evidence should match the claim.
| Claim | Evidence that raises or lowers confidence |
|---|---|
| “This compound binds the target” | Biophysical binding with concentration-response behavior, appropriate protein construct, controls, and reproducibility |
| “This compound engages the target in cells” | CETSA, NanoBRET, occupancy assay, chemoproteomics, or another cellular target-engagement assay in the relevant cell type |
| “This compound is selective” | Broad biochemical or chemoproteomic profiling, inactive analogs, orthogonal chemical series, and off-target interpretation |
| “This phenotype is target-mediated” | Target knockdown or knockout, rescue, pathway readout, inactive analog, and dose-time coherence |
| “This compound is a probe” | Potency, selectivity, cellular activity, known exposure range, negative control compound, and clear use guidance |
| “This degrader works through the intended mechanism” | Target degradation, proteasome dependence, E3 dependence, ternary-complex evidence, degradation kinetics, and loss of downstream signaling |
| “This molecular glue has a defined target” | Induced interaction evidence, target depletion or functional modulation, genetic resistance or rescue, and proteomic selectivity |
| “This docking result is useful” | Prospective assay success, chemical novelty, physical pose validity, selectivity data, and comparison against standard docking baselines |
The evidence ladder should be pre-specified before a campaign starts. Otherwise, the program drifts toward post hoc mechanism stories after a phenotype appears.
Failure Modes
Invalid docking confidence. Docking scores are not binding constants, and low-RMSD poses are not automatically chemically plausible. DUD-E and related benchmarks are useful, but benchmark enrichment does not equal target engagement in cells (Mysinger et al., 2012).
Phenotype equals mechanism. A compound that clusters with a known perturbagen may share mechanism, stress response, toxicity, pathway convergence, or imaging artifact. Profile similarity should lead to target-engagement work, not replace it.
Probe label inflation. A weak, promiscuous, or poorly characterized compound can generate publishable-looking biology and still be a bad probe. Probe quality is an evidence status, not a branding term.
No negative control. Without an inactive analog, orthogonal chemotype, or genetic rescue, many mechanism claims remain fragile. Negative controls are not decorative. They are what make the interpretation readable.
Degrader oversimplification. PROTAC and molecular-glue campaigns often fail when target and E3 expression, ternary complex geometry, permeability, efflux, linker effects, and degradation kinetics are ignored.
Assay mismatch. A biochemical assay may show binding to purified protein while the cellular system lacks target accessibility, correct localization, active conformation, or sufficient exposure.
Data leakage. Random splits in chemical ML may place close analogs on both sides of evaluation. Apparent generalization may reflect memorized series behavior rather than transfer to a new scaffold.
What should researchers, biotech teams, funders, and program leaders do with this?
- Define the biological question before selecting a computational or experimental method: target engagement, probe quality, MoA inference, degrader campaign, or phenotypic triage.
- Require a target-engagement plan before treating any computational hit as a mechanism claim.
- Use inactive analogs, orthogonal chemotypes, and genetic perturbation as standard controls.
- Separate biochemical binding, cellular engagement, phenotypic response, and therapeutic relevance in milestone reviews.
- Report denominators: compounds proposed, selected, made, tested, inactive, promiscuous, toxic, and advanced.
- Treat docking and virtual-screening scores as triage, not evidence of mechanism.
- For degraders, measure target loss, kinetics, E3 dependence, proteasome dependence, and downstream pathway change.
- For molecular glues, require induced-interaction evidence and proteomic selectivity.
- For phenotypic screens, confirm that the phenotype is not assay artifact, toxicity, confluence, edge effect, staining change, or batch effect.
- Keep the final claim modest unless direct target engagement and causality are shown.
FAQ
Is a high docking score evidence of target engagement?
No. A docking score is a hypothesis about pose or binding plausibility. Target engagement requires measured evidence in a biochemical or cellular system, preferably with concentration-response behavior and controls.
When is a compound a chemical probe?
A compound becomes a chemical probe when it is potent enough for the intended assay, selective enough for interpretation, active in the relevant cellular context, paired with controls, and documented so other researchers understand its limits.
Does Cell Painting identify the target of a compound?
Not by itself. Cell Painting measures morphology. Similar profiles may suggest shared mechanism, but target identity requires orthogonal evidence such as chemoproteomics, genetics, target engagement, or rescue.
What makes PROTAC design different from ordinary inhibitor design?
PROTAC activity depends on a three-part system: target binder, E3 ligase binder, and linker. The active species is a ternary complex, and degradation depends on cellular context, cooperativity, kinetics, and proteasome routing.
What is the practical role of AI in chemical biology today?
The practical role is prioritization: choosing compounds, assays, controls, and follow-up experiments more intelligently. The claim becomes stronger only when target engagement, selectivity, and mechanism evidence are measured.