Structure-based drug design (SBDD) begins with a protein structure and ends — if everything works — with a clinical candidate. Between those endpoints lies a series of decision points where computational methods can accelerate the path or, used carelessly, send a program in the wrong direction. This walkthrough covers each stage with practical guidance on where physics-based simulation adds the most value.
Stage 1: Target preparation
A crystal structure in the PDB is a starting point, not a ready-to-use simulation input. Before any docking or FEP calculation, the target requires preparation that significantly affects prediction quality.
Protonation states. X-ray crystallography doesn't place hydrogen atoms — protonation states of active site residues must be assigned by prediction. For histidines in the binding site (common in kinases, proteases, and GPCRs), the protonation state at physiological pH depends on the local electrostatic environment and can differ from the pKa of free histidine. PropKa or MCCE-based assignments are standard; DrugSynq applies protonation state estimation as part of target preparation. Errors here propagate into every downstream calculation.
Water molecules. Crystallographic waters in the binding site are not simply deleted — they participate in protein-ligand interactions through hydrogen bond networks. Conserved waters (appearing in multiple crystal structures of the same target) should be retained unless the ligand displaces them. Displacement of a water molecule from a hydrophobic pocket releases its hydrogen bonds to the protein, contributing a favorable entropic term to binding. This is one mechanism where FEP captures something docking misses.
Missing loops and disordered regions. Many crystal structures have disordered loops, particularly in kinase activation loops and protein-protein interaction interfaces. Missing regions must be modeled (homology modeling or loop modeling) before calculation. If the missing region is adjacent to the binding site, predictions in that region should be treated with additional uncertainty.
Which crystal structure to use. For targets with multiple crystal structures available (CDK2 alone has 400+), the choice of structure matters. Select the structure co-crystallized with the most similar reference ligand. If you're optimizing a hinge-binding kinase inhibitor with a sulfonamide head group, use a structure where the reference ligand's binding mode is well-defined at the hinge. Resolution alone is insufficient — a 1.8 Å structure with a distant reference ligand may produce worse FEP predictions than a 2.4 Å structure with a closer reference ligand.
Stage 2: Hit generation — docking at scale
Given a prepared target and a defined binding site, docking screens virtual libraries to identify candidate hits. Modern rigid docking (Glide XP, AutoDock-GPU, PLANTS) can process 10,000–1,000,000 compounds in hours, making it an effective pre-filter before more expensive calculations.
What docking does well: generating binding poses for congeneric series where the binding mode is known. What it does poorly: predicting absolute binding affinity, handling induced-fit binding sites, and distinguishing actives from decoys in novel chemical space where no training data exists. Docking enrichment factors of 10–20× over random are reasonable expectations; enrichment factors of 100× are usually the result of benchmarking on artificially easy datasets.
Practical docking output: keep the top 500–1,000 poses by docking score, not just the top 50. Docking scores are noisy enough that the true top 10 actives may rank anywhere in the top 500. This poses a problem: you can't synthesize 500 compounds. This is where FEP becomes essential — it dramatically narrows the prioritized list while providing accuracy that docking cannot.
Stage 3: Hit-to-lead — FEP ranking of the docking output
The docking top-500 enters FEP ranking. This step converts a long list of candidates ordered by an approximate scoring function into a shorter list ordered by a physics-grounded free energy estimate.
For a virtual screen input, the first FEP round is often an "absolute" binding free energy calculation — transforming each ligand from a reference compound to measure its relative binding. More typically at the hit-to-lead stage, you have a cluster of structurally similar hits from the docking run and want to know which cluster member to prioritize for synthesis. Relative FEP within clusters answers this directly.
Expected output at this stage: a ranked list of 10–30 synthesis candidates, with ΔΔG values referenced to a shared anchor compound. The ADMET scoring layer runs in parallel. Compounds with predicted favorable binding but unfavorable ADMET scores (hERG flag, poor solubility) are ranked down in the MPO score, reducing the synthesis queue to the most de-risked candidates.
Stage 4: Lead optimization — iterative analog design
Once a lead series is identified with confirmed bioassay activity, the lead optimization phase begins. This is where FEP delivers the most value relative to alternatives. You now have experimental activity data to calibrate against — allowing you to validate FEP predictions against your own assay results and identify if the method has systematic errors for your target class.
The iterative workflow per cycle:
- Analyze SAR from the previous synthesis round. Identify the primary optimization hypothesis (improve hERG, improve solubility, improve metabolic stability, improve potency at a specific binding sub-site).
- Design an analog set that tests the hypothesis — typically 15–25 compounds per cycle. Constrain the scaffold; vary substituents at one or two positions per cycle to generate interpretable SAR data.
- Submit analog set to DrugSynq FEP. Receive ranked ΔΔG + ADMET output within 48 hours.
- Select top 3–5 analogs for synthesis based on combined FEP rank and ADMET profile. Priority goes to analogs that improve the target property without introducing new flags.
- Assay the synthesized compounds. Compare experimental results to FEP predictions. Identify outliers — compounds where FEP was substantially wrong provide structural information about the binding site that isn't captured in the model.
The goal is to complete 6–10 productive design cycles in the time that previously required 3. FEP makes this possible by substituting overnight compute for 3-week synthesis-assay cycles for the 80% of analogs that would have been deprioritized anyway.
Stage 5: ADMET gating — when to gate, when to optimize concurrently
One practical question: should ADMET screening be a sequential gate (pass all ADMET criteria before advancing) or a concurrent optimization objective (balance potency and ADMET profile simultaneously)?
Sequential gating made more sense when ADMET data was expensive and slow. With computational ADMET available on every analog simultaneously with FEP binding affinity prediction, concurrent optimization is now the default approach. The MPO score combines both dimensions — you can set threshold cutoffs (e.g., no compounds with hERG AUROC > 0.7 in priority synthesis) while still advancing the highest-ranked compounds by composite score.
The argument for concurrent optimization: some ADMET liabilities are tightly coupled to potency. For certain kinase inhibitors, the groups that achieve selectivity over kinase family members (e.g., a bulky gatekeeper substituent) also decrease Caco-2 permeability. Treating permeability as a hard gate would kill all the selective compounds. Treating it as a weighted objective allows the medicinal chemist to decide where on the potency-permeability Pareto frontier the program should sit.
What SBDD cannot tell you
Structure-based design requires a known binding mode. For targets without validated crystal structures, or where the compound binds at an allosteric site not co-crystallized with a reference ligand, the accuracy of docking and FEP degrades substantially. Cryogenic electron microscopy has expanded the structural repertoire for GPCRs and other difficult-to-crystallize targets, but the resolution of cryo-EM structures is often insufficient for FEP (2.5–4.0 Å range doesn't adequately resolve water positions and side chain rotamers).
Additionally, SBDD doesn't account for cell permeability, efflux transport, metabolic clearance, or protein binding in vivo. Computational predictions for these parameters (covered in the ADMET modeling layer) complement structure-based binding predictions but address different failure modes. A compound with excellent predicted binding affinity and clean computational ADMET profile may still fail in cellular assays due to off-target binding not visible in the structure, or in vivo due to pharmacokinetic parameters not captured by the in vitro models.
SBDD accelerates the path from target to candidate by improving the information available at each design decision point. It does not guarantee success — it improves the probability of success per synthesis cycle. In a field where the probability of a candidate advancing from discovery to Phase I is roughly 5%, improvements at the early stages compound significantly.
