This is a walkthrough of how a kinase inhibitor lead optimization program ran through eight design cycles using DrugSynq FEP predictions and ADMET scoring. The compound series and specific molecular structures are anonymized. The SAR observations, prediction accuracy data, and workflow decisions are real — drawn from a program that ran on DrugSynq in 2024. The target is a serine/threonine kinase involved in oncology signaling.
Note: this case study describes an anonymized program. Compound structures, the specific target identity, and the sponsoring organization are not disclosed. The purpose is to illustrate where FEP predictions were accurate, where they were wrong, and how ADMET co-prediction affected synthesis decisions.
Program background
Starting point: an aminopyrimidine-based hinge binder with IC50 = 380 nM against the target kinase, identified from a commercial fragment library screen. The compound was Lipinski-compliant (MW 342, cLogP 2.8, 2 HBD) but showed high metabolic instability in human liver microsome (HLM) assay — half-life of 12 minutes, predicting rapid clearance in vivo. CYP3A4 inhibition IC50 was 2.1 µM, a moderate liability.
The program goal: improve metabolic stability to HLM t½ > 30 minutes while maintaining IC50 < 100 nM against the kinase. Secondary goal: resolve the CYP3A4 liability (IC50 > 10 µM) without increasing hERG risk.
Cycles 1–2: Identifying the metabolic soft spot
Before generating an analog series, we submitted the lead compound for metabolic soft spot prediction using the DrugSynq ADMET engine. The model predicted high oxidation probability at two positions: the unsubstituted pyrimidine C-5 position and the terminal phenyl ring para-position. Both are known CYP3A4 substrates when unsubstituted.
The medicinal chemistry hypothesis: block the predicted metabolic soft spots with fluorine (bioisosteric substitution with minimal steric change, favorable for electronic deactivation of CYP3A4 oxidation) or methyl groups.
Cycle 1 analog set: 18 compounds. Fluorine at pyrimidine C-5, fluorine at phenyl para-, methyl at phenyl para-, and combinations. FEP predicted binding affinities (ΔΔG vs. lead) for all 18 before synthesis.
FEP prediction accuracy, cycle 1: 14 of 18 compounds ranked correctly in the top/bottom tercile by FEP vs. experimental IC50. The 4 mispredictions were all single-fluorine changes at the pyrimidine C-5 position — FEP underestimated the electronic effect of fluorination at that position on hinge hydrogen bond geometry. RMSE for the 18-compound set was 0.84 kcal/mol, which is within expected range for well-behaved congeneric series.
Synthesized the top 4 compounds by combined FEP + ADMET score. Key experimental result: the 5-F pyrimidine analog improved metabolic stability to HLM t½ = 28 minutes while maintaining IC50 = 190 nM. Not yet at target, but a meaningful improvement from a single modification.
Cycles 3–4: Addressing hinge binding and potency
With a partial solution to the metabolic liability, cycles 3 and 4 focused on improving potency. The binding mode from a published homologous kinase structure suggested a potential hydrogen bond acceptor in the gatekeeper-adjacent region that the lead wasn't exploiting.
Design hypothesis: introduce a hydrogen bond acceptor (carbonyl or sulfonamide) at the end of the aminopyrimidine, directing it toward the catalytic lysine. Simultaneously, maintain the fluorine at the 5-position of pyrimidine from cycle 1.
Cycle 3 analog set: 14 compounds testing sulfonamide, urea, and amide linker variants at the C-4 position of the aminopyrimidine. FEP perturbation network was built around the cycle 1 winner as the new reference compound.
ADMET flag from cycle 3: The sulfonamide variants were predicted clean across all ADMET endpoints. The urea variants were flagged for metabolic instability (urea as a metabolic hot spot) — consistent with known medicinal chemistry limitations of urea linkers in kinase inhibitors. This ADMET prediction alone removed 5 compounds from synthesis consideration before any lab work was done.
FEP accuracy, cycles 3–4: RMSE = 0.71 kcal/mol across 14 compounds. The FEP model correctly predicted that a primary sulfonamide (not secondary) at C-4 would improve binding by ~1.8 kcal/mol vs. the reference. Experimental result: IC50 = 48 nM — confirming the FEP prediction within 0.4 kcal/mol.
Cycles 5–6: The hERG incident
By cycle 4, we had a compound with IC50 = 48 nM and HLM t½ = 35 minutes — both targets met. The CYP3A4 liability improved to IC50 = 6.8 µM. The remaining issue: the compound's cLogP had crept up to 3.8 as potency-improving lipophilic groups were introduced in cycles 3–4.
Cycle 5 focused on reducing lipophilicity while maintaining potency. The medicinal chemistry approach: replace the terminal phenyl with a pyridine, reducing cLogP by approximately 0.5 units. FEP predicted the phenyl-to-pyridine change as nearly isosteric from a binding perspective (ΔΔG = +0.2 ± 0.3 kcal/mol — not significantly different from zero).
What FEP didn't predict: The phenyl-to-2-pyridine substitution introduced a predicted hERG liability that wasn't present in the phenyl compound. DrugSynq's ADMET engine flagged hERG inhibition as MODERATE RISK for the 2-pyridine analog (IC50 predicted 3.1 µM). The 3-pyridine regioisomer was predicted clean.
This is an example where ADMET prediction redirected synthesis before any lab work. Cycle 5 synthesis excluded the 2-pyridine analog and included the 3-pyridine and 4-pyridine regioisomers. Experimental hERG result: 2-pyridine analog IC50 = 2.4 µM (ADMET model predicted 3.1 µM — close), 3-pyridine analog IC50 = 18 µM (predicted clean, confirmed clean), 4-pyridine analog IC50 = 11 µM.
The 3-pyridine compound advanced as the preferred compound: IC50 = 51 nM (slightly worse than the phenyl compound, consistent with FEP ΔΔG = +0.2 kcal/mol), HLM t½ = 38 minutes, CYP3A4 IC50 = 8.2 µM, hERG IC50 = 18 µM. All targets met.
Cycles 7–8: Selectivity and finishing work
Cycles 7 and 8 focused on selectivity profiling against the closest kinase family members. FEP was run against a panel of three off-target kinases using crystal structures from the PDB. The perturbation networks covered the most selective analogs from cycle 6.
FEP selectivity predictions were less accurate than potency predictions — RMSE of 1.4 kcal/mol vs. experimental ΔΔG against off-targets, compared to 0.71–0.84 kcal/mol against the primary target. This is expected: the off-target crystal structures had lower resolution and were co-crystallized with different reference ligands, reducing the accuracy of the binding mode assumption. We used FEP selectivity data as directional guidance rather than quantitative prediction in these cycles.
The final candidate from cycle 8: IC50 = 44 nM at primary target, >200× selectivity over two of three off-target kinases, HLM t½ = 42 minutes, CYP3A4 IC50 = 9.1 µM, hERG IC50 = 21 µM, aqueous solubility 28 µg/mL (acceptable for early profiling), Caco-2 Papp = 18 × 10−6 cm/s.
What we learned about prediction accuracy
Aggregating across 8 cycles, 134 FEP-predicted analogs (of which 41 were synthesized), and experimental data against 41 synthesized compounds:
- FEP rank-ordered the top quartile (best predicted compounds) correctly in 78% of cases — meaning the synthesized top-quartile FEP compounds were experimentally better than the average of the other synthesized compounds.
- Median |error| between predicted and experimental ΔΔG: 0.68 kcal/mol.
- ADMET flags correctly identified 4 hERG concerns in advance, 2 metabolic instability concerns, and 1 CYP3A4 inhibition concern. One ADMET false positive: a compound was flagged for hERG inhibition that assayed clean (IC50 > 30 µM).
- Synthesis effort reduction compared to design-all-synthesize-all: 41 compounds synthesized vs. 134 predicted. If the program had synthesized all 134, the expected hit rate at target criteria is ~18% based on the ratio of compounds meeting all targets. Approximately 109 wasted synthesis slots were avoided.
Where FEP was wrong — and what we learned
The systematic failure mode was fluorine substitution adjacent to the hinge-binding aminopyrimidine nitrogen. FEP underestimated the effect of electron withdrawal on the hinge hydrogen bond by approximately 0.4–0.6 kcal/mol in four predictions in cycles 1 and 2. This error was systematic and traceable: the binding mode places the aminopyrimidine NH in direct contact with the hinge carbonyl, and electron-withdrawing substituents on the pyrimidine reduce NH donation more than the OPLS4 partial charges captured.
After identifying this systematic error in cycle 2, we added a correction term to the FEP predictions for this scaffold class: compounds with electron-withdrawing groups directly adjacent to the hinge NH received a +0.4 kcal/mol correction to the FEP output. Cycles 3–8 predictions with this correction applied showed no further systematic bias in this region.
This is the correct way to use FEP in a real program: not as a black-box prediction system, but as a starting point that gets calibrated against your experimental data as the program progresses. A FEP model that can't be corrected for systematic errors against your own assay results is a less useful tool than one that can.
