Designing drugs around the proteins they must spare

Three pan-FGFR inhibitors are approved, and Table 9 shows what is wrong with them. All three are as potent on FGFR1 as on the target isoforms, and FGFR1 inhibition is the hypothesised cause of the hyperphosphataemia that is their commonest toxicity. All three also collapse on the gatekeeper mutants that drive clinical resistance: erdafitinib loses 35-fold on FGFR3 V555L, pemigatinib 220-fold. They bind through a 3,5-dimethoxyphenyl group that reaches past the gatekeeper, and a larger gatekeeper residue leaves it nowhere to go. A different chemotype was needed, not a better version of the same one.

Screening hit, already mutant-tolerant

A docking model said an sp3 ring at the ether could reach an asparagine, so the isopropyl was cyclised.

The designed hydrogen bond: (S)-tetrahydrofuran

A scaffold hop for permeability, then N-CD3 to shut down the metabolism that inhibited CYP3A4.

Lead compound, 82 per cent oral in rat

A hydrogen bond drawn on a model, then measured in the crystal.

hingegatekeeper

Paper 1 / 4

Sheet12/31PaperFGFR3Cpd30PDB8E1X

Each edit moved one property without breaking the others.

+15×methoxy · FGFR3

4→5

+49×oxetane · FGFR3

9→16

+2.7×(S)-THF · FGFR3

16→19

+2×N-CD3 · human CL

27→29

Paper 1 / 4

Sheet13/31PaperFGFR3Cpd·PDB·

A methoxy for fifteen-fold potency: Adding a methoxy group at the carbon adjacent to the pyridazine nitrogen took FGFR3 from 68 nM to 4.6 nM while holding FGFR1 selectivity at about 19-fold. This position, the ether vector, became the campaign's main potency lever and everything in Tables 2 and 3 explores it.

Cyclised onto an oxetane: Docking compound 9 into FGFR3 placed the isopropyl beside Asn562, suggesting a heteroatom there could engage the asparagine N-H. Cyclising to an oxetane gave 1.1 nM, a fifty-fold gain, and lifted FGFR1 selectivity to 65-fold.

The active enantiomer: The racemic tetrahydrofuran 17 gave 2.4 nM. Resolving it showed the whole of the activity sits in one enantiomer: the (S) isomer 19 at 0.9 nM against the (R) isomer 18 at 41 nM, a 45-fold difference between mirror images. 19 was the first sub-nanomolar compound in the series.

Three deuteriums to close a metabolic door: If CYP3A4 inhibition came from the amide carbonyl or its N-demethylated metabolite, blocking N-demethylation should remove it. The N-CD3 congener cleared CYP3A4 past 25 uM and halved intrinsic clearance from 1.0 to 0.5 L/h/kg while keeping potency at 0.5 nM. The N-ethyl analogue 28 cleared CYP3A4 too, but lost potency and selectivity, so the deuterium is doing something the extra carbon cannot.

The screening hit 4 was already unusual: 68 nM on FGFR3 and 22 nM on the V555L mutant, so mutant tolerance came free with the chemotype. A methoxy at the ether vector took potency to 4.6 nM. Growing it to ethoxy then cost four-fold, which normally means the pocket is full. The informative result is compound 9: the bulkier but branched isopropoxy lost nothing at all. A pocket that rejects a linear group and accepts a branched one is not small, it is shaped, and that is what suggested a ring would fit.

FGFR1 falls away while the gatekeeper mutants stay inhibited.

gatekeeperhinge

FGFR3 vs FGFR1 IC50, every compound (nM, log)

6.1 nMV555M mutant

2.7 nMV555L mutant

Paper 1 / 4

Sheet14/31PaperFGFR3Cpd30PDB8E1X

Every polar contact PLIP measured

PDB	type	residue	Å	role
8E1X	hydrogen bond	Leu487 (bb)	4.10	protein_acceptor
8E1X	hydrogen bond	Ala567 (bb)	3.03	protein_donor
8E1X	hydrogen bond	Asn571	3.01	protein_donor

KLIFS (ok, from 8E1X): gatekeeper Val564; hinge Glu565, Tyr566, Ala567; DFG in; αC in.

Three deuterium atoms closed the last metabolic door.

0.5 nMFGFR3 IC50

>25 µMCYP3A4 IC50

Rat oral bioavailability

Paper 1 / 4

Sheet15/31PaperFGFR3Cpd29PDB·

The screening hit was a better JAK2 inhibitor than CDK2.

CDK2 607 nMJAK2 <1 nM

CDK2 3.2 nM

CDK2 0.29 nM

My role: co-crystal structure 8UV0, co-author

Paper 2 / 4

Sheet16/31PaperCDK2Cpd17PDB·

Screening an internal collection returned compound 1: submicromolar against CDK2 at 607 nM, but subnanomolar against JAK2, which was its real target. The pyrazolyl aminopyrimidine core was worth keeping. The anilino methanesulfonamide was not, because that end of the molecule sits in the hinge region, and hinge substitution is what decides kinase selectivity in this scaffold class. Ebvaciclib, carrying an aminopiperidine there, spares JAK2; zotiraciclib, carrying a substituted phenyl, inhibits both. So the hinge was the first thing to change.

Screening hit, and a JAK2 inhibitor

Swapping the aniline for an sp3 aminopiperidine erased JAK2 and found CDK2.

Core scaffold, C5 trifluoromethyl

Property work at the pyrazole and the sulfonyl bought solubility without losing isoform selectivity.

Lead candidate, crystallised in 8UV0

Sulfonamide oxygens reach Asp86; the aminopyrimidine holds the hinge.

hingegatekeeper

Paper 2 / 4

Sheet17/31PaperCDK2Cpd17PDB8UV0

An sp3 swap erased JAK2; everything after was trade-offs.

+38×sp3 amine · CDK2

1→2

+5×CF3 · CDK2

2→3

+280×diol · SGF

8→9

+3×cyclopropyl · WB

9→17

Paper 2 / 4

Sheet18/31PaperCDK2Cpd·PDB·

The swap that changed target: The anilino methanesulfonamide of the hit was replaced by an sp3 1-(methylsulfonyl)piperidin-4-ylamine at the pyrimidine C2 position. Precedent in this hinge region (ebvaciclib against zotiraciclib, and an aminopiperidine analogue of milciclib) associates sp3 hybridisation here with general kinase selectivity, and that is how the authors rationalise the result.

Chlorine to trifluoromethyl at the gatekeeper: Exchanging the C5 chloro for a trifluoromethyl fills the space against the Phe80 gatekeeper. CDK2 potency improved five-fold and CDK1 selectivity went from 64-fold to 130-fold, which established C5 as the isoform selectivity lever for the whole series.

Fluorinated tail to geminal dimethyl alcohol: Exchanging the trifluoroethyl for a 2-hydroxy-2-methylpropyl group gives subnanomolar potency at 0.42 nM, 150-fold CDK1 selectivity, solubility of 842 ug/mL and clearance below 0.5 L/h/kg, at the cost of CDK6 potency at 160 nM.

The cyclopropyl that finished it: With the pyrazole alcohol of 9 fixed, a cyclopropyl on the sulfonyl gives 0.29 nM potency, 220-fold CDK1 selectivity, gastric solubility above 400 ug/mL, clearance of 0.5 L/h/kg and Caco-2 of 13. This is compound 17, the lead, and the ligand in 8UV0.

What follows in Tables 1 and 2 is not an optimisation so much as a negotiation. An isoelectronic N-methylpiperidine cleared 400 ug/mL of gastric solubility and dropped CDK1 margin from 130-fold to 29-fold. A nitrile lowered clearance and took solubility below 1 ug/mL. The one clean move was a nitrogen walk from N-methylpyrazole to N-methylimidazole on the sulfonyl group, which raised solubility from below 1 ug/mL to 322 ug/mL and cost nothing. Applying the C5 nitrile to that compound then undid the solubility again.

One trifluoromethyl at the gatekeeper and CDK1 falls away.

gatekeeperhinge

CDK2 vs CDK1 IC50, every compound (nM, log)

Paper 2 / 4

Sheet19/31PaperCDK2Cpd17PDB8UV0

Every polar contact PLIP measured

PDB	type	residue	Å	role
8UV0	hydrogen bond	Leu83 (bb)	2.73	protein_acceptor
8UV0	hydrogen bond	Leu83 (bb)	3.26	protein_donor
8UV0	hydrogen bond	Asp86 (bb)	3.04	protein_donor
8UV0	hydrogen bond	Lys89	3.43	protein_donor
8UV0	hydrogen bond	Asp145	3.66	protein_acceptor
8UV0	salt bridge	Asp86	4.96	protein_negative
8UV0	water bridge	Glu12	3.70	protein_donor
8UV0	water bridge	Gln85	3.06	protein_donor

KLIFS (ok, from 8UV0): gatekeeper Phe80; hinge Glu81, Phe82, Leu83; DFG in; αC out.

Potent, isoform-selective, soluble, and exposed by mouth.

0.29 nMCDK2 IC50

220×over CDK1

Rat oral bioavailability

Paper 2 / 4

Sheet20/31PaperCDK2Cpd17PDB·

Inhibit JAK2 and you take the blood counts with you.

JAK1 1.6 nMJAK2/JAK1 19×

JAK1 0.25 nM

JAK1 11 nMJAK2/JAK1 46×

My role: co-author

Paper 3 / 4

Sheet21/31PaperJAK1Cpd22PDB·

The approved dual JAK1/JAK2 inhibitors work in inflammatory disease, and their JAK2 activity brings anaemia and other cytopenias with it, because JAK2 carries the signals that maintain haematopoiesis. Autoimmune and inflammatory signalling runs predominantly through JAK1 heterodimers, so a JAK1-selective inhibitor should keep the benefit and leave blood cell production alone. The difficulty is that the four JAK kinases share a highly conserved ATP site, and the campaign's own starting candidate, itacitinib, managed only 21-fold.

The isopropyl amide that bought selectivity

An ortho-fluorine folds the amide into an internal hydrogen bond, and permeability and exposure follow.

Excellent in rat and monkey, hopeless in dog

Shrinking the hinge binder cost 44-fold of enzyme potency and fixed the dog.

Povorcitinib: 100 per cent oral in dog

Povorcitinib holds the JAK1 hinge with two hydrogen bonds.

hingegatekeeper

Numbering: JAK2 = JAK1 − 27 at hinge, − 26 at P-loop.

Paper 3 / 4

Sheet22/31PaperJAK1Cpd22PDB10PI

One fluorine folded the amide; a smaller hinge binder fixed the dog.

+2×ortho-F · Rat F

9→15

+37×CF3 amide · Cyno AUC

15→16

−1,256×pyrazole · JAK1

18→20

+100×methyl · Dog F

21→22

Paper 3 / 4

Sheet23/31PaperJAK1Cpd·PDB·

One fluorine, folded into a hydrogen bond: An ortho-fluorine lets the amide N-H fold back into an intramolecular hydrogen bond, masking the polar surface that was blocking permeability. Caco-2 rises three-fold, rat exposure four-fold and rat bioavailability two-fold. The tactic came from two earlier in-house programmes on entirely different targets.

The trifluoromethyl only works with the fluorine: Adding the trifluoromethyl to the ortho-fluorinated amide gave Caco-2 of 4.7, rat exposure of 28,766 and a 52-fold increase in cynomolgus exposure over compound 9. The same trifluoromethyl without the ortho-fluorine (14 against 9) did almost nothing, which is what makes this a real interaction between two changes rather than two independent gains.

Shrinking the hinge binder, and paying for it: The 7H-pyrrolo[2,3-d]pyrimidine is ubiquitous in kinase inhibitors and, being electron-rich, is prone to CYP-mediated oxidation. A bare pyrazole still binds the hinge, at 314 nM: a 1250-fold loss. The two pyrazole nitrogens make the same contacts, to the main-chain amide of Leu959 and the carbonyl of Glu957 (the paper calls these Leu932 and Glu930, using JAK2 numbering), but the smaller ring has poorer geometry and less surface to work with.

The second methyl, and the drug: The 3,5-dimethylpyrazole recovers 28-fold overall to 11 nM. Against 18 it is 44-fold weaker on the enzyme and nearly twice as isoform-selective, at 46-fold, with whole blood selectivity up 2.7-fold. The prize is elsewhere: dog clearance falls from 82 to 4 per cent of hepatic blood flow, and dog oral bioavailability rises from 1 per cent to 100 per cent.

An ortho-fluorine lets the amide N-H fold back into an intramolecular hydrogen bond, masking polar surface and lifting Caco-2 three-fold with a four-fold gain in rat exposure. Adding the trifluoromethyl on top gave compound 16: Caco-2 up five-fold and cynomolgus exposure up 52-fold against compound 9. The trifluoromethyl alone, without the ortho-fluorine, did essentially nothing, so the two changes only work together. The tactic was not new here: the same trick had produced an approved c-Met inhibitor and a clinical 11-beta-HSD1 inhibitor in earlier in-house programmes.

Same ligand, same hinge, in both isoforms.

JAK1 10PIJAK2 10PJ

JAK1 vs JAK2 IC50, every compound (nM, log)

Numbering: JAK2 = JAK1 − 27 at hinge, − 26 at P-loop.

Why selective: authors' P-loop hypothesis

Paper 3 / 4

Sheet24/31PaperJAK1Cpd22PDB10PI + 10PJ

The pyrrolo[2,3-d]pyrimidine hinge binder is electron-rich and prone to CYP oxidation, so it was replaced by a 3,5-dimethylpyrazole: 44-fold weaker on the enzyme, nearly twice as isoform-selective at 46-fold, and in beagle dogs clearance fell from 82 to 4 per cent of hepatic blood flow with oral bioavailability rising from 1 to 100 per cent. Povorcitinib was then crystallised in both JAK1 and JAK2, the only true twin structure in this set. The hinge contacts are identical in both: Glu957 and Leu959 in JAK1, the same two positions numbered Glu930 and Leu932 in JAK2, which is the pair of numbers the paper uses for both. The authors attribute the selectivity to JAK1's P-loop sitting about 2 A higher, opening a lipophilic groove for the trifluoromethyl, with JAK2's Asn859, Phe860 and Glu1015 closing that space, and they state plainly that this is speculative.

The authors attribute JAK1 selectivity to a P-loop sitting about 2 A higher in JAK1, opening a lipophilic groove for the trifluoromethyl, and say plainly this is speculative. PLIP agrees on the hinge: Glu957/Leu959 in JAK1 at 2.8/2.9 A, Glu930/Leu932 in JAK2 at 2.9/3.0 A (after the pyrazole tautomer fix, DECISIONS D5).

Every polar contact PLIP measured

PDB	type	residue	Å	role
10PI	hydrogen bond	Gly884 (bb)	3.46	protein_donor
10PI	hydrogen bond	Gly887 (bb)	3.44	protein_donor
10PI	hydrogen bond	Glu957 (bb)	2.81	protein_acceptor
10PI	hydrogen bond	Leu959 (bb)	2.93	protein_donor
10PI	water bridge	His885	3.48	protein_donor
10PJ	hydrogen bond	Phe860 (bb)	2.82	protein_donor
10PJ	hydrogen bond	Gly861 (bb)	2.71	protein_donor
10PJ	hydrogen bond	Glu930 (bb)	2.91	protein_acceptor
10PJ	hydrogen bond	Leu932 (bb)	2.96	protein_donor
10PJ	water bridge	Asp994	4.04	protein_acceptor

KLIFS (fallback_by_uniprot, from 3EYH (KLIFS structure 1462, same UniProt P23458)): gatekeeper Met956; hinge Glu957, Phe958, Leu959; DFG in; αC in.

Forty-four-fold weaker on the enzyme, fully absorbed in dog.

11 nMJAK1 IC50

46×over JAK2

Dog oral bioavailability

Paper 3 / 4

Sheet25/31PaperJAK1Cpd22PDB·

Aspartate is not cysteine: no warhead, so shape must do it.

Exchange 8 µM

SPR <1 pM

NHP F 42%

My role: co-crystal structures 9E5F and 9E5D, co-author

Paper 4 / 4

Sheet26/31PaperKRAS G12DCpd23PDB·

KRAS G12C fell to covalent drugs: a warhead reacts with the mutant cysteine and the inhibitor never leaves. G12D is the commoner mutation, 37 per cent of pancreatic tumours, and the swap from cysteine to aspartate removes that option. Aspartate is a far weaker nucleophile, and making a warhead reactive enough to compensate raises its own selectivity and safety problems. So the inhibitor has to be reversible, and it has to distinguish a mutant from a wild-type protein present in every healthy cell. The only starting point was a patent compound with its warhead removed: 8 uM.

A patent hit with its warhead removed

A rigid bicyclic amine, then a two-carbon linker reaching the Asp12 backbone, took it from micromolar to picomolar.

Picomolar, and it cannot leave the gut

A thousand-fold of binding potency was traded away, deliberately, for rigidity and balanced polarity.

INCB159020: oral exposure in non-human primate

A rigid cage holds the amine where Tyr96 can reach it.

P-loopswitch Iswitch IIG12D

No contact with the Asp12 side chain

Paper 4 / 4

Sheet27/31PaperKRAS G12DCpd3PDB9E5F

It grips the backbone beside Asp12, never the side chain.

5a · 9E5D3 · 9E5FG12D

55×over wild-type KRAS

Paper 4 / 4

Sheet28/31PaperKRAS G12DCpd5aPDB9E5D + 9E5F

Every polar contact PLIP measured

PDB	type	residue	Å	role
9E5F	hydrogen bond	Gly10 (bb)	2.96	protein_acceptor
9E5F	hydrogen bond	Glu63 (bb)	3.82	protein_acceptor
9E5F	hydrogen bond	Tyr96	2.77	protein_donor
9E5F	hydrogen bond	Arg102	3.69	protein_donor
9E5F	salt bridge	Glu62	2.87	protein_negative
9E5F	water bridge	Gly10	2.85	protein_donor
9E5F	water bridge	Asp12	3.72	protein_donor
9E5F	hydrogen bond	Gly10 (bb)	2.96	protein_acceptor
9E5F	hydrogen bond	Glu63 (bb)	3.82	protein_acceptor
9E5F	hydrogen bond	Tyr96	2.77	protein_donor
9E5F	hydrogen bond	Arg102	3.69	protein_donor
9E5F	salt bridge	Glu62	2.87	protein_negative
9E5F	water bridge	Gly10	2.85	protein_donor
9E5F	water bridge	Asp12	3.72	protein_donor

KRAS motifs from UniProt P01116 and Pfam PF00071: P-loop 10-17, switch I 30-38, switch II 60-76, G12D, Gly13, Gln61.

SPR discrepancy, compound 23: graphical abstract 2.2 nM, Table 3 22 nM. Compound 10 whole-blood pERK: graphical abstract 451 nM, Table 1 410 nM. Table values shown throughout; neither averaged.

A thousand-fold weaker, and finally absorbed by mouth.

22 nMG12D SPR

331 nMwhole-blood pERK

Monkey oral bioavailability

Table values shown; graphical abstract prints 2.2 nM (T3: 22) and 451 nM (T1: 410).

Paper 4 / 4

Sheet29/31PaperKRAS G12DCpd23PDB·