Not in this copy. Fetch them from podium.mdeller.com.
Marc C. Deller, D.Phil.
The pathNine stops in twenty years, from oncostatin M at Oxford to Incyte's platform.
The threadBuild what makes structures arrive faster, then use them to choose what to make.
The problemApproved pan-FGFR drugs also hit FGFR1, and collapse on the gatekeeper mutants.
The resultA chemotype that avoids both: 42-fold over FGFR1, mutants still inhibited.
The problemA patent deadline fourteen days away, needing a bound conformation to claim.
The resultThe Fab complex solved in ten days. First-in-class, and no public coordinates.
The problemThe screening hit was a better JAK2 inhibitor, and CDK1 is pan-lethal.
The resultOne sp3 swap at the hinge changed the target; 220-fold over CDK1.
The problemG12D has no cysteine, so the covalent trick that beat G12C is unavailable.
The resultA rigid cage binding reversibly beside switch II: 55-fold over wild type.
The problemJAK2 carries the signals that keep red cells and platelets coming.
The resultPovorcitinib: 11 nM on JAK1 against 505 nM on JAK2.
Anti-targetsEach drug is defined by the protein it must spare: FGFR1, CDK1, wild-type KRAS, JAK2.
Measured contactsEvery bond on screen is one PLIP measured in the deposited crystal structure, never drawn by hand.
Pocket motifsKLIFS regions for the kinases and the KRAS switches show where selectivity is won.
Property trade-offsPotency, permeability, metabolism and oral exposure, traded one edit at a time.
SpeedAn automated platform: 100+ co-crystal structures in under three years; a patent structure in ten days.
Twenty yearsFrom cytokine receptors at Oxford to kinase and KRAS drugs at Incyte, and now enzymes at Elora.
On the filmRings, and the shadow of the beam stop.
Why it mattersIt looks like nothing. Every structure here starts as something that does.
What he readRepeat distances in angstroms, and a different set once the fibre was stretched.
Why it mattersA shape, a change, and a measurement. The whole job, stated in 1931.
What changedFibres became crystals and film became detectors. The question did not move.
Why it mattersThe rest of this deck asks it again: what shape, and what binds there?
The press got the other half of it: the same man spun groundnut protein into a fibre and wore the coat to the BBC. A structure you can read is a material you can make.




What the papers hadHe talked on the BBC about the fibres of the future and left a groundnut-protein overcoat in the cloakroom. He had worn it two years.
Where it is written downKersten T. Hall, The Man in the Monkeynut Coat, Oxford University Press: the Astbury biography, and the book the coat gets its name from.
Why it mattersSame science as the film beside it: read a protein well enough and it becomes a material.









Where fromJCSG structural genomics, then co-crystals at Agouron and Incyte.
SourceScopus, across structural biology, structural genomics and medicinal chemistry.
SourceScopus, over twenty years of published structures and methods.
What it meansTwenty-five papers, each cited at least twenty-five times.
FiledIncluding the mutant CALR antibody, whose structure made the claim specific.
What it meansProgrammes that reached the point of filing to test in humans.
The span1991 to now: cytokine receptors at Oxford, to kinase and KRAS drugs.
gp130 · OSMRWhat it isOncostatin M and the receptor chains it signals through. Structure, 2000.
Why it mattersThe pathway starts at this receptor, and so did I.

Env · E2 HIV · TopoWhat it isA cleaved HIV-1 Env trimer, Science 2013, the gp120 glycan supersite, and HCV E2 with AP33.
Why it mattersAn antibody stops the signal before the membrane. HIV and topoisomerase at Agouron and Pfizer came first.
mutCALRWhat it isThe Fv on the mutant calreticulin neoepitope, solved in ten days against a fourteen-day deadline.
Why it mattersFirst in class, and the filing could claim a bound conformation. WO 2023/107994.
FGFR1-4What it isA pan-FGFR chemotype that spares FGFR1 and still holds the gatekeeper mutants.
Why it matters42-fold over FGFR1, where the approved drugs collapse on the mutants.
PETaseWhat it isPETase engineered at Elora to break down microplastics in the body.
Why it mattersThe protein is the drug rather than the target. No paper yet.
GLP1RWhat it isAI-driven design against GLP-1R with DeepCovalent: docking, co-folding, FEP, MD.
Why it mattersThe structures are there to check the predictions. Current, unpublished.

JAK · STATWhat it isThe pathway's kinases twice over: JAK and STAT biology at Yale, then povorcitinib at Incyte.
Why it matters11 nM on JAK1 against 505 on JAK2, because JAK2 carries the red cells and the platelets.
ChEM-HWhat it isThree targets from the Stanford ChEM-H service: DYRK1, a polyketide synthase, STING.
Why it mattersNothing alike. The service existed so groups with no crystallography could get a structure.

Grb2 · SosWhat it isNine years at JCSG, 70+ novel structures, then a lab built from nothing at Stanford.
Why it mattersThe adaptors carry the signal sideways, and the service solved structures for seventy groups.
CDK2What it isOne sp3 swap at the hinge moved a screening hit off JAK2 and onto CDK2.
Why it matters220-fold over CDK1, which is pan-lethal and the reason it needed the margin.

KRAS · ERKWhat it isA rigid cage binding reversibly beside switch II of KRAS G12D, 55-fold over wild type.
Why it mattersG12D has no cysteine, so the covalent trick that beat G12C is unavailable.
NANOGWhat it isThe NANOG homeodomain on the OCT4 promoter, PDB 4RBO, and the L122A variant it gave.
Why it mattersThe end of the pathway, and the start of a cell deciding what to be.
Incyte platform, before and after
What it wasDesigned and built at Incyte, 2018 to 2022.
What it countsDistinct proteins taken from construct design through to a solved structure.
What it countsAntibodies and antibody-antigen complexes, the least obliging crystals in the set.
What it countsDistinct collaborators across the whole publication record.
What it countsDrug discovery programmes the Incyte platform served in under three years.
Who they wereUsers from ten countries on the structure service founded at Stanford.
What they areLive scientific web apps built with AI assistance, this deck among them.
What it isHuman KRAS isoform 2B, residues 1 to 169 with the G12D mutation, codon optimised and cloned into pET3a+ behind a 6xHis tag and a TEV site.
Why it mattersNothing downstream fixes a bad construct. Stopping at 169 drops the disordered tail that would never crystallise, and the TEV site means the tag can come off before it does.
What it isE. coli BL21(DE3) carrying that plasmid, grown overnight at 37 °C in LB with magnesium and carbenicillin, and never induced.
Why it mattersNo induction at all. The promoter leaks just enough, and letting it trickle overnight beats forcing it, which for this construct gives inclusion bodies instead of protein.
What it isNi-NTA resin holds the tag while the rest washes off at 10 mM imidazole, then 250 mM lifts it. TEV protease cuts the tag off overnight at 4 °C.
Why it mattersThe clever part is the second nickel pass, which now catches the tag and lets the protein through. Then sizing in GDP and magnesium, because the switch loops are only ordered when the nucleotide is there.
What it isA Formulatrix NT8 sets 200 nL of protein at 25 mg/mL against 200 nL of reservoir, by vapour diffusion at 20 °C, and the plates sit for five days.
Why it mattersNucleation cannot be forced, so you buy attempts instead. This crystal came from one well of a sparse-matrix screen: ammonium citrate, 20 per cent PEG 3350 and the compound itself at 1 mM.
What it isKRAS grown from one of those drops, caught in a nylon loop and cooled to 100 K, turning on the spindle the way it does while the data are collected.
Why it mattersOne crystal is enough. Everything the rest of this deck claims about the pocket comes from objects this size.
What it isEach spot is a reflection from a plane of atoms repeated through the crystal. The crystal turns while the detector counts, so the reflections arrive in crescent-shaped lunes. Their spacing gives the cell, the outer edge gives the resolution.
The catchA spot says how strong a wave is and nothing about where its crest falls. That missing half is the phase problem, and it is why this takes a specialist.
What it is2Fo-Fc electron density at 1.5 sigma, cut around the inhibitor. The phases came from an earlier KRAS structure with its ligand deleted, so nothing here was assumed about this compound.
Why it mattersAt 1.36 Å with R-free 0.193 the atoms are separated. You are not inferring where the molecule sits; you are looking at it.
What it isCompound 5a, ligand code A1BEI, placed atom by atom into that density, with GDP and its magnesium still in the site next door.
What it buysEvery contact the rest of this deck draws comes from a map like this one. That is the whole argument: measured, not modelled.
Protein folding, calcium, cell adhesion and integrin signalling.
DEEQRTRRMMRTKMRMRRMRR: eleven arginines in twenty-one residues, found nowhere healthy.
The second most common driver after JAK2, the pathway three sheets back.
A new C-terminus is a surface, not a cleft. Nothing for a small molecule to bind.
My partThe structure of the Fv on the neoepitope, so the filing could describe a bound conformation.
Why the sprintA 14-day deadline, solved in just over 10. A projection, not revenue.
A hydrogen bond to a backbone carbonyl, not a side chain.
A salt bridge into heavy chain CDR1.
Hydrogen bond and salt bridge, from heavy chain CDR3.
The shortest contact here, and its only one to the light chain.
The other half of the clamp, from the heavy chain.
The peptide's last residue, held by light chain CDR3.
Six residues read, five of them arginines the frameshift made.
BiologyFGFR2/3 fusions and mutations drive bladder cancer and cholangiocarcinoma.
LandscapeApproved pan-FGFR drugs (erdafitinib, pemigatinib, futibatinib) also hit FGFR1 and lose to gatekeeper mutants.
BiologyCyclin E1-amplified ovarian cancer, and breast cancers escaping CDK4/6 drugs, lean on CDK2.
LandscapeCDK4/6 inhibitors are approved; CDK2-selective inhibitors are now in the clinic.
BiologyG12D is the commonest KRAS mutation, dominant in pancreatic cancer.
LandscapeCovalent G12C drugs are approved; aspartate cannot be hit that way, so G12D needs reversible binders.
BiologyJAK1 carries inflammatory cytokine signals; JAK2 carries erythropoietin and thrombopoietin.
LandscapeJAK1/2 and JAK1-selective drugs are approved; povorcitinib is in late-stage trials.
Upregulation of the fibroblast growth factor receptor (FGFR) signaling pathway has been implicated in multiple cancer types, including cholangiocarcinoma and bladder cancer. Consequently, small molecule inhibition of FGFR has emerged as a promising therapy for patients suffering from these diseases. First-generation pan-FGFR inhibitors, while highly effective, suffer from several drawbacks. These include treatment-related hyperphosphatemia and significant loss of potency for the mutant kinases. Herein, we present the discovery and optimization of novel FGFR2/3 inhibitors that largely maintain potency for the common gatekeeper mutants and have excellent selectivity over FGFR1. A combination of meticulous structure-activity relationship (SAR) analysis, structure-based drug design, and medicinal chemistry rationale ultimately led to compound 29, a potent and selective FGFR2/3 inhibitor with excellent in vitro absorption, distribution, metabolism, excretion (ADME), and pharmacokinetics in rat. A pharmacodynamic study of a closely related compound established that maximum inhibition of downstream ERK phosphorylation could be achieved with no significant effect on serum phosphate levels relative to vehicle.
The paperAn inhibitor that still works on the gatekeeper mutants that defeat the approved pan-FGFR drugs, while sparing FGFR1.
My roleco-crystal structure 8E1X, co-author
BiologyFGFR2/3 fusions and mutations drive bladder cancer and cholangiocarcinoma.
LandscapeApproved pan-FGFR drugs (erdafitinib, pemigatinib, futibatinib) also hit FGFR1 and lose to gatekeeper mutants.
A hydrogen bond drawn on a model, then measured in the crystal.
The anchor nearly every ATP-competitive inhibitor makes. Generic, so it buys no selectivity.
Val555 in FGFR3, the resistance site. This chemotype never reaches past it.
DFG-in, aC-in. A type I binder competing directly with ATP.
In, with DFG in. No type II pocket is opened or used here.
Docked, built on purpose, then measured at 3.0 A. FGFR1 carries the same asparagine.
A contact through an ordered water. PLIP's own category, drawn dotted.
Most of the buried surface, almost none of the selectivity. Shape, not a bond.
Designed, built and measured at 3.0 A, and still not the reason selectivity improved.
DesignDocking put the isopropyl group beside Asn562, so the ether was cyclised to deliver an oxygen there and engage the asparagine N-H. The one contact here that was predicted before it was measured.
Why it matters49-fold on FGFR3 and FGFR1 selectivity up to 65-fold, from one ring closure. The crystal structure later confirmed the bond, at 3.0 Å to the equivalent asparagine in FGFR2.
ProblemThe polar core that spared CYP3A4 would not cross a membrane, so cell potency lagged far behind the enzyme. The two requirements pulled against each other.
Why it mattersN-isopropyl on the pyrazole lifted Caco-2 sevenfold, and moving the bridgehead nitrogen took cell potency from 969 to 29 nM. Permeability was the gate, not affinity.
CauseThe CYP3A4 liability was traced to N-demethylation of the amide, so the N-methyl became N-CD3. A carbon-deuterium bond is harder to break, which slows that one step.
Why it mattersInhibition moved past 25 µM and human clearance halved, with potency held at 0.5 nM. A three-atom change that cost nothing on target.
Why it mattersAn FGFR inhibitor is taken by mouth for years, so bioavailability decides whether any of the rest counts.
Result82 per cent in rat, and serum phosphate stayed flat: the FGFR1-driven hyperphosphataemia that limits the approved drugs did not appear.
FGFR3 vs FGFR1 IC50, every compound (nM)
Cell potency along the series (BaF3, nM)
Four edits built the potency; the cells followed.
What changedA methoxy for fifteen-fold potency
Why it mattersAdding 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.
What changedCyclised onto an oxetane
Why it mattersDocking 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.
What changedThe active enantiomer
Why it mattersThe 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.
What changedBoth trends on one ring
Why it mattersCombining the meta methyl with the para fluorine gave compound 23 at 0.7 nM with more than 100-fold selectivity over all four problem kinases: KIT 1657 nM and PDGFRB 1386 nM.
CDK2 is a critical regulator of the cell cycle. For a variety of human cancers, the dysregulation of CDK2/cyclin E1 can lead to tumor growth and proliferation. Historically, early efforts to develop CDK2 inhibitors with clinical applications proved unsuccessful due to challenges in achieving selectivity over off-target CDK isoforms with associated toxicity. In this report, we describe the discovery of (4-pyrazolyl)-2-aminopyrimidines as a potent class of CDK2 inhibitors that display selectivity over CDKs 1, 4, 6, 7, and 9. SAR studies led to the identification of compound 17, a kinase selective and highly potent CDK2 inhibitor (IC50 = 0.29 nM). The evaluation of 17 in CCNE1-amplified mouse models shows the pharmacodynamic inhibition of CDK2, measured by reduced Rb phosphorylation, and antitumor activity.
The paperA screening hit that was really a JAK2 inhibitor, turned into a CDK2 inhibitor selective over five other CDKs.
My roleco-crystal structure 8UV0, co-author
BiologyCyclin E1-amplified ovarian cancer, and breast cancers escaping CDK4/6 drugs, lean on CDK2.
LandscapeCDK4/6 inhibitors are approved; CDK2-selective inhibitors are now in the clinic.
Sulfonamide oxygens reach Asp86; the aminopyrimidine holds the hinge.
A two-point clamp on the hinge. Kept unchanged across all seventeen compounds.
The CF3 packs here. That one swap took CDK1 selectivity 64- to 130-fold.
DFG-in, so this competes with ATP rather than binding an inactive form.
Out, because CDK2 needs cyclin to pull it in. The protein's doing, not the ligand's.
What the sp3 piperidine bought: 607 to 16 nM, and JAK2 activity gone.
Lys89 direct, Gln85 only through a water. Real, and the first to go.
Gln85 is reached only through a water, which PLIP reports separately.
Most of the buried surface, almost none of the selectivity. Shape, not a bond.
A piperidine moved the target and a CF3 bought the margin. Both explained after the fact.
What changedThe swap that changed target
Why it mattersThe 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.
What changedChlorine to trifluoromethyl at the gatekeeper
Why it mattersExchanging 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.
What changedFluorinated tail to geminal dimethyl alcohol
Why it mattersExchanging 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.
What changedThe cyclopropyl that finished it
Why it mattersWith 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.
CDK2 vs CDK1 IC50, every compound (nM)
Solubility against selectivity, every compound
Every fix traded solubility against selectivity; 17 won both.
The start607 nanomolar on CDK2, read off the paper's Figure 2, and the only number the paper gives this compound.
What came nextThree orders of magnitude of potency, after which the whole argument moves to the margin over CDK1.
A thousand-fold0.42 nanomolar on CDK2 and 150-fold over CDK1, with solubility in simulated gastric fluid at 842 micrograms per millilitre.
Still to doSelectivity over CDK1 is what the rest of the campaign buys, and what the trifluoromethyl at the gatekeeper is for.
The candidate0.29 nanomolar and 220-fold over CDK1, 51 per cent orally bioavailable in rat.
What carried overThis is the compound the follow-up paper on sheet 14 begins from, where replacing its pyrazole takes the CDK1 margin past 500-fold.
CDK2 is a critical regulator of cell-cycle progression and an attractive therapeutic target for cancers characterized by dysregulated cyclin E1 signaling. Despite this biological rationale, CDK2 has historically been challenging to target due to the difficulty of achieving selectivity over related CDK isoforms, particularly CDK1, where off-target inhibition may limit the therapeutic window. Building on our previously reported (4-pyrazolyl)-2-aminopyrimidine series, we pursued a multiparameter optimization strategy aimed at improving selectivity for CDK2 over CDK1 while balancing pharmacokinetic and physicochemical properties. Computationally guided replacement of the pyrazole with an imidazole led to a (4-imidazolyl)-2-aminopyrimidine series and culminated in the discovery of INCB127443, a potent and orally bioavailable CDK2 inhibitor.
The paperThe follow-up to the pyrazole series: swapping one ring collapsed the molecule's torsional choices, and the crystal structure found it bound in the shape that was predicted.
My roleco-author; the series was designed from 8UV0, which I solved
CalculatedA torsional scan put the pyrazole in two low-energy wells and the imidazole in one, at 180 degrees. The imidazole's nitrogen sits beside the pyrimidine's and the two repel; the pyrazole has a carbon there and pays nothing.
Measured13GK finds the bond at 174 degrees, read off the coordinates rather than off a figure. The molecule arrives already in the shape it was designed to prefer.
Selectivity over CDK1, five matched pairs
Three hours on CDK2, fifty-three seconds on CDK1.
What changedOne ring, and nothing else on the molecule. The imidazole has a nitrogen where the pyrazole has a carbon, which raises one of the two conformations the pyrazole could adopt and leaves a single low-energy shape.
Why it mattersSelectivity over CDK1 went from 221-fold to 568-fold on this pair. The ligand now arrives folded the way the pocket wants rather than paying to fold once it is there.
What changedThe tertiary alcohol on the imidazole nitrogen was replaced by a 2-chloro-3-cyanophenyl ring. Removing the ortho chlorine cost both potency and selectivity, so the twist that substituent forces is doing real work.
Why it mattersCaco-2 went from 1 to 9.5 and CDK2 from 0.84 to 0.28 nanomolar at the same time. Solubility stayed the problem: FaSSIF was still 76 micrograms per millilitre.
What changedThe nitrile became a benzylamine and then an azetidine: more sp3 character, a flatter molecule made less flat, and crystal packing disrupted on purpose.
Why it mattersSolubility went from 76 to 895 micrograms per millilitre with CDK2 potency and the CDK1 margin held. This is the compound in the crystal structure, INCB125496.
What changedOne fluorine on the aryl ring, between the chlorine and the azetidine. Everything else is INCB125496.
Why it mattersSolubility past 1000 micrograms per millilitre, Caco-2 8, 94 per cent orally bioavailable in rat and 64 in monkey, with a half-life of 10 hours and a flat profile. CDK2 at 0.99 nanomolar, 316-fold over CDK1.
The inhibition of mutant KRAS proteins has emerged as a promising approach for treating KRAS-driven cancers, as evidenced by the clinical success of KRAS G12C inhibitors. KRAS G12D, the most common mutant, promises significant expansion of the addressable patient population; however, the reduced nucleophilicity of aspartate compared to cysteine poses significant challenges in balancing sufficient potency with ADME properties to support oral exposure. Herein, we describe the discovery of KRAS G12D inhibitor 23 (INCB159020), which achieves oral exposure in nonhuman primate (NHP). Starting from a weakly potent hit, structure-based drug design was utilized to drive significant potency. Focus on molecular rigidity and balanced polarity then allowed for successful optimization of properties required for oral exposure.
The paperAspartate is not cysteine, so the covalent trick that beat KRAS G12C does not work here. This one binds without a warhead.
My roleco-crystal structures 9E5F and 9E5D, co-author
BiologyG12D is the commonest KRAS mutation, dominant in pancreatic cancer.
LandscapeCovalent G12C drugs are approved; aspartate cannot be hit that way, so G12D needs reversible binders.
A rigid cage holds the amine where Tyr96 can reach it.
Bridged to the P-loop through one water, not bonded directly to it.
The effector region. A KRAS drug must not simply block it.
Aspartate cannot be alkylated, so this had to be reversible and shape-driven.
Backbone N-H through a water. The mutated side chain is never touched.
Drawn, and never touched by the inhibitor.
The shortest contact here at 2.8 A, and what the rigid cage exists to position.
Two of them, in a pocket with very few contacts.
Most of the buried surface, almost none of the selectivity.
One strong bond and a rigid cage, with the mutated side chain never touched.
What changedA screen of rigid amines gave compound 3 at 671 nM, and its structure at 1.35 Å explains why nearly every replacement for that group failed.
Why it mattersThe azabicyclohexane holds its protonated amine directly below the plane of the core. That one conformation is what the rest of the campaign was built around, so the scaffold was never really open to change.
What changedGrowing a two-carbon linker from the imidazole 2-position towards Asp12 bought a fifty-fold gain, and the structure of 5a shows how.
Why it mattersThe amide carbonyl takes the backbone N-H beside Asp12, not the aspartate side chain. The gain is real and the mechanism is not the one the name suggests.
What changedBelow 1 pM in the binding assay, and unusable: no measurable Caco-2 flux at all, clearance above 1.1 L/h/kg, and four separate liabilities.
Why it mattersEvery liability was worth more than tenfold of potency, and removing any one of them cost 20-fold or more. Potency had been bought with the properties needed to make a drug.
What changedCompound 23 binds about a thousand times more weakly than compound 10 and is the one that went forward, at 42 per cent oral bioavailability in primate.
Why it mattersThis is the whole argument of the slide. Shape had to do the work a warhead does for G12C, and the campaign's decisive moves were giving potency back to get a drug.
Whole-blood potency along the series (nM)
Permeability along the series (Caco-2)
A thousand-fold weaker, and finally absorbed by mouth.
The start8000 nanomolar in the nucleotide exchange assay, off the paper's Figure 2.
The targetKRAS G12D, where the aspartate the mutation installs is the thing to reach and the thing every later compound is shaped around.
Four thousand-fold2.1 nanomolar in exchange, under the SPR assay's own floor, and 410 nanomolar on pERK in whole blood.
The gapPotency in hand, exposure not. Everything after this is permeability and clearance rather than affinity.
The candidateINCB159020. Whole-blood pERK at 331 nanomolar, Caco-2 3.6, and 42 per cent orally bioavailable in monkey.
What carried overPredicted human clearance of 1 litre per hour per kilogram, which is the number the paper on sheet 17 sets out to halve.
The discovery of the clinical KRAS G12D inhibitor INCB161734 is described. Starting from a previously disclosed orally bioavailable tool compound 1, a concerted effort to improve the predicted human dose via modulation of potency and pharmacokinetics was undertaken. Careful consideration of the metabolic profile was key to designing low clearance compounds. In addition, the molecular size and fraction sp3 character were carefully controlled to improve permeability and crystalline solubility. From this effort, INCB161734 was identified as a balanced compound with excellent oral exposure, potent antiproliferative activity in xenograft tumor models, and excellent selectivity against wild-type K/H/NRAS. INCB161734 is currently undergoing clinical trials in patients with KRAS G12D mutations.
The paperA clearance campaign, not a potency one: blocking one metabolic route opened another, and the structure showed that a ring in the right place was what closed it.
My roleco-author
The shuntOxidation of the morpholinone was the main clearance route, so it was designed out. The metabolism moved to the caged bicyclic amine and cleaved that instead, and it tolerates almost no substitution.
The structureNo dealkylative cleavage was seen in the pyrrolidine series at all. The structure shows the ring sitting close to the C-H beside that nitrogen, occluding the CYP, which the authors say potentially rather than certainly.
Permeability against clearance, every compound
One residue carries the whole RAS family.
Where it beganThis paper's compound 1 is the compound the last two sheets end on: compound 23 of the first KRAS paper, with moderate oral exposure in monkey and a projected human dose in grams twice a day.
The problemNot potency. Oxidation of the morpholinone was the main clearance route, and across 79 profiled compounds almost everything that cleared slowly was too polar to cross a membrane.
What changedDesigning out the morpholinone oxidation moved the metabolism to the caged bicyclic amine, which was cleaved instead and tolerates almost no substitution. Replacing the morpholinone with a pyrrolidine restored sp3 character.
Why it mattersNo dealkylative cleavage of the caged amine was seen in this series at all. The paper reads the shunt as operating wherever the mutation region is less hindered, which turns a metabolism result into a design rule.
What changedA single methyl, placed to raise the energy of the radical the CYP has to form next to the ring nitrogen.
Why it mattersPredicted human clearance crossed the programme's objective, from above 1.1 to below 0.5 litres per hour per kilogram, on one carbon. Whole-blood potency was the price: 516 to 1725 nanomolar.
What changedThe methyl became a cyclopropane fused across the pyrrolidine, which adds the same steric block with more strain and less flexibility.
Why it mattersClearance stayed below 0.5 and the potency came back: whole-blood pERK 364 nanomolar, Caco-2 2.3, 76 per cent orally bioavailable in monkey. It is in clinical trials.
Aberrant JAK2 signaling is a key driver in cancers such as myeloproliferative neoplasms (MPNs), whereas JAK1 acts as a central mediator of autoimmune and inflammatory diseases. Since JAK2 supports blood cell production, its inhibition can lead to anemia and other cytopenias, making selective JAK1 inhibition a potentially safer therapeutic strategy opposed to dual JAK1/JAK2 inhibition. We initially identified novel scaffold 9, which selectively inhibited JAK1 over JAK2. Iterative breakthroughs led to the discovery of povorcitinib (INCB054707, 22), which showed improved pharmacokinetics due to intramolecular hydrogen bonding of the amide moiety and a reduced ring size in the hinge binding motif. Protein crystallography suggested that JAK1 selectivity was conferred by the amide's alkyl group accessing a small lipophilic groove formed by the P-loop conformation of JAK1 that was distinct from JAK2. In murine arthritis models, povorcitinib demonstrated dose-dependent efficacy and reduced inflammatory signaling, supporting its therapeutic potential in immunological disorders.
The paperSparing JAK2 is the whole point: inhibit it and you take the patient's blood counts with you.
My roleco-author
BiologyJAK1 carries inflammatory cytokine signals; JAK2 carries erythropoietin and thrombopoietin.
LandscapeJAK1/2 and JAK1-selective drugs are approved; povorcitinib is in late-stage trials.
Povorcitinib holds the JAK1 hinge with two hydrogen bonds.
Two hinge bonds at 2.8 and 2.9 A. PLIP missed one until the tautomer was fixed.
Methionine in JAK1 and in JAK2, with the ligand sitting the same way in both.
Two weak backbone contacts under the most mobile part of the pocket.
One bridging water, and nothing here that explains the isoform margin.
Resolved and real, but waters rarely survive a change of ligand or crystal.
Most of the buried surface, almost none of the selectivity. Shape, not a bond.
A clean, ordinary type I mode. The 46-fold margin is not visible here.
What changedOne fluorine, folded into a hydrogen bond
Why it mattersAn 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.
What changedThe trifluoromethyl only works with the fluorine
Why it mattersAdding 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.
What changedShrinking the hinge binder, and paying for it
Why it mattersThe 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.
What changedThe second methyl, and the drug
Why it mattersThe 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.
JAK1 vs JAK2 IC50, every compound (nM)
Oral bioavailability in three species
Selectivity from a P-loop 2 Å higher; forty-four-fold on the enzyme bought full absorption.
The startJAK1 at 1.5 nanomolar and already 21-fold over JAK2, so the isoform margin came with the chemotype rather than being designed into it.
The problemNot potency. Caco-2 was 0.9, and oral bioavailability 6 per cent in dog and 9 in monkey. Everything after this is an absorption campaign.
The most potent0.25 nanomolar on JAK1, six times compound 9, with projected human clearance under 0.5 and bioavailability over 60 per cent.
What it did not fixCaco-2 only 1.4 and monkey still 28 per cent, and the whole-blood window narrowed: TPO over IL-6 fell from 11-fold to 5.
The candidatePovorcitinib. Fully absorbed in rat, dog and monkey, Caco-2 4.5, the whole-blood window back out to 16-fold and hERG over 20 micromolar.
The tradeForty-four-fold weaker than 19 on the enzyme at 11 nanomolar, and 46-fold over JAK2. The enzyme number was the wrong one to optimise.
How it was wonA chemotype that does not reach past the gatekeeper, so the V555M and V555L mutants stay inhibited where the approved drugs lose 35- to 220-fold.
What it costThe designed asparagine bond does not explain it: FGFR1 has the same asparagine. Earned empirically, and the paper says so.
How it was wonAn sp3 piperidine at the hinge removed JAK2 entirely; a trifluoromethyl at C5 took the CDK1 margin from 64-fold to 220.
What it costCDK6 was given up at 160 nM to buy solubility and clearance. Every fix here traded one property for another.
How it was wonA rigid cage shaped to the mutant pocket, binding reversibly: aspartate cannot be alkylated the way G12C drugs alkylate cysteine.
What it costMeasured by SPR, and no contact with the Asp12 side chain. It is shape, not recognition of the mutation.
How it was won11 nM on JAK1 against 505 on JAK2. Sparing JAK2 keeps red cells and platelets, the toxicity that limits the JAK1/2 drugs.
What it costNothing in the pocket explains it: the same ligand binds JAK2 the same way, so the margin sits outside the contacts drawn here.
How it was wonA frameshift writes a new C-terminus, twenty-one residues found in patients and nowhere healthy. An antibody reads a surface; a small molecule has no pocket to bind.
What it costThe epitope is the whole target, so there is nothing to optimise for potency. The structure had to be solved inside the 14-day filing deadline, and was.
Why it worksA selectivity margin is a difference between two pockets. Measuring it tells you how big it is; seeing it tells you which atoms to keep.
What it does not doThree of these five margins are not explained by the contacts drawn. The structure says where to look, and says plainly when the answer is not there.
Co-founder and CSO
Engineered enzyme medicines for microplastic-associated disease.
What I ownI co-founded it and I run the science: which enzyme, whether it survives serum, and how any of it gets measured. The whole published PETase field optimises above PET's glass transition at about 70 degrees, and a medicine has to work at 37 in blood.
What it rests onA catalogue of 1,348 characterised hydrolases with 150 that were expressed, assayed and released nothing. That measured negative class is what showed the field's best structural predictor to be an artefact of which enzymes get into a database, and it is why we are not engineering a thermostable industrial enzyme. It is early validation: there is no in vivo data, and no claim that microplastics cause the disease.
Scientific technical adviser
De novo small molecules designed in a single pass.
What I ownThe physics and the experiment: structure-based pipelines, in silico screening, the validation strategy, and the CROs who have to reproduce any of it. Throughput up by half.
What it rests onA generative engine is worth what the thing that can refuse it is worth. OS3D designs in one forward pass; what I add is the gate each design has to survive before anyone spends money making it. It is validated in silico against approved drugs: no wet data is public and no target is named.
Building the ARC Engine
Programmable induced proximity: designing the ternary complex.
What I ownThe ARC Engine: eight phases that take an E3 ligase and a target with nothing to bind and design the peptide glue that holds them together, scored by cofolding, DockQ, PAE and MM/GBSA rather than by eye.
What it rests onA deposited ligase and not a model. CRBN comes from 9FJX at 2.00 angstroms, wild type, zinc and ligand kept, and the open conformer excluded on purpose because a glue designed against the wrong state is designed against nothing. It is pre-data: the pipeline is built and has not yet produced a molecule.
Calculate the TPSA for CN(Cc1ccc(F)cc1)S(=O)(=O)c1ccc(C)cc1.
from rdkit import Chemfrom rdkit.Chem import rdMolDescriptorsmol = Chem.MolFromSmiles('CN(Cc1ccc(F)cc1)S(=O)(=O)c1ccc(C)cc1')print(round(rdMolDescriptors.CalcTPSA(mol), 1))
TPSA = 37.4 Ų. This compound meets the Veber oral absorption threshold (TPSA ≤ 140 Ų).
round 5 test set, unedited · 9.65 s
What it isA 32B model fine-tuned on his own laptop to answer chemistry by running RDKit rather than by recalling it. The exchange on screen is one unedited answer from the round five test set.
Why it mattersAlmost all of the jump from v4 to v5 is tool executability, 71% to 95%. What was fixed was the habit of computing, not the amount of chemistry. Code to quote fidelity is still 61%, its weakest score: it can round in prose what it printed in code.
What it isOpenMM set up without writing the script. This is the verification run: a class B GPCR with its agonist in a lipid bilayer, graded against its own starting pose.
Why it mattersA short run that reports its own equilibration, its own grade and the limit of its own energy term is worth more than a long one that reports a single number. MM/GBSA ranks ligands against each other and is not an affinity.
What it isESMFold predicts the structure, a simple physics model relaxes the chain on to it, and every new residue contact becomes a note. The charts are the simulation's own and the music is this fold's, recorded from the app.
Why it mattersIt is the one on the list built for the pleasure of it, and it is the one that gets a room to sit through a secondary structure plot. It is a relaxation on to a known fold, not a prediction of the folding pathway.



Hit this. Spare that.
numerous times, over twenty years
Leeds · Oxford · Yale · Pfizer · Scripps · Stanford · Incyte
The patternEvery campaign on this deck is a pair: the protein to hit and the protein that has to be left alone. The number between them is the margin that was designed from a structure and then measured.
mutCALRThe row with no number. The mutant calreticulin epitope is twenty-one residues that exist in patients and nowhere healthy, so there is no wild-type counterpart to spare and nothing to express as a ratio.