Multimodal small-molecule screening for human prion protein binders

Prion disease is a rapidly progressive neurodegenerative disorder caused by misfolding and aggregation of the prion protein (PrP), and there are currently no therapeutic options. PrP ligands could theoretically antagonize prion formation by protecting the native protein from misfolding or by targeting it for degradation, but no validated small-molecule binders have been discovered to date. We deployed a variety of screening methods in an effort to discover binders of PrP, including 19F-observed and saturation transfer difference (STD) nuclear magnetic resonance (NMR) spectroscopy, differential scanning fluorimetry (DSF), DNA-encoded library selection, and in silico screening. A single benzimidazole compound was confirmed in concentration-response, but affinity was very weak (Kd > 1 mM), and it could not be advanced further. The exceptionally low hit rate observed here suggests that PrP is a difficult target for small-molecule binders. While orthogonal binder discovery methods could yield high affinity compounds, non-small-molecule modalities may offer independent paths forward against prion disease.


INTRODUCTION
Prion disease is a rapidly progressive neurodegenerative disorder caused by misfolding and aggregation of the prion protein, or PrP 1 . No effective therapeutics currently exist for prion disease, but PrP is a genetically and pharmacologically validated drug target 2 . PrP-lowering antisense oligonucleotides (ASOs) are in preclinical development [3][4][5] , and PrP-binding antibodies have been tested preclinically 6 as well as clinically in a compassionate use context 7 . Here, we sought to augment the therapeutic pipeline by discovering small molecules that bind PrP.
In principle, small molecules could prevent or treat prion disease by protecting PrP from misfolding or by lowering its abundance. By sterically blocking interactions with misfolded PrP, or simply through the free energy of binding, a chaperone might stabilize PrP against misfolding, following precedents in transthyretin amyloidosis 8,9 and cystic fibrosis 10,11 . Proofs of principle for this approach include the efficacy of monoclonal antibodies to PrP to clear prion infection in cell culture 12,13 and in peripheral tissues of animals 14 as well as the stability of PrP "stapled" with non-native disulfide bonds 15 . Alternatively, small-molecule binding events can sometimes directly lead to protein degradation 16,17 , and if not, a binder could serve as a starting point for engineering a bifunctional molecule to specifically target PrP for degradation 18,19 . Although at present most bifunctional degrader strategies are best suited to intracellular targets owing to reliance on cytoplasmic E3 ubiquitin ligases, recent studies suggest alternate routes to targeted degradation of cell surface proteins 20 such as PrP.
Decades of effort have not yet yielded a small-molecule PrP binder suitable for advancement as a drug candidate 21 . The development of phenotypic screening for antagonists of misfolded PrP accumulation in cultured cells 22 enabled the identification of several compounds effective in vivo [23][24][25][26] , but advancement of these compounds has been hindered by lack of activity against human prion strains and unclear mechanisms of action [26][27][28][29] . Meanwhile, several compounds shown to interact with PrP through biophysical assays have demonstrated antiprion activity in a range of experimental systems [30][31][32][33][34] . However, these compounds likewise appear to lack clinical promise as none are simultaneously specific 35 , potent, and drug-like. Certain metallated porphyrins 36,37 interact with PrP with affinity values comparable to their effective concentrations in cell culture 36,37 , and some exhibit in vivo activity in certain paradigms 38 . Similarly, a range of anionic polymers 33,[39][40][41][42] also bind PrP and show in vivo antiprion activity in certain contexts 41,43,44 . However, these binding events may not be monomeric 45 nor specific to PrP 35,46 . Still other compounds with demonstrated antiprion activity exhibit interaction with PrP only at concentrations orders of magnitude above their effective concentration in cell culture 47,48 .
Because PrP's biology does not lend itself to enzymatic or activity assays, we chose to apply several nonactivity-based screening modalities: 19 F-observed and saturation transfer difference (STD) NMR fragment screening, differential scanning fluorimetry (DSF), DNA-encoded library (DEL) selection, and an in silico screen. We selected a fragment-based drug discovery paradigm as a starting point based on PrP's small size and lack of obvious binding pockets 49 , and the success of this method in identifying ligands for targets refractory to other approaches 50 . Our campaign utilized 19 F NMR and STD NMR because these approaches are sensitive to weak affinity binders 51 , have low false positive rates 52 and allow searching a large swath of chemical space through small, highly soluble fragments that can later be optimized into larger, higher affinity molecules 53 . DSF was employed to find compounds that directly influence the thermal stability of PrP. This technique quantifies protein thermal stability by measuring fluorescence of a solvatochromic dye (SYPRO Orange) as a function of temperature as it binds to unfolded regions of a protein 54 . In principle, DSF hits should have the desired property of stabilizing the target protein. DNA-encoded library (DEL) selection uses a pool of thousands to trillions of individually DNA-barcoded molecules that are added to an immobilized recombinant protein. Nonspecific molecules are washed away, and putative binders are eluted, PCR amplified, and subjected to next-generation DNA sequencing for identification. We used a DEL of peptide macrocycles 55 , which display better stability and have a lower entropic cost of binding compared to linear peptides, making them suitable for targeting surfaces of proteins 56,57 . Finally, we employed the artificial intelligence-based in silico screening method, AtomNet® 58 , which uses a protein structure-based, convolutional neural network to predict molecular binding affinities. This technique was recently used to discover a selective binder and degrader of Miro1 59 .

Fragment-based drug discovery through NMR screening
From five commercial and internal sources ( Figure 1A) we selected 6,630 low molecular weight, high solubility fragments for a fragment-based drug discovery campaign. The compounds in these libraries were mostly small (<300 Da) and had a range of cLogP values and hydrogen bond donor and acceptor sites that mostly fell within the Rule of Three 60 for fragment-based screening ( Figures 1B-D). Fragments were screened against either HuPrP23-231 or HuPrP90-231 using pooled, 19 F ligand-observed or STD NMR methods as a primary screen, singleton STD NMR for re-testing, and protein-observed 1 H-15 N TROSY NMR for validation (Table 1). 19

F or STD NMR. C and D) Number of hydrogen bond donors (HBD) or hydrogen bond acceptors (HBA).
Orange denotes compliance with the "Rule of Three" 60 and black denotes noncompliance.

Validated hits
Broad Institute 19 F library  785  14  14  5  0  Broad Institute 1st gen STD library  1,116  16  15  4  0  Broad Institute 2nd gen STD library  1,823  43  55  34  1  Schreiber chiral collection  381  1  5  0  0  ChemBridge High Solubility Subset  2,525  31  149  37  0  Total  6,630  105  238  80  1  Table 1. Summary of NMR fragment screening. "Compounds screened" lists the total number of compounds in a given collection of molecules. "Pools with hits" indicates the number compound pools with observed hits. "Retested by STD" indicates the number of individual compounds that were retested for STD signal from each hit pool. "Retested by TROSY" shows the number of compounds that gave an STD signal and were therefore advanced to 2D TROSY NMR.
Of 6,630 compounds tested, 238 initial hits were re-tested as singletons by STD NMR, of which 80 were further tested by TROSY NMR. A single compound 5,6-dichloro-2-methyl-1H-benzimidazole (1) gave a reproducible STD signal in the presence of PrP ( Figure 2A) and induced chemical shift perturbations (CSPs) in the TROSY spectrum of HuPrP90-231 (Figures 2B-C, S1, and S2A). Mapping these residues onto an NMR structure of HuPrP (PDB 1HJM) 61 revealed no discernable pocket, with shifts scattered across the structure ( Figure 2D). Nonetheless, we observed similar resonance shifts in the full-length protein HuPrP23-231, suggesting that this binding is not an artifact of using a truncated construct ( Figure S2A), and CSPs were confirmed to be doseresponsive for several residues ( Figures 2E and S2B), and the compound caused a small (~0.2°C) but significant decrease in melting temperature by DSF ( Figure S2D).
Because the CSPs caused by 1 are so small, we wanted to be certain that this compound was not perturbing HuPrP due to nonspecific colloidal aggregation. To test whether 1 is causing CSPs due to aggregation, 15 N-HuPrP90-231 was incubated with 1 in the presence or absence of nonionic detergents Triton X-100 or Tween-20. The CSPs resulting from 1 were preserved in the presence of detergent, suggesting that 1 is not an aggregator ( Figure 2F). To assess compound aggregation by an orthogonal method, we used the well-established AmpC β-lactamase inhibition assay 62 . AmpC is inhibited by small molecules that form aggregates, and these aggregates can be disrupted by addition of detergent. No significant inhibition of AmpC was observed with 1 even at 500 µM concentrations, while the positive control compounds rottlerin and anacardic acid (AA) showed inhibition at 10 µM that could be relieved upon detergent addition ( Figure 2G). Analogs of 1 (compounds 2-10) were also tested; the majority of them did not inhibit AmpC, and none inhibited AmpC as well as the positive controls. Collectively, these data argue that 1 is not an aggregator and does not cause PrP CSPs via aggregation.
The small magnitude of CSPs and lack of saturation at concentrations up to 0.75 mM suggested that 1 has a Kd for PrP in the millimolar range, too weak to interrogate by many non-NMR orthogonal biophysical assays. In an attempt to find a stronger binder, we tested 54 analogs of 1 from commercial sources and the Broad Institute's internal library by STD NMR and TROSY (Table 2, Table S1, and Figure S3). Of the 20 compounds most similar to 1, 11 demonstrated positive STD and TROSY signal (Table 2), however, none of the analogs had TROSY CSPs larger than 1 by visual inspection, and thus they were not subjected to further biophysical assays. Despite efforts to soak unliganded PrP crystals with 1 and 20 of its analogs, no electron density attributable to a compound was identified (Table S1).  61 . E) Concentration dependent CSPs of residues Q212, M129, and Y218 upon addition of 1. F) 1 H-15

Thermal shift screening
We tested a library of 30,013 compounds from the Novartis Screening Set for External Collaborations (SSEC) in singleton for thermal stabilization of PrP using DSF (Figures 3 and S5A-B). An ideal DSF screen would possess a high signal-to-baseline fluorescence ratio, tightly distributed melting temperatures in the apo condition, and a ≥10:1 molar ratio of soluble ligand to protein 54 . We varied a range of assay parameters including protein concentration, dye concentration, assay volume, use of HuPrP90-231 or HuPrP23-231, and buffer conditions including buffering agent, metals, and DMSO concentration ( Figure S4). We obtained acceptable melt curves only at high protein concentrations ( Figure S4). Our final screening conditions achieved a ~5:1 signal-to-baseline fluorescence and a 0.06 °C median absolute deviation (MAD) with 30 µM HuPrP90-231; median Tm was 68.5 °C ( Figure 3A). Compounds were screened at 100 µM for a >3:1 ligand-to-protein ratio, though we lack empirical data on their solubility over the temperature ramp. We chose hit compounds that either positively or negatively affected the melting temperature (Tm) of PrP based on separate criteria (see Methods). An internally developed pipeline was used to perform Boltzmann fitting of the fluorescence data and call Tm values. Irregular melt curves were automatically flagged and discarded. Two hit criteria were chosen for positive Tm shifters: 1) a statistical cutoff of greater than 3*MAD of the DMSO control wells (0.17 °C) and 2) an initial fluorescence of less than 6 to eliminate compounds that have intrinsic fluorescence or distort the melt curve. Because there were so many negative shifters, stricter hit cutoffs of −0.7 °C > ΔTm > −9 °C were applied. Even though negative shifters are predicted to destabilize PrP, one hypothesis is that such compounds bind a partially folded or destabilized form of PrP and could exhibit antiprion properties 63 . Compounds that passed our hit criteria were retested by DSF in triplicate ( Figure 3C). Here, we applied stricter hit cutoffs (see Methods) due to throughput limitations of our orthogonal heteronuclear single quantum coherence spectroscopy (HSQC) NMR assay. The 183 reproducible positive hits were passed through frequent hitter and PAINS filtering 64 , narrowing the list to 117 compounds, of which we were able to test 93 for PrP binding by HSQC using 15 N-HuPrP90-231. Even though the melt curves of PrP with these compounds were often very robust and reproducible ( Figure 3D), none of the compounds tested by HSQC led to PrP CSPs at 100 µM concentration ( Figures 3E-F). This suggested that the observed thermal shifts were not mediated by binding PrP, and were likely artifacts. In support of this interpretation, when we tested eight of the validation screen hits by an orthogonal thermal shift method, differential scanning calorimetry, we were unable to reproduce the change in melting temperature (ΔTm) observed in DSF (Figures S5D-E).

DNA-encoded library selection
We performed a selection using HuPrP90-231 with a DEL library of 256,000 macrocycles. Barcode rank abundance in the unenriched library was plotted against enrichment observed in the PrP condition versus a noprotein control condition, revealing enriched compounds across three structural scaffolds ( Figure 4A). The *KRD scaffold was judged to be a likely covalent binder and was not pursued further. Representative compounds from the CC*S and *CJS series were resynthesized off-DNA ( Figures 4B-C), as both cis and trans isomers, for validation. None of these compounds produced appreciable CSPs against 15 N-HuPrP90-231 using TROSY at 200 µM, suggesting that these hits were either false positives or have affinities too weak to be detected by NMR (Figures 4D-E).

In silico screening
We used Atomwise's AtomNet® convolutional neural network method 58 to search for compounds that bind PrP at a specific site. Since there are no reported structures of human PrP bound to a lead-like ligand, we instead used the reported structure of mouse PrP bound to promazine (PDB 4MA7) 65 to create a homology model of human PrP also bound to promazine between helix 2 (α2) and the two beta strands (β1 and β2) ( Figure 5A). The regions that were modeled share a high degree of sequence identity with only 12 amino acid differences over residues A117-R230 (mouse PrP numbering). The promazine binding site was screened against 6,922,894 molecules. After additional filtering, the top 81 compounds (Table S2) were selected as predicted binders and assayed for binding using both DSF and STD NMR. By DSF, none of the compounds (90 µM) increased the Tm of PrP more than three standard deviations (0.79 °C), and none of the compounds (100 µM) showed an appreciable STD signal in the presence of HuPrP90-231 ( Figures 5B-C). Compounds that resulted in PrP shifts of less than one degree were plotted at −1. Error bars represent standard deviation of three measurements made on the same day using the same batch of protein-dye mix. SD of the DMSO conditions (n = 32 technical replicates per plate) was 0.26 °C. C) STD NMR of selected compounds that had a ΔTm > 0.05 °C by DSF. STD spectra are scaled to 5X the reference spectrum intensity.

DISCUSSION
We pursued four different screening modalities aimed at discovering binders of the human prion protein.
Despite the large number of molecules tested and complementary approaches used, we were unable to identify any hits suitable for advancement into medicinal chemistry. Our fragment screening campaign identified compound 1 (5,6-dichloro-2-methyl-1H-benzimidazole) and several analogs that weakly bind PrP and were validated with orthogonal NMR assays. However, the poor affinity of these compounds (> 1 mM) coupled with the absence of improved binding of chemical analogs effectively precluded their validation through non-NMR methods, and none were pursued further. Meanwhile, our thermal shift, DNA-encoded library, and in silico screening approaches yielded no validated hits at all.
A variety of target-specific technical challenges may have contributed to our inability to identify binders by the approaches employed here. Some reports indicate that the transfer of NMR saturation is weak for smaller proteins, which may have produced false negatives in our STD NMR screens 66,67 . SYPRO Orange dye fluorescence in the presence of unfolded PrP was weak, which necessitated DSF screening at 30 µM protein concentration; compounds were accordingly screened at 100 µM, but solubility limitations may have prevented saturable binding with a maximum thermal shift. Our in silico screen utilized a homology model based on a crystal structure of promazine bound to mouse PrP, but promazine has not been shown to bind human PrP in solution, and promazine analogs that exert antiprion activity in cells appear to do so through an orthogonal mechanism 48 . In general, without a positive control available, it is difficult to guide the optimization of screening assays. Taken as a whole, our experimental screens cannot be considered definitive given their modest scale. But considering the diverse methods and compound sets employed, our results may hint toward relative rarity of PrP binders in chemical space.
Alternative screening approaches might also improve the probability of discovering a high affinity PrP binder. We used recombinantly expressed PrP from E. coli for our experiments, which lacks post translational modifications (PTMs) including two N-linked glycosylations and a GPI anchor. Purification of PrP from mammalian cells 68 and insertion into nanodiscs 69 or micelles might more faithfully recapitulate PrP's PTMs and endogenous membrane environment, potentially yielding binding sites not present on recombinant PrP. Encouragingly, DEL screening has been used successfully with nanodisc immobilized proteins 70 . We could also extend our fragment-based drug design strategy by using chemoproteomics 71 to directly assess PrP ligandability on the cell surface. Multiple approaches may be necessary, because targets with low NMR and thermal shift hit rates are reported to, on average, also have lower hit-to-lead development success rates 72 .
Overall, despite various technical limitations, our inability to identify even weak binders through multiple orthogonal screening modalities is striking. The absence of obvious binding pockets on PrP's structure, together with the predominance of indirect mechanisms of action revealed in phenotypic screening campaigns, have led to the perception that PrP is a difficult target for small-molecule discovery 31 . Our data may provide some support for this conclusion. On balance, our results motivate an emphasis on non-small-molecule technologies, such as oligonucleotide therapeutics, as means for targeting PrP, but do not rule out the possibility that small-molecule binders could be discovered through an expanded screening effort.

Log(P) and H-bond donor/acceptor calculations
SMILES strings were parsed to yield molecular weight, ALogP, and hydrogen bond donor and acceptor counts using RCDK 73 .

NMR data acquisition and analysis
Spectra were acquired on a 600 MHz Bruker Avance III NMR Spectrometer equipped with a 5 mm QCI cryoprobe using 3 mm sample tubes (Bruker Z112272) and a SampleJet for sample handling. Spectra were analyzed in TopSpin version 4.0.2 and MestreNova version 10.0.1. Hit identification was performed by visual inspection of the data.

F NMR screening
HuPrP23-231 (final concentration 9 µM) or matched dialysis buffer alone (20 mM HEPES pH 6.8, no-protein control condition) was combined with 10% D2O and 4.5% DMSO containing a pool of 10 19 F fragments per NMR tube (45 µM each) with a total volume of 200 µL for each sample. A 19 F NMR spectrum was obtained for each sample using a standard 1 H-decoupled one-pulse experiment with 64 scans and a spectral width of 237 ppm with the carrier frequency at -100 ppm. The sample temperature was 280 K. Fragment hits were identified by comparison of both the peak position and peak width between the control (no protein) sample and the protein-containing sample. For this screen minimal line-broadening was observed; fragment hits were identified by visual review of chemical shifts, with perturbation of ≥0.005 (3 Hz) as an approximate threshold. Hit peaks from pools were compared to reference spectra to identify the likely hit fragment. Resupplied fragments were tested by STD NMR and/or 1 H-15 N TROSY NMR as described below.

STD NMR screening
HuPrP90-231 was buffer exchanged into 20 mM HEPES-d18 pH 7.4, 150 mM NaCl in ~99% D2O using a 5 kDa MWCO centrifugal concentrator (Millipore C7715). Pre-plated fragment pools were thawed in a dessicator at room temperature. For the Broad Institute 1 st and 2 nd generation STD libraries, 180 µL of HuPrP90-231 (11 µM) in deuterated buffer was added to each well, mixed with the fragments (1.6% DMSO-d6, v/v), and transferred immediately to a 3 mm NMR tube. The ChemBridge High Solubility Subset and Schreiber chiral fragment collection screens used HuPrP90-231 at 10 µM, 2 % DMSO-d6 (v/v), in 20 mM HEPES-d18 pH 6.8, 25 mM NaCl in ~99% D2O. Fragments were pooled with eight (Broad 1 st and 2 nd Gen STD and ChemBridge High Solubility Subset) or five (Schreiber chiral fragment collection) fragments per tube, always at a final concentration of 200 µM each. The Schreiber chiral fragment collection consisted of n = 381 compounds synthesized in-house, many of which have been described previously [77][78][79] . In silico screening hits (100 µM) were mixed with HuPrP90-231 (11 µM) in 20 mM HEPES-d18 pH 7.4, 150 mM NaCl in ~99% D2O with 1% DMSO-d6 (v/v). Ligand-observed screening was done using STD NMR. On-resonance irradiation of the protein was done at -0.25 ppm and off-resonance irradiation at 30 ppm. To saturate the protein, a 2 s train of 50 ms gaussian pulses separated by 1 ms delays was used. A 27 ms spin-lock pulse was used to suppress protein signals, and water suppression was accomplished using the excitation sculpting with gradients pulse scheme. The sample temperature was 280 K. Hit pools were identified by visual inspection and fragment hits were confirmed as singletons using the same experimental conditions as above. Compound 1 (5,6-dichloro-2methyl-1H-benzimidazole) was purchased from two different vendors (Combi-Blocks HC-3145 and Key Organics PS-4319) and retested for an STD signal, which was reproducible across vendors. The compound from Key Organics was deemed more pure than other sources by NMR, thin-layer chromatography, and LC-MS and was used for the majority of experiments presented here.  AmpC aggregation counter-screen Compounds were tested for colloidal aggregation using an AmpC β-lactamase counter-screen 62 . Recombinant E. coli AmpC was expressed in Rosetta cells and purified using a published protocol 80 . The enzymatic assay was performed in 50 mM sodium phosphate pH 7.0 ± 0.01% Triton X-100 (v/v) in clear UVtransparent 96-well half-area microplates (Corning 3679) in 150 μL final reaction volumes. Final concentration of DMSO was 1.0% (v/v). Compounds were incubated with 5 nM AmpC in 143.5 μL reaction solution for 5 min at r.t., followed by the addition of 1.5 μL of nitrocefin substrate (Cayman 15424) dissolved in DMSO (100 μM initial substrate concentration). Reaction solutions were gently mixed by multichannel pipette. Reaction progress was continuously monitored by absorbance at 482 nm for 5 min at r.t. on a SpectraMax M3 plate reader. Percent activity was calculated from reaction rates (slope) and normalized to DMSO-only controls after background subtraction with an enzyme-free reaction. Anacardic acid (AA), rottlerin, and 3',3'',5',5''tetraiodophenolphthalein (TIPP) (Combi-Blocks QE-5474) were used as positive controls. The significance difference between detergent and non-detergent tests was defined as p < 0.01 (after correction for multiple comparisons). Four intra-run technical replicates were performed on same microplate.

DSF screening
All concentrations are final assay concentrations unless otherwise indicated. All reagents were diluted in 20 mM HEPES pH 7.4, 150 mM NaCl, 1 mM EDTA. Protein was thawed and centrifuged (4500 g, 10 min, 4 °C) to pellet precipitate. HuPrP90-231 (30 μM) was mixed with SYPRO Orange dye (10X) (ThermoFisher S6651) and centrifuged (4500 g, 10 min, 4 °C) then decanted into a new bottle and covered in foil. Compound stocks dissolved in DMSO:H2O (90:10, v/v) were dispensed into 384-well barcoded plates (200 nL, 5 mM stock concentration, 100 μM final concentration) (4titude 4Ti-0381). 10 μL of protein-dye mix were added to each well with a Multidrop™ Combi Reagent Dispenser (ThermoFisher), shaken (2 min, r.t.), and centrifuged (1 min, r.t.). All Combi lines were covered to block ambient light. A Roche Lightcycler II was used for fluorescence measurements with filter sets 465/580. A temperature ramp from 30-90 °C, rate of 0.07 °C/s, and 6 acquisitions per second (6.5 min run) was used to collect the data. At the beginning of each day, 4 plates with no compounds were run to equilibrate the system. Melting temperatures (Tm) were calculated by fitting the fluorescence data to a Boltzmann curve. ΔTm values were calculated by taking the Tm with compound and subtracting it from the average of the DMSO control wells on each plate (n = 32 DMSO wells per plate). After removing error code flagged wells, positive ΔTm hits were picked using the following criteria: ΔTm > 3*MAD (0.17 °C), initial fluorescence intensity < 6. Negative ΔTm hits were picked using the following criteria: −0.7 °C > ΔTm > −9 °C, initial fluorescence intensity < 6. 1,129 hit compounds were tested in triplicate, all wells flagged with error codes 3 and 4 were removed, and hits were chosen based on the following criteria: ΔTm sign must be the same as the primary screen, 19.93 > Tm > 0.3, initial fluorescence < 4, Δfluorescence of 2-17.1. Hits from the triplicate validation screen were filtered for frequent hitters and PAINS and 84 compounds were tested by HSQC NMR. DSF data were analyzed using Tibco Spotfire and RStudio.

Differential scanning calorimetry
HuPrP90-231 (30 µM) was mixed with buffer or compound (100 µM) in 20 mM HEPES pH 7.4, 150 mM NaCl, 1 mM EDTA, 2% DMSO (v/v) (400 µL total volume per well). All concentrations are final assay concentrations. Data was acquired using MicroCal VP-Capillary DSC instrument from Malvern Panalytical and data was analyzed using Origin software provided by the vendor. A temperature ramp from 20-100 °C was conducted at a rate of 200 °C/h. Four buffer-only runs preceded the sample runs to equilibrate the system. n = 4 for DMSO controls and n = 1 for each compound. 3*SD was used as a hit cutoff.
Macrocycle DEL screening DNA-encoded library selection was performed as described using a 256,000-compound macrocycle library 55 . Briefly, 40 µg of HuPrP90-231, purified as described above, was loaded onto His Dynabeads (Invitrogen 10103D), relying on the intrinsic metal-binding properties of untagged PrP. Beads were washed, blocked, incubated with 50 µL of DNA-encoded library (60 min, 4 °C), washed three times, and protein was eluted with 300 mM imidazole. Barcodes were sequenced on an Illumina MiSeq and enrichment was calculated against a no protein condition run in parallel.

In silico screening
The homology model of human PrP with bound ligand promazine was built on the template of the mouse PrP structure 65 (PDB ID: 4MA7, chain A) using SWISS-MODEL 81 .The binding site is surrounded by residues V122, G124, L125, G126, Y128, Y162, I182, Q186, V189, T190 on the human PrP homology model.
The virtual screen was carried out using the AtomNet neural network, the first deep convolutional neural network for structure-based drug design 58,59 . A single global AtomNet model was employed to predict binding affinity of small molecules to a target protein. The model was trained with experimental Ki, Kd, and IC50 values of several million small molecules and protein structures spanning several thousand different proteins, curated from both public databases and proprietary sources. Because AtomNet is a global model, it can be applied to novel binding sites with no known ligands, a prerequisite to most target-specific machine-learning models. Another advantage of using a single global model in prospective predictions is that it helped prevent the so-called model overfitting. The following three-step procedure was applied to train the AtomNet model. The first step is to define the binding site on a given protein structure using a flooding algorithm 82 based on an initial seed. The initial starting point of the flooding algorithm may be determined using either a bound ligand annotated in the PDB database or crucial residues as revealed by mutagenesis studies, or identification of catalytic motifs previously reported. The second step is to shift the coordinates of the protein-ligand cocomplex to a three-dimensional Cartesian system with an origin at the center-of-mass of the binding site. Data augmentation was performed by randomly rotating and translating the protein structure around the center-ofmass of the binding site to prevent the neural network from memorizing a preferred orientation of the protein structure. The third step is to sample the conformations or poses of a small-molecule ligand within the binding site pocket. For a given ligand, an ensemble of poses was generated, and each of these poses represented a putative co-complex with the protein. Each generated co-complex was then rasterized into a fixed-size regular three-dimensional grid, where the values at each grid point represent the structural features that are present at each point. Similar to a photo pixel containing three separate channels representing the presence of red, green, and blue colors, our grid points represent the presence of different atom types. These grids serve as the input to a convolutional neural network and define the receptive field of the network. A network architecture of a 30×30×30 grid with 1Å spacing for the input layer, followed by five convolutional layers of 32×3 3 , 64×3 3 , 64×3 3 , 64×3 3 , 64×2 3 (number of filters × filter-dimension), and a fully connected layer with 256 ReLU hidden units was used. The scores for each pose in the ensemble were combined through a weighted Boltzmann averaging to produce a final score. These scores were compared against the experimentally measured pKi or pIC50 (converted from Ki or IC50) of the protein and ligand pair, and the weights of the neural network were adjusted to reduce the error between the predicted and experimentally measured affinity using a mean-squareerror loss function. Training was done using the ADAM adaptive learning method 83  For each small molecule, we generated a set of 64 poses within the binding site. Each of these poses was scored by the trained model, and the molecules were ranked by their scores. Due to a lack of a well-defined small molecule binding pocket on the human prion protein structure, there was low confidence in the predicted binders. Regardless, the top 50,000 ranking compounds were clustered based on chemical similarity and filtered for CNS drug-like properties using the Lipinski's CNS rules 85 with MW ≤ 400, clogP ≤ 5, number of hydrogen bond donors ≤ 3, and number of hydrogen bond acceptors ≤ 7. The final set of 81 compounds containing diverse chemical scaffolds were selected and sourced from Mcule.

DSF of in silico hits and compound 1
Compounds were diluted to 1 mM in DMSO/HBS (20 mM HEPES pH 7.4, 150 mM NaCl) (20:80, v/v). 1 μL of compound was added to each well of a 384-well plate (90 μM final concentration). Mixtures of HuPrP90-231 (30 μM) were prepared with SYPRO Orange dye (10X) in 20 mM HEPES pH 7.4, 150 mM NaCl, 1 mM EDTA buffer then centrifuged (4000 g, 10 min, 4 °C) and decanted into a new tube. 10 μL were added to each compound and mixed (2% DMSO final concentration, v/v). A Roche Lightcycler II was used for fluorescence measurements with filter sets 465/580. A temperature ramp from 30-90 °C, rate of 0.07 °C/s, and 8 acquisitions per second (~18 min per plate) was used to collect the data. Three independent plates (using the same protein-dye mix) were measured with n = 32 for DMSO controls per plate and n = 1 for each compound per plate. ΔTm values were calculated by subtracting the average apo Tm from the Tm with compound. Three ΔTm values per compound were averaged and plotted. Error bars represent standard deviation (SD = 0.26 °C). For DSF of compound 1 with HuPrP90-231, the same experimental parameters were used as above except compound 1 was used at 500 μM concentration and n = 8 intra-run technical replicates performed on the same assay plate.

Intact protein LC-MS
Purified HuPrP90-231 was diluted to 2 µM in 20 mM HEPES pH 7.4, 50 mM NaCl. 1 µL of diluted protein was injected onto a Waters BioAccord LC-ToF (composed of an ACQUITY I-Class UPLC and RDa detector with ESI source). Mobile phase A consisted of 0.1% formic acid (Millipore LiChroPur) in LC-MS grade water (JTBaker) and mobile phase B consisted of 0.1% formic acid in LC-MS grade acetonitrile (JTBaker). Protein was trapped on a C4 column (ACQUITY UPLC Protein BEH, 300Å, 1.7 µm, 2.1 X 50 mm) held at 80 °C for the entire analysis. The protein was desalted for one minute before elution with a gradient of 5% to 85% mobile phase B in 2.5 min. Ionization was performed with 55 V cone voltage and 550 °C ionization temperature. The instrument scan rate was 0.2 scans/s over 50 to 2000 m/z. PrP eluted at an observed retention time of 2 min. The PrP charge envelope was deconvoluted into the intact mass using the MaxEnt1 function using UNIFI software (Waters).

Code and data availability
Raw data and source code will be made available in a public GitHub repository: http://github.com/ericminikel/binder_screening

ACKNOWLEDGMENTS
This work was supported by the National Institutes of Health (F31 AI122592 to EVM, R35 GM118062 to DRL supporting AIC and DLU, and T32 HL007627 to JLD), Prion Alliance, BroadIgnite, and an anonymous organization. Atomwise provided effort and reagents in kind as part of an Artificial Intelligence Molecular Screen (AIMS) Award. Novartis provided reagents, expertise in screening sciences and efforts to perform the work as part of a research collaboration agreement. We thank Dr. Kirk Clark (Novartis) for guidance on the DSF screen and related NMR validation experiments.

DECLARATION OF INTERESTS
EVM has received consulting fees from Deerfield Management and Guidepoint and has received research support in the form of unrestricted charitable contributions from Charles River Laboratories and Ionis Pharmaceuticals. SLS serves on the Board of Directors of the Genomics Institute of the Novartis Research Foundation ("GNF"); is a shareholder and serves on the Board of Directors of Jnana Therapeutics; is a shareholder of Forma Therapeutics; is a shareholder and advises Kojin Therapeutics, Kisbee Therapeutics, Decibel Therapeutics and Eikonizo Therapeutics; serves on the Scientific Advisory Boards of Eisai Co., Ltd., Ono Pharma Foundation, Exo Therapeutics, and F-Prime Capital Partners; and is a Novartis Faculty Scholar. SMV has received speaking fees from Illumina and Biogen and has received research support in the form of unrestricted charitable contributions from Charles River Laboratories and Ionis Pharmaceuticals. DRL is a consultant for and co-founder of Exo Therapeutics, which uses DNA-encoded libraries for drug development. DC, DSA, OKS, and SKW are employees of Novartis. KTN is an employee of Atomwise.