Novel GPU Engines for Virtual Screening of Giga-Sized Libraries Identify Inhibitors of Challenging Targets

Abstract

To accelerate virtual ligand screening (VLS) and identify potent drug leads from massive chemical libraries, we developed two GPU-accelerated methods: Rapid Docking GPU Engine (RIDGE) for receptor-based screening and Rapid Isostere Discovery Engine (RIDE) for ligand-based 3D similarity screening. RIDGE performance surpassed or was as good as previously described methods when tested on 102 proteins from the Directory of Useful Decoys, Enhanced (DUD-E). We used RIDGE and RIDE to screen ultralarge virtual libraries against challenging cancer targets, PD-L1 and K-Ras G12D. This led to the discovery of novel inhibitors with single-digit to submicromolar affinities (five for PD-L1, three for K-Ras G12D). Docking scores from our methods were better predictors of binding than conventional VLS. These novel GPU-accelerated methods expand screenable chemical space and successfully identify active hits, even for challenging targets. Further optimization and libraries with higher-molecular-weight cutoffs could further improve targeting of nondruggable proteins.

Introduction

Virtual screening has become an indispensable tool in drug discovery, enabling the identification of potential lead compounds for target inhibition or activation. This approach significantly reduces the need for expensive and environmentally harmful chemical synthesis in the search for active leads and can handle compound libraries containing millions of synthesizable molecules, which represent only a tiny fraction of the astronomically vast number of possible drug-like compounds in the universe, estimated at 10⁶³. − As the synthesizable chemical space continues to expand, the ability to screen ultralarge libraries of billions of compounds and beyond becomes crucial. Performing virtual screening with such massive libraries will accelerate drug development by uncovering superior compound leads and enhancing success rates through access to greater chemical diversity. Recent advancements have improved the virtual screening of ultralarge libraries, including acceleration via artificial intelligence and machine learning, − application of Bayesian optimization techniques, large-scale screens through CPU parallelization, hierarchical screening of synthons or building blocks, , optimal utilization of GPU resources, and advanced parallel computing strategies.

This work evaluates two new GPU-accelerated methods for structure-based and ligand-based virtual screening. Rapid Docking GPU Engine (RIDGE) is a structure-based virtual screening engine that runs completely on GPUs to rapidly dock approximately 100 compounds per second with modern hardware. Rapid Isostere Discovery Engine (RIDE) is a ligand-based virtual screening method that uses atomic property fields to perform a 3D molecular similarity search. Using this method, a single GPU can search approximately 5,00,000 compound conformers per second, making it possible to screen multibillion compound databases.

To validate the accuracy of these methods and compare them with existing platforms, we applied RIDGE to 102 targets from the Directory of Useful Decoys, Enhanced (DUD-E). RIDGE performed as well as, and often better than, previously tested methodologies for this set of targets. However, all DUD-E targets are considered “easily druggable”. Accessing large chemical spaces is particularly beneficial for “nondruggable targets”, which lack well-defined binding pockets and pose significant challenges for traditional drug discovery methods. This “undruggable” class accounts for over 85% of all proteins, representing a substantial gap in targeting the majority of known disease-related proteins. We tested our GPU-based virtual screening methods against two challenging proteins, programmed death-ligand 1 (PD-L1) and K-Ras G12D. PD-L1, in conjunction with its partner protein, PD-1, − plays a role in immune regulation in the tumor microenvironment. PD-L1 is also expressed by tumor cells to escape immune responses of the host, making it an attractive target for inhibition. K-Ras is a signaling protein in the RAS/MAPK pathway that converts between an active GTP-bound state and an inactive GDP-bound state. More than 15% of all human cancers are associated with aberrant activity of Ras proteins due to somatic mutations, of which 85% are K-Ras mutations. , One of the most prevalent and difficult-to-target K-Ras mutations has been G12D, making it an attractive selective inhibition target. Although PD-L1 and K-Ras G12D lack well-defined binding sites, small-molecule ligands have been identified and crystal structures of their ligand-bound complexes have been determined to less than 2 Å resolution. , Thus, the ligand-bound complexes of PD-L1 and K-Ras G12D are good models for evaluating RIDGE and RIDE, and known inhibitors serve as positive controls for our experimental validation. By applying RIDGE and RIDE to these especially challenging targets, we show that the efficient screening of ultralarge chemical spaces, enabled by our GPU-accelerated tools, can successfully identify hits for challenging targets.

Results and Discussion

Technical Benchmarks

RIDGE is a structure-based virtual screening method that docks compounds to a known protein binding site. Its exceptional speed stems from several key optimizations. First, the core docking engine is fully implemented on the GPU, minimizing data transfer between the CPU and GPU, which is a common bottleneck. Second, RIDGE employs massive parallelization, an area where GPUs excel, to simultaneously handle ligand posing, interaction calculations, and scoring for numerous compounds. Third, it uses optimized memory access and highly compressed conformer databases as input, enabling the GPU to rapidly load and process molecular information. Finally, RIDGE utilizes a hybrid CPU/GPU workload; while core docking is GPU-driven, low-level CPU multithreading handles complementary tasks such as reading conformers, assigning charges, and initial data preparation, ensuring a balanced workflow.

We evaluated RIDGE’s screening speed and its dependence on GPU hardware by testing it with conformers of 10,000 compounds randomly selected from a 10-million-compound vendor library. Average speeds on consumer-grade cards were: NVIDIA GeForce RTX 3090 (49.1 molecules/s), NVIDIA GeForce RTX 4090 (101.5 molecules/s), and NVIDIA GeForce RTX 5090 (112.8 molecules/s). Some of the data center GPUs yielded even higher speeds: NVIDIA A100 80GB PCIe (98.0 molecules/s), NVIDIA H100 80GB HBM3 (143.4 molecules/s), and NVIDIA H200 140GB (165.9 molecules/sec). Given that data center GPUs can be up to 5–10 times more expensive, it is remarkable that consumer-grade cards can be used with only a modest speed penalty. Notably, Uni-Dock, previously reported as the fastest GPU-accelerated docking program, screened 38.2 million molecules in 12 h using 100 NVIDIA V100 GPUs. This performance, equating to approximately 9 molecules per second per GPU, is approximately 10-fold slower than that of RIDGE.

We used Directory of Useful Decoys, Enhanced (DUD-E) to benchmark the performance of RIDGE and RIDE. DUD-E consists of 102 proteins and the corresponding sets of active and decoy compounds for each target, 22,886 actives and 14,11,214 decoys total. It has been applied for the benchmarking of various virtual screening methods in the past. − Two metrics are commonly used to evaluate the performance of the screens, ROC’s (receiver operating characteristic’s) area under the curve (AUC) and Enrichment Ratio (ER). AUC evaluates the method’s overall ability to distinguish between binders versus nonbinders. ER is calculated using the following formula: ER = (Decoys_total/Decoys_hits)/(Ligands_total/Ligands_hits) and determines hits enrichment at a score cutoff that results in a certain fixed false positive rate, e.g., 1%, defined as percentage of Decoys_hits from Decoys_total. RIDGE achieved mean and median AUC values of 76.9 and 78.8, respectively, across 102 targets. Mean and median ER_1% were 29.2 and 24.0, respectively. Chaput et al. evaluated the performance of four popular virtual screening programsGold, Glide, Surflex, and FlexX on DUD-E benchmarks reporting BEDROC (Boltzmann-enhanced discrimination of receiver operating characteristic) metrics. BEDROC is a generalization of ROC that specifically addresses the “early recognition” problem. Table compares RIDGE’s BEDROC values with these published results. While performance varies significantly across targets, RIDGE performed competitively against established methods. It outperformed Gold, Glide, Surflex, and FlexX for 33 out of 102 targets, while Gold surpassed other methods for 34, Glide for 27, Surflex for 4, and FlexX for 5. RIDGE outperformed Gold for 45 targets, Glide for 50, Surflex for 71, and Flexx for 72, while being the worst for only 8 targets.

1. BEDROC Metrics of five VLS Methods across 102 Targets in DUD-E.

	method
target	Gold ,	Glide ,	Surflex ,	FlexX ,	RIDGE
aa2ar	0.29	0.13	0.34	0.17	0.378
abl1	0.48	0.32	0.35	0.51	0.37
ace	0.3	0.1	0.09	0.03	0.25
aces	0.37	0.16	0.12	0.03	0.07
ada	0.35	0.46	0.56	0.05	0.29
ada17	0.32	0.12	0.19	0.18	0.12
adrb1	0.43	0.31	0.25	0.18	0.30
adrb2	0.43	0.5	0.41	0.36	0.24
akt1	0.42	0.24	0.05	0.11	0.11
akt2	0.62	0.41	0.24	0.27	0.36
aldr	0.4	0.39	0.14	0.21	0.62
ampc	0.04	0.09	0	0.04	0.47
andr	0.04	0.37	0.08	0.01	0.16
aofb	0.29	0.15	0.15	0.1	0.23
bace1	0.43	0.26	0.27	0.19	0.27
braf	0.43	0.56	0.15	0.51	0.35
cah2	0.29	0.08	0.04	0.17	0.10
casp3	0.53	0.56	0.37	0.49	0.12
cdk2	0.3	0.47	0.07	0.24	0.30
comt	0.68	0.71	0.05	0.17	0.18
cp2c9	0.12	0.04	0.05	0.19	00.05
cp3a4	0.21	0.17	0.13	0.08	0.04
csf1r	0.36	0.32	0.05	0.45	0.39
cxcr4	0.08	0.01	0.27	0.01	0.01
def	0.49	0.18	0.62	0.01	0.42
dhi1	0.17	0.15	0.13	0.02	0.24
dpp4	0.29	0.18	0.09	0.21	0.31
drd3	0.18	0.04	0.15	0.06	0.06
dyr	0.66	0.4	0.41	0.2	0.81
egfr	0.42	0.4	0.24	0.35	0.59
esr1	0.41	0.79	0.45	0.5	0.67
esr2	0.32	0.75	0.43	0.51	0.69
fa10	0.74	0.31	0.31	0.74	0.80
fa7	0.8	0.73	0.79	0.92	0.74
fabp4	0.62	0.45	0.12	0.37	0.26
fak1	0.51	0.34	0.12	0.31	0.56
fgfr1	0.8	0.26	0.3	0.31	0.40
fkb1a	0.23	0.62	0.22	0.01	0.39
fnta	0.3	0.11	0.05	0.12	0.04
fpps	0.96	0.01	0.04	0.08	0.05
gcr	0.13	0.21	0.3	0.18	0.34
glcm	0.89	0.48	0.93	0.22	0.41
gria2	0.24	0.44	0.13	0.12	0.57
grik1	0.42	0.31	0.19	0.36	0.67
hdac2	0.31	0.17	0.25	0.3	0.28
hdac8	0.2	0.09	0.08	0.27	0.08
hivint	0.19	0.03	0.22	0.05	0.02
hivpr	0.3	0.14	0.1	0.05	0.01
hivrt	0.42	0.37	0.13	0.19	0.26
hmdh	0.4	0.66	0.38	0.04	0.37
hs90a	0.23	0.03	0.02	0.03	0.05
hxk4	0.3	0.34	0.04	0.08	0.49
igf1r	0.49	0.52	0.2	0.49	0.31
inha	0.51	0.27	0.21	0.18	0.23
ital	0.12	0.1	0.02	0.03	0.13
jak2	0.5	0.41	0.17	0.19	0.73
kif11	0.55	0.59	0.12	0.08	0.55
kit	0.26	0.75	0.6	0.07	0.15
kith	0.24	0.12	0.05	0.17	0.74
kpcb	0.49	0.6	0.43	0.45	0.65
lck	0.26	0.41	0.19	0.39	0.54
lkha4	0.29	0.54	0.47	0.16	0.09
mapk2	0.4	0.45	0.04	0.2	0.66
mcr	0.04	0.3	0.06	0.01	0.08
met	0.64	0.44	0.17	0.35	0.77
mk01	0.46	0.26	0.03	0.39	0.23
mk10	0.39	0.28	0.03	0.29	0.18
mk14	0.24	0.34	0.06	0.29	0.34
mmp13	0.4	0.33	0.3	0.09	0.19
mp2k1	0.39	0.23	0.03	0.16	0.25
nos1	0.4	0.16	0.35	0.14	00.49
nram	0.34	0.69	0.42	0.28	0.5
pa2ga	0.32	0.58	0.09	0.3	0.20
parp1	0.24	0.65	0.34	0.28	0.82
pde5a	0.16	0.3	0.19	0.05	0.14
pgh1	0.22	0.24	0.14	0.09	0.26
pgh2	0.32	0.48	0.27	0.21	0.40
plk1	0.48	0.53	0.11	0.23	0.51
pnph	0.45	0.15	0.22	0.21	0.68
ppara	0.32	0.19	0.37	0.09	0.14
ppard	0.23	0.11	0.16	0.04	0.14
pparg	0.32	0.15	0.12	0.04	0.23
prgr	0.07	0.05	0.07	0.1	0.16
ptn1	0.78	0.32	0.36	0.32	0.45
pur2	0.97	1	0.98	0.99	0.94
pygm	0.17	0.02	0.03	0.12	0.06
pyrd	0.5	0.57	0.27	0.48	0.71
reni	0.44	0.53	0.46	0.18	0.16
rock1	0.2	0.4	0.05	0.35	0.63
rxra	0.53	0.79	0.54	0.27	0.74
sahh	0.73	0.99	0.97	0.8	0.91
src	0.24	0.24	0.12	0.3	0.52
tgfr1	0.59	0.66	0.28	0.44	0.53
thb	0.54	0.54	0.29	0.26	0.68
thrb	0.54	0.66	0.58	0.62	0.38
try1	0.65	0.67	0.75	0.8	0.75
tryb1	0.6	0.47	0.44	0.5	0.40
tysy	0.76	0.65	0.46	0.32	0.36
urok	0.66	0.85	0.29	0.77	0.84
vgfr2	0.36	0.26	0.07	0.52	0.34
wee1	0.92	0.99	0.82	0.72	0.99
xiap	0.49	0.8	0.45	0.57	0.86

^aData from ref .

RIDE is a ligand-based method that focuses on 3D similarity to a query molecule. Its acceleration is also heavily dependent on GPU utilization. RIDE is built upon Atomic Property Fields (APF), which are 3D grids representing the physicochemical properties of molecules. The comparison of these computationally intensive fields for similarity is ideally suited for GPU acceleration due to their grid-based nature. Furthermore, RIDE performs rapid 3D molecular superposition, a process of aligning molecules in 3D space to maximize their similarity. This particular task, especially when dealing with large databases of conformers, benefits immensely from the parallel execution capabilities of GPUs. RIDE can search approximately 0.5 million conformers per second per GPU, with speeds reaching up to 1.5 million conformers/second on the RTX 4090. Like RIDGE, RIDE uses preprepared, highly compressed conformer databases. This efficient data management is critical for quickly feeding data to the GPU and preventing input/output (I/O) from becoming a bottleneck, ensuring sustained high-speed screening.

We also used DUD-E to evaluate the performance of RIDE in 3D ligand-based screening. For each of the 102 targets, the DUD-E benchmark provides a single ligand from an X-ray structure that can be used as a query. A recently published study used DUD-E to evaluate the performance of a wide range of 3D ligand-based virtual screening algorithms. We compare RIDE to the published results in Table .

2. Performance of RIDE and Other Ligand-Based Screening Methods.

method	ROC AUC (mean and SD)	EF 1% (mean and SD)
ROCScolorscore	0.620 ± 0.139	13.20 ± 13.74
Phase shape_pharm	0.692 ± 0.160	16.03 ± 16.53
Shape-it	0.541 ± 0.133	4.73 ± 4.99
Align-it	0.659 ± 0.137	12.44 ± 13.79
ShaEPbest	0.658 ± 0.122	10.28 ± 10.73
SHAFTS	0.733 ± 0.144	19.13 ± 16.96
WEGA	0.645 ± 0.143	12.14 ± 12.45
LIGSIFT	0.718 ± 0.133	16.22 ± 15.29
LS-align	0.699 ± 0.126	15.88 ± 14.39
RIDE	0.817 ± 0.131	29.79 ± 19.78

^aData from ref. Highest-performing flavors of ROCS, Phase, and ShaEP are listed. EF1% is an enrichment factor at 1% of the screening library. SD refers to standard deviation.

Although some EF_1% values, when considering their standard deviations, are statistically indistinguishable, ROC AUC shows the superiority of RIDE compared to other methods. This is consistent with findings in an earlier study evaluating the performance of molecular 3D alignment methods, where the algorithm based on Atomic Property Fields (the underlying methodology of RIDE) similarly outperformed other techniques.

Virtual Screening with RIDE and RIDGE

We investigated the performance of GPU-based virtual ligand screening (VLS) methods, RIDE and RIDGE, in hits discovery. We tested these methods against two notoriously challenging cancer targets: programmed death-ligand 1 (PD-L1), a protein crucial for immune regulation, and K-Ras, a key signaling protein in the RAS/MAPK pathway. To assess the suitability of protein structures for virtual screening, we initially employed CPU-based ICM-VLS to screen ChemSpace’s on-the-shelf chemical library, comprising nearly 7 million compounds. Subsequently, we utilized GPU-based RIDGE to screen a 48 million compound diversity set from the Enamine REAL database. Furthermore, GPU-based RIDE was applied to identify analogues of known inhibitors from a vast database that exceeds 14.35 billion compounds. This extensive database integrated Synthetically Accessible Virtual Inventory (SAVI), Enamine’s REAL database, Enamine’s MADE building blocks, ChemSpace’s Freedom, and WuXi’s GalaXi. These methods have enabled the identification of submicromolar inhibitors that can be further optimized.

Discovering PD-L1 Inhibitors Using GPU-Based VLS Methods

All three methods were applied to screen the dimer structure of PD-L1. The majority of known small-molecule inhibitors of PD-L1 act by causing protein dimerization, which leads to the formation of a long cleft. However, complex formation is characterized by induced fit, meaning that structures of the protein bound to different inhibitors do differ. Out of 23 structures of PD-L1 complexed to small-molecule inhibitors, we chose PDB: 5N2F, the structure of the protein bound to compound BMS-200. We chose this structure because not only does this compound, which exhibits an IC₅₀ = 3.1 nM, redock with highly favorable scores, but other known PD-L1 dimerizers, like INCB086550, also dock well into the structure. The pocket occupied by the compound is large, with a volume = 1208 A³, and it is elongated with nonsphericity = 2.05 and hydrophobic. A summary of the compounds identified by the virtual screen and ordered for testing is presented in Table . Two docking scores were used to evaluate the predicted binding of each compound. The ICM-VLS score is a physics-based energy scoring function that considers several energy terms of the interaction. Generally, the lower the score, the better the prediction of binding, with −32 being considered a good threshold to predict binding versus nonbinding. The Radial Topological Convolutional Neural Net (RTCNN) score is trained to recognize favorable interaction patterns observed in experimental crystal structures in contrast to decoy non-native complexes. The lower the RTCNN score, the better predicted binding. However, there is not an empirically derived threshold for binding as for the ICM-VLS score. From each screen, compounds were selected based on a combination of low ICM-VLS and RTCNN scores and structural diversity among ordered compounds.

3. Summary of Ordered Compounds Derived from Virtual Screening and Experimental Results for Inhibition of the PD-1-PD-L1 Interaction that was Measured by Time-Resolved Fluorescence Resonance Energy Transfer (TR-FRET) .

^a100% activity is activity in the absence of inhibitors. Values above 100% indicate even stronger interaction between PD-1 and PD-L1 than the positive control under those experimental conditions.

Initially, we screened the ChemSpace on-the-shelf chemical library, comprising nearly 7 million compounds, by utilizing a conventional CPU-based virtual ligand screening (VLS) approach. The screen was completed with a thoroughness of 1, and top hits were redocked with a thoroughness of 300. After manual refinement that was performed to refine the poses and to allow for guided scoring interpretation, seven of the top compounds were selected for testing with RTCNN scores ranging from −38 to −55 and ICM-VLS scores ranging from −37 to −45. One compound (5n2f-4) could not be tested due to its poor solubility.

RIDGE enables the exploration of chemical spaces that are practically inaccessible to traditional CPU-based methods due to time and resource constraints, leading to the potential for screening orders of magnitude more compounds in the context of giga-sized libraries. We screened 48 million compounds from the Enamine REAL database against the PD-L1 dimer. This resulted in 6921 compounds with a RTCNN score ≤ −45. These compounds were redocked with a thoroughness of 300. After manual redocking and inspection, eight of the top compounds were selected for testing, with RTCNN scores ranging from −46 to −56 and VLS scores from −38 to −43. One compound (5n2f-13) could not be tested due to its insufficient solubility.

We then used RIDE to search for 3D analogs of a known potent inhibitor of PD-L1, INCB086550 that has been tested in clinical trials but discontinued due to high toxicity. The RIDE search encompassed our SAVI (Synthetically Accessible Virtual Inventory) database, Enamine REAL database, ChemSpace Freedom database and WuXi GalaXi database, totaling about 14.35 billion compounds. The screen was run on 4 GPUs and took 94 h, yielding 6,29,138 hits with an APF cutoff of 0.35. It should be noted that the time needed for the RIDE screen depends on the conformational flexibility of the reference compound. INCB086550 has 10 rotatable bonds, which is unusually high. It leads to multiple conformations that need to be tested, slowing RIDE screens significantly. All RIDE hits have been subjected to docking in ICM-Pro and top-scoring compounds manually evaluated for pose, their associated scores, and structure similarity. Since the goal was to minimize synthetic efforts, only two hits, 5n2f-16 and 5n2f-17 (Table ), have been selected for the synthesis.

Experimental Validation of PD-L1 Inhibitors

The interaction between PD-L1 and its binding partner on T cells, programmed cell death protein 1 (PD-1), is a critical checkpoint mechanism of the immune system that prevents T cells from attacking healthy cells and tissues that express PD-L1. However, tumor cells can upregulate PD-L1 expression to prevent killing by T cells. This has led to the development of immune checkpoint inhibitors that target the PD-1–PD-L1 interaction and enhance the immune response against cancer cells. , Small-molecule inhibitors of the PD-1–PD-L1 interaction have been previously shown to block the interaction by binding with high affinity to PD-L1, inducing its dimerization. , INCB086550 has been shown to potently bind human PD-L1 at the same site as anti-PD-L1 therapeutic antibodies with an IC₅₀ of 3.1 nM.

We analyzed the ability of GPU-based VLS methods to identify novel hits for the inhibition of PD-L1. Based on the VLS methods described above, we acquired 17 compounds in addition to a known inhibitor, INCB086550, for biophysical testing. Two of the small-molecule hits were determined to have insufficient solubility in aqueous conditions and therefore are not presented with the results. Compounds were assayed for inhibition of the PD-L1–PD-1 interaction by time-resolved fluorescence resonance energy transfer (TR-FRET), which combines a time-resolved measurement of fluorescence with fluorescence energy transfer. This experiment monitors the transfer of energy between europium-labeled PD-1 ([Eu]-PD-1) and a dye-labeled acceptor that binds to [biotin]-PD-L1. Upon excitation at 337 nm, emission is measured at 620 nm (europium) and 665 nm (red fluorophore acceptor). In the case of a [Eu]-PD-1–[biotin]-PD-L1 binding event, the emission at 665 nm is expected to increase due to the energy transfer between the proximal Eu and the dye-labeled acceptor. Activity is calculated based on the Em₆₆₅/Em₆₂₀ ratio, where this ratio of the positive control ([Eu]-PD-1 + dye-labeled acceptor + [biotin]-PD-L1) represents 100% activity and the ratio of the negative control ([Eu]-PD-1 + dye-labeled acceptor) represents 0% activity. A decrease in Em₆₆₅/Em₆₂₀ relative to the absence of the ligand indicates an increased distance and loss of PD-1–PD-L1 binding. TR-FRET was measured at 5 μM and 25 μM of each ligand. The known inhibitor, INCB086550, shows a complete loss of activity at 5 μM (Figure A). Eight out of 15 novel compounds inhibited PD-L1 activity by at least 20% at a 25 μM concentration (Figure A). Three of these compounds were derived from CPU-based VLS (5n2f-1, 5n2f-6, and 5n2f-7). Four inhibitors were derived from RIDGE screens (5n2f-8, 5n2f-9, 5n2f-10, and 5n2f-12). The most potent novel inhibitor (5n2f-17) was discovered using RIDE, which was not surprising since the RIDE search covered the largest chemical space, 14.35 billion compounds. 5n2f-17 reduced PD-L1 activity to 6 ± 9% at 5 μM concentration (Figure A). Some compounds showed greater inhibition at a lower concentration. This may be due to aggregation of the compound at higher concentrations, effectively lowering its concentration in solution, interference of the compound with the assay, or nonspecific binding at higher concentrations. Although 5n2f-17 was found by RIDE using INCB086550 as a template, 2D similarity between the two compounds was low with Tanimoto distance between them equal to 0.58, thus suggesting a different scaffold for PD-L1 inhibitors development. RIDGE was also able to find an inhibitor with a novel scaffold and, importantly, a different predicted binding mode. 5n2f-10 reduced PD-L1 activity to 61 ± 12% at 5 μM (Figure A).

1.

PD-1–PD-L1 TR-FRET assay. The assay was performed per the manufacturer’s protocol. (A) Each compound was used at 5 μM (white bars) or 25 μM (gray bars) in PBS at room temperature. The Pos Ctrl sample contains [Eu]-PD-1 and [biotin]-PD-L1 in the absence of an inhibitor. The Neg Ctrl sample contains only [Eu]-PD-1. % Activity is the TR-FRET ratio normalized to that of the controls. Error bars represent the standard deviation from triplicate measurements. Red arrows indicate those compounds that show activity of 80% (black dashed line) or less at 25 μM. Lower inhibition at a higher concentration could have been caused by compound aggregation, assay interference at higher concentrations, or nonspecific binding. (B–D) Dose–response curves for (B) INCB086550, (C) 5n2f-10, and (D) 5n2f-17. Compounds were assayed in a 5:1 serial dilution series in PBS at room temperature. For each dilution point, % Activity is the TR-FRET ratio normalized to that of the controls. Error bars represent the standard deviation from triplicate measurements. IC₅₀ values were calculated in GraphPad Prism based on fitting a nonlinear regression. Values in parentheses are estimated ranges of IC₅₀s.

We then aimed to determine IC₅₀ values for the most potent compounds, INCB086550, 5n2f-10, and 5n2f-17, by monitoring % activity with increasing ligand concentration. Known inhibitor, INCB086550, was found to have an IC₅₀ of 5.6 nM with a range of 1.6–10.2 nM (Figure B). This is consistent with the IC₅₀ value of 3.1 nM previously published. Novel compounds 5n2f-10 and 5n2f-17 gave IC₅₀ values of 2.6 μM (1.6–5.0 μM) (Figure C) and 0.19 μM (0.11–3.6 μM) (Figure D), respectively.

To evaluate the structural novelty of our newly identified PD-L1 inhibitors, we performed similarity searches against known antagonists from the ChEMBL and PubChem databases. We focused on eight compounds with confirmed inhibitory properties: 5n2f-1, 5n2f-6, 5n2f-7, 5n2f-8, 5n2f-9, 5n2f-10, 5n2f-12, 5n2f-17, and 5n2f-16, which showed no inhibition. Using a 40% similarity cutoff, our search revealed no known analogs for eight of these nine compounds. However, for inhibitor 5n2f-17, we found eight similar compounds reported in the literature on PD-L1 inhibitors with Tanimoto scores ranging from 0.43 to 0.54 (Table S3). All of these shared a 3-(2,3-dihydrobenzo­[b]­[1,4]­dioxin-6-yl)-2-methylaniline moiety, which is also a key component of BMS1166, a drug developed by Bristol-Myers Squibb. The low similarity scores are likely due to the unique structure of the second half of 5n2f-17, specifically the 3-chloro-7-(3-(methoxycarbonyl)­bicyclo[1.1.1]­pentane-1-carbonyl)-5,6,7,8-tetrahydroimidazo­[1,2-a]­pyrazine-2-carboxylic acid block. This distinctive part of the molecule is also likely responsible for its extensive interactions with PD-L1.

Predicted Interactions of PD-L1 Inhibitors

ICM-Pro docking predicts INCB086550 to dock well to the dimer interface, with an RTCNN score of −53.50 and a VLS score of −32.89 (Table ). INCB086550 spans the entire β-sheet sandwich, making stabilizing hydrogen bonds on both ends of the β-sheet with N63 and R125 of chain B and with R125 of chain A (Figure A,B). Residues Y123, Y56, M115, and D122 of chain A and Y123, Y56, M115, D122, A121, and I54 of chain B also make significant contacts with the molecule (Figure A,B). These are consistent with contacts observed in X-ray crystal structures of other PD-L1 small-molecule inhibitors that induce dimerization. ,− 5n2f-17 identified by RIDE, docks with the best RTCNN score of the tested compounds, −64.41, and a favorable LE score of −40.41 (Table ). The predicted interaction of 5n2f-17 with the PD-L1 dimer spans only three β-strands of the chain A β-sheet and makes a hydrogen bond with R125 of chain A or N63 of chain B (Figure C,E). Additional contacts are with chains A Y123, D122, M115, and A121 and chain B Y56, M115, I54, and A121 (Figure C,E). This inhibitor could potentially be expanded toward the N-terminal β-strands of chain A to enhance its potency. RIDGE-derived compound 5n2f-10 is much smaller than INCB086550 or 5n2f-17; however, it still inhibits with an IC₅₀ of 2.6 μM. 5n2f-10 is predicted to make contacts with chain A Y56 and M115 and chain B Y123, M115, and D122 (Figure D,F). Thus, interaction with this central region of the β-sandwich is sufficient for inhibition. However, expansion of the molecule to extend across the entirety of the β-sheets enhances the potency dramatically. This highlights the need for the generation of virtual libraries of larger small molecules (>500 g/mol) in virtual databases that focus on protein–protein interaction inhibitors’ discovery. In the case of PD-L1, to obtain submicromolar inhibition, 5n2f-17 is 577 g/mol and still has potential to be expanded and optimized for better inhibition. INCB086550, which does span the length of the β-sandwich, is 694 g/mol. Inclusion of larger small molecules in the virtual databases is likely to speed up the optimization process needed to build upon smaller molecules that are not as potent in large pockets often found in undruggable proteins.

2.

PD-L1 inhibitor docking models. Dimeric PD-L1 (PDB: 5n2f, chain Ayellow, chain Bgray) complexed with (A, B) INCB086550, (C) 5n2f-17, and (D) 5n2f-10. Panels (B–D) are zoomed in to highlight the interactions between the PD-L1 dimers and inhibitors. Hydrogen bonds are indicated by black lines with hydrogen-bonding atom distances in Å. (E, F) 2D ligand–receptor contact maps of (E) 5n2f-17 and (F) 5n2f-10 with the PD-L1 dimer. Different types of atomic contacts are indicated in color (redpi-pi interactions, bluehydrogen bonds, greenhydrophobic interactions, and grayvan der Waals interactions).

Enamine recently launched the REAL Protein–protein Interactions (PPIs) modulators library, a collection of 1.2 billion compounds with molecular weights ranging from 400 to 700 Da. To evaluate its effectiveness, we used RIDE to screen this new library against INCB086550 and compared the results to a screen of the entire 5.5-billion-compound REAL library. For both screens, we set a cutoff of a sum score below −80 to identify potential binders. The full REAL library yielded 114 compounds that met this criterion, with the lowest score being −94.8. In contrast, the PPI library produced 94 hits, and its lowest score was −89. Additionally, the molecular weight of the hits from the full REAL library ranged from 317 to 501 Da, while hits from the PPI library ranged between 400 and 499 Da. These findings suggest that specialized libraries like Enamine’s PPI library could significantly reduce computational effort when searching for ligands that target larger binding sites. However, the slightly less favorable scores from the PPI library screen indicate that removing smaller compounds might have also reduced the overall diversity of the library, potentially limiting the discovery of some high-scoring binders.

Activity Shows Good Correlation with RTCNN Score

Based on the ligands tested for inhibition of PD-L1, we asked whether the docking score is correlated with the activity of the ligands. Figure presents plots of % PD-1–PD-L1 binding activity in the presence of 5 μM of each ligand vs RTCNN, LE, or a Combined score (Figure A–C). We find that both RTCNN score and the Combined score are positively correlated to activity with Pearson coefficients (r) greater than 0.5 and p-values less than 0.05. The LE score is not significantly correlated with % activity in this system, most likely, because of high flexibility of the receptor and induced fit-type interaction with the ligands. Thus, for this system, the RTCNN score is the better predictor of receptor binding. On average, those molecules discovered by first screening using RIDGE (−51.09; Figure , blue) or RIDE (−49.55; Figure , green) had more favorable RTCNN scores than those discovered using VLS alone (−46.60; Figure , red). For this system, the compounds that showed the most PD-L1 inhibition tended to have more negative RTCNN and/or combined docking scores. Figure D is a plot of Combined score vs RTCNN score for each tested compound. Large points are those compounds that showed inhibition. It was observed that all compounds with an RTCNN score < −52 or a Combined score < −93 showed inhibition. Only two inhibitors, which were both derived from VLS screens had higher RTCNN and Combined scores. All compounds that showed no inhibition have both high RTCNN and Combined scores. Thus, for this system, the RTCNN and Combined scores provide a good predictor for protein binding, especially when screening ultralarge libraries with RIDGE or RIDE prior to docking with VLS.

3.

Score-activity analysis. (A–C) Correlation of docking scores vs activity for all tested compounds. % Activity for 5 μM of each compound in the PD-1:PD-L1 TR-FRET assay is plotted against (A) RTCNN score, (B) ICM VLS (LE) score, or (C) a combined score, which is the sum of the RTCNN and ICM VLS scores. A linear trendline is shown in each with corresponding Pearson correlation coefficients (r), t-statistics, and p-values. Compounds that were derived from different screening methods are shown in different colors. (D) RTCNN score vs Combined score for each compound. Color corresponds to the screening method used to derive each compound. Large circles are those compounds that show activity according to the PD-1:PD-L1 TR-FRET assay.

Identification of K-Ras G12D Inhibitors Using GPU-Based VLS Methods

K-Ras G12D is another highly sought after undruggable protein. K-Ras is an archetypically challenging target due to its smooth surface with no identifiable druggable pockets. One compound, MRTX1133, was identified as the first noncovalent, selective inhibitor of a mutant K-Ras. Unfortunately, clinical trials have been terminated after Phase I due to formulation challenges. We evaluated RIDE against the K-Ras G12D by using MRTX1133 as the template. By performing a RIDE search over the 14.35 billion compounds in Enamine REAL, SAVI, WuXi GalaXi, and ChemSpace Freedom, we found 175538 hits with an APF cutoff of 0.65. The screen that was run on 4 GPUs took 41 h. The hits were further docked into K-Ras G12D structure (PDB: 7rpz) using ICM-Pro software run on NIH’s Biowulf cluster and individually redocked on a PC to arrive at three best-scoring compounds (Table ) that were synthesized and tested experimentally. All three compounds were from Enamine’s REAL database.

4. Hits Identified by Virtual Screening and Experimental Results for K-Ras G12D Inhibition.

To assess the binding of the RIDE-derived compounds to K-Ras G12D, we used nano-differential scanning fluorimetry (nanoDSF). This technique reports on the conformational stability of a protein by monitoring the intrinsic fluorescence over a range of temperatures. The method relies on changing the environments of tyrosine and tryptophan residues upon unfolding of the protein. The peak of an emission spectrum of a protein will change when these residues are oriented away from (folded) or exposed to (unfolded) aqueous solution. Thus, by monitoring the ratio of emission at 350/330 nm at increasing temperature, we can track unfolding of the protein. K-Ras contains eight tyrosine residues that collectively contribute to the overall fluorescence. The ratio of Em_350 nm/Em_330 nm as a function of temperature for K-Ras G12D in the absence or presence of compound is shown in Figure A. The melting point is defined as the inflection point of the unfolding curve, which is determined by taking the first derivative (Figure B). It is apparent that the addition of any of the three identified hits to K-Ras G12D increases the melting point of the protein significantly, suggesting that the compounds bind to K-Ras and stabilize the folded protein conformation. The largest melting temperature shift is observed with the addition of 7rpz-2, which increases the melting temperature by 12.2 °C. Addition of 7rpz-1 and 7rpz-3 increases the melting temperature by 7.3 and 7.0 °C, respectively. This substantial change in protein stability induced by the compounds proves that they all interact with K-Ras G12D directly, with the interaction of 7rpz-2 having the largest impact on protein stability.

4.

Experimental validation of K-Ras G12D compounds. (A) Melting curves were measured for K-Ras G12D (20 μM) by nanoDSF in the absence of compounds (black) or in the presence of 50 μM 7rpz-1 (blue), 7rpz-2 (red), or 7rpz-3 (green) from 20 to 90 °C. Fluorescence plotted is the ratio of Em_350 nm/Em_330 nm. (B) Melting temperatures were measured as the first derivatives of the melting curves in (A). Dashed vertical lines guide the eye from the minimum of the first derivative to the inflection point of the melting curve. Measured melting temperatures are given in the legend. (C) MTT assay of LS 180 colorectal cells in the presence of putative K-Ras G12D inhibitors. Cells were subjected to varying concentrations of MRTX1133 (black), 7rpz-1 (blue), 7rpz-2 (red), or 7rpz-3 (green) for 48 h. % of surviving cells relative to in the absence of compound was measured by an MTT cell toxicity assay. Dashed lines are shown at 50% and 0% cell growth for reference.

To evaluate the biological activity of compounds, we assessed the cytotoxicity on LS 180 colorectal adenocarcinoma cells bearing the K-Ras G12D mutation (Figure C). MRTX1133, used as a positive control, had a submicromolar GI₅₀ in the MTT assay. All three RIDE-derived compounds also showed toxicity, with 7rpz-2 being the most toxic, having a GI₅₀ near 6 μM. This is consistent with 7rpz-2 also inducing the largest thermal shift in nanoDSF (Figure A,B). In general, the order of toxicity of the three compounds is in agreement with the changes in melting temperature measured by nanoDSF. Thus, the binding of the compounds to K-Ras G12D correlated with their biological activity.

Predicted Interactions of K-Ras G12D Inhibitors

Although the three 7rpz compounds demonstrate direct interaction with K-Ras G12D, they are less potent than MRTX1133, which was used as a template for their discovery by RIDE. This could be predicted by their increased docking scores relative to MRTX1133 and fewer contacts observed in the docked complexes. MRTX1133 makes hydrogen bonds to D12, G60, H95, R68, and D69 of K-Ras with additional stabilizing contacts to E62, Y96, M72, and Y64 (Figure A–C,F). The most potent of the novel compounds, 7rpz-2, has a binding mode similar to that of MRTX1133, as it is also predicted to make hydrogen bonds to D12, G60, H95, and R68 (Figure D,E,G). This compound also makes contacts with Y64,Y96, M72, E62, and Q99 (Figure D,E,G). The difference in potency between MRTX1133 and the novel compounds can be attributed to their insufficient ability to fill the pocket to the same extent as MRTX1133 due to their small size. Figure B,E highlights the pocket filled by MRTX1133. 7rpz-2 fits the internal part of this pocket (Figure E); however, it does not sufficiently fill the entirety of the pocket at the surface. All three compounds form an ionic interaction with the carboxyl of D12 that is critical for the selectivity toward mutated K-Ras. Stronger inhibitors can be envisioned by optimizing lead compounds by extending their structures to make contacts at the surface of K-Ras, with 7rpz-2 being particularly promising. This further supports the need for larger molecules in virtual databases to better fill the large pockets often found in undruggable proteins.

5.

K-Ras G12D inhibitor docking models. Docked models of K-Ras G12D (PDB: 7rpz) in complex with (A–C) MRTX1133 and (D, E) 7rpz-2. K-Ras G12D is shown in gray ribbon and surface representations, and the compounds are colored. ADP and an Mg2+ ion are shown in ball and stick representation. (A) Full view of K-Ras G12D in complex with MRTX1133. (B) Surface representation of K-Ras G12D in complex with MRTX1133 viewed from the top and highlighting the binding pocket filled with MRTX1133 (yellow). (C) Zoomed-in view of the contacts MRTX1133 makes with K-Ras G12D. (D) Zoomed-in view of the contacts 7rpz-2 makes with K-Ras G12D. (E) Surface representation of K-Ras G12D in complex with 7rpz-2, viewed from the top, highlighting the insufficiently filled binding pocket (yellow). Hydrogen bonds are indicated by black lines. (F, G) 2D ligand–receptor contact maps of (F) MRTX1133 and (G) 7rpz-2 with K-Ras G12D. Different types of atomic contacts are indicated in color (redpi-pi interactions, bluehydrogen bonds, greenhydrophobic interactions, and grayvan der Waals interactions).

Conclusions

Two new techniques, RIDGE and RIDE, use GPUs to accelerate structure-based or ligand-based screening of giga-sized virtual libraries. RIDGE accelerates ligand docking by ∼1000× compared to conventional docking on CPU cores and shows high precision in virtual screening benchmarks. 3D molecular similarity tool, RIDE based on atomic property fields, allows screening at a rate of 25–160 million compounds per hour, depending on the database and flexibility of the template, on a single GPU. Both methods allowed us to access orders of magnitude more molecules, up to multibillions, in ultralarge virtual libraries that are not accessible by CPU-based VLS. By applying these tools to two challenging cancer targets, PD-L1 and K-Ras G12D, we obtain hit rates over 50%, testing only 15 and 3 novel compounds, respectively. In both cases, we identify compounds with single-digit micromolar to submicromolar inhibition. This demonstrates that these methods can reliably identify active ligands in these challenging cases. Identified hits, however, do not fill large target pockets completely due to their small size that originates from the database prefiltering for drug-like properties. The data suggest that large databases optimized for nondruggable targets and containing larger small molecules with molecular weights >500 g/mol (“beyond-rule-of-5”) could facilitate drug discovery for difficult cases. However, the data also shows that even currently available billions-sized databases contain hits with novel scaffolds that can be quickly identified using GPU-accelerated methods such as RIDE and RIDGE.

Methods

In Silico Screening

Docking screens were conducted using ICM-Pro (Molsoft L.C.C., San Diego, CA) software.

RIDE (Rapid Isostere Discovery Engine)

RIDE is a GPU-accelerated version of the Atomic Property Field (APF)-based ligand superposition and ligand-based virtual screening method. The template ligand is represented by the seven-component generalized pharmacophoric property fields that describe atoms in terms of being hydrogen bond donors and acceptors, and five other descriptors that reflect lipophilicity, Sp2 hybridization, size, charge, and electronegativity. Each atom’s descriptors are projected into the surrounding 3D space using a Gaussian function. The APF field of a molecule is defined as a total of all its atoms’ Gaussian fields. Best superposition is achieved by sampling position and conformation of the second molecule and locating minima of the APF pseudoenergy:

$$E_{\mathrm{APF}}=-\sum \sum {{\varphi }_i}^j{P}_i({\mathit{r}}^j)$$

where ϕ _i is a property vector of atom j, P _i ( r ^j ) is the i-th property field at r _j , and summations are over all atoms j and their properties i. High accuracy of APF molecular alignments and its virtual screening efficiency have been independently demonstrated. In the original, CPU-based APF superposition method, Monte Carlo (MC) sampling with gradient minimization in internal coordinates is used to locate the global minimum pose of a flexible ligand. While RIDE searches minima of the same 3D chemical similarity function, it uses pregenerated ligand conformers and systematic positional sampling instead of the Monte Carlo technique to exploit massively parallel GPU computing most efficiently. A GPU is capable of simultaneously executing thousands of threads, e.g., Nvidia RTX4090, a “gamer” grade GPU available since 2022, has 16,384 CUDA cores, hundreds to thousands of times more than the regular CPU core count. However, CPU cores can operate essentially independently, often making the coarse-grain parallelization of screening algorithms trivial. On the other hand, for a GPU, computations need to be organized in a way that keeps as many cores as possible synchronously performing essentially identical tasks. Typically, this task has to be a fairly short subroutine (the so-called CUDA kernel).

In the RIDE implementation of the APF search, each GPU thread/core operates on a small subset of possible superpositions of the query and a single conformation of a database molecule, executing a kernel that evaluates APF scores for each superposition. A batch of simultaneously launched threads performs the same operation on different subsets of possible superpositions and different conformations of the database molecules. Moreover, while the GPU computes APF scores for one batch, CPU loads and prepares the next batch of data. The computational pipeline is further optimized using a highly compressed conformer database storage format based on internal coordinates, which eliminates the conformer input data transfer bottleneck.

RIDGE (RapId Docking on GPU with Energy Minimization)

RIDGE identifies putative docked poses of a ligand as the minima of the ligand–receptor interaction energy and ligand internal energy. Docked poses are searched using systematic positional sampling and pregenerated ligand conformers as starting points. Initial systematic sampling is followed by the two-stage gradient minimization. During the first stage, only position and orientation variables are optimized in the interaction potentials of the receptor. A subset of lowest interaction energy poses proceeds to the second stage where the ligand is fully flexible and the receptor interaction energy is minimized together with the internal force-field energy of the ligand (using MMFF94 force field). Receptor potentials that represent van der Waals, electrostatic, hydrogen bond, and lipophilic interactions are precalculated on grids as in the original ICM-dock docking protocol. , Grid representation of the receptor avoids expensive, explicit pairwise potential calculations across hundreds of receptor atoms. Final poses are scored using two scoring functions: (1) a neural network-based RTCNN (Radial and Topological Convolutional Neural Network) score; and (2) a physics-based ICM VLS scoring function that includes hydrogen-bonding, van der Waals, hydrophobicity, desolvation, solvation electrostatics, and entropy terms, with weights optimized for true ligand discrimination. RTCNN is used to select the top pose for each ligand, and a combination (the sum) of the two scores is used for the final ligand ranking. Both stages of gradient minimization and the RTCNN rescoring are implemented as a series of CUDA kernels that evaluate force-field energy terms and grid receptor interactions, perform iterations of the Conjugate Gradient minimizer, and evaluate RTCNN. The implementation allows calculations for thousands of poses and across multiple ligands to be performed simultaneously on a corresponding number of GPU cores. As a result, GPU affords a dramatic speedup over the original sequential Monte Carlo sampling in the standard ICM-dock.

Benchmarking RIDGE and RIDE Using Directory of Useful Decoys-Enhanced (DUD-E)

Conformers of DUD-E’s ligands and decoys have been generated using Graph Internal-coordinate Neural network conformer Generator (GINGER) run on NVIDIA RTX4090 GPUs. Receptor pockets for all targets have been defined by selecting receptor atoms within 5 Å of the corresponding ligand. Receptor structures from DUD-E were processed using a standard ICM-Pro protein preparation procedure. Default settings were used in the RIDE and RIDGE screening runs.

VLS Screens

Up to 1000 parallel processes were run on the Biowulf cluster supercomputer of the National Institutes of Health (NIH) as “swarm” jobs. Identification of pockets was done by selecting residues up to 5.5 Å from ligands in the crystal structures used, which were removed before screening libraries. PDB: 5N2F structure was used for identification of (PD-L1) hits, and PDB: 7RPZ for K-Ras G12D inhibitors. Initial fast screens were performed on ChemSpace in stock screening library containing 69,85,997 compounds. Each job screened 5000 compounds with a thoroughness of 3 and used 1 CPU per job. The hits from the fast screens were subjected to a second screen using a thoroughness of 200–300 and screening of 5 compounds per job. The top 20–30 compounds from the second screen were manually redocked to refine the pose and provide for visual inspection. Finally, manual redocking scores for individual poses and structural diversity were considered when selecting compounds for ordering for subsequent experimental testing.

Conformer Generation for RIDGE and RIDE

Conformers for ultralarge databases have been generated in compressed .molt format from SMILES files provided by database creators using ICM-Pro software run on the NIH supercomputer cluster Biowulf. Enamine’s REAL contained 6.5 billion compounds, ChemSpace’s Freedom4.95 billion, WuXi’s GalaXi1.6 billion, SAVI1.2 billion and a diversity set of the Enamine REAL database contained 48 million compounds.

RIDGE

RIDGE screens were performed using the Biowulf cluster supercomputer of the NIH. The pocket of interest was identified as described above. The diversity set of the Enamine REAL database was split into 8 parts, and each of these 8 sets was screened against the PD-L1 binding pocket in parallel using 1 NVIDIA V100-SXM2 GPU each and using an RTCNN cutoff of −45 for PD-L1 and −40 for K-Ras G12D. Hits were further screened against the PD-L1 binding pocket using CPU-based VLS screens as described above. RIDGE hits were directly screened using a thoroughness of 300 and screening of 5 compounds per job. The top compounds were manually redocked to determine which compounds to order and test experimentally.

RIDE

RIDE searches were performed using known inhibitors, INCB086550 for PD-L1 and MRTX1133 for K-Ras G12D, as templates. The screens were performed on an Exxact Linux server equipped with 4 NVIDIA RTX A6000 GPUs, 36 CPUs, 8 8TB SSDs, and 125GB RAM. For both ligands, over 14.35 billion compounds in the SAVI database, Enamine REAL database, ChemSpace Freedom database, and WuXi GalaXi database were screened. For INCB086550, an APF cutoff of 0.35 was used. For MRTX1133, an APF cutoff of 0.65 was used. Potential hits identified by these searches were then docked using CPU-based VLS screens, as described above. RIDE results were docked into their respective receptor structures by first performing an initial fast screen, in which each job screened 50–1000 compounds with a thoroughness of 5–50. Resulting hits were then run through secondary screens using a thoroughness of 200–300 and screening 5 compounds per job. Top compounds from the more thorough screens were manually redocked to determine compounds to purchase and test.

Compounds

Compounds 5n2f-1–15 and 7prz-1–3 were synthesized by Enamine (Kiev, Ukraine). Compounds 5n2f-16 and 5n2f-17 were prepared by WuXi AppTec (Shanghai, China). All compounds had a purity of more than 95% as analyzed by LC/MS.

PD-1 - PD-L1 TR-FRET Assay

The PD-1–PD-L1 TR-FRET assay kit (BPS Bioscience, San Diego, CA) was used to assess the inhibition of the PD-1–PD-L1 interaction by the compounds listed in Table . The assay was performed per the manufacturer’s protocol. All components of the kit were used as suggested by the manufacturer. All measurements were performed in triplicate in white, nonbinding 384-well microtiter plates. Compounds were stored as 5 mM DMSO stocks and were diluted to 4× the working concentration in PBS. Five μL [biotin]-PD-L1 (11 μg/mL in 1× Immuno Buffer 1) were added to 5 μL 4× compound solution (single dose assays: 100 μM or 20 μM; dose–response assays: INCB086550- 1 μM, 0.2 μM, 40 nM, 8 nM, and 1.6 nM, 5n2f-10- 100, 20, 4, 0.8, and 0.16 μM, 5n2f-17- 4 μM, 0.8 μM, 0.16 μM, 32 nM, and 6.4 nM) in corresponding Test wells and incubated at room temperature in the dark for 10 min. Before incubating, 5 μL of [biotin]-PD-L1 (11 μg/mL) and 5 μL of PBS were added to Positive Control wells, and 5 μL of 1× Immuno Buffer 1 and 5 μL of PBS were added to Negative Control wells. Dye-labeled acceptor was diluted 100-fold in 1× Immuno Buffer 1. [Eu]-PD-1 was diluted to 0.8 μg/mL in 1× Immuno Buffer 1. A 1:1 mixture of diluted dye-labeled acceptor and [Eu]-PD-1 was prepared, and 10 μL was added to every well. The plate was then incubated at room temperature in the dark for 1.5 h. Fluorescence intensity was read on a BMG Labtech CLARIOstar plate reader (BMG Labtech) capable of performing TR-FRET measurements. For an excitation wavelength of 337 nm, emission was measured at 620 and 665 nm. Data analysis was performed using the TR-FRET ratio (665 nm emission/620 nm emission × 10,000). % Activity was calculated as (Positive Control ratio – Sample ratio)/(Positive Control ratio – Negative Control ratio) × 100.

Expression and Purification of Recombinant K-Ras G12D

The plasmid encoding K-Ras G12D was generated by performing site-directed mutagenesis using the K-Ras G61H plasmid (residues 1–169) from Addgene (#25153) as the template. The plasmid was transformed into Escherichia coli BL21 (DE3)-Codon Plus RIL cells. Cell cultures were grown in 500 mL of Terrific broth at 37 °C with shaking and allowed to reach an OD600 of 0.6–0.8. Expression was induced with 1 mM IPTG at 37 °C for 3 h, and cell pellets were harvested for purification. Cells were lysed by resuspending in fresh 10 mM Tris, pH 7.5, 500 mM NaCl, 5 mM imidazole, 1 mM PMSF, 5% CHAPS, and 1X Roche Complete Protease Inhibitor EDTA-free solution. The resuspension was sonicated on ice at 30% amplitude continuous power, cycling on 30 s, off 20 s, three times. The sonicated resuspension was centrifuged at 20,000 rpm for 45 min at 4 °C to remove cell debris. The cell lysate was passed over a His60 Ni gravity column (Clontech Laboratories, Inc.) equilibrated with 10 mM Tris, pH 7.5, 500 mM NaCl, and 5 mM imidazole. Three washing steps with increasing concentrations of imidazole were used (5, 30, and 50 mM). K-Ras G12D was eluted from the column with 100 mM imidazole. Purest fractions of K-Ras G12D (as analyzed by SDS-PAGE) were exchanged into the working buffer (20 mM HEPES, pH 7.4, 250 mM NaCl, 5 mM MgCl₂, 0.5 mM TCEP) using Amicon Ultra-15 filters with 10 kDa molecular weight cutoff. Protein purity and molecular weight were determined by HPLC-electrospray mass spectrometry. To load GDP into K-Ras G12D, EDTA was added to a final concentration of 10 mM and GDP was added to a final concentration of 1 mM to the protein solution. The mixture was incubated at 37 °C for 10 min, and nucleotide exchange was stopped by adding 100 mM MgCl₂. The GDP-loaded K-Ras G12D was exchanged into the working buffer with 10 μM GDP.

NanoDSF

Melting temperature was measured by nanoDSF for K-Ras G12D protein (20 μM in 20 mM HEPES, pH 7.4, 250 mM NaCl, 5 mM MgCl₂, and 0.5 mM TCEP) in the absence or presence of each compound (50 μM) listed in Table . Compounds were diluted from 10 mM DMSO stocks. NanoDSF experiments were performed on a Prometheus NT.48 instrument (NanoTemper Technologies, Munich, Germany). All measurements were repeated in triplicate. All solutions were incubated at room temperature for 1 h prior to measurement. Approximately 10 μL samples were loaded into standard capillaries for Prometheus NT.48. Melting scans were run with 100% excitation power over a thermal gradient of 20–90 °C, increasing at a rate of 1 °C per minute. Upon excitation at 285 nm, emission at 330 and 350 nm was recorded every 0.046 °C. At the end of the experiment, the instrument was returned to room temperature without a deramping phase. Melting temperatures were determined in ThermControl software (NanoTemper) by plotting the emission ratio (350/330 nm) and deriving the inflection point from the first derivative.

Cell Toxicity Assay

LS 180 cells (ATCC, CL-187) were cultured in Dulbecco’s modified Eagle’s medium (DMEM) containing 10% fetal bovine serum (FBS) at 37 °C and 5% CO₂. For the assay, 5000 cells/well were seeded in DMEM containing 5% FBS in 96-well plates and allowed to attach for 24 h at 37 °C and 5% CO₂. Compounds were added to cells at 10, 4, 1.6, or 0.64 μM final concentration. Each condition was performed in sextuplicate. The concentration of DMSO was kept constant at 0.2% across all wells. After 48 h incubation with compounds, MTT (0.35 mg/mL final concentration) was added, and cells were incubated for 4 h. Stop solution (40% DMF, 10% SDS (w/v), 25 mM HCl, and 2.5% acetic acid in H₂O) was then added, and cells were incubated overnight at room temperature. The next day, absorbance at 570 nm was measured using a CLARIOstar plate reader (BMG Labtech, Ortenberg, Germany).

RIDE and RIDGE are proprietary software modules available as part of the Molsoft ICM-Pro + VLS package. Licensing information can be obtained directly from Molsoft LLC. (www.molsoft.com). The raw and processed data generated from the RIDE virtual screening campaigns and associated scoring data along with PDB files of the proteins’ complexes with the most active hits, 5n2f-17 and 7rpz-2, are available in the Supporting Information accompanying this article.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.5c01335.

• Top 50 hits from RIDGE screen of a 48 million diversity set of the Enamine REAL database against PD-L1 structure (Table S1) (XLSX)

• Top 50 hits from RIDE search of over the 14.35 billion compounds in Enamine REAL, SAVI, WuXi GalaXi, and ChemSpace Freedom database using G12D mutant K-Ras inhibitor MRTX1133 as a template with an APF cutoff of 0.65 (Table S2) (XLSX)

• Hits from a compound similarity search on PD-L1 inhibitor, 5n2f-17, in ChEMBL (Table S3) and references (PDF)

• Complexes of the most potent hits (PDB)

• Hits from a compound similarity search for PD-L1 inhibitor, 5n2f-17 (PDB)

§.

E.R. and C.L.H contributed equally. E.R. and M.T. designed the software and wrote the scripts. C.L.H., K.M.K., and R.B. performed virtual screens and characterized activity of compounds. E.R., M.T., and N.I.T. performed benchmarking of the software. M.T. and N.I.T. developed the concepts and planned the studies. The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

This project was funded in part with Federal funds from the Intramural Research Program of the NIH, NCI, Center for Cancer Research, and Cancer Innovation Laboratory and from Federal funds from the National Cancer Institute, National Institutes of Health, under Contract No. 75N91024F00011.

The authors declare the following competing financial interest(s): Authors Eugene Raush and Maxim Totrov are employed by Molsoft LLC.

Published as part of Journal of Chemical Information and Modeling special issue “Chemical Compound Space Exploration by Multiscale High-Throughput Screening and Machine Learning”.