The crystal structures of the IDH1^R132H in complex with an inhibitor were downloaded from the PDB database (http://www.rcsb.org), including 4UMX, 5L57, 5L58, 5LGE, 5SUN, 5SVF, 5TQH, 6ADG, and 6B0Z. For each complex, the Protein Preparation Wizard module in Schrödinger 2015 (Schrödinger, LLC, New York, NY, 2015) was applied to add hydrogen and missing side chains, remove all water molecules, assign protonation states and partial charges through OPLS2005 force field (Jorgensen et al., 1996), and minimize all heavy atoms until the root-mean-square deviation (RMSD) was reached ≤0.3 Å.

To evaluate the virtual screening capability of different crystal structures, 423 actives were directly extracted from the PubChem database and served as a validation data set (https://pubchem.ncbi.nlm.nih.gov/bioassay/1344832#section=Top), and their decoys, generated by DUD•E (Mysinger et al., 2012), were considered as a decoy data set. In total, 23,900 decoys were generated.

In order to discover the favorable crystal structure for virtual screening, the docking performance of each IDH1^R132H structure was systematically evaluated. All actives and decoys were preprepared using the LigPrep (LigPrep, Schrödinger, LLC, New York, NY, 2015) module in the Schrödinger package. The possible ionized states of each compound were calculated by using Epik (Shelley et al., 2007) at pH= 7.0 ± 2.0. The chirality of the IDH1^R132H inhibitors with 3D structures were preserved, while the chirality of the decoys was determined from 3D structures based on the different combinations. The stereoisomers for each ligand generated, at most, 32, and the other parameters were set to default values. Subsequently, a grid box of each complex was generated by using the Receptor Grid Generation module of Schrödinger software, which was centered at the native ligand of the complex and defined as a similar size to the native ligand space. Finally, all chemicals in the validation set and decoy set were docked into the binding site of each IDH1^R132H complex in turn and evaluated by using the standard precision (SP) and extra precision (XP) scoring function of Glide. In order to choose the best crystal structure of IDH1^R132H for virtual screening, the enrichment factor (EF) (Halgren et al., 2004) was used to evaluate the virtual screening capability of each model, which was defined as the following Equation:

where Hits_set is the number of actives in the selected subset n of the ranked database and Hits_all is the total number of actives in the database. The model with the highest EF value was reserved and used to screen potential IDH1^R132H inhibitors.

All compounds in the ChemDiv and Specs database were first preprocessed according to the method of the above step, and then screened by docking-based virtual screening against two IDH1^R132H crystal structures (PDB ID: 5TQH and 6B0Z). After the possible ionized states and tautomer were calculated at pH=7.0 ± 2.0 by using Epik module, the chirality of each compound was determined from 3D structures; the stereoisomer for each ligand generated, at most, 32. The final virtual screening library was generated to include approximately 2 million compounds, and then initially filtered by Lipinski’s Rule, removing ligands with reactive functional groups. Finally, docking-based virtual screening was employed by use of the high throughput virtual screening (HTVS) scoring function, SP scoring function, and XP scoring function of Glide in sequence. In the screening process at each step, 10% of the best compounds were reserved for further analysis.

The reserved compounds after docking-based virtual screening were structurally clustered into 30 clusters by using K-means clustering on the MACCS structural keys in Canvas (Canvas, Schrödinger, LLC, New York, NY, 2015), and the compound in each cluster with the lowest docking score was selected. In the end, 60 chemicals were eventually submitted to purchase from Topscience Co., Ltd (https://www.tsbiochem.com).

The primary assay was carried out in 10 μL of base buffer (10 mM MgCl2, 20 mM Tris pH7.5, 150 mM NaCl, 0.05% (w/v) bovine serum albumin) containing 2.5 μL of the test compound, 5 μL of an enzyme solution (0.3 ng/μL mutant IDH1^R132H), and 2.5 μL of a substrate solution (4 mM α-KG, 16 μM NADPH). This assay added into a 384-well blank plate and then incubated at room temperature for 60 min. The secondary assay, with 5 μL of base buffer containing 15 μM resazurin and 0.01 unit diaphorase, was added to the entire plate and incubated at room temperature for 10 min. Florescence was read on a SYNERGY^H1 microplate reader (BioTek) at Ex 540 Em 590. Curve fitting for dose response IC₅₀ was done using GraphPad Prism.

As a significant indicator of the docking reliability, docking power was used to reveal the binding pose of the experiment between small molecules and proteins, which was mainly evaluated after redocking with the RMSD value of the docking pose and native pose of the small molecule in the IDH1^R132H complex. For each IDH1^R132H complex, after the native inhibitor was separated from the corresponding complex and preprepared, it was redocked into the original binding site. The RMSD value between the native conformation of the inhibitor and the docked pose for each crystal structure was respectively computed, and RMSD ≤ 2.0 Å served as the evaluation standard to verify the docking reliability. It can be seen from Table 1 that Glide docking could identify the near-native pose of most inhibitors in IDH1^R132H crystal structures by using the XP or SP scoring function in Glide.

Next, screening power of glide docking was used to identify the reported inhibitors from the decoys in nine IDH1^R132H complexes, and these were compared and calculated. In contrast with the docking power of glide docking, the screening power of each crystal structure is a more important index for the docking-based virtual screening process. Herein, we performed student’s t test to evaluate the significant difference between the means of the two distributions of the Glide XP or SP scores for the known actives and decoys. It can be seen from Table 2 that molecular docking of Glide can efficiently discriminate the IDH1^R132H inhibitors from the decoys in nine complexes of IDH1^R132H based on the relatively low p value. The area under the receiver operating characteristic curve (AUC-ROC), EF, and Robust Initial Enhancement (RIE) were also employed to comprehensively evaluate the screening capabilities of each crystal structure. As shown in Figure 3 , the best screening power (p value = 2.42x10^-135, AUC-ROC=0.96, and RIE= 12.88) was acquired by using SP scoring function and 6B0Z was reserved as the screening template. However, 5TQH exhibited the best screening power (p value = 6.09x10^-102, AUC-ROC=0.96, and RIE= 15.76) in XP scoring function, which was also retained as a screening complex. Our results suggest that that it is necessary to compare the performance of different complexes in the process of virtual screening.

The overall workflow of structure-based virtual screening was shown in Figure 2 . The Specs and ChemDiv database, which consisted of more than 2,100,000 compounds, have been used for virtual screening of small molecule databases. Firstly, Lipinski’s rules of five was employed to filter compounds that did not meet the criteria, and then these compounds containing PAINS substructures were also removed. A total of 1.46 million compounds were retained. According to the MACCS structural fingerprint, residual chemicals were structurally clustered in 30 clusters via K-means clustering in Canvas, and the chemical with the lowest docking score in each cluster was retained was retained (see Table 3 ). Finally, a total of 57 chemicals were purchased and tested based on the docking-based virtual screening on two IDH1 complex (5TQH and 6B0Z).

To verify the inhibitory activity of screening compounds targeting IDH1^R132H, enzyme activity assay was performed. As shown in Figure 4 , we found that 12 compounds (7, 8, 14, 23, 25, 33, 44, 46, 47, 52, 53, and 57) exhibited over 50% inhibition at 50 μM. These 12 ligands were submitted to determine the IC₅₀. It can be seen from Table 4 that 10 of them show IC₅₀ ≤ 50 μM. Molecular structures of the 10 selected compounds of IDH1^R132H are exhibited in Figure 5 . The enzymatic curves and docking score for these 10 compounds against IDH1^R132H are depicted in Figure 6 . Tanimoto coefficient (Tc) (Willett and Winterman, 1986; Willett et al., 1986), in terms of the ECFP4 fingerprint, was calculated to compare the structural similarity between 10 compounds and reported inhibitors. As shown in Figure S1 , we can find that putative hits have low similarity with reported inhibitors (Tc < 0.2). Therefore, these compounds are structurally new and have the potential to be promising leads for further optimizations.

The binding pocket of IDH1 lies on the dimer interface and most of the reported compounds have been shown to bind to this allosteric site. In order to gain insight into the structural basis of the identified IDH1^R132H inhibitor, the binding mode of the three compounds was compared with AG-120. As shown in Figure 7 , three molecules that could be docked into the binding pocket consisted of hydrophobic residues in a similar manner to AG-120, and formed intermolecular hydrogen bonds with key residues, which stabilized the complex. L806-0255 and V015-1671 form a key hydrogen bond with ILE128, which is consistent with AG-120. In addition, V015-1671 and AQ-714/41674992 also form a key hydrogen bond with ALA111. Moreover, the hydrophobic contacts formed between surrounded residues, such as VAL276, SER278, SER287, ILE128, PRO118, and compounds also contribute to enhanced binding of the small molecule inhibitor to IDH1 ^R132H. Therefore, the above results suggested that L806-0255, V015-1671, and AQ-714/41674992 could bind to IDH1^R132H.