2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Technology and Procedures, Assays, Imaging, Omics Technologies, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Combination Therapy, Drug Development, Biological Therapies, Antibody Therapy, Immunotherapy, Monoclonal Antibody Therapy, Non-Biological Therapies, Pharmacology, Artificial Intelligence (AI), Machine Learning, Biological Processes, Apoptosis, Metabolism, Molecular Biology, Diseases, Lymphoid Malignancies, Myeloid Malignancies, Acute Myeloid Malignancies, Research, Translational Research, Study Population, Animal Model
PAPER

2559 Machine learning-based structural prediction studies inform the identification of “druggable” pockets in previously understudied myeloma-preferential dependencies.

Evangelos Papadopoulos, Senior Scientist, Ricardo De Matos Simoes, PhD2, Olga Dashevsky, PhD3, Huihui Tang, PhD4, Shizuka Yamano5, Eleni Anastasakou6, Vikas Gupta, MD, PhD7, Jonathan Licht, MD8, Benjamin Gabriel Barwick, PhD9, Lawrence H Boise, PhD10, Bertal Aktas11, Yuefang Deng12, Constantine Mitsiades, MD, PhD13
1 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
2 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
3 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
4 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
5 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
6 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States
7 Emory Winship Cancer Institute, Atlanta, Georgia, United States
8 The Van Andel Family Pavilion, Ada, Michigan, United States
9 Emory Winship Cancer Institute, Atlanta, Georgia, United States
10 Emory Winship Cancer Institute, Atlanta, Georgia, United States
11 Brigham & Women's Hospital, Boston, Massachusetts, United States
12 Stony Brook University, Stony Brook, New York, United States
13 Broad Institute of MIT & Harvard, Cambridge, Massachusetts, United States, Dana-Farber Cancer Institute, Boston, Massachusetts, United States, Harvard Medical School, Boston, Massachusetts, United States, Ludwig Center at Harvard, Boston, Massachusetts, United States

We previously applied genome-scale CRISPR knockout studies across various cell lines to identify genes that are highly potent and recurrent genetic dependencies for specific cancer types (de Matos Simoes et al., Nature Cancer 2023). Through this effort, we identified 116 genes that are preferentially essential for multiple myeloma (MM) cells compared to other hematologic malignancies or solid tumors. These MM-preferential dependencies (MM-PrefDeps) largely lack known pharmacological ligands and most of them have not been formally examined in drug screens. To accelerate therapeutic development, the current study sought to characterize "druggable" pockets within these MM-PrefDeps and establish an in silico framework to nominate candidate small-molecule inhibitors or degraders.

Protein sequences for MM-PrefDeps were extracted from UniProt, and predicted structures were obtained using the AlphaFold database and a local implementation of AlphaFold2 for cross-validation. Each predicted structure was annotated based on AlphaFold confidence scores and visualized through pseudo-GenBank files created with custom Biopython scripts and Geneious software to differentiate structured from unstructured protein regions. Of the 116 MM-PrefDeps, 82 had at least one experimental PDB structure, while 34 had none. Across 83,258 amino acids analyzed, 50,046 (60.1%) exhibited high-confidence AlphaFold pLDDT scores of ≥ 70, with 23,377 of these residues lacking experimental PDB coverage across 78 proteins. For regions lacking experimental PDB coverage, we developed a quality-filtering script to retain high-confidence AlphaFold regions. Short flanking overhangs were included to preserve local structural continuity and support accurate pocket identification. To reliably identify binding sites in these regions, we used three complementary pocket-identification methodologies: AutoSite, FPocket, and the Difference-of-Gaussians–based method DoGSite. This consensus approach was designed to maximize the reliability of identified pockets by filtering for agreement across different algorithms and a common-residue overlap fraction of more than 50% for at least one constituent pocket. We identified a prioritized set of approximately 100 consensus pockets across 65 proteins, with an expanded analysis encompassing 330 selected pockets across 103 proteins. Notably, among the top 100 prioritized consensus pockets, 46 lacked residue-level PDB coverage. Of the latter, 27 consisted entirely of residues with pLDDT ≥70, while another 18 contained a mixture of high-confidence residues and flanking overhang residues retained for structural continuity.

To further benchmark our pipeline, we compared the MM-PrefDep consensus pockets with AF2BIND predictions at the residue level. AF2BIND provided predicted pocket clusters for 79 of the 116 MM-PrefDeps. Of these, 39 showed residue-level overlap with a top-100 MM-PrefDep pocket and 61 with at least one pocket in the expanded 330-pocket set. Conversely, our pipeline identified top-100 pockets in 18 genes (increasing to 33 in the expanded analysis) that lacked an AF2BIND pocket. Importantly, 10 top-100 pockets had no residue overlap despite AF2BIND predictions for the same proteins, illustrating the distinction between prediction availability and residue-level concordance.

Target-binding regions derived from high-score consensus pockets were used for small-molecule docking via AutoDock Vina. We screened ~300,000 molecules from the NCI/DTP Open Chemicals Repository and initiated expanded docking of ~3.6 million commercially available drug-like compounds. As a representative validation, our pipeline identified an EP300 (p300) consensus pocket that overlaps the experimentally determined binding cavity of the p300 inhibitor inobrodib (PDB 8FVF), supporting the reliability of our unbiased in silico pipeline.

This study establishes a structural “druggability” atlas for MM-PrefDeps, identifying critical primary-sequence regions and three-dimensionally defined binding sites. By integrating machine learning-based structural prediction with consensus pocket identification and large-scale docking, this framework provides essential groundwork for the development of inhibitors against previously unexplored MM dependencies. We envision that this pipeline can be applied broadly to nominate potential pharmacological modulators for critical dependencies in other neoplasias beyond MM.

Evangelos Papadopoulos, Senior Scientist: No relevant conflicts of interest to declare.