Biomedical research is rapidly adopting artificial intelligence (AI). Yet the inherent complexity of biomedical data preparation requires implementing actionable, robust criteria for ethical and explainable AI (XAI) at the "pre-model" stage, encompassing data acquisition, detailed transformations, and ethical governance. Simple conformance to FAIR (Findable, Accessible, Interoperable, Reusable) Principles is insufficient. Here, we define criteria and practices for reliable AI-readiness of biomedical data, developed by the NIH Bridge to Artificial Intelligence (Bridge2AI) Standards Working Group across seven core dimensions of dataset AI-readiness: FAIRness, Provenance, Characterization, Ethics, Pre-model Explainability, Sustainability, and Computability. Conformance to these criteria provides a basis for pre-model scientific rigor and ethical integrity, mitigating downstream risks of bias and error prior to AI modeling. We apply and evaluate these standards across all four Bridge2AI flagship datasets, spanning functional genomics to clinical medicine, and encode them in machine-actionable metadata bound to the datasets. This framework sets a benchmark for preparing ethical, reusable datasets in biomedical AI and provides standardized methods for reliable pre-model data evaluation.

Abstract Introduction: AML is a molecularly heterogeneous disease that is classified by recurrent cytogenetic abnormalities and gene mutations. Recent studies have shown divergent frequencies of several genetic aberrations depending on genetic ancestry and self-reported race/ethnicity, highlighting the need to broaden sequencing efforts to include more diverse pts. Methods: We performed paired tumor/normal whole exome sequencing (WES) and transcriptome sequencing on 271 ancestry and/or ethnically diverse pts [including 100 African American (AA) and 71 self-identified Hispanic pts; CALGB/Alliance], and a validation cohort of 45 AA pts (University of Pennsylvania). Results: We identified >20 genes to be mutated in 3-8% of pts that were not seen in previous sequencing efforts of predominantly non-Hispanic White/European ancestry (EA) pts. Notably, variants in genes encoding Rho-GTPase regulatory proteins (ARHG family, belonging to the RAS gene superfamily) were identified in 12% of pts, placing these genes in the top 5 of recurrently mutated genes in this pt cohort. This frequency was confirmed in the second cohort of AA pts (n=7/45, 15%). In contrast, analysis of 805 EA adults with WES data (BeatAML 2022) and 877 pediatric AML pts (TARGET) found ARHG gene family variants in 26/805 (3%) and 17/877 (1.9%) of EA AML pts, respectively. With a median age of 41y, ARHG-mutated(m) pts tended to be younger (P=.16) and more often diagnosed with core-binding factor AML (39% vs 19%, P=.02). ARHG mutations frequently co-occurred with NRAS and FLT3 mut (each found in 35% of ARHGm pts). Notably, survival of ARHGm pts was poor, with a median overall survival (OS) of <12 months, thereby mirroring OS of the 2022 European LeukemiaNet (ELN) Adverse risk group. Within the 2022 ELN Favorable risk group in the ancestry diverse cohort, ARHGm pts had shorter OS than ARHGwt pts (P=.02). The clinical outcome was especially poor in young adolescents and adults (AYA, 18-39y) (mut vs wt; 3y disease-free survival, 10% vs 50%, P<.001; 3y OS, 20% vs 55%, P=.008). RNAseq of 17 ARHGm pts showed transcriptomic RAS pathway activation, with 52% displaying a RAS-associated signature, also in the absence of other RAS mutations. Furthermore, bulk transcriptomic analyses of 1250 AML pts identified 120 predicted RAS signature genes, with the upregulated genes being enriched in metallopeptidases, MAP kinase phosphatases, and ARHG genes. Conclusion: We identified mutations in ARHG family genes as frequent yet thus far unrecognized RAS pathway activators in AML associated with poor survival that are not yet included in clinical testing panels. Their lack of recognition is likely due to the heterogeneity of mutationally affected ARHG family genes, enrichment in AYA pts and the high frequency in pts of non-European ancestry, both of which are pt populations that were underrepresented in previous sequencing efforts. Citation Format: Ethan Hamp, Lorenz Oelschläger, Bailee N. Kain, Deedra Nicolet, Krzysztof Mrozek, Katherine E. Miller, Audrey Bollas, Michael C. Walker, Christopher J. Walker, Jill Buss, Andrea Laganson, Andrew J. Carroll, William G. Blum, Bayard L. Powell, Geoffrey L. Uy, Wendy Stock, Marina Y. Konopleva, Richard M. Stone, John C. Byrd, Martin Carroll, Tanmoy Sarkar, Akmaljon Salimov, Benjamin J. Kelly, Electra D. Paskett, Jesse J. Plascak, Shannon McWeeney, Jeffrey W. Tyner, Jeffery Klco, Nathan Salomonis, H. Leighton Grimes, Elaine R. Mardis, Ann-Kathrin Eisfeld. Novel Rho-GTPase regulatory protein gene family variants are frequent and associate with poor survival in patients (pts) with acute myeloid leukemia (AML) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 7897.

SPI1 knockdown does not disrupt dual FLT3/LSD1 inhibition synergy.

Regulon Enrichment Analysis of Primary AML Blast Bulk RNA-seq from 681 Patient Samples

Efficacy of dual FLT3/LSD1 inhibition in MOLM13 cells.

West Nile virus (WNV) infection can lead to a wide range of clinical outcomes, from asymptomatic to self-limiting febrile and serious neuro-invasive disease. Knowledge of the genetic factors contributing to the heterogeneity of this disease can inform pathogenic mechanisms, help guidethe development of cell based therapeutics, vaccines and inform lifestyle choices. Yet this knowledge is incomplete in humans. Here we present data from a large-scale experiment aimed at identifying quantitative trait loci (QTL) associated with adaptive immune cellular phenotypes observed in response to infection with WNV. This data was generated using the Collaborative Cross (CC) mouse model, previously demonstrated as a representative model for human WNV infection and homeostatic immune states. In addition, due to challenges of QTL mapping with the large number of intermediate, highly coordinated immune and virologic phenotypes, we also provide a computational pipeline for the prioritization of gene candidates designed to leverage those characteristics for downstream mechanistic studies.

Transcription Factor Enrichment Analysis of Primary AML Bulk RNA-Seq (24 Hours After Drug Treatment)

Frontline use of the BCL2 inhibitor, venetoclax, for acute myeloid leukemia (AML) has resulted in broad improvements in patients' outcomes. A major remaining challenge is the development of venetoclax resistance, frequently driven by compensatory transcriptional programs that promote cell survival and differentiation. These changes reduce dependence on BCL2 in favor of alternative anti-apoptotic BCL2 family members such as MCL1 or BCL2L1 (BCL-XL). Using CRISPR-based genome-wide perturbation screens, we investigated the genetic dependencies of venetoclax and the BCL2/BCL2L1 dual inhibitor AZD4320. We identified the N6-methyladenosine (m6A) writer RBM15, and the nucleosome remodeling and deacetylase (NuRD) complex interactor ZMYND8 as novel mediators of resistance to both venetoclax and AZD4320. Loss of RBM15 or ZMYND8 induced drug resistance, concurrent with alterations in BCL2 family expression and monocytic differentiation. Accordingly, in AML patients' samples we found reduced expression of the respective m6A or NuRD complexes was significantly associated with monocytic differentiation and ex vivo resistance to the same drugs. These findings provide critical insights into previously undescribed mechanisms of BCL2 family inhibitor resistance in AML.

Supplementary Materials and Methods

Excess over Bliss of Primary AML Blasts from 72 Patient Samples

Given the highly aggressive and heterogeneous nature of metastatic triple-negative breast cancer, molecular subtypes have been evaluated for their utility in patient stratification and therapeutic selection. Leveraging both our unique longitudinal multimodal analysis of serial tumor biopsies, as well as existing public reference cohorts, we refined clinically relevant molecular subtypes through de-novo network-based approaches. A plasma/B-cell related co-expression module emerged as a robust predictor of clinical response. Refinements of this module were significantly associated with pathological complete response and survival in the CALGB and METABRIC cohorts, as well as dramatically improving the call rate in a CLIA setting. We explored patient-specific networks to monitor individual adaptive responses to therapy, allowing for dynamic adjustments in treatment strategies. Our work supports the shift from traditional molecular subtyping towards a more integrated view that includes the tumor microenvironment and immune landscape in a network-based context.

Correlation of Regulon Enrichment Signatures with MYC Gene Expression in Primary AML

Super-Enhancer-Type Analysis of STAT5 ChIP-Seq Peaks (24 Hours After Drug Treatment).

ABSTRAT:MDM2 inhibitors are promising therapeutics for acute myeloid leukemia (AML) with wild-type TP53. Through an integrated analysis of functional genomic data from primary patient samples, we found that an MDM2 inhibitor, idasanutlin, like venetoclax, is ineffective against monocytic leukemia (French-American-British [FAB] subtype M4/M5). To dissect the underlying resistance mechanisms, we explored both intrinsic and extrinsic factors. We found that monocytic leukemia cells express elevated levels of CEBPB, which promote monocytic differentiation, suppress CASP3 and CASP6, and upregulate MCL1, BCL2A1, and the interleukin (IL-1)/tumor necrosis factor alpha (TNF-α)/NF-κB pathway members, thereby conferring drug resistance to a broad range of MDM2 inhibitors, BH3 mimetics, and venetoclax combinations. In addition, aberrant monocytes in M4/M5 leukemia produce elevated levels of IL-1 and TNF-α, which promote monocytic differentiation and upregulate inflammatory cytokines and receptors, thereby extrinsically protecting leukemia blasts from venetoclax and MDM2 inhibition. Interestingly, IL-1β and TNF-α only increase CEBPB levels and protect M4/M5 cells from these drugs but not M0/M1 leukemia cells. Treatment with venetoclax and idasanutlin induces compensatory upregulation of CEBPB and the IL-1/TNF-α/NF-κB pathway independent of the FAB subtype, indicating drug-induced compensatory protection mechanisms. The combination of venetoclax or idasanutlin with inhibitors that block the IL-1/TNF-α pathway demonstrates synergistic cytotoxicity in M4/M5 AML. As such, we uncovered a targetable positive feedback loop that involves CEBPB, IL-1/TNF-α, and monocyte differentiation in M4/M5 leukemia and promotes both intrinsic and extrinsic drug resistance and drug-induced protection against venetoclax and MDM2 inhibitors.

Even though head and neck squamous cell carcinoma (HNSCC) is the seventh most common cancer worldwide, there are only two PD-1 targeted immunotherapies (pembrolizumab and nivolumab) and one tumor intrinsic EGFR targeted therapy (cetuximab) that are FDA approved for treatment of HNSCC. Taking advantage of a high throughput inhibitor assay and computational tools originally showing success in leukemia, we designed and employed HNSCC-specific inhibitor panels that capture the diversity of aberrational pathways in HNSCC to test viable cells derived from patients' HNSCC tumors. This provides a functional context to the multi-omic readouts conducted on these samples (mutations, protein expression and copy number alterations). In addition to generating these deeply characterized functional genomics datasets, we also developed additional visual analytics that have the potential to provide greater insight into HNSCC drug response patterns and potentially aid precision oncology tumor boards in evaluation and assessment of effective targeted therapeutic agents.

GFI1 knockdown weakens the effects of dual FLT3/LSD1 inhibition.

Background: A water extract of the Ayurvedic plant Centella asiatica (L.) Urban, family Apiaceae (CAW), improves cognitive function in mouse models of aging and Alzheimer’s disease and affects dendritic arborization, mitochondrial activity, and oxidative stress in mouse primary neurons. Triterpenes (TT) and caffeoylquinic acids (CQA) are constituents associated with these bioactivities of CAW, although little is known about how interactions between these compounds contribute to the plant’s therapeutic benefit. Methods: Mouse primary cortical neurons were treated with CAW or equivalent concentrations of four TT combined, eight CQA combined, or these twelve compounds combined (TTCQA). Treatment effects on the cell transcriptome (18,491 genes) and metabolome (192 metabolites) relative to vehicle control were evaluated using RNAseq and metabolomic analyses, respectively. Results: Extensive differentially expressed genes (DEGs) were seen with all treatments, as well as evidence of interactions between compounds. Notably, many DEGs seen with TT treatment were not observed in the TTCQA condition, possibly suggesting CQA reduced the effects of TT. Moreover, additional gene activity seen with CAW as compared to TTCQA indicates the presence of additional compounds in CAW that further modulate TTCQA interactions. Weighted Gene Correlation Network Analysis (WGCNA) identified 4 gene co-expression modules altered by treatments that were associated with extracellular matrix organization, fatty acid metabolism, cellular response to stress and stimuli, and immune function. Compound interaction patterns were seen at the eigengene level in these modules. Interestingly, in metabolomics analysis, the TTCQA treatment saw the highest number of changes in individual metabolites (20), followed by CQA (15), then TT (8), and finally CAW (3). WGCNA analysis found two metabolomics modules with significant eigenmetabolite differences for TT and CQA and possible compound interactions at this level. Conclusions: Four gene expression modules and two metabolite modules were altered by the four treatment types applied. This methodology demonstrated the existence of both negative and positive interactions between TT, CQA, and additional compounds found in CAW on the transcriptome and metabolome of mouse primary cortical neurons.

Epigenetic impact of dual FLT3/LSD1 inhibition.

Efficacy of dual FLT3/LSD1 inhibition in primary AML samples.

Transcription Factor Enrichment Analysis of MOLM13 Bulk RNA-Seq (24 Hours After Drug Treatment)