Fig S2. A-C. Recursive feature elimination analysis. Mean squared error for each model is shown for a range of total number of features included in a given model. Optimal model size is indicated by a dashed red line. D-E. Pearson correlation matrices of features respective to germline and somatic models.

Fig S9. Kaplan-Meier curve of overall survival of responders only in MHC-II reliant patients vs MHC-I reliant patients for discovery group with Van Allen cohort removed.

Fig S10. A. Univariate analysis of potential checkpoint inhibitors in the discovery samples. B. Proportion of patients responding in each MHC reliance group when discovery samples are partitioned according to low versus high expression of 7 different checkpoint genes. C. Univariate analysis of potential checkpoint inhibitors in the validation samples. D. Proportion of patients responding in each MHC reliance group when validation samples are partitioned according to low versus high expression of 7 different checkpoint genes

Fig S6. Difference in feature importance rankings of composite model features in linear (left) versus nonlinear (right) models.

Fig S11. A. Checkpoint expression and neoantigen levels by MHC Reliance category in discovery samples. B. Checkpoint expression and neoantigen levels by MHC Reliance category in validation samples. C. Pearson correlation between various checkpoint genes across discovery samples. D. Pearson correlation between various checkpoint genes across validation samples.

Supplementary Figure 5: the orthotopic mouse model

Fig S5. (A-C). Measures of TIME infiltrates stratified by high vs low composite IC-Index scores. T-tests were used to compare means between groups. (D-E) Confusion matrices of somatic and germline IC-Indices (cutoff >=5) vs TIME score (cutoff=median).

Fig S4. A-C. Pearson correlation of three different IC-Index scores with TMB. D. Purity vs ploidy across all patients, colored by response with pearson correlation shown. E-G. ROC plots comparing the performance of composite IC-Index to TMB and clinical predictors of ICB response. H-I. ROC comparing the performance of composite IC-Index to transcriptomic predictors of ICB response J. Histogram of composite IC-Index across patients, colored by ICB response status.

PURPOSE:Malignancies have complex and distinct molecular profiles that may not segregate by tumor type. However, most precision oncology treatments are matched to a single biomarker. We aimed to optimize therapy for advanced cancers using individually dosed drug regimens customized to cotarget multiple molecular alterations. METHODS:Investigation of Profile-Related Evidence Determining Individualized Cancer Therapy (I-PREDICT; NCT02534675) is a prospective, investigator-initiated, multidepartment/pan-cancer trial for aggressive advanced/metastatic malignancies. Patients had tissue and/or blood next-generation sequencing (NGS; Foundation Medicine). A molecular tumor board made suggestions. Degree of biomarker matching to drugs given was calculated by a matching score (MS; broadly, number of pathogenic alterations targeted divided by total pathogenic alterations). RESULTS:Overall, 210 evaluable patients (n = 456 consented) received ≥1 US Food and Drug Administration-approved drug (mostly off label) after NGS. Median number of pathogenic alterations/tumor was five (range, 0-20); approximately 95% of patients had unique molecular landscapes. Consistent with I-PREDICT's objective to optimize/tailor treatment for each patient, we administered 157 different regimens (including 103 personalized combinations without established safety/dosing data). For previously unstudied combinations, starting doses were reduced and titrated to tolerance (intrapatient dose-finding); only 6.5% experienced Grade 3/4 drug-related toxicities (v 15.5% of those receiving established regimens). Higher disease control rate (stable disease ≥6 months/objective response), and longer progression-free survival and overall survival correlated significantly/independently/linearly with greater degrees of drug matching to alterations (higher MS), but did not vary by drug number or dosages. CONCLUSION:The I-PREDICT strategy of maximizing personalized biomarker matching with individually dosed customized drug combinations enabled safe and active N-of-1 matched treatment, including regimens previously unstudied in Phase I trials. I-PREDICT represents a blueprint for a new personalized precision oncology paradigm, which merits validation via additional prospective trials.

Fig S3. A-C. SHAP derived nonlinear feature importance beeswarm plots of germline, somatic, and composite models. D-E. Correlation of immunogenicity features with TMB. F-G. Coefficients of features from linear regression-based germline and somatic models

Fig S7. A. Waterfall plot of ratio of well-presented MHC-II to MHC-I neoantigens across discovery patients. B. Response rates by reliance groupings across discovery patients. Chi squared tests were performed between groups. C. MHC-I presentation pathway damage as a proportion of total number of mutations split by MHC Reliance grouping. T-tests were used to compare groups. D. Proportion of patients with any damage to MHC-I antigen presentation pathway split by MHC Reliance grouping. Chi squared tests were performed between groups. E. Immunoediting in MHC-II Reliant patients measured by SOPRANO immune dN/dS. T test was used to compare means between patients with and without MHC-I pathway damage. F. TFH cell infiltration estimates stratified by MHC-reliance grouping and ICB response. Mann-Whitney U tests were used to compare means between groups.

Supplementary Figure 4: Synergistic effects of Metformin and Dichloroacetate (DCA) in HNSCC cell lines

Supplementary Figure 3: Metabolic profiling of CAL27 Control and HRAS mutant cells reveals changes in ATP production and oxygen consumption rates.

Abstract Precancers represent a diverse collection of lesions that occupy an intermediate state between normal tissues and malignant tumors, yet their biological and clinical definition remains unsettled. To provide a genomic foundation for refining this continuum, we analyzed 1,495 precancers across 17 tissue types using uniformly processed whole-genome and whole-exome sequencing data. This pan-tissue resource reveals striking heterogeneity in the mutational and structural features of precancers, ranging from genomically quiet lesions with minimal alterations to highly aberrant genomes resembling those of invasive cancers. Mutational signature analysis identifies 27 single-base substitution signatures, all previously observed in cancer, indicating that the mutational processes active in malignant cancer are often active at the earliest stages of tumorigenesis. Driver-gene analyses revealed 100 unique genes under positive selection, with TP53 and CDKN2A emerging as the most frequently inactivated, often via biallelic “double-hit” events preceding invasion. Comparative analyses with corresponding cancers showed a progressive enrichment of copy-number alterations, driver mutations, and mutational signatures associated with defective DNA repair and exogenous carcinogens. Random-forest models highlight that cancers converge on a state of extensive genomic disruption, irrespective of the particular mutational processes involved, indicating that canonical tissue-specific driver phenotypes emerge from a backdrop of generic accumulation of genomic alterations that distinguish malignant transformation across tissues. Collectively, this unified atlas of human precancers clarifies the genomic transitions from benign to malignant states, and suggests that many canonical “cancer drivers” may act primarily as general fitness-enhancing alterations detectable well before invasion. These findings provide a genomic framework to support emerging efforts toward molecularly informed early detection and precision prevention strategies in oncology. Citation Format: Christopher D. Steele, Yudou He, Scott M. Lippman, Ludmil B. Alexandrov. The genomic landscape of likely human precancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(8_Suppl):Abstract nr LB401.

Fig S8. A. MHC-I reliant patients in TCGA with melanoma, non-small cell lung cancer or renal cell carcinoma stratified by CD4/CD8 T Cell infiltration ratio. B. MHC-II reliant patients in TCGA with melanoma, non-small cell lung cancer or renal cell carcinoma stratified by CD4/CD8 T Cell infiltration ratio. P Values generated via log-rank test. C. Immune infiltrate adjusted HLA DRB1 expression is associated with prolonged survival in MHC-II Reliant TCGA patients. D. Immune infiltrate adjusted HLA DRB1 expression is not protective in MHC-I Reliant TCGA patients.

Supplementary Figure 1: cancer-associated HRAS mutations in HNSCC using data from The Cancer Genome Atlas (TCGA) and Represetive images of the clonogenic assays.

Head and neck squamous cell carcinoma (HNSCC) is among the 10 most common cancers worldwide and is associated with high morbidity and poor survival. Diminished HNSCC outcomes are often related to delayed diagnosis and treatment of occult progression of premalignant lesions, underscoring the need for effective and low-risk chemoprevention strategies. In this regard, metformin has shown promising clinical activity for HNSCC prevention. In this study, we performed a genome-wide CRISPR/Cas9 screen of metformin-treated HNSCC cells and identified the activation of PKA signaling as the top resistance pathway. We show that metformin mediates PKA activation in HNSCC cells and that PKA inhibition, when combined with metformin treatment, synergistically inhibits HNSCC growth. We found that metformin-induced PKA activation is mediated by a prostaglandin E2 autocrine loop, which can be blocked using cyclooxygenase-2 (COX2) inhibitors. Importantly, COX2 inhibition using nonsteroidal anti-inflammatory drugs (NSAID) combined with metformin treatment synergistically inhibits HNSCC cell growth and prevents the progression of oral premalignant lesions into invasive HNSCC in a model of tobacco-driven oral carcinogenesis. Together, these findings demonstrate that metformin and NSAID combination therapy may represent a promising therapeutic strategy for HNSCC chemoprevention. PREVENTION RELEVANCE:Our findings reveal that using metformin for head and neck cancer chemoprevention leads to compensatory activation of a PKA-driven resistance mechanism that can be blocked by cotreatment with NSAIDs. These findings provide a rationale for combining metformin with NSAIDs as a precision head and neck cancer chemoprevention strategy.

Fig S1. A. eQTL-score schematic. SNPs alleles are oriented so that they are all in alignment in terms of direction of effects on gene expression. Number of alleles affecting gene expression are then summed into a continuous score B. Pearson correlation of Gene level eQTL-scores in TCGA and Discovery cohorts. P<=0.05 is indicated by an X.