Peter Winter

Principal Investigator

Broad Institute
Country flagUnited States
ApplyKite TwinLearns as you apply

Make your next move.

Prepare your application

Review drafts before you send them.

Explore related searches

Articles (12)

The vision of a “virtual cell”—a computational model that simulates biological function across modalities and scales—has become a defining goal in computational biology. While powerful unimodal foundation models exist, the lack of large-scale paired data prohibits the joint training of multimodal approaches. This scarcity favors compositional foundation models (CFMs): architectures that fuse frozen unimodal experts via a learned interface. However, it remains unclear when this multimodal fusion adds task-relevant information beyond the strongest unimodal representation and when it merely aggregates redundant signal. Here, we introduce the Synergistic Information Score (SIS), a metric grounded in partial information decomposition (PID), that quantifies the information gain achievable only through cross-modal interactions. Extending theoretical results from self-supervised learning, we show that standard alignment-based fusion objectives on frozen encoders inherently collapse to detecting linear redundancies, limiting their ability to capture nonlinear synergistic states. This distinction is directly relevant for tasks aiming to link tissue morphology and gene expression. Benchmarking ten fusion methods on spatial transcriptomics datasets, we use SIS to demonstrate that tasks dominated by linear redundancies are sufficiently served by unimodal baselines, whereas complex niche definitions benefit from synergy-aware integration objectives that enable cross-modal interactions beyond linear alignment. Finally, we perform a scaling analysis which highlights that fine-tuning a dominant unimodal expert is the most sample-efficient path for standard tasks, suggesting that the benefits of multimodal frameworks only emerge when tasks depend on information distributed across modalities. Together, these results establish that building towards a virtual cell will require a fundamental shift from alignment objectives that emphasize shared structure to synergy-maximizing integration that preserves and exploits complementary cross-modal signal.

Year:

2026

While machine learning models trained on single-cell transcriptomics data have shown great promise in providing biological insights, existing tools struggle to effectively model underrepresented and out-of-distribution cellular features or states. We present a generalizable Adaptive Resampling (AR) approach that addresses these limitations and enhances single-cell representation learning by resampling data based on its learned latent structure in an online, adaptive manner concurrent with model training. Experiments on gene expression reconstruction, cell type classification, and perturbation response prediction tasks demonstrate that the proposed AR training approach leads to significantly improved downstream performance across datasets and metrics. Additionally, it enhances the quality of learned cellular embeddings compared to standard training methods. Our results suggest that AR may serve as a valuable technique for improving representation learning and predictive performance in single-cell transcriptomic models.

Year:

2025

Accurately annotating cell types is essential for extracting biological insight from single-cell RNA-seq data. Although cell types are naturally organized into hierarchical ontologies, most computational models do not explicitly incorporate this structure into their training objectives. We introduce a hierarchical cross-entropy loss that aligns model objectives with biological structure. Applied to architectures ranging from linear models to transformers, this simple modification significantly improves out-of-distribution performance (12–15%) without added computational cost.

Year:

2025

Pancreatic ductal adenocarcinoma (PDAC) has been classified into classical and basal-like transcriptional subtypes by bulk RNA measurements. However, recent work has uncovered greater complexity to transcriptional subtypes than was initially appreciated using bulk RNA expression profiling. To provide a deeper understanding of PDAC subtypes, we developed a multiplex immunofluorescence (mIF) pipeline that quantifies protein expression of six PDAC subtype markers (CLDN18.2, TFF1, GATA6, KRT17, KRT5, and S100A2) and permits spatially resolved, single-cell interrogation of pancreatic tumors from resection specimens and core needle biopsies. Both primary and metastatic tumors displayed striking intratumoral subtype heterogeneity that was associated with patient outcomes, existed at the scale of individual glands, and was significantly reduced in patient-derived organoid cultures. Tumor cells co-expressing classical and basal markers were present in > 90% of tumors, existed on a basal-classical polarization continuum, and were enriched in tumors containing a greater admixture of basal and classical cell populations. Cell–cell neighbor analyses within tumor glands further suggested that co-expressor cells may represent an intermediate state between expression subtype poles. The extensive intratumoral heterogeneity identified through this clinically applicable mIF pipeline may inform prognosis and treatment selection for patients with PDAC. Significance: A high-throughput pipeline using multiplex immunofluorescence in pancreatic cancer reveals striking expression subtype intratumoral heterogeneity with implications for therapy selection and identifies co-expressor cells that may serve as intermediates during subtype switching.

Year:

2025

The success of transformer-based foundation models on natural language and images has motivated their use in single-cell biology. Single-cell foundation models have been trained on increasingly larger transcriptomic datasets, scaling from initial studies with 1 million cells to newer atlases with over 100 million cells. This study investigates the role of pre-training dataset size and diversity on the performance of single-cell foundation models on both zero-shot and fine-tuned tasks. Using a large corpus of 22.2 million cells, we pre-train a total of 400 models, which we evaluate by conducting 6,400 experiments. Our results show that current methods tend to plateau in performance with pre-training datasets that are only a fraction of the size of current training corpora.

Year:

2024

Supplemental Table 7A: Clinical evaluation and characterisitcs of KPC mice treated until clinical endpoint. Supplemental Table 7B: Clinical evaluation and characterisitcs of KPC mice treated for 3 days.

Year:

2024

Supplemental Table 5A: Drug sensitivity metrics from MRTX1133-treated KRAS G12D patient-derived organoids. Supplemental Table 5B: MRTX1133 dose response across KRAS G12D mutated patient-derived organoids.

Year:

2024

Supplemental Figure 1: Acquired resistance to KRAS G12C inhibition in PDAC and other GI cancers. Supplemental Figure 2: MRTX1133 sensitivity across KRAS G12D mutant in vitro models of PDAC. Supplemental Figure 3: Isogenic models of acquired resistance to MRTX1133. Supplemental Figure 4: In vivo treatment and tumor monitoring for the KPC PDAC mouse model. Supplemental Figure 5: Genomic characterization of KPC tumors. Supplemental Figure 6: snRNA-seq quality metrics and description of the tumor microenvironment. Supplemental Figure 7: Identification and characterization of malignant cell populations. Supplemental Figure 8: Characterization of malignant metaprograms in KPC tumors. Supplemental Figure 9: Treatment with MRTX1133 induces modest changes in the immune microenvironment following tumor regression. Supplemental Figure 10: Treatment of 6694C2-LM tumors with MRTX1133 reduces granulocytes but has little effect on T cells. Supplemental Figure 11: Neoadjuvant and adjuvant therapy in a metastatic model of PDAC.

Year:

2024

KRAS inhibitors demonstrate clinical efficacy in pancreatic ductal adenocarcinoma (PDAC); however, resistance is common. Among patients with KRAS G12C -mutant PDAC treated with adagrasib or sotorasib, mutations in PIK3CA and KRAS , and amplifications of KRAS G12C , MYC , MET , EGFR , and CDK6 emerged at acquired resistance. In PDAC cell lines and organoid models treated with the KRAS G12D inhibitor MRTX1133, epithelial-to-mesenchymal transition and PI3K-AKT-mTOR signaling associate with resistance to therapy. MRTX1133 treatment of the Kras LSL-G12D/+ ; Trp53 LSL-R172H/+ ; p48- Cre (KPC) mouse model yielded deep tumor regressions, but drug resistance ultimately emerged, accompanied by amplifications of Kras , Yap1 , Myc , Cdk6, and Abcb1a/b , and co-evolution of drug-resistant transcriptional programs. Moreover, in KPC and PDX models, mesenchymal and basal-like cell states displayed increased response to KRAS inhibition compared to the classical state. Combination treatment with KRAS G12D inhibition and chemotherapy significantly improved tumor control in PDAC mouse models. Collectively, these data elucidate co-evolving resistance mechanisms to KRAS inhibition and support multiple combination therapy strategies. Significance: Acquired resistance may limit the impact of KRAS inhibition in patients with PDAC. Using clinical samples and multiple preclinical models, we define heterogeneous genetic and non-genetic mechanisms of resistance to KRAS inhibition that may guide combination therapy approaches to improve the efficacy and durability of these promising therapies for patients. See related commentary by Marasco and Misale, p. 2018

Year:

2024

Supplemental Table 4A: Differential gene expression analysis between MRTX1133 sensitive and resistant KRAS G12D cell lines. Supplemental Table 4B: Analysis of Copy number and RPPA datasets between MRTX1133 sensitive and resistant KRAS G12D cell lines.

Year:

2024

Year:

2024

Year:

2024

Collaborators (22)

Juha Väyrynen

University of Oulu

FINLAND
View details

Alexander Jordan

Assistant Professor

Yale University

UNITED STATES
View details

Stephanie Dougan

Associate Professor

Dana-Farber Cancer Institute

UNITED STATES
View details

Jiping Wang

Research Director, Division of Surgical Oncology, Brigham and Women's Hospital

Brigham and Women's Hospital

UNITED STATES
View details

Eliezer Van Allen

Professor

Harvard University

UNITED STATES
View details

Ben Stanger

University of Pennsylvania

UNITED STATES
View details

Fabian Theis

Director of the Institute of Computational Biology and Professor

Technical University of Munich

GERMANY
View details

Lorin Crawford

-

UNITED STATES
View details

Paola A. Guerrero

Research Group Leader

The University of Texas MD Anderson Cancer Center

UNITED STATES
View details

Srivatsan Raghavan

Assistant Professor

Dana-Farber Cancer Institute

UNITED STATES
View details

Richard Dunne

Associate Professor of Medicine

University of Rochester Medical Center

UNITED STATES
View details

David S. Hong

The University of Texas MD Anderson Cancer Center

UNITED STATES
View details

Daniel Chang

University of Michigan

UNITED STATES
View details

Chen Yuan

Assistant Professor

Cedars-Sinai Medical Center

UNITED STATES
View details

Simona Cristea

Dana-Farber Cancer Institute

UNITED STATES
View details

Anirban Maitra

Laura and Isaac Perlmutter Professor of Pathology

NYU Langone Health

UNITED STATES
View details

Michael P. Kim

Associate Professor

The University of Texas MD Anderson Cancer Center

UNITED STATES
View details

David C. Linehan

Dean, CEO, Sr. VP Health Sciences

University of Rochester Medical Center

UNITED STATES
View details

Shubham Pant

Professor

The University of Texas MD Anderson Cancer Center

UNITED STATES
View details

Aram F. Hezel

University of Rochester Medical Center

UNITED STATES
View details

Alex K. Shalek

J. W. Kieckhefer Professor

Massachusetts Institute of Technology

UNITED STATES
View details

Jennifer Roth

Broad Institute

UNITED STATES
View details