Flagship Projects

FLAGSHIP 1

From genes to cell states: controlling cellular programs

While each human cell contains only about 20,000 protein-coding genes, it is the complex regulatory networks among these genes that create the wide variety of cell types and functions. Genetic perturbation is a powerful method for analyzing the function of individual genes and their interactions with one another. Recent technological advances have enabled genome-wide perturbation screens, where individual genes are ablated in single cells, with imaging and sequencing readouts. These advances, when paired with the appropriate computational paradigms, open a unique window to catalog all cellular programs in a multimodal fashion (leveraging imaging and sequencing data) to understand the processes underlying cell state transitions, and design precise perturbations to control these processes and induce any desired cell state transition.

Towards this, novel theoretical, algorithmic and computational paradigms are needed for:

integrating diverse, unpaired, multi-modal data to enable the prediction of missing modalities and translating between different cell lines and species;
guiding the experimental design by predicting the outcome of unseen perturbations and identifying, through active learning, the most informative perturbations to test experimentally;
identifying regulatory (causal) networks within and among cells.

We also launched our inaugural machine learning challenge in 2023, the Cancer Immunotherapy Machine Learning Challenge.

To measure our progress on these problems, we have chosen to focus on the following biological challenges aligned with ongoing initiatives at the Broad:

Determining cell intrinsic factors to induce desired cell state changes ex vivo: In collaboration with Evan Rosen’s, Melina Claussnitzer’s, Paul Blainey’s and Devavrat Shah’s groups, we aim to define the transcription factor combinations that, when perturbed in human adipose-derived mesenchymal stem cells (AMSCs) ex vivo, give rise to the adipocyte subtypes observed in vivo. Leveraging large-scale single-cell RNA-seq/multiome, and LipocyteProfiler single-cell imaging data on patient-derived AMSCs throughout the ex vivo differentiation period, we will generate a multi-modal catalog of cellular programs involved in adipocyte (trans-)differentiation to gain an understanding of the paths that can be taken by cells to move between AMSCs and different adipocyte subtypes.

Determining cell intrinsic factors to induce desired cell state changes in vivo: In collaboration with Nir Hacohen’s, Ramnik Xavier’s and Munther Dahleh’s groups, we will map the full regulatory landscape of CD8 T cell subtypes/states. To achieve this, we are using single-cell Perturb-seq to introduce single-gene knockouts in isolated T cells, re-infused in vivo, before transcriptomic readout of the effects of the mutation. Our goal is to identify the optimal combination of genetic perturbations to obtain the T cell population ratio needed for the resolution of different diseases, with an initial focus on enhancing cancer immunotherapy.

To support and complement the two projects described above and test the generalizability of new algorithms across a wide variety of perturbations and contexts, there is a need for systematic experiments in cell lines. In collaboration with DepMap, Jason Buenrostro, Xiao Wang, and the Carpenter/Singh lab, we are generating high-content transcriptomic, epigenetic and phenotypic data collected at the single-cell level and at different time points on cell lines from the Cancer Cell Line Encyclopedia exposed to single-drug treatments. These data will provide the basis for comprehensively cataloging cellular programs in a multi-modal fashion.

ML competitions: To benchmark the status of the field in terms of available algorithms, we launched our first ML challenge in 2023, the Cancer Immunotherapy Machine Learning Challenge. Participants were asked to develop methods and algorithms to predict the effects of gene perturbations on the distribution of T cell subpopulations and identify perturbations predicted to shift the T cell composition to a specified ratio. For the training we provided an in-house generated single-cell Perturb-seq dataset. This dataset was obtained by isolating naive T cells and subjecting them to single-gene knockouts for 70 genes, and then re-injecting them in a melanoma mouse model for ten days before sequencing. We have experimentally validated an additional 60 gene knockouts to evaluate the submitted algorithms. Some predictions were promising and biologically relevant, but the results also highlighted the complexity of the problem. While we can achieve measurable progress, solving these prediction problems requires iterative rounds of new experiments and new ML developments by an interdisciplinary team of researchers.

Like the prominent ImageNet Challenge for the problem of object recognition, we aim to expand our first ML challenge to solve the problem of predicting the effect of genetic perturbations. Towards this, we will develop a large-scale genetic perturbation database for both training and testing of ML algorithms. Following the example of CASP, which was instrumental in the protein structure prediction revolution, every year we will perform new experiments aligned with our flagships, thereby providing the means to objectively test and benchmark algorithms for predicting the effect of new perturbations. This Cell Perturbation Prediction Challenge (CPPC) will establish the state-of-the-art in perturbation prediction, identify what progress has been made, and highlight where future research may be most productively focused.

FLAGSHIP 2

From cells to tissues: controlling cellular networks

New technologies enable us to assay individual cells while retaining their spatial coordinates. 3D atlases are being developed of whole organs (e.g., gut, liver, brain) and even complete organisms (e.g., Drosophila, mouse). These datasets, paired with the appropriate computational paradigms, are the gateway to understanding the biology of tissues by cataloging all cell intrinsic and extrinsic programs in a multimodal fashion, understanding the processes underlying transitions in tissue architecture, and designing precise perturbations to control these processes and induce desired tissue architecture transitions.

Towards this, novel theoretical, algorithmic and computational paradigms are needed for:

characterizing the conditional probability of a cell’s state given its neighbors;
learning which neighbors carry the most information; and
understanding how this probability changes under perturbations on neighbors.

To support this flagship project, we also launched our second machine learning competition in 2024, the Autoimmune Disease Machine Learning Challenge.

With the rise of powerful attention-based models, there is a unique opportunity to develop a multimodal foundation model of cells in tissues, where attention between cells can be interpreted causally to map out regulatory networks of cells in tissues. To measure our progress on these problems, we have chosen to focus on the following biological challenges aligned with ongoing initiatives at the Broad:

Determining cell intrinsic and extrinsic programs underlying tissue zonation in vivo: In collaboration with Jonathan Weissman’s and Philippe Rigollet’s groups, we will determine the cell intrinsic and extrinsic programs defining tissue zonation in the liver and the onset of tissue-restricted diseases such as liver steatosis, which are difficult to model in vitro. The Weissman group together with Xiaowei Zhuang’s lab recently developed Perturb-Multi, a platform for large-scale multimodal in vivo pooled genetic screens with 3D transcriptomic and phenotypic readout. Since the genetic perturbations are inducible, it is possible to layer perturbations at different time points, for example, after exposing mice to starvation or a high-fat diet to analyze how genetic perturbations affect the cellular responses to steatosis-inducing factors.

Determining cell intrinsic and extrinsic programs in diseases affecting multiple tissues: Follicular lymphomas (FL) are a heterogeneous group of B-cell neoplasms with unpredictable clinical outcomes arising from functionally distinct clonal precursor cells (CPCs). Previous studies indicate that the type of CPC from which the tumor originates, along with abnormal stromal and immune cells in their niche, determine the FL progression and responsiveness to treatments. In collaboration with Todd Golub, Fei Chen, and Ari Melnick, we will generate a catalog of FL phenotypes and dependencies based on time-resolved multi-omic spatial data on human samples and genetic animal models engineered to carry human-relevant FL mutations. Connecting this high-resolution functional view of the disease to standard diagnostic assays such as H&E images will enable predictions on patients’ prognosis and responsiveness to therapies.

Spatial ML competition: The main goal of Flagship 2 is to integrate and translate multimodal spatial information and ultimately be able to use simple pathology images to predict data usually acquired through more complex and expensive modalities, such as gene expression information. To support this flagship project, our current Autoimmune Disease Machine Learning Challenge, in collaboration with Ramnik Xavier and the Klarman Cell Observatory, is testing the state of the field in terms of ML algorithms that could allow disease progression prediction from pathology images. Given an in-house generated spatial transcriptomic dataset on 460 genes, matching pathology images from intestinal bowel disease (IBD) samples, and single-cell RNA seq data from related samples, participants are asked to predict the spatial expression of unseen genes. Patients with IBD have a two-fold increased risk of developing colorectal cancer, but this connection is not understood functionally. While colorectal cancer is an aggressive tumor, it is highly treatable if diagnosed early. Towards overcoming these challenges, given a pathology image from an IBD patient with signs of dysplasia (pre-cancerous state), matching spatial transcriptomic data on 460 genes and single-cell transcriptomic data from related samples, the participants are asked to rank all protein-coding genes based on how well they are expected to discriminate between dysplasia and noncancerous tissue regions. The submitted algorithms will be evaluated based on experiments with new gene panels.

Building on our first Spatial ML Challenge, our aim is to define an ML Curriculum for spatial foundation models: a carefully crafted series of progressively harder tasks that make use of pathology image data, spatial transcriptomic, and other medical imaging modalities to train and benchmark foundation models for spatial biology. Importantly, such an ML Curriculum will enable the identification of emergent abilities, referring to a foundation model's capability to perform tasks that it wasn’t explicitly trained on. This effort will establish the state-of-the-art spatial biology modeling, identify what progress has been made, and highlight where future research may be most productively focused.

Flagship Projects

The Eric and Wendy Schmidt Center has convened investigators in biomedical science and machine learning from Broad, MIT, and Harvard to understand the landscape of opportunities and define flagship projects focused on important biological challenges where progress can be clearly measured.
‍

FLAGSHIP 1

From genes to cell states: controlling cellular programs

FLAGSHIP 2

From cells to tissues: controlling cellular networks

Get Involved

Flagship Projects

The Eric and Wendy Schmidt Center has convened investigators in biomedical science and machine learning from Broad, MIT, and Harvard to understand the landscape of opportunities and define flagship projects focused on important biological challenges where progress can be clearly measured.‍

FLAGSHIP 1

From genes to cell states: controlling cellular programs

Learn More

FLAGSHIP 2

From cells to tissues: controlling cellular networks

Learn More

Get Involved

Connect with us on social media:

Contact Us

The Eric and Wendy Schmidt Center has convened investigators in biomedical science and machine learning from Broad, MIT, and Harvard to understand the landscape of opportunities and define flagship projects focused on important biological challenges where progress can be clearly measured.
‍