Technology Archives | LSP

Research Reproducibility

The LSP is strongly committed to making our research findings FAIR (Findable, Accessible, Interoperable and Reusable). 

LSP investigators thoughtfully consider the factors that influence research reproducibility and advocate for solutions. Whenever feasible, we make our software open-source and release our data available under Creative Commons licenses. We co-host a seminar series that features speakers at the cutting edge of data and knowledge management. 

LSP investigators have studied the irreproducibility of preclinical drug response and pharmacodynamic data in detail (Niepel, 2019) and developed multiple methods to address the problem (Hafner, 2016; Mills, 2022). We have also commented on the importance of public data release for reproducibility (AlQuraishi, 2016) and developed methods to make survival data from clinical trials available to the community (cancertrials.io, Plana, 2022).

LSP investigators have also created open-source software and standards for the emerging field of multiplexed imaging. These include MCMICRO, an open-source data processing pipeline that increases the reliability of complex data analysis (Schapiro, 2022a), MITI, a metadata scheme for tissue images (Schapiro, 2022b), and Cylinter, the first quality control software for high-plex imaging data (Baker, 2024). 

In addition, we developed Minerva, a lightweight software that makes it possible to view whole-slide multiplex images online without download, which can be a barrier to sharing large image files (Hoffer, 2020; Rashid, 2022). We’ve partnered with external organizations like cBioPortal (Wala, 2024) and the Data Coordinating Center of the Human Tumor Atlas Network (HTAN) (De Bruijn, 2024) to incorporate the Minerva image viewer into existing data repositories.

Systems Pharmacology

By pairing computational and experimental approaches, LSP investigators enable a rational, systems-level approach for evaluating preclinical compounds.  

Despite the high costs associated with drug development, only 10% of preclinical compounds pass clinical trials. This drop-off is partially because most preclinical studies reduce a disease to its simplest components, focusing on a few specific targets in a limited number of disease models. Although this reductionist approach provides useful results, it overlooks the fact that within an organism, most drugs interact with multiple targets at the same time— a phenomenon known as polypharmacology. 

To address this gap, LSP researchers use quantitative systems pharmacology, combining traditional drug development approaches with engineering and computation to model how drugs will behave in the body.

Since 2014, LSP investigators have developed numerous systems pharmacology methods for evaluating preclinical compounds. These efforts help clarify how polypharmacology influences drug action, uncover sources of drug sensitivity and resistance, and inform clinical trial design. Ultimately, these efforts aim to provide a quantitative basis for identifying preclinical compounds with the greatest potential for clinical success.

Image Analysis & Visualization

The LSP develops open-source software tools that improve the reproducibility, accuracy, and accessibility of spatial biology datasets.

Spatial biology methods like highly-multiplexed tissue imaging make it possible to visualize the precise locations of 60-100 proteins within human tissues. The resulting images contain deep biological insights, but their size – which can be 200 GB to 1 TB – makes them challenging to analyze and share. 

LSP investigators have developed numerous open-source software solutions that address these persistent challenges. These include tools for identifying single-cell boundaries (a process known as segmentation), performing quality control, analyzing the single-cell data, and visualizing tissue images online (Minerva). Many of these tools have been integrated into MCMICRO, a customizable software pipeline that is used by labs across the globe. 

LSP investigators are collaborating with external communities (like nf-core and the NCI Imaging Data Commons) to integrate our software packages into standardized infrastructures. We aim for our software tools to ensure precise and reproducible tissue imaging datasets so that the deep molecular insights within spatial datasets can be fully realized. 

AI & Machine Learning

LSP investigators leverage cutting-edge ML/AI approaches to enhance biological understanding of diseases, identify therapeutic targets, and improve clinical decision-making. 

LSP investigators harness the latest machine learning and artificial intelligence (ML/AI) methods to unravel complex biology and derive novel insights that can improve therapeutic development. The LSP approaches scientific questions from a systems biology framework. We develop and use novel imaging and high-throughput screening methods that yield rich biological datasets that can be challenging to interpret. Machine learning methods can help derive insights from these datasets, allowing LSP investigators to follow up on specific hypotheses for more rigorous evaluation.

For our tissue imaging datasets, we use human-in-the-loop ML/AI methods to derive sets of features associated with particular outcomes. We have leveraged these models to identify potential prognostic biomarkers for cancer progression and to understand (and predict) how tumors will respond to immunotherapy. 

We also use ML/AI tools to understand polypharmacology – the phenomenon of drugs interacting with multiple proteins – which plays an essential role in drug action, especially for multi-enzyme families like kinases and deubiquitinases. Our open-source methods can predict protein structure (OpenFold) and protein-drug docking interactions (KinCo), as well as extract and assemble mechanistic biological knowledge from published literature (INDRA). These efforts have been useful for understanding therapeutic resistance and sensitivity in cancer, developing novel analgesics, and identifying new approaches to treating neurodegenerative disease.

Computational Pathology

LSP researchers merge traditional histopathology with high-plex imaging and machine learning to generate biological insights that enable personalized patient care. 

Developing a precise, individualized, patient-centered approach to medicine relies on the ability to accurately diagnose diseases and determine which patients are most likely to benefit from a specific therapy.

Despite a growing understanding of the genetic, cellular, and environmental drivers of human disease, the practice of histopathology—the primary method for diagnosing cancer—remains largely pre-digital, relying on visual examination of tissue slides under a microscope.

To address this gap, LSP researchers combine traditional histopathology with high-plex tissue imaging, spatial transcriptomics, and machine learning algorithms to analyze tissue samples at a much higher level of detail. We recently used multimodal data from patients with colorectal cancer to develop machine learning models that could predict disease progression. Building on this work, we are partnering with clinicians to analyze large cohorts of patients from numerous disease states. Ultimately, we aim to develop robust methods for computational pathology that can be readily applied to a wide variety of clinical settings to improve the diagnosis and management of human disease.

Tissue Profiling & Imaging

LSP researchers are developing high-plex methods that combine spatial imaging and transcriptomics to advance our understanding of the tumor-immune microenvironment. 

The LSP is pioneering new methods for highly-multiplexed tissue profiling that capture precise spatial molecular information from 2D and 3D human tissue specimens. LSP researchers invented tissue-based cyclic immunofluorescence (CyCIF), which can stain and image over 60 proteins at single-cell resolution in human tissue. CyCIF has revealed insights into tumor biology, immunology, and the role of inflammation in disease. 

We recently built upon this work with 3D-CyCIF, a method that enables a 3-dimensional view of the sub-cellular arrangement of organelles within cells of the tumor microenvironment. 3D CyCIF has revealed novel biology, showing that 3D tissue architecture impacts cell-type identification and neighborhood analysis. 

We are also working to integrate spatial biology with clinical workflows. We recently developed the Orion method, which can collect rapidly high-plex immunofluorescence and diagnostic-grade H&E images from the same tissue section. We see Orion as a critical step towards bringing precision diagnostics to the clinic. 

To learn more about our methods and view our tissue profiling data, visit the Harvard Tissue Atlas.