Open data as scientific infrastructure: from one sample to a transcriptomicfoundation model.
Open Research objectives/practices
-
Making the outputs of research freely available
-
Increasing the transparency of research processes and methodologies
-
Using open collaborative methods and tools
Introduction
My research has been built on a simple Open Research practice: reuse publicly available molecular datasets to ask new questions, and return the resulting methods, code, data products and tools to the community for further reuse.
I began applying this principle with TACNA (transcriptional adaptation to copy-number alterations), where we reanalysed 34,494 publicly available gene-expression profiles to develop a platform-independent method for studying how genomic alterations reshape transcription. Applying this approach to more than 28,000 patient-derived tumours generated new hypotheses about genomic instability and tumour immune evasion.
Since then, public transcriptomic data have become the foundation of a connected research programme rather than a resource for a single project. We have used large collections of existing transcriptomes to predict gene functions, study replication stress, develop interpretable machine-learning approaches, and investigate clinically relevant transcriptional programmes in colorectal and ovarian cancer. In immunotherapy research, we used 14,085 publicly available whole-blood transcriptomes to construct a reference landscape against which a local patient cohort could be interpreted. Most recently, we integrated 294,159 public transcriptomic profiles to identify ranscriptional programmes linking copy-number alterations, tumour immunity and immunotherapy response.
Our Open Research practice therefore operates in both directions. We reuse data that other researchers have made available, and we return research outputs through open-access publications, openly available analysis code, archived datasets and results, reusable transcriptional components and interactive resources.This creates a cycle in which one research output can become the starting material for another scientific question.
Motivation
My motivation for working openly comes directly from experiencing what openness makes possible. Much of this research could not have been performed by a single research group generating its own data. The biological signal needed to understand heterogeneous diseases such as cancer is distributed across studies, platforms, tumour types and patient populations. Public data allow these individual contributions to become collectively much more powerful.
The benefits occur at several levels. For researchers, combining existing datasets provides statistical power, biological diversity and opportunities for independent validation while avoiding unnecessary duplication of expensive data generation. For patients and research participants, it means that a sample contributed to one study can continue generating knowledge beyond its original research question. This is particularly important for rare cancers and other data-scarce diseases, where assembling sufficiently large cohorts locally may be impossible.
Our ovarian-cancer work illustrates this potential particularly clearly. By integrating publicly available transcriptomes and subsequently examining public spatial-transcriptomic datasets, we identified transcriptional patterns associated with differences in patient survival, including a previously unrecognised synaptic-signalling programme in tumour regions. Open data therefore enabled us not only to stratify retrospective patient groups but also to extract new biological hypotheses from information that was already present in the scientific community.
This experience now motivates our next step. We are developing an interpretable foundation model for transcriptomics trained primarily on large collections of publicly available transcriptomic data. The goal is to learn reusable biological representations that could ultimately generate realistic transcriptomes, improve research in data-scarce settings such as rare cancers, and predict candidate effects of gene perturbations for experimental testing.
Crucially, the intention is again to return what we build. The planned outputs include open versioned code and models, harmonised data products, containerised workflows, downloadable results, a web portal and API access. For me, this is the cumulative impact of Open Research: data shared by others enabled our discoveries; those discoveries become open resources for others; and collectively those resources make scientific questions possible that none of us could address alone.