Meet the Team: Madhumita, Senior Data Scientist at COSMIC
How can data science help researchers uncover meaningful signals within vast and complex cancer datasets? In the latest instalment of our blog series spotlighting the COSMIC team, Senior Data Scientist Madhumita shares her journey into computational biology, the challenges of working with large-scale cancer genomics data and how a new hotspot dataset combining COSMIC and GENIE could help researchers identify promising areas for further investigation.
Could you tell us a little about your academic journey and what first drew you towards data science and computational biology?
If I trace it back, what has consistently driven me is a belief that the biggest problems in biology today aren't going to be solved by observation alone, they need rigorous quantitative reasoning behind them. I was drawn to mathematics and statistics because they give you a language precise enough to make sense of complexity, and I was drawn to biology because it's one of the richest, most complex systems we have to study. Computational Biology, to me, was where those two things could genuinely meet, not as separate skills, but as a single way of thinking about the natural world.
That conviction shaped my education deliberately. My integrated BSc BEd in Odisha, India gave me a solid grounding in the life sciences, but even then, I was already thinking about how quantitative frameworks could be brought into biological questions. That's why I pursued an MSc in Life Sciences with a specialisation in Bioinformatics at the Central University of Punjab. I wanted formal training in applying statistical and computational methods directly to biological data, rather than treating them as two separate disciplines.
That direction became the foundation of my PhD at IIT Jodhpur in the Department of Bioscience and Bioengineering, where I spent five years developing machine learning and deep learning methods for multi-omics data integration at Computational Biology Laboratory, working on how to bring together different layers of molecular data, such as genomic and transcriptomic information, to uncover clinically meaningful cancer subtypes. For me, that work represents exactly the kind of science I want to keep doing: using quantitative methods not just to analyze biology, but to genuinely advance how we understand and treat disease.
What attracted you to join the COSMIC team, and what appealed to you about working with cancer genomics data?
COSMIC appealed to me because it sits at the intersection of exactly what I'd spent years working towards, using quantitative methods to make sense of complex cancer biology, but at a scale and level of impact that goes far beyond any single research project. During my PhD, I was working with multi-omics data to identify cancer subtypes, and I understood first-hand how much a resource like COSMIC, with its curated, comprehensive catalogue of somatic mutations, underpins that kind of research globally. So joining the team felt like an opportunity to work on the very foundation that so much cancer genomics research depends on, rather than just building on top of it.
What particularly drew me in was the nature of the data itself. Cancer genomics data is messy, heterogeneous, and constantly evolving as new sequencing studies come in, and there's a real intellectual challenge in figuring out how to structure, harmonise, and extract meaningful signals from it in a way that's useful to researchers worldwide. That's precisely the kind of problem I find most engaging: not just applying a model, but thinking carefully about what the data actually represents biologically, and making sure the methods respect that.
And beyond the technical side, there's something genuinely motivating about working on a resource that has real downstream impact, researchers around the world use COSMIC to inform how they understand and treat cancer. That combination of rigorous data science and tangible relevance to patient outcomes is what made this role stand out to me.
COSMIC brings together a huge amount of complex cancer mutation data. What are some of the biggest challenges, and opportunities, in working with datasets of this scale?
Working at this scale brings a distinct set of challenges. The first is heterogeneity, data comes in from thousands of publications and sequencing studies, each with different platforms, annotation standards, and levels of curation quality. Reconciling all of that into a single, consistent, queryable resource is genuinely difficult, because you're not just merging numbers, you're merging different scientific conventions and sometimes ambiguous nomenclature for the same mutation.
The second challenge is scale itself, in a more literal sense, as sequencing becomes cheaper and more widespread, the volume of somatic mutation data keeps growing rapidly, and a lot of it now comes from large-scale whole-genome and whole-exome studies rather than smaller, targeted ones. That changes the computational demands considerably; you need pipelines and infrastructure that can keep pace with that growth without compromising the quality or interpretability of what's curated.
There's also the challenge of separating signal from noise, distinguishing driver mutations that are functionally meaningful from passenger mutations that occur incidentally. At scale, this becomes both a statistical and a biological problem, and getting it right really matters, because researchers and clinicians downstream rely on that distinction to inform real decisions.
But the scale is also where the biggest opportunities lie. With enough data, you can start to detect patterns that are invisible in smaller cohorts, subtle mutational signatures, rare but recurrent alterations across cancer types, or associations between specific mutations and treatment response. It also opens the door to more powerful machine learning approaches, since these methods genuinely benefit from large, diverse datasets rather than small, narrow ones. And because COSMIC is used so widely, improvements in how we structure or interpret this data don't stay contained, they ripple out into research and clinical practice globally.
You're one of the people driving work on a new hotspot dataset that combines information from COSMIC and GENIE. Could you explain what this resource is designed to do?
At the heart of it is a fairly intuitive idea: within a cancer gene, not every mutation carries equal weight. KRAS is a well-known driver gene, but really it’s mutations at codons 12 and 13 that matter, the rest of the protein contributes comparatively little on its own.
What we've built is a systematic statistical framework that scans every amino acid position in every Cancer Gene Census gene. For each position, we ask a simple question: is this position mutated more often, across independent patients, than the gene's overall mutation burden would predict by chance? We run that test in parallel across two independently collected pan-cancer cohorts, COSMIC and GENIE, and treat agreement between them as an added layer of confidence.
The result is a tiered catalogue of positions, ranked from strongest to more provisional evidence, which we hope can serve as a useful starting point for researchers, for example, in prioritising variants for follow-up, or as one input among others when thinking about relevance.
For researchers who may not be computational specialists, what problem does this new resource solve?
I think the core problem it solves is prioritising under uncertainty. If you're a researcher working on a cancer gene, you might see a long list of mutations reported across patients, but very little immediate guidance on which of those are likely to actually matter biologically versus which are just incidental noise. Without a computational background, it's hard to independently work out which positions in a gene are recurrently and significantly mutated, especially across large public datasets.
What this resource does is take that statistical heavy lifting and turn it into something simple to interpret. It creates a ranked list at the level of individual amino acid positions, that says: here's where the evidence for biological importance is strongest, and here's where it's more provisional. So instead of needing to run your own analysis across thousands of samples, a researcher can look up a gene of interest and immediately get a sense of where to focus, whether that's for designing a follow-up experiment, or narrowing down candidates worth investigating further.
Ultimately, it replaces a fairly common but limited approach, picking hotspots based on an arbitrary mutation count threshold, with one grounded in statistical significance, using recurrence patterns in COSMIC data to determine which positions are genuinely mutated more often than chance would predict, rather than just more often than some fixed cutoff.
What makes this dataset different from existing hotspot resources?
A few things. First, rather than borrowing an external or global background mutation rate, the statistical framework conditions on each gene's own observed mutation burden, which means it works correctly even for rarely-mutated genes where approaches relying on external rate estimates can break down.
Second, we have an iterative caller that progressively unmasks weaker signals hidden behind dominant hotspots. In a gene like EGFR, for example, once you account for the dominant L858 signal, 92 additional positions become detectable, those would be invisible in a single-pass analysis.
Third, and I think this is really important, every position gets two independent annotations: an evidence tier reflecting statistical significance, and a robustness label reflecting how stable that signal is when you reduce the data. So you can immediately see not just whether a position is significant, but how much confidence to place in that call. And finally, we analyse both COSMIC and GENIE with identical methods and report cross-dataset concordance, giving researchers a level of independent, cross-cohort validation that most existing single-dataset catalogues don't provide.
Were there any particularly interesting technical or scientific challenges in integrating data from COSMIC and GENIE?
Yes, a few, and the more interesting ones were scientific rather than technical. The one that stuck with me most was that COSMIC and GENIE have very different "personalities" in terms of background noise. COSMIC is built from two decades of published, literature-curated studies, so it disproportionately captures the mutations that were interesting enough for someone to write a paper about, which sounds like an advantage, but it actually concentrates the recorded mutation spectrum artificially around a small number of already-famous positions. GENIE, being clinical panel sequencing, captures a broader and more representative spread per gene, simply because it isn't filtered by publication interest. So the same gene, same true biology, can look statistically quite different depending on which dataset you're looking at, and we had to build a framework that could be applied identically to both without either dataset's biases distorting the comparison, then treat the differences between them as informative rather than noise to average away.
On the more practical side, the two databases don't always number amino acid positions against the same canonical transcript for a given gene, so cross-dataset comparisons had to be restricted to the subset where COSMIC and GENIE agreed on a canonical transcript. Standard bioinformatics housekeeping, but getting it wrong would have silently corrupted every downstream concordance result, so it needed care rather than cleverness.
A smaller but very concrete challenge: deduplication and filtering had to be handled differently for each mutation class, and in COSMIC's case specifically, we also had to guard against a single contributing study dominating a position's count, since COSMIC aggregates many independent submissions, a heavily-sampled cohort could otherwise masquerade as broad pan-cancer recurrence rather than genuine independent replication.
Can you give an example of how you imagine researchers using this resource in practice? Who do you think will benefit most?
A very concrete example: imagine a researcher looking at a somatic mutation at a position they don't immediately recognise as a well-known hotspot. Right now, answering "is this actually interesting, or is it just background noise" often means either relying on institutional memory of the literature or running your own recurrence analysis from scratch. With this resource, they can look up the gene and position directly and see whether it's been called a hotspot, at what confidence tier, and importantly whether that signal holds up independently in both COSMIC and GENIE, rather than being an artefact of one particular cohort. That turns a fairly open-ended judgement call into a quick, evidence-backed first pass.
The users I'd expect to get the most out of it would be researchers trying to make sense of a mutation of uncertain significance in a patient sample, where this gives them a quick, evidence-backed starting point rather than a guess. Computational or functional cancer researchers are looking for a prioritised shortlist of positions worth deeper investigation, particularly the ones that are statistically robust but don't already have an obvious mechanistic story attached, since those are exactly the positions that are easy to overlook and interesting to actually go and test in the lab.
What kinds of analyses or discoveries do you hope this resource will enable?
The thing I'm most hopeful about is functional follow-up on positions that don't already have an obvious story attached to them. We found a meaningful number of statistically robust positions where standard conservation-based predictors don't flag anything unusual at all, and yet the recurrence signal holds up independently across both cohorts. Those are exactly the cases that are easy to overlook, because they don't fit the pattern you'd expect from evolutionary conservation alone: the site looks structurally unremarkable, but cancer appears to be selecting for a change there anyway. If wet-lab groups pick even a handful of those up and test what they actually do to the protein, that would be a genuinely satisfying outcome for what is, on our end, a purely statistical piece of work.
More broadly, I'd also like to see this catalogue used as a stepping stone into structural and mechanistic questions, where in three-dimensional space do these recurrent positions cluster, do they sit at interaction interfaces or catalytic sites, is there a shared mechanism across genes that look superficially unrelated. The statistics can tell you where to look; I'd love for this to be the thing that gets someone to actually look.
Cancer genomics and data science continue to evolve rapidly. What developments are you most excited about over the next five to ten years?
Two things genuinely excite me here, and neither is really about my own project, they're about where the whole discipline seems to be heading.
The first is convergence. For a long time, cancer genomics has progressed as a set of parallel, largely separate lines of evidence, sequence recurrence, protein structure, functional assays, evolutionary conservation, clinical outcomes, each generating its own signal, each requiring its own expert to interpret. I think we're approaching a point where a data scientist won't be choosing between these lenses, but building models that reason across all of them simultaneously, the way a truly excellent human curator already does in their head, just at a scale no single person could sustain across the entire genome. Large language models are a genuine unlock here, not because they replace expertise, but because they can hold and cross-reference far more context than any one of us can, and do it fast enough to keep pace with how quickly the underlying evidence is now being generated.
The second is scale, finally catching up to ambition. Resources like COSMIC have spent two decades accumulating exactly the kind of deep, curated evidence this field needs, and cohorts like GENIE keep expanding the breadth of what's captured, as sequencing itself gets cheaper and more standardised across institutions worldwide. For most of my career, "statistically well-powered" and "comprehensive" have been in tension, you could have deep evidence on a handful of famous genes and cancer types, or broad but shallow evidence everywhere else. I think that tension is genuinely starting to dissolve. The kinds of analyses that are only just becoming possible today for the best-studied genes, the hotspot work we've discussed is one small example, should, within a decade, extend to the cancer types and genes that have historically been too data-poor to support any confident conclusions. That's not a small thing. It's the difference between cancer genomics serving the diseases we already understand well, and starting to genuinely serve the ones we don't.
Finally, what advice would you give to students or early-career scientists who are interested in combining biology with data science?
Get comfortable being genuinely bilingual in the languages of biology and data science, rather than fluent in one discipline’s language and just polite in the other. That means understanding the nuances of how biologists describe and interpret what they see, not just knowing the terminology. You need enough biology to recognise when a result that looks impressive on paper is actually biologically uninteresting, and enough data science to recognise when a biologically exciting-looking pattern won't hold up once you test it properly. It's a slower skill to build than either half on its own, and early on it can feel like you're never quite expert enough in either field, but that discomfort is basically the job, and it's also what makes your work trustworthy rather than just impressive-looking on a slide.
The other piece of advice, more practical: don't wait for a "real" dataset to start building intuition. I learned as much from wrestling with messy, incomplete data during my Master's as I did from any formal course. Biological data is going to be messy your entire career, and the sooner you're comfortable with that, the sooner you can actually start asking interesting questions instead of waiting for clean conditions that never arrive.
We hope you’ve enjoyed getting to know Madhumita and learning more about the role data science plays in turning complex cancer genomics data into meaningful insights. Throughout this series, we’ll continue to introduce you to the curators, scientists, developers and commercial team behind COSMIC, sharing the expertise and perspectives that keep the Knowledgebase evolving. Stay tuned for more conversations, and don’t forget to follow us on LinkedIn or subscribe to our newsletter to keep up with the latest updates.
