An Ideal State For Bioinformatics Data

By Darren Ames, Head of Solutions Science

Being a bioinformatician is like being a forensic accountant for a chaotic-neutral billionaire who keeps his receipts in a series of wet shoeboxes and occasionally writes key clinical outcomes in crayon on the back of a Sizzler menu.

We spend our lives in a state of perpetual triage. Most days, our pipelines are held together by legacy Perl scripts, excessive caffeine, and pure, concentrated spite. We pray that the next batch of WGS data doesn't have a library prep bias.

What if the data actually behaved? What if pharma, clinical diagnostics, and AMCs stopped treating data like a private collection of rare, uncatalogued stamps and started treating it like a high-performance substrate for AI?

The Challenge Of Semantic Interoperability

Currently, 80% of our science is just digital janitorial work. We spend our lives mapping sex: 1 to sex: male and wondering why a clinical site in Des Moines decided to record BMI as a string containing the word "beefy." We’ve all felt the soul-crushing despair of seeing MARCH1 auto-corrected to March-01 in Excel—a biological tragedy that has probably set back cancer research by a decade.

In a world of harmonized, AI-ready data, the semantic interoperability problem vanishes. We move from the current hellscape of mismatched ontologies to a unified, graph-based representation. Imagine a world where every patient entity is pre-aligned to a common data model (like OMOP on steroids) across the entire ecosystem.

Multi-omic Latent Spaces: Navigating The Manifold

If the data were truly organized, we could stop squinting at univariate correlations like we're reading tea leaves and start living in the latent space. Instead of being buried by the curse of dimensionality—where every new multi-omic layer just adds more noise and fewer degrees of freedom—we could deploy multimodal variational autoencoders (MVAEs) at an industrial scale.

Think of an MVAE as a high-dimensional biological trash compactor with a PhD. It takes the disparate, screaming signals of single-cell expression, longitudinal EHR tokens, and digital pathology pixels and compresses them into a lower-dimensional, continuous mathematical landscape known as a latent manifold. This matters because biology doesn't happen in a single spreadsheet; it happens in the complex interactions between these layers.

In the precision medicine community, this is the holy grail. It allows us to find the shared signal across different modalities, even when some data is missing. We can finally identify why two patients with the same ICD-10 code follow completely different clinical trajectories by observing where they sit on this latent manifold. We move from comparing raw, noisy technical artifacts to navigating a smooth mathematical representation of human health. We’d finally have the power to decorrelate batch effects from actual biology, finding the disease vector that actually predicts drug response instead of just predicting which version of the Illumina chemistry was used before the lab's air conditioning broke.

This isn't just science fiction—this is exactly how forward-thinking institutions are deploying on DNAnexus today to turn the fever dream into actual clinical workflows. Take City of Hope’s POSEIDON platform (Precision Oncology Software Environment Interoperable Data Ontologies Network), built on the DNAnexus technology stack. By unifying comprehensive germline/somatic genomic profiling, longitudinal EHRs, and imaging data across more than 670,000 patients, they’ve industrialized their insight loop, slashing complex biomarker analysis workflows from three grueling days down to less than two hours. Similarly, Emory University’s Winship Cancer Institute launched WALI (Winship Analytics for Learning & Investigation) on the DNAnexus Trusted Research Environment to natively blend clinical registries, clinical trial data, imaging, and molecular diagnostics.

The outcomes of navigating these multi-omic latent spaces are entirely tangible: these environments allow researchers to spot biomarker-driven treatment patterns, generate hyper-accurate synthetic control arms, and dramatically optimize clinical trial matching. By moving beyond technical artifacts, these precision medicine workflows empower teams to isolate genuine biological signals that drive drug discovery and patient care.

Biological gaslighting: "don't confuse me with the facts"

In our current fragmented reality, we often engage in a sort of collective biological gaslighting. We run a GWAS, find a signal that’s basically a statistical hiccup, and build an entire R&D strategy around it because the alternative—admitting the data is too noisy to tell us anything—is too depressing to contemplate. When the validation study fails, the internal mantra is often: "Don't confuse me with the facts." We’ve already committed to the target; the data is just being difficult.

But when data is harmonized and AI-ready, the facts become impossible to ignore. We move from chasing p-values that are thinner than a CVS receipt to in silico phase 0 trials. We could simulate the physiological response of a virtual cohort of 50,000 diverse digital twins before we even synthesize a molecule. We stop trying to force the biology to fit our narrative and start letting the high-fidelity multimodal data do the talking. It turns out, when you have organized data, the facts are actually quite helpful.

The Role Of Federated Learning

The biggest bottleneck isn't just messy data; it's the fact that institutions guard their data like Gollum and his Precious.

With structured, AI-ready data, we can implement federated learning at scale. We stop trying to move the data (which is heavy, legally radioactive, and prone to breaking during transit) and move the gradient instead. If every AMC and pharma silo spoke the same language, we could train a global transformer model on patient trajectories without a single byte of PHI ever crossing a firewall.

The DNAnexus reality check

We aren't just building a storage bucket. We’re building the Biological OS. When the data is organized, harmonized, and AI-ready, we stop being the people who fix the files and start being the people who find the cures.

The infrastructure is ready. The math is solid. The only thing left is to stop slipping on that banana peel and actually start walking toward the future.