On April 13, 2026, Gladstone Institutes posted a deceptively simple question on X: “What if we could finally understand which tiny changes in our DNA actually cause disease?” Behind that question sits a decade of frustration. We can sequence an entire human genome in hours for under $1,000. We still cannot reliably interpret most of it.
The Human Genome Project finished in 2003. Twenty-three years later, roughly 98–99% of our DNA — the non-coding regulatory regions that control when, where, and how genes switch on — remains functionally opaque. Millions of genetic variants in those regions are classified as variants of uncertain significance (VUS): we know they differ between people, but we cannot say whether they matter.
Gladstone’s Keck Center for Machine-Guided Functional Genomics just changed the terms of that problem. Their platform — detailed in a major March 2026 feature — isn’t a faster sequencer or a better database. It’s a closed-loop system where AI predicts, wet-lab hardware validates, and single-molecule readouts feed back into the models. The genome’s dark matter finally has a decoder.
The Sequencing Era Created a New Bottleneck, Not a Solution
Cheap sequencing surfaced the problem in full. Whole-genome sequencing now reveals millions of variants per individual. Most are benign. A tiny fraction drive disease. The challenge is separating signal from noise — specifically in the non-coding regions that traditional methods can’t test at scale.
Katie Pollard, Director of the Gladstone Institute of Data Science and Biotechnology, frames the gap plainly: “We can run thousands of experiments on the computer in one day that would take years in a traditional lab.” The implied problem is that those years still represent the field’s actual throughput — one edit, one assay, one answer. At the scale of millions of variants, that’s not a workflow, it’s a queue with no end.
Non-coding variants are particularly hard because they don’t change protein sequences — they alter the regulatory grammar encoded in enhancers, silencers, and topological chromatin domains. A single nucleotide swap in an enhancer can reshape gene expression in one cell type while leaving another unaffected. No amount of protein-structure modeling catches that. You need to read the chromatin context.
Gladstone President Deepak Srivastava, MD, puts the clinical stakes bluntly: “We’re about to enter a world where everybody will have their DNA sequenced, but most of the information we get from a genome sequence is still not interpretable.”
A Three-Layer Bio-Compute Loop That Actually Closes
The Keck Center platform is structured as three interlocking layers — each one making the others more powerful.
Layer 1 — AI Prediction (Pollard Lab). Deep learning models trained on population-scale genomic and epigenomic datasets learn the regulatory logic embedded in DNA sequences. Bioinformatics fellow Zhirui Hu’s systems can simulate trillions of virtual experiments, prioritizing which of millions of variants are most likely to be causal. Related work from Christina Theodoris — her Geneformer transformer model for single-cell gene activity prediction — provides a foundation model that can be fine-tuned for variant-effect scoring at scale.
Layer 2 — High-Throughput Validation (Shipman Lab). Top AI candidates don’t wait for sequential bench work. Seth Shipman’s team deploys editrons — engineered plasmids that combine CRISPR-Cas9 with bacterial retrons — to test hundreds or thousands of variants simultaneously in pooled living cells. “Now, we can effectively do one single experiment that actually tests thousands of genome edits simultaneously,” Shipman says. Team members Jihoon Han and Alejandro González-Delgado have built the multiplexing pipeline that makes this practical rather than theoretical.
Layer 3 — Single-Molecule Chromatin Readout (Ramani Lab). Edited cells are analyzed with SAMOSA (Single-molecule chromatin mapping with long-read sequencing), developed by Vijay Ramani and collaborators Jose Nunez and Hani Goodarzi. Unlike short-read sequencing, SAMOSA preserves long-range chromatin architecture and 3D folding on individual DNA molecules — capturing how a non-coding edit propagates into changes in gene regulation, accessibility, and cell-type-specific behavior. That data loops directly back to retrain and refine the AI prediction layer.
The result is a self-improving machine. Not metaphorically — literally: real experimental outcomes continuously update the model’s priors, narrowing the candidate space with each cycle.
What a “Million-Fold” Speed Increase Actually Unlocks
Pollard’s claim — “increase the speed of discovery a millionfold” — sounds like marketing. It isn’t. The arithmetic is straightforward: traditional one-edit-at-a-time approaches test tens of variants per experiment. Editron-based multiplexing handles thousands per run. AI pre-filtering eliminates millions of low-probability candidates before a single cell is touched. Compound those across an iterative loop and the throughput gain is structural, not incremental.
What that unlocks, clinically:
- Rare developmental disorders: Conditions like congenital heart defects and certain autism spectrum subtypes — where only ~20% of cases have a clear genetic explanation — become diagnosable from non-coding variant profiles.
- Organ-specific phenotypes: The platform can explain why a variant expressed ubiquitously causes defects in one tissue only — a question that stumped mechanistic biology for decades.
- Pre-symptomatic intervention: Ramani’s vision of a “lookup table” — submit a patient’s variant, receive a disease probability and tissue-specific effect — becomes technically achievable rather than aspirational.
- Target validation at speed: Srivastava estimates the path from variant discovery to clinic could compress to under 5 years in some disease areas.
Pollard’s framing is precise: “By creating a platform that allows us to decode the language of life as it’s written in DNA, we will ultimately be able to predict disease before it happens and intervene.”
The Genome Is a Program. We Just Got a Debugger.
The Gladstone platform proves something that matters beyond genomics: the most productive biological systems aren’t pure wet-lab or pure silicon. They’re closed loops where AI and molecular hardware operate as a single integrated process — each layer amplifying the resolution of the other.
The non-coding genome has been called dark matter because it was invisible to the tools we had. It was never inert. Every developmental decision, every tissue-specific expression pattern, every disease susceptibility encoded in a regulatory region was always there — just unreadable.
We didn’t need more sequencing. We needed a better interpreter. Gladstone built one.
References
- Gladstone Institutes. (2026, March 23). How AI Is Pinpointing the Genetic Cause of Disease. Gladstone News. https://gladstone.org/news/how-ai-pinpointing-genetic-cause-disease
- Gladstone Institutes. (2026, March 23). This Is AI Decoding the Genetic Cause of Disease (video). Gladstone News. https://gladstone.org/news/ai-decoding-genetic-cause-disease
- Gladstone Institutes. Keck Center for Machine-Guided Functional Genomics. https://gladstone.org/science/keck-center-for-machine-guided-functional-genomics
- Theodoris, C.V. et al. (2023). Transfer learning enables predictions in network biology. Nature. https://doi.org/10.1038/s41586-023-06139-9
Related: What Is a Biocomputer in 2026? · AI-Biology Convergence
Feature image: AI-generated using Grok.