The convergence of artificial intelligence and genomic sequencing is reshaping the life sciences industry. AI models are quickly moving beyond optimizing individual steps in the sequencing workflow to now connecting genomic sequence to function. Genomic interpretation that once required painstaking, variant-by-variant analysis is now tractable at scale. For drug developers, that means a more direct path from raw genomic data to functionally validated targets, which have a stronger track record in the clinic.
While the field advances rapidly, investors and leaders in the industry should be careful to distinguish where AI is delivering near-term value from where expectations are running ahead of evidence. It’s a fine line, says Dr. Michael Gandal, a faculty member at the University of Pennsylvania and Children's Hospital of Philadelphia and an expert in the GLG Network. We spoke with Dr. Gandal, whose research focuses on the genetics and genomics of brain disorders like autism and schizophrenia, to get his view on where AI is already delivering practical value across the sequencing workflow, challenges and opportunities in his field, and what life sciences companies should prioritize as they build AI and genomics strategies.
In clinical genomics, Dr. Gandal says the bottleneck is deciding who to sequence and making sense of the results reliably. AI can help on both ends, but autonomous clinical interpretation isn’t ready to replace the molecular pathologists, geneticists, and genetic counselors who do that work. Dr. Gandal views precision medicine in brain disorders similarly, where the realistic near-term payoff is biological stratification and pathway discovery rather than a decisive diagnostic test. For life sciences companies, the highest-value investments are in robust biology, diverse and well-consented clinical datasets, and applications where the human genetic signal is already strong, like oncology, rare disease, and pharmacogenomics.
Read the full Q&A with Dr. Gandal below. Learn how GLG supports companies in the life sciences industry here.
For someone who hasn’t been following genomics closely, how would you describe what AI is now making possible that wasn’t five years ago?
Five years ago, AI in genomics made existing steps easier - sharper basecalling, more accurate variant calling, and smarter prioritization of candidate genes. Now, models are starting to connect genomic sequence to function: how a variant affects gene regulation and protein folding, which cell types it acts in, and when in development it matters.
Genomic interpretation used to be painstaking; now it's tractable at scale. For drug development, targets with human genetic support have a better track record in the clinic, and AI now helps prioritize functional targets directly from raw genomic and multi-omic data.
2. Next-generation sequencing generates large volumes of data. Where in that workflow can AI have an immediate practical impact, and where is it still more promise than reality?
In clinical genomics, you must decide who to sequence and interpret the data quickly and reliably. AI can help upstream by mining the EHR - labs, developmental history, clinical notes - to flag patients who may benefit from genetic evaluation.
After sequencing, AI supports basecalling and variant calling, quality control, phenotype-driven variant prioritization, integration of genomic and clinical data, literature synthesis, and report triage.
Approach autonomous clinical interpretation with caution. AI isn’t a replacement for molecular pathologists, clinical geneticists, genetic counselors, or physicians, especially when many variants come back of uncertain significance.
3. The brain has historically been one of the hardest organs to study, partly because disorders like autism and schizophrenia are defined by symptoms and behavior rather than clear biological markers. How are genomics and AI’s ability to analyze genomic data beginning to change that?
From my perspective as a psychiatrist and genomics researcher, this is one of the most consequential shifts in the field. Genomics, single-cell profiling, multi-omic data integration, and AI are beginning to connect genetic risk to specific brain cell types, developmental windows, regulatory programs, and molecular pathways. We now ask not only which genes are associated with a disorder, but what those genes are doing, in which cells and stage of development, and how mechanisms cut across diagnoses.
It’s important to not oversell this as an imminent diagnostic blood test for brain disorders. The realistic near-term payoff is biological stratification: identifying subgroups and shared mechanisms that point to better targets, biomarkers, and smarter ways to design and enrich clinical trials. For pharma, the value is upstream: building a better biological map of disorders historically defined at the symptom level to develop improved targets and trials.
4. There’s concern about bias in AI models and that genomic datasets historically skew toward certain populations, which can create inequitable outcomes. How serious is that problem today, and what’s being done to address it?
Global genetic diversity is still badly underrepresented, and datasets skew heavily toward European populations.
This makes variant interpretation, polygenic prediction, and functional prediction less accurate and less generalizable. Models trained on a small slice of genetic diversity learn a narrower map of allele frequencies, linkage patterns, and rare variants which leads to poor clinical interpretation for underrepresented patients and weakens the science for everyone by missing genetic variation to clarify mechanisms, improve fine-mapping, and reveal biology that would otherwise be invisible.
This is beginning to be addressed through expanded global biobanks, local reference panels, and multi-ancestry methods. Genetic diversity is not only an equity issue but a validity issue. AI models must be tested across ancestries and clinical settings.
5. Interpretability has emerged as a barrier to clinical adoption. Clinicians need to understand why an AI model reached a conclusion and not just the result. What steps can the medical community take to overcome this?
A model must show its work in terms physicians use, and clinical interpretation should trace back to evidence that can be evaluated and challenged.
Part of this is cultural. Genetics is probabilistic - a variant is rarely cleanly "pathogenic" or "benign," and a risk score is a distribution, not a verdict. So, interpretability must mean calibrated uncertainty, not false precision. Clinicians need an honest probability they can act on - not a certainty that doesn't exist.
We must demand prospective validation, calibration, performance reported separately across ancestry and disease groups, version control, and audit trails and keep humans in the loop for high-stakes clinical decisions. A tool behaving reproducibly in the clinic on actual patients will produce trust.
6. What are the most significant medical privacy concerns when training AI on genomic and clinical data?
Genomic data require a higher privacy standard because they are permanent, identifiable, and familial. An exposed genome cannot be “unshared” and can reveal patient, ancestry, disease risk, biological relationships, and relatives who never consented to testing.
Risks include re-identification of “de-identified” data and leakage or memorization through model outputs, clinical notes, imaging, or consumer data unanticipated by patients. Data collected for care or research could not only be stolen, but used to train commercial models, infer risks in family members, or support decisions patients didn’t agree to.
Data minimization, purpose limitation, strict access controls, audit trails, and privacy-preserving methods like federated learning are important safeguards. Patients and families must believe their data will be used how they were told. Vague consent around AI development, secondary use, or commercial partnerships is where trust breaks.
7. How will precision medicine advance with the rise of AI in genomics?
Precision medicine works best where the biology is decisive. It reads several layers of a patient's biology at once and interprets the combination. The near-term wins stay in oncology, rare disease, and pharmacogenomics. In complex brain disorders, progress will come through stratification and pathway discovery rather than one decisive variant.
Genomics is becoming a therapeutic design platform, not just a diagnostic one. When the causal gene, transcript, or pathway is known, you can design sequence-defined therapies. AI helps with target selection, construct design, and off-target prediction. The hard parts remain delivery, safety, durability, manufacturing, and proving clinical benefit.
The value for companies is upstream: better targets, sharper indication selection, and finding patients likely to respond in a trial. Models that predict noncoding regulatory variants are valuable, as it’s where most disease-associated variation lives. The need to understand the biology remains, but it makes the evidence broader, more integrated, and faster to interrogate.
8. For a life sciences company thinking about where to invest in AI and genomics over the next three to five years, what would you tell them to prioritize and what would you tell them to be cautious about?
Prioritize robust biology, human genetic variation, and well-consented, longitudinally phenotyped, diverse clinical datasets that will matter after this generation of models are obsolete. High-value uses are target discovery grounded in human genetics, functional validation of disease mechanisms and biomarkers, patient stratification and trial enrichment, and real-world evidence linking genotype to longitudinal clinical outcomes.
Invest in strong data and biology rather than a model trained on public data with weak clinical labels and limited generalizability.
Do not overclaim diagnostic performance, build on biased or poorly consented data, or mistake a strong retrospective benchmark for proof that a tool improves clinical decisions. Successful programs will pair robust biology, diverse human genetics, functional validation, regulatory planning, prospective validation, and expert human review from the start and will be honest about which diseases and use cases are ready.
Contact Us
Enter your contact information below and a member of our team will reach out to you shortly.