GPT-Rosalind に新機能導入を発表

We’re introducing a new model update to our GPT‑Rosalind series purpose-built for life sciences research at enterprise scale. It combines GPT‑5.5’s agentic coding and tool-use capabilities with stronger model intelligence in core drug-discovery domains such as medicinal chemistry and genomics, while advancing performance across broader life sciences analysis, design, and experimental workflows.

Progress in life sciences depends on synthesizing data and evidence across scales and modalities: molecules, genes, pathways, and living systems. In our evaluations, the updated GPT‑Rosalind shows broad performance gains on research tasks from biology experts, complex medicinal chemistry queries, quantitative biology, and wet lab troubleshooting.

GPT‑Rosalind is now available in research preview to eligible organizations globally through our trusted-access deployment structure.

Improving performance on scientifically-valuable tasks

In order to measure and continuously improve the real-world impact of GPT‑Rosalind, we designed LifeSciBench, an externally expert-judged benchmark focused on foundational aspects in life sciences research. Unlike existing benchmarks that evaluate a single component of model performance or biological domain in isolation, LifeSciBench takes an end-to-end view of scientifically valuable work by drawing tasks from six workflow areas central to life sciences research: evidence handling, analysis, design and optimization, scientific reasoning, validation and operations, and translation and communication. We use this benchmark to align progress with the needs and realities of life sciences research.

Extracting, reconciling, and auditing scientific evidence from papers, figures, tables, and experimental records.

Candidate Response

Bottom line: As presented, this package is not strong enough to support accelerated approval on micro-dystrophin expression as a surrogate endpoint reasonably likely to predict clinical benefit. A skeptical FDA reviewer would likely view the biomarker, functional, durability, safety, and generalizability evidence as materially under-supported.

Package Item

Key Failure Mode

What Would Be Needed

Western blot quantification

MANEX1A binds an N-terminal epitope shared by endogenous full-length dystrophin and the micro-dystrophin transgene, so the assay does not cleanly distinguish transgene from residual/revertant dystrophin. Quantifying a 138 kDa micro-dystrophin against a healthy full-length dystrophin standard is also invalid.

Use a recombinant micro-dystrophin standard and an orthogonal method that distinguishes transgene from endogenous dystrophin, such as targeted mass spectrometry or a transgene-specific/epitope-specific assay.

Immunofluorescence

The C-terminal polyclonal antibody is poorly suited because the 138 kDa construct lacks the C-terminal domain. Many DMD patients have revertant fibers, and revertant dystrophin can retain C-terminal epitopes. Revertant fibers may expand clonally with age, biasing IF signal, especially in older boys.

Repeat IF with an antibody against an epitope present in the transgene but absent from revertant dystrophin. Quantify transgene-positive fibers separately from revertant fibers.

Surrogate endpoint validity

The package conflates protein amount with clinical function. “38% of healthy-control protein mass” does not mean 38% of normal dystrophin function because micro-dystrophin is structurally truncated.

Empirically validate the relationship between micro-dystrophin mass-percent, sarcolemmal localization, downstream functional restoration, and clinical benefit before treating expression as a surrogate endpoint.

Biopsy design

Pre- and post-treatment contralateral vastus lateralis biopsies introduce left-right and intramuscular spatial variability. Disease progression and fibro-fatty replacement can also change total-protein-normalized signal.

Standardize biopsy site using consistent anatomical landmarks, normalize to muscle-specific proteins, and measure fibro-fatty composition in parallel.

NSAA comparator/statistics

An external natural-history cohort is not a randomized concurrent control. Trial eligibility, supportive care, participation effects, baseline NSAA, steroid regimen, age, and exon class can all bias the comparison. An unpaired t-test is not sufficient. Also, a +1.4 NSAA change is within test-retest variability for this age group.

Run a randomized concurrent placebo-controlled study, or at minimum use adjusted analyses accounting for baseline NSAA, age, steroid regimen, exon class, and other confounders.

Age-window confounding

Boys age 4–7 are in a developmental window where untreated ambulatory DMD patients may gain motor function before decline dominates. A 48-week NSAA change mixes developmental gain, disease progression, and possible treatment effect.

Use a concurrent randomized control with age stratification to separate developmental trajectory from treatment effect.

Prior clinical precedent

Open-label micro-dystrophin functional signals have not reliably predicted confirmatory benefit; published precedent includes micro-dystrophin gene therapy confirmatory trials failing to reproduce open-label NSAA improvements.

Do not rely on open-label NSAA change as decisive support. Require controlled functional evidence.

Structural limits of the construct

The 138 kDa construct deletes spectrin repeats R16/17, which contain nNOS-binding sites. Loss of nNOS recruitment can impair functional sympatholysis and ischemia protection during exercise, creating a mechanistic ceiling on rescue independent of expression level.

Add mechanistic studies showing whether this specific construct restores relevant dystrophin-associated complex function, nNOS localization, exercise physiology, and muscle protection.

AAV durability

Vector genomes at 12 weeks do not establish durable expression. AAV9 genomes are largely non-integrating episomes and may decline over time. Vector-genome persistence is not the same as persistent protein expression.

Measure longitudinal transgene protein expression and functional biomarker durability beyond 12 weeks.

Immune/safety profile

Transaminitis in 8/12 patients is consistent with immune response to AAV-transduced cells, but the mechanism is not established. One myocarditis case is concerning given AAV9 cardiac tropism.

Provide deeper immune monitoring, liver/cardiac safety characterization, and intensified cardiac follow-up.

Patient selection/generalizability

Excluding anti-AAV9 neutralizing-antibody-positive patients limits generalizability. Excluding exon-44 deletions limits applicability to that DMD subgroup. n=12 is too small to characterize safety and efficacy across the broader DMD population.

Broaden eligibility where possible or pre-specify stratified analyses by antibody status, genotype/exon class, age, and baseline function before using the result to support broad approval.

Regulatory conclusion: The package may show biological activity, but it does not yet establish that the measured micro-dystrophin expression is a reliable surrogate reasonably likely to predict clinical benefit. The main gaps are assay specificity, invalid quantification standards, possible revertant-fiber confounding, lack of a randomized control, age-related NSAA confounding, uncertain durability, and unresolved safety/generalizability issues.

To close the gap, the program would need a controlled, age-stratified clinical design with transgene-specific expression assays, orthogonal protein quantification, tissue-composition controls, longitudinal durability data, mechanistic functional assays for the truncated construct, and stronger safety monitoring, especially hepatic and cardiac.

Rubric Criteria & Grades

Criterion

Points

Identifies assay/measurement problems in micro-dystrophin quantification, including MANEX1A epitope sharing, invalid full-length dystrophin standards, and need for recombinant or orthogonal transgene-specific measurement.

+24

Explains why micro-dystrophin expression level is not automatically a valid surrogate for functional clinical benefit.

+22

Flags biopsy-site, tissue-composition, and age-window confounding that weaken expression and NSAA interpretation.

+19

Critiques the NSAA comparator/statistics, especially reliance on external natural-history controls.

+12

Addresses AAV durability, immune response, transaminitis, myocarditis, and need for longer-term expression/safety follow-up.

+15

Notes patient-selection/generalizability gaps, including anti-AAV9 exclusion, exon-44 exclusion, and small sample size.

+8

Stronger scientific reasoning

GPT‑Rosalind achieves industry-leading performance in medicinal chemistry, a field focused on turning molecules into useful drugs. We designed MedChemBench to reflect realistic medicinal chemistry workflows, evaluating multimodal chemical structure understanding; structure-activity relationship (SAR); prediction of drug potency, toxicity, and absorption, distribution, metabolism, excretion (ADME); multiparameter lead-optimization decision-making; and retrosynthesis. GPT‑Rosalind out-performs GPT‑5.5 at 27.5% vs. 25.1% on MedChemBench, while using 7.2% fewer tokens.

On GeneBench, our agentic evaluation on long horizon, end-to-end analysis in genomics and quantitative biology, GPT‑Rosalind uses 31% fewer tokens than GPT‑5.5 while achieving a higher accuracy of 21.6% vs. 20.4%. GeneBench assesses agentic performance on long-horizon quantitative tasks: based on realistic scientific data, can an agent plan valid analysis, QC, modeling, and corrections to arrive at decision-relative answers? Included problems span a variety of domains, including functional genomics, spatial transcriptomics, proteomics, epigenomics, and applied genetics.

We introduce a new evaluation to test GPT‑Rosalind’s ability to help scientists conducting lab work in the real world. LabWorkBench tests the model's ability to link perturbations to experimental outcomes in real wet lab protocols used by scientists, for the purposes ranging from troubleshooting to optimization. The data used by LabWorkBench are proprietary and thus uncontaminated. GPT‑Rosalind scores 63.2% vs. GPT‑5.5 at 55.8%, while using 5.3% fewer tokens.

From reasoning to executed workflows

We built the Life Sciences Research⁠(opens in a new window) and Life Sciences NGS Analysis⁠(opens in a new window) plugins to extend the increased intelligence of GPT‑Rosalind with a practical execution layer for repeatable scientific workflows. Together, these plugins bring sourced evidence retrieval, biological interpretation, and bioinformatics execution into the same workspace, helping researchers connect external evidence with internal omics analyses while preserving artifacts and provenance. All users can now access both plugins through Codex. Qualified GPT‑Rosalind enterprise users can additionally use GPT‑Rosalind to power these plugins.

To better leverage Codex as a dynamic workbench for scientists, we added interactive viewers for biologically native file types. The initial set of sequence, alignment, and structure viewers are designed to keep scientists close to the evidence as GPT‑Rosalind reasons across a workflow and directly answer follow-up questions using the active viewer in-context.

scRNA-seq QC & Annotation

Bulk RNA-seq FASTQ QC

Expanded access for trusted organizations

We are expanding access to the GPT‑Rosalind series to eligible organizations globally. GPT‑Rosalind will be available in research preview through our trusted-access deployment structure for organizations that are conducting legitimate scientific research with clear public benefit, have strong governance and safety oversight, and controlled access with enterprise-grade security.

As part of this global expansion, we’re excited to help support Novo Nordisk’s mission of bringing innovative treatment options to patients faster by helping scale their medical research with GPT‑Rosalind. Novo Nordisk is leveraging frontier AI capabilities to help researchers analyze complex datasets, uncover useful patterns, and test hypotheses more quickly. GPT‑Rosalind’s stronger biological understanding will help teams connect evidence across literature, genomics, transcriptomics, sequence, structure, and experimental results, making it easier to move from data to clearer research decisions.

“Life sciences research is complex, data-rich, and interdisciplinary. To deliver meaningful value for researchers, advanced AI models must be grounded in trusted scientific data, connected to validated tools, and integrated into the real-world workflows researchers use every day. We’re pleased with our partnership with OpenAI and the opportunity to explore how GPT‑Rosalind can support more rigorous, practical approaches to drug discovery.”

Mishal Patel, Group Vice President, AI & Digital Innovation, R&D - Novo Nordisk

We are also now offering an OpenAI managed workspace for qualified organizations without an Enterprise account.

What’s next

The updated GPT‑Rosalind is the next step in our broader commitment to building AI systems that can help accelerate scientific discovery while ensuring that advanced biological capabilities are deployed with appropriate safeguards. We will continue improving the model’s biological reasoning, expanding support for tool-heavy and long-horizon research workflows, and working with qualified organizations across regions to evaluate real-world impact.

This also means applying life sciences AI to high-impact public-benefit work, from drug discovery and translational medicine to public health, preparedness, and biodefense. Through Rosalind Biodefense and our trusted-access deployment model, we aim to put frontier biological capabilities in the hands of the researchers, institutions, and defenders working to improve human health and strengthen societal resilience.

We will continue building GPT‑Rosalind to become a more capable partner across the full life cycle of scientific research, helping scientists move more quickly from the right questions to clearer evidence, better experiments, and ultimately new treatments for patients.