Noetik (Emily Corse)

Noetik is an AI-native company founded in 2022 with a mission to leverage advanced AI/ML methods to discover and develop cancer therapies. Our current focus is predicting therapeutic response in oncology to help identify patients most likely to benefit from a particular treatment. To do that, we build foundation ‘world models’ of human biology. ‘World models’ are generative AI models built to predict the complete or future state of a system given only information about its current or partially hidden state. Our models aim to understand the full complexity of a patient’s tumor tissue well enough to predict which drugs will work and why, prior to initiating treatment.

At Noetik, we train proprietary foundation models using custom-generated multimodal data from thousands of human cancer samples, including diverse hematoxylin and eosin (H&E) stains and high-plex spatial transcriptomics data. At inference time, these models require only a digital whole slide image of an H&E stained tumor tissue slide as input. As a first step, using this input, the model constructs a spatial representation of the tumor microenvironment - a biological representation that includes, but is not limited to, inferred spatial transcriptomics (iSpT) expression values for more than 18,000 genes at single-cell resolution. In a second step, a separate locked machine learning classifier works on inferred biological representation to predict response for a particular treatment. This approach can be used to develop treatment-response predictors for investigational agents and established therapies.

In Noetik’s methodological paradigm, we develop, train, evaluate and maintain both our own GenAI foundation models as well as the machine learning classifier that work on the output of the foundation model to predict treatment response. Both models will be locked to produce deterministic outputs. Combining these features, we are able to fully lock down the behavior of the end-to-end computation performed within the GenAI-enabled medical device, and maintain full control of all the steps in the prediction chain. As Noetik moves into clinical translation and the design of GenAI-enabled medical devices around this technology, we offer recommendations on how the proposed framework could be applied to devices built on custom-built biological foundation models.

We appreciate the Agency’s initiative in soliciting this feedback on this discussion paper. The examples in the paper, however, focus primarily on the use of large-language-foundation models and conversational GenAI. Biological foundation models, such as models that infer biological state from routine pathology images to support treatment selection, are not specifically addressed, nor are companion or complementary diagnostic applications. In addition, there are several risks that are central to the paper, like open-ended outputs, non-deterministic behavior, and dependence on foundation models developed by third parties, that are absent or substantially reduced in a vertically integrated GenAI model architecture like Noetik’s. Within this context of possible reduced risk, we acknowledge that several of the principles that the paper cites apply directly to this new category, and we appreciate this opportunity to contribute to this discussion as part of our responses to questions 7 & 17, 10, 22 and 24. Please see the attached pdf for our responses.

Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

Public Comment on Docket FDA-2026-N-7874

Considerations for the Regulation of Generative AI-Enabled Medical Devices —
Discussion Paper and Request for Feedback

Introduction:

Noetik is an AI-native company founded in 2022 with a mission to leverage advanced
AI/ML methods to discover and develop cancer therapies. Our current focus is
predicting therapeutic response in oncology to help identify patients most likely to
benefit from a particular treatment. To do that, we build foundation ‘world models’ of
human biology. ‘World models’ are generative AI models built to predict the complete or
future state of a system given only information about its current or partially hidden state.
Our models aim to understand the full complexity of a patient’s tumor tissue well enough
to predict which drugs will work and why, prior to initiating treatment.

At Noetik, we train proprietary foundation models using custom-generated multimodal
data from thousands of human cancer samples, including diverse hematoxylin and
eosin (H&E) stains and high-plex spatial transcriptomics data. At inference time, these
models require only a digital whole slide image of an H&E stained tumor tissue slide as
input. As a first step, using this input, the model constructs a spatial representation of
the tumor microenvironment - a biological representation that includes, but is not limited
to, inferred spatial transcriptomics (iSpT) expression values for more than 18,000 genes
at single-cell resolution. In a second step, a separate locked machine learning classifier
works on inferred biological representation to predict response for a particular
treatment. This approach can be used to develop treatment-response predictors for
investigational agents and established therapies.

In Noetik’s methodological paradigm, we develop, train, evaluate and maintain both our
own GenAI foundation models as well as the machine learning classifier that work on
the output of the foundation model to predict treatment response. Both models will be
locked to produce deterministic outputs. Combining these features, we are able to fully
lock down the behavior of the end-to-end computation performed within the
GenAI-enabled medical device, and maintain full control of all the steps in the prediction
chain. As Noetik moves into clinical translation and the design of GenAI-enabled
medical devices around this technology, we offer recommendations on how the
proposed framework could be applied to devices built on custom-built biological
foundation models.

We appreciate the Agency’s initiative in soliciting this feedback on this discussion paper.
The examples in the paper, however, focus primarily on the use of
large-language-foundation models and conversational GenAI. Biological foundation
models, such as models that infer biological state from routine pathology images to
support treatment selection, are not specifically addressed, nor are companion or

​ Page 1
Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

complementary diagnostic applications. In addition, there are several risks that are
central to the paper, like open-ended outputs, non-deterministic behavior, and
dependence on foundation models developed by third parties, that are absent or
substantially reduced in a vertically integrated GenAI model architecture like Noetik’s.
Within this context of possible reduced risk, we acknowledge that several of the
principles that the paper cites apply directly to this new category, and we appreciate this
opportunity to contribute to this discussion as part of our responses to questions 7 & 17,
10, 22 and 24.

Questions 7 and 17: 7. Is the competency-based approach described above, i.e.,
device benchmarking followed by clinical confirmation, a useful and appropriate
framework for evaluating GenAI-enabled devices? 17. CDRH is exploring whether
a competency-based approach could be applied to devices that incorporate a
variety of underlying model architectures, such as multimodal vision-language
models and generative or predictive world models.

We support the competency-based approach combining device benchmarking with
clinical confirmation. In the next few paragraphs, we describe a proposed framework
that draws upon the fundamentals of the competency-based approach and also
adaptations that we believe are required for biological multimodal foundation models.

For foundation models that we develop, and the devices built using their outputs, the
competency based framework should clearly define competencies according to the
biological or clinical function being performed. To recap, our model takes routine H&E
pathology images as input and includes iSpT as an output, which is an internal
intermediate representation, not reported to the end-user. Benchmarking of this
generated representation establishes its biological competence and supports its reuse
across applications. In particular, we recommend that the evaluation of iSpT layer use
competency tests of biological competence:

●​ Spatial agreement with tissue morphology. Do epithelial, stromal, and immune
markers localize to the corresponding tissue compartments visible in H&E?

●​ Preservation of biological relationships. Do genes belonging to the same
biological program exhibit coherent spatial expression, and do programs
associated with distinct compartments remain appropriately separated?

●​ Preservation of reference-derived gene modules. Do multi-gene programs
identified in measured spatial transcriptomics retain their expected relationships
in inferred outputs?

●​ Tissue and disease specificity. Do inferred expression patterns preserve
expected differences between tissue types and cancer indications?

​ Page 2
Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

●​ Characterization of limitations. Can the evaluation identify biological contexts,
gene programs, or spatial scales for which the outputs are less reliable?

We also acknowledge that these competency tests require appropriate controls. Tissue
density, for example, can cause many genes to increase together, inflating apparent
agreement without demonstrating gene-specific biological fidelity. Therefore, this
component should be controlled for before assessing gene-specific relationships.

Section V.D.1 raises an important question about whether the comparator should be a
clinician, a clinician panel, or another reference standard. For a model that generates
spatial molecular information, clinician performance is generally not a sufficient primary
comparator: clinicians do not directly infer thousands of spatial gene-expression values
from H&E. Relevant standards could include measured molecular data,
reference-derived biological relationships, and expert assessment of tissue morphology.
Pathologists and other appropriately qualified scientific experts could help define and
adjudicate these benchmarks, while molecular measurements provide evidence that
visual assessment alone cannot establish. These performance benchmarks and
thresholds should be prespecified and justified using biological reference data,
measurement variability, the intended population, and the consequences of error.

Now, coming to the topic of clinical confirmation, in our device, a separate model is used
to analyze the biological representation to predict treatment response, which is the final
output presented to the clinician. Clinical validation of the complete device, from H&E
input to treatment-response prediction, should establish the safety and effectiveness for
the intended use.

To summarize, we recommend that the Agency explicitly recognize biological
competencies within its proposed framework of competency-based approach combining
device benchmarking with clinical confirmation. Such a framework of foundation model
evidence could support multiple applications within its demonstrated scope, while each
device based on the output of the model would require clinical validation of the
assembled device for its unique intended use.

Question 10: CDRH recognizes that publicly available benchmarking assets may
be subject to data contamination, saturation, and limited real-world
representativeness. How should a sponsor establish that performance on a given
benchmark predicts safe and effective real-world behavior for the device’s
intended use?

What evidence should support the construct validity of a benchmark used to gate
device evaluation, and what role should sponsor-developed benchmarks play
given potential concerns around independence and optimization to the test?

​ Page 3
Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

Noetik recommends that the Agency evaluate benchmarks according to the biological or
clinical capabilities they measure, the independence of their evaluation data, and the
evidence connecting those capabilities to the device’s intended use.

For specialized applications such as treatment-response prediction like ours, publicly
available benchmarks are unlikely to provide all the biological data, treatment-specific
outcomes, and clinical context needed to evaluate performance for our specific intended
use. Public benchmarks may support assessment of individual capabilities, but
fit-for-purpose, sponsor-developed benchmarks will likely be required to address these
gaps.

To illustrate this further, consider pathology foundation models (PFMs) that are trained
on datasets of histopathology images that can learn robust patient representations that
transfer to downstream disease modeling and drug response prediction tasks. While
PFMs are trained on large amounts of unlabeled histopathology data, their evaluation is
typically restricted to smaller labeled cohorts, reflecting the scarcity of meaningful and
validated clinical labels in publicly available benchmarking assets.

To mitigate these difficulties, we believe that the PFM evaluation benchmark suite
should have rich biological diversity, carefully measured labels, and biological
replicates. Such a benchmark suite would allow for evaluation of models on their ability
to represent a range of phenotypic factors, including retrieval of key validated labels,
such as genotype, treatment response, perturbation, or tumor microenvironment
features measured with orthogonal assays.

With this context, we also believe such sponsor-developed benchmarking systems
should have certain safeguards. These benchmarks should be assessed on
methodological quality, relevance to the particular biological and clinical capabilities,
and their evaluation independence, rather than ownership alone. Such a benchmark
suite can provide a value platform for device evaluation. We also acknowledge that
such a benchmark would be a necessary, but not a sufficient criteria for clinical
readiness.

Question 22. CDRH envisions that a manufacturer’s premarket
competency-based assessment might serve as a baseline against which
post-deployment modifications could be re-evaluated. How might the extent of
re-benchmarking or other evidence be scaled to the nature and expected impact
of a given modification? For example, are there categories of change that might
not significantly affect safety or effectiveness of a GenAI-enabled device, or
might be appropriately managed within a sponsor’s quality management system
versus requiring FDA premarket review and authorization before implementation?
Are there categories of changes that might be appropriate for
inclusion in a PCCP?

​ Page 4
Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

In a post-deployment setting, the competency-based framework can create a
predictable path for improving and updating the biological foundation models while
preserving the evidence supporting their clinical applications. Such change-control
considerations in a post-market setting are more straightforward for a device like
Noetik’s where, as a device manufacturer, we develop, train, evaluate and maintain the
foundation models as well as the response prediction models that use their outputs.

In Noetik’s architecture terms, this would mean evaluating whether an update preserves
the relevant biological competencies and whether any changes propagate into clinically
meaningful differences in downstream predictions that affect the safety and
effectiveness of the GenAI-enabled device.

We see real-world evidence (RWE) collection as an opportunity to strengthen validation,
as the device reaches more institutions and encounters greater biological and technical
diversity. We support a practical and least-burdensome approach that combines
periodic re-benchmarking against the premarket baseline with data collected through
RWE channels, operating within the tiered structure as follows:

●​ Changes manageable within the quality management system: Changes that do
not alter device outputs beyond prespecified tolerances, such as infrastructure,
software dependency, or computational performance changes, verified by
regression testing against a locked reference dataset.
●​ Changes appropriate for a PCCP: retraining or fine-tuning of the foundation
model and the treatment-response classifier to improve robustness to
pre-analytical and technical variation (e.g., new staining protocols, scanners, or
sites) within the authorized intended use. The PCCP would prespecify
re-benchmarking of biological competencies, concordance of downstream
treatment-response calls on a locked reference set, and acceptance criteria for
both. Because intended use remains unchanged, the relationship to treatment
outcomes would continue to be assessed through postmarket evidence.
●​ Changes requiring FDA review: changes to intended or indication for us,
foundation-model or response model changes that do not meet PCCP
acceptance criteria.

Question 24. For devices built on third-party foundation models, changes to the
underlying model may be initiated by the third-party model developer rather than
the device manufacturer. How can a manufacturer detect, evaluate, and respond
to such changes in a timely manner? What mechanisms—for
example,contractual, technical, or through a PCCP—could provide reasonable
assurance that third-party developer-initiated changes do not compromise the
safety or effectiveness of the device?
Many of the risks described in Question 24 arise when a device manufacturer depends
on a foundation model that is developed and updated by a third party. Where the device
manufacturer develops, trains, and maintains the foundation model itself, as Noetik

​ Page 5
Docusign Envelope ID: 631A9088-FB33-8067-8097-498398961DD0

does, these risks can be addressed through the manufacturer’s design controls and
change management: no foundation-model change can occur outside the
manufacturer’s quality system. We recommend that FDA recognize full sponsor control
of the foundation model as a mitigation that reduces the need for additional assurance
mechanisms, such as master files or third-party change notification, for such devices.
Conclusion
Noetik appreciates the opportunity to comment and supports FDA’s effort to develop a
predictable, risk-based framework for these technologies. We would welcome the
opportunity to discuss these recommendations with the Digital Health Center of
Excellence.
For questions regarding this submission, please contact Emily Corse. Thank you for the
opportunity to contribute to this important initiative.

Sincerely,

Emily Corse Lacey Padron ​ Ron Alfa
EVP, Therapeutics ​ Chief Technology Officer ​ ​ Chief Executive Officer

Date: Oct 19, 2026

​ Page 6