SOW.pdf

Statement of Work

1. Background

Nitrosamines are organic compounds that exist in low levels in our everyday water and foods, and they can also form in drugs during manufacturing from chemical reactions. In people exposed to nitrosamine above acceptable levels and over long periods of time the risk of cancer may be increased. The discovery of nitrosamines in some types of drug products led FDA to issue guidances on control of nitrosamine impurities in human drugs and on acceptable intake limits for nitrosamine drug substance related impurities (NDSRIs).1,2 FDA recognized reformulation of affected drug products by addition of antioxidants and/or pH modifiers as possible mitigation strategies to control the risk induced by nitrosamines. However, there is a knowledge gap regarding the impact of the added excipients (e.g., antioxidants and/or pH modifiers) on the bioavailability (BA) and bioequivalence (BE) of the reformulated drug products. Additional pharmacokinetic (PK) BE studies may be needed for the reformulated drug products, especially for Biopharmaceutics Classification System (BCS) Class IV drugs substances that have low solubility and low permeability. Considering the potential regulatory impact on both the applicants and the Agency, it is critical to have a database with comprehensive drug BCS class information, especially for BCS Class IV drugs.

However, there is currently no comprehensive database that contains the BCS class information of most drug substances. To obtain the BCS classification information for one drug, FDA reviewers usually need to manually search and extract this information from multiple internal and external documents which include, but are not limited to, related submission(s) and review(s) documents and any relevant scientific and technical literatures. Such process is labor-intensive and time-consuming.

Large language models (LLMs) can understand and generate human-like text, as advanced artificial intelligence systems. LLMs can perform a wide range of natural language processing tasks, from text completion to translation and summarization, by learning patterns and gaining knowledge from vast datasets. For example, in a previous BAA contract (#75F40119C10106), LLMs have been used to develop a tool (i.e., PharmBERT) to learn/comprehend texts in drug labeling and, specifically, to automatically extract the absorption, distribution, metabolism, and excretion (ADME) information from drug labeling to facilitate the development of product-specific guidances.3

It would therefore be extremely useful to leverage LLMs to i) search/extract BCS classification information from relevant data sources, ii) automatically designate the BCS class of a drug substance and then iii) construct a BCS database that contains the BCS class of drug substances

1 https://www.fda.gov/media/141720/download 2 https://www.fda.gov/media/170794/download 3 ValizadehAslani, T. et al., PharmBERT: a domain-specific BERT model for drug labels. Briefings in Bioinformatics 24(4),1-10, (2023).

in most of the currently approved new drug application (NDA) products. Upon successful completion of the project, such newly constructed BCS-database automating labor-intensive tasks would save significant time/efforts for FDA reviewers. It will allow BE reviewers to quickly identify the appropriate regulatory pathway for the reformulated products with nitrosamine reducing excipients.

2. Objective

OGD aims to construct a database that contains the BCS class of drug substances, particularly BCS Class IV drug substances, found in most of the currently approved NDA products. This database will provide the Agency a critical tool for determining appropriate regulatory pathways for a product reformulated with nitrosamine reducing excipients.

3. Scope

The contractor shall construct a user-friendly database that contains the BCS class information of drug substances found in most of the currently approved NDA drug products. The constructed database will facilitate the selection of the appropriate regulatory pathway for drug products reformulated by adding nitrosamine reducing excipients. Only FDA-approved NDA drug products will be included to construct the database. BCS IV class drug substances must be identified while other BCS classes (I, II, or III) may not need to be specified. The delivered BCS class database shall be accurate and can be executed and maintained inside the FDA computer/firewall system.

4. Tasks

4.1 Data Collection and Process Algorithmic Framework Establishment

The contractor shall identify available data sources that contain information useful for BCS classification such as drug substance solubility and permeability information, including but not limited to, NDA submission and review documents, drug labeling, OrangeBook, and scientific and technical literature, etc. The contractor shall establish an algorithmic framework to duplicate/simulate current process and criteria used for classifying the BCS class of a drug substance. The contractor shall collect, preprocess, and organize the identified available data sources in a structural format for the downstream LLM model building and validation.

4.2 LLMs Building and Validation

The contractor shall build, train, and fine-tune LLMs to the extent that the model would be able to extract relevant information for BCS classification from the collected data sources efficiently and accurately. The contractor shall establish reasonable acceptance criteria for model evaluation and validation and shall validate the model.

4.3 Drug Substance BCS Class Database Construction

The contractor shall apply the validated model to all currently FDA-approved NDA drug products and automatically designate the BCS class of the drug substances according to the established algorithms based on current regulatory process and criteria, as described above.

The contractor shall construct a BCS class database that can be easily understood and accessed by FDA reviewers and that can be easily maintained and updated by FDA data experts to include future NDA drug products. The constructed BCS class database shall indicate whether a drug substance belongs to BCS class IV and shall cover more than 80% of the currently approved NDA drug products for completeness. The contractor shall train FDA data experts to apply the validated LLM model to designate BCS class for future NDA approvals and shall provide an instruction manual and/or other training material for maintaining and updating the BCS class database.

5. Deliverables

The deliverables needed for this requirement are as follows:

1. A database that contains the BCS class information (especially class IV) of over 80% of the currently approved NDA drug products.

2. LLMs as the backend of the database, including model building (e.g., pre-training and fine-tuning), validation, maintenance and updating, as well as sufficient training materials.

6. Delivery

The contractor shall electronically deliver the BCS class database, the code, and the instruction manual/materials to the COR at least two weeks prior to the end of the contract. The delivered BCS class database must be executed and maintained inside the FDA computer/firewall system.

7. Inspection and Acceptance Criteria

The COR/designee will inspect and verify that the delivered BCS class database is complete, accurate, and user-friendly and that the code is operational, functional, and compatible with FDA hardware and software environment. The COR/designee will provide written notification of acceptance or rejection within 10 days of receipt to the contractor via email. The contractor shall correct any deficiencies within thirty (30) calendar notice. If the Contractor cannot correct the deficiencies within this time frame, the Contractor shall immediately notify the COR/designee of the reason for the delay and provide a proposed corrective action plan within ten (10) business days.

8. Period of Performance

The period of performance shall be twelve (12) months from the date of contract award.

9. Place of Performance

10. Government Furnished Equipment

FDA will provide a government owned laptop to the contractor with the prerequisite that the contractor passes the background investigation.

11. Contract Type

Firm-fixed price

12. Government Points of Contact - COR

Fenggong Wang

FDA/CDER/OGD/ORS

10903 New Hampshire Avenue FDA WO Building 75, Room 4734 Silver Spring MD 20903 240-402-2244 Fenggong.Wang@fda.hhs.gov