Harnessing large language models’ zero-shot and few-shot learning capabilities for regulatory research

Abstract

Large language models (LLMs) are sophisticated AI-driven models trained on vast sources of natural language data. They are adept at generating responses that closely mimic human conversational patterns. One of the most notable examples is OpenAI's ChatGPT, which has been extensively used across diverse sectors. Despite their flexibility, a significant challenge arises as most users must transmit their data to the servers of companies operating these models. Utilizing ChatGPT or similar models online may inadvertently expose sensitive information to the risk of data breaches. Therefore, implementing LLMs that are open source and smaller in scale within a secure local network becomes a crucial step for organizations where ensuring data privacy and protection has the highest priority, such as regulatory agencies. As a feasibility evaluation, we implemented a series of open-source LLMs within a regulatory agency’s local network and assessed their performance on specific tasks involving extracting relevant clinical pharmacology information from regulatory drug labels. Our research shows that some models work well in the context of few- or zero-shot learning, achieving performance comparable, or even better than, neural network models that needed thousands of training samples. One of the models was selected to address a real-world issue of finding intrinsic factors that affect drugs' clinical exposure without any training or fine-tuning. In a dataset of over 700 000 sentences, the model showed a 78.5% accuracy rate. Our work pointed to the possibility of implementing open-source LLMs within a secure local network and using these models to perform various natural language processing tasks when large numbers of training examples are unavailable.

Introduction

The progress of artificial intelligence (AI) has evolved with revolutionary advances in natural language processing (NLP) in recent years, with large language models (LLMs) at the forefront of this revolution [1]. These models are pre-trained on a huge amount of natural language data and subsequently undergo supervised fine-tuning or reinforcement learning from human feedback (RLHF) to mimic our everyday interactions using language [2, 3]. The core concept of LLMs is found in the Transformer architecture, a deep learning structure designed specifically for the purpose of processing sequential data, such as text [3]. The Transformer employs a unique ‘attention mechanism’ that allows the model to capture long-range relationships within the text, enabling it to analyze the context and meaning of words and phrases in a more comprehensive way than traditional NLP models [4]. OpenAI's ChatGPT (Generative Pre-trained Transformer) [5] and Google's Gemini [6] are famous examples of such LLMs that have achieved great success in a variety of areas.

NLP is also ushering in a new era in drug discovery and development. In these fields, NLP algorithms excel at searching through vast libraries of medical literature and patient records, properly extracting critical insights on drug effectiveness, adverse effects, and potential interactions, which are fundamental aspects of pharmacokinetics/pharmacodynamics (PK/PD) studies [7–9]. However, relying on ChatGPT or equivalent online models poses a risk of unintentional disclosure of sensitive data [10]. A promising alternative is the internal deployment of open-source LLMs within a secure network [11]. This approach connects the substantial benefits of these models while addressing privacy concerns, which are of paramount importance to organizations such as regulatory agencies. By deploying LLMs on a local network, it is possible to effectively reduce the risk of data breaches and improve operational efficiency [1, 12, 13].

One significant advantage of LLMs over conventional machine learning methodologies lies in their proficiency in few-shot and zero-shot learning. Few-shot learning enables these models to learn and execute tasks with only a handful of examples for guidance, whereas zero-shot learning allows them to undertake tasks without any prior examples [1–3]. This attribute marks a substantial contrast to traditional models that necessitate extensive datasets for effective training. For example, in our previous work, we implemented a BioBERT-based NLP model that can automatically extract directional drug–drug interaction (DDI) from the natural text (drug labels and literature) to identify the change in one drug’s clinical exposure due to the pharmacokinetic (PK) DDI of other drugs [8]. While achieving a high accuracy, the training (fine-tuning) and validation processes utilized over thirty thousand sentences from hundreds of manually annotated drug labels. Such a high training cost prevented the model from being quickly adapted to cover other tasks. For instance, identifying the change in one drug’s clinical exposure due to the intrinsic factors of the patients (e.g. liver and/or kidney impairment) is a closely related task. However, due to the lack of large, annotated datasets containing positive and negative examples of such interaction, so far, no published work has attempted to develop an NLP model to automatically extract such information.

Here we reported the implementation and application of open-source LLMs in a local network of a regulatory agency. We demonstrated that some of these models can use few-shot or even zero-shot learning to achieve superior performance over our previous model, which was dependent on an extensive training dataset. Building on this observation, we directly applied zero-shot learning and used such models to process all FDA drug labels and extract information about patients’ intrinsic factors altering drug’s clinical exposure of a selected list of drugs. Our method can be quickly adapted to cover other tasks without large quantities of task-specific training data.

Material and methods

Datasets

For identifying PK-DDI sentences (i.e. PK-DDI classification), we employed a pre-existing dataset originally assembled from the TAC 2019 DDI track used by Zirkle et al. [8]. This dataset contains 21 593 sentences from a set of FDA labels as training data, and 10 634 sentences as validation data. For few-shot learning, a few sentences were randomly selected from the training dataset (more details below). For zero-shot learning, no training sentences were used. The validation sentences were used to evaluate the performance of various models in the context of zero- and few-shot learning. Of note, for the original TAC 2019 DDI track, the DDI interaction type was categorized by FDA and National Library of Medicine (NLM) experts into one of four classes (Pharmacokinetic DDI, Pharmacodynamic DDI, Unspecified interaction, and No interaction) by following guidelines established before annotation [14]. Subsequently, we combined three classes (Pharmacodynamic DDI, Unspecified interaction, and No interaction) into a ‘Negative’, or ‘non-PK-DDI’ class, and the remaining class (Pharmacokinetic DDI) as ‘Positive’, or ‘PK-DDI’ class. Further minor adjustments were performed to correct some miscategorized classes as identified from our previous study [8] such as clarifying ambiguities in food–drug interactions versus drug–drug interaction, to ensure the reliability of our dataset.

For identifying patient’s intrinsic factors affecting drug’s pharmacokinetics, we downloaded all FDA prescription drugs’ labels from the DAILYMED website [15]. As in our previous model, we are interested in a list of 28 drugs from the Comprehensive in vitro Proarrhythmia Assay (CiPA) Initiative. This resulted in a total of 708 024 sentences being extracted from the labels. Out of the extracted sentences, 500 were selected to form a dataset. To categorize sentences about intrinsic factors, any sentences that describe a change in a drug’s clinical exposure due to patient factors such as gender, age, weight, genetics, organs’ function (e.g. liver or kidney problems), and other comorbidities were categorized into ‘Positive’, while the rest into ‘Negative’. Two FDA experts independently assessed the categorized sentences, and the agreement rate was found to be very high (>95%). This small dataset contains 198 positive and 302 negative sentences and was used for model performance evaluation. The rest of the 708 024 sentences were used as a real-world application to extract information about intrinsic factors affecting CiPA drugs’ clinical exposure from all FDA labels.

Pre-trained transformer models

We utilized a selection of pre-trained transformer models:

1. Flan-T5-xxl where Flan stands for ‘Fine-tuned Language Net’ [16, 17] and T5 is a pretrained encoder-decoder transformer model that can convert all tasks into a text-to-text task [18]. With 11 billion parameters, this model is particularly adept at understanding and executing instructions [13]. Flan-T5-xxl has been pre-trained on a diverse range of datasets, including general text, biomedical literature, and domain-specific corpora [19]. This extensive pre-training enables the model to excel in both general language understanding and specific biomedical contexts.

2. Tk-instruct is built by fine-tuning T5 with Super-Natural Instructions (SUP-NATINST), a large collection of NLP tasks with expert written instructions [20, 21]. The emphasis on understanding and following instructions is considered a key element to our initial assessment, as we intended to evaluate LLMs’ zero-shot, in-context learning capabilities.

3. T0pp is part of the T0 series, which is a variant of the T5 model (T5 + LM) [22] trained on a variety of tasks to generalize its ability to understand and respond to a wide range of language processing tasks through zero-shot learning [21, 23]. T0pp model are capable of being fine-tuned for any specific tasks, and it has recently been utilized for various biomedical applications [24, 25].

4. Vicuna (version vicuna-13b-v1.5) is a model with 13 billion parameters, developed by supervised instruction fine-tuning of LLaMA 2 [26], a decoder-only transformer model pretrained on trillions of tokens [27]. Having been fine-tuned on a vast number of human conversations, it has the capability to interact with users in a coherent and contextually appropriate manner [28].

5. BioBERT is a domain-specific adaptation of the BERT model (Bidirectional Encoder Representations from Transformers) [29] for biomedical text mining [7, 30]. Due to the small size (345 million parameters), BioBERT and BERT needed task-specific fine-tuning on a large training dataset, as done in our previous model development [8].

6. SetFit, Sentence Transformer Fine-tuning, is a small (110 million parameters) but efficient model specially designed for few-shot learning [31]. While other LLMs such as T5 above utilized in-context learning (no need to modify any model parameters), SetFit uses a more traditional training (fine-tuning) step by updating parameters based on the limited number (few-shot) of training examples [32].

Most of these models have a size of 10 to 20 billion parameters, about the maximum size our local computing resources can accommodate at the time of writing. Of the two smaller models, BioBERT was chosen because it provided state-of-the-art performance on the PK-DDI classification task after fine-tuning [8], and SetFit was chosen because it was specially designed to perform few-shot learning, which can serve as a benchmark for this technique [32]. These models were readily accessible through the Hugging Face website [33], a platform for accessing and utilizing pre-trained language models [34].

In order to setup our models, we utilized the transformer package [3] and PyTorch along with DataLoader [35, 36] to handle deep learning tasks. All computations were carried out using the NVIDIA Tesla V100. Key components like the Tokenizer and Model are initialized with specific parameters, such as the tokenizer being adjusted to enable padding from the left and truncation from the right. Additionally, we configured the maximum length for tokenized sentences to 512 for zero-shot and 750 for few-shot learning. The model’s max_new_token was set to 10 with the exception for the Vicuna LLM, which was set to 3.

Few-shot learning

As shown in Fig. 1, few-shot learning enables us to feed the LLMs a small amount of training data to effectively guide its predictions. We used four training sentences, two from each class. For the PK-DDI classification, we randomly selected two positive and two negative PK-DDI sentences from the 21 593 sentences of training dataset. These four randomly selected sentences form a training sample. For the validation dataset, the PK-DDI group had 358 sentences, while the non-PK-DDI group had 10 276. In order to address the imbalanced dataset’s concern, we tried using an undersampling method [37]. In this method, the validation sentences (10 630 in total) were randomly sampled to form a balanced validation set that comprised of 356 PK-DDI and 356 non-PK-DDI sentences. Each training sample (four selected training sentences) was used to guide the models to classify the sentences in the balanced validation set, and the performance was recorded. Such process was repeated 100 times for each training sample. To ensure comprehensive validation and model robustness, the process of selecting training sentences was repeated 100 times too. As a result, a total of 10 000 samples (100 sets of training samples × 100 balanced validation set) were evaluated under few-shot learning, and the average and standard deviation of the evaluation metrics were reported as the results of our study. For further details on the imbalanced dataset, please refer to the supplementary material.

Figure 1.

Schematic procedure for large language models: We utilized four labeled sentences as our few-shot learning examples, comprising two PK-DDI and two non-PK-DDI samples; Each of these examples, along with the validation sentence set, was accompanied by a prompt question to feed the model; For zero-shot learning, only the prompt along with the validation sentence, without any provided examples, were fed to the model.

For the sentence classification task focused on intrinsic factors that affect drug clinical exposure, a set of four sentences (two positive and two negative) was randomly selected from the manually constructed dataset (198 positive and 302 negative sentences, respectively) as a training sample. All remaining sentences were used for model validation/evaluation without undersampling.

Zero-shot learning

In parallel to our few-shot Learning scenario, we also explored zero-shot learning (Fig. 1). For zero-shot learning, the LLMs are directly prompted to predict or categorize the provided sentences, without the need to form a training sample. Nevertheless, we implemented the same undersampling technique in our PK-DDI classification task, by forming 1000 balanced validation sets through random sampling. For the task of finding intrinsic factors, when zero-shot learning was applied to the manually annotated dataset, no undersampling was needed because the dataset was balanced (198 positive and 302 negative sentences). When zero-shot learning was applied to the full dataset, no undersampling was performed as this is a real-world application to find intrinsic factors for a list of drugs, without the need to calculate performance metrics.

Prompt

As shown in Table 1, for the first and second tasks, we used a similar style of prompt for the same LLM. Since the prompt was primarily selected based on the first task, the performance on the second task serves as a validation of the generalizability of the selected prompt. The model’s publisher served as the source of guidance for these prompts. We then tried to modify these prompts, aiming to enhance the performance of our LLMs. This included testing a variety of words and phrases through trial and error to optimize the prompts.

Table 1.

A unique prompt was utilized to feed each LLM, obtained as recommended by the model’s developers.

Model	Task 1: Identifying PK-DDI sentences
Flan-T5	Is it true that this sentence relates to how the pharmacokinetics of one drug, namely absorption, distribution, metabolism, or excretion, change when co-administered with another drug, indicating a pharmacokinetic drug–drug interaction? ‘The co-administration of AFINITOR with long-acting octreotide increased octreotide C min by ~50%’. Options: - Yes - No Yes
Tk-instruct	In this task, you are given two sentences. Indicate if the first sentence clearly entails the second sentence (i.e. one can conclude the second sentence by reading the first one). Indicate your answer with 1 if the first sentence entails the second sentence, otherwise answer with 0. Input: Sentence 1: The coadministration of AFINITOR with long-acting octreotide increased octreotide C min by ~50%.”. Sentence 2: The sentence is about pharmacokinetic drug–drug interaction, which relates to how the pharmacokinetics of a drug, namely absorption, distribution, metabolism, or excretion, change when co-administered with another drug. Output: 1
T0pp	The following sentence relates to how the pharmacokinetics of a drug, namely absorption, distribution, metabolism, or excretion, change when co-administered with another drug, indicating pharmacokinetic drug–drug interaction. ‘The co-administration of AFINITOR with long-acting octreotide increased octreotide C min by ~50%’. True, or False? True
Vicuna	Given ‘The co-administration of AFINITOR with long-acting octreotide increased octreotide C min by ~50%’. Is it true that this sentence relates to how the pharmacokinetics of a drug, namely absorption, distribution, metabolism, or excretion, change when co-administered with another drug, indicating a pharmacokinetic drug–drug interaction? exclusively answer with Yes or No. Yes.
Model	Task 2: Identifying patient’s intrinsic factors altering drugs’ clinical exposure sentences
Flan-T5	Is it true that the given sentence relates exclusively to the drug’s pharmacokinetics as altered by patients’ intrinsic factors, such as gender, age, weight, genetics, organ’s function, additional diseases, and so forth? Given sentence: ‘Levetiracetam C max and AUC were 20% higher in women ( N  = 11) compared to men ( N  = 12)’. Options: - Yes - No Yes
Tk-instruct	In this task, you are provided with two sentences. Determine whether the first sentence clearly implies the second sentence. It means the second sentence can be determined only from the first sentence. Provide your answer as either 1 if the first sentence implies the second sentence, or 0 if it does not. Input: Sentence 1: ‘Levetiracetam C max and AUC were 20% higher in women ( N  = 11) compared to men ( N  = 12)’. Sentence 2: the given sentence relates exclusively to the drug’s pharmacokinetics as altered by patients’ intrinsic factors, such as gender, age, weight, genetics, organ’s function, additional diseases, and so forth? 1
T0pp	The given sentence relates exclusively to the drug’s pharmacokinetics as altered by patients’ intrinsic factors, such as gender, age, weight, genetics, organ’s function, additional diseases, and so forth? Given sentence: ‘Levetiracetam C max and AUC were 20% higher in women ( N  = 11) compared to men ( N  = 12)’. True or False? True
Vicuna	Given ‘Levetiracetam C max and AUC were 20% higher in women ( N  = 11) compared to men ( N  = 12)’. Is it true that this sentence relates exclusively to the drug’s pharmacokinetics as altered by patients’ intrinsic factors, such as gender, age, weight, genetics, organ’s function, additional diseases, and so forth? Exclusively answer with Yes or No Yes

The sentence highlighted in italics was taken directly from an FDA drug label; the question provided in the prompt is indicated in bold, while the answers categorized as labels are in plain text.

Results

Compare few-shot and zero-shot learning capabilities of locally implemented LLMs

In our earlier study [8], a fine-tuned BioBERT model (BioBERT_directionalDDI) was trained on 21 593 sentences and evaluated on 10 592 sentences for their accuracy in identifying pharmacokinetic (PK) drug–drug interaction (DDI) sentences. We utilized this model and dataset to benchmark various LLMs implemented locally. In contrast to the traditional model BioBERT_directionalDDI, our LLMs, including Flan-T5-xxl [20], Tk-Instruct [23], T0pp [27], and Vicuna [32], were subjected to a training procedure that used either 4 (for few-shot learning) or 0 training sentences and then evaluated using the full test dataset with over 10 000 sentences. As a comparison, we also put the baseline BioBERT model, the pre-trained model that our BioBERT_directionDDI model was based on, through the same few-shot or zero-shot procedure. Because the use of very small datasets may incur instability [38, 39], and both the training and validation datasets have considerable imbalance, with a predominance of non-PK-DDI sentences, we employed undersampling techniques, repeatedly sampling an equal number of PK-DDI and non-PK-DDI sentences to perform many runs of training and validation (see few-shot learning in the Material and methods section) [40–43]. This approach created more balanced datasets and reported not only the average but also the standard deviation of measures across runs for a robust analysis of the prediction performance.

As shown in Table 2, for both few-shot and zero-shot learning for identifying PK-DDI sentences, Tk-instruct, Flan-T5, and Vicuna are among the top-performing models. These models consistently achieved precision and specificity levels exceeding ~85%, showcasing their reliability and accuracy. T0pp shows better performance in zero-shot learning, achieving precision rates of ~70%. However, its performance in few-shot learning, as reflected in sensitivity and F-score metrics, fell below other top-performing models (Table 2). Conversely, the baseline BioBERT model showed lower performance and is not suitable choices for zero- or few-shot learning, as indicated by our evaluation results (Table 2). However, as shown in the last rows of Table 2, the results from the fine-tuned bidirectional BioBERT model implemented in our previous study [8] are very similar to those obtained with the LLMs in the current study. Notably, in certain aspects, such as Flan-T5 using zero-shot learning, the performance exceeds that of our earlier BioBERT NLP model. (Table 2).

Table 2.

Performance of the sentence classification between PK (positive) and non-PK-DDI (negative) sentences.

Zero-shot learning	Precision	Sensitivity	F1-score	Specificity
Tk_Instruct	0.85 ± 0.02	0.92 ± 0.00	0.88 ± 0.01	0.83 ± 0.02
Flan-T5	0.94 ± 0.01	0.85 ± 0.00	0.89 ± 0.01	0.95 ± 0.01
Vicuna	0.91 ± 0.01	0.84 ± 0.00	0.88 ± 0.01	0.92 ± 0.01
T0pp	0.73 ± 0.02	0.84 ± 0.00	0.78 ± 0.01	0.68 ± 0.02
BioBERT (nonfine-tuned)	0.49 ± 0.02	0.55 ± 0.02	0.52 ± 0.02	0.42 ± 0.03
BioBERT (fine-tuned) [8]	0.83	0.81	0.82	0.99
Few-shot learning	Precision	Sensitivity	F1 score	Specificity
Tk_Instruct	0.97 ± 0.01	0.76 ± 0.08	0.85 ± 0.06	0.98 ± 0.01
Flan-T5	0.96 ± 0.01	0.80 ± 0.02	0.88 ± 0.01	0.97 ± 0.01
Vicuna	0.88 ± 0.03	0.93 ± 0.02	0.90 ± 0.01	0.87 ± 0.04
T0pp	0.58 ± 0.12	0.69 ± 0.34	0.57 ± 0.22	0.48 ± 0.31
BioBERT (nonfine-tuned)	0.63 ± 0.03	0.22 ± 0.02	0.32 ± 0.02	0.87 ± 0.02
BioBERT (fine-tuned) [8]	0.83	0.81	0.82	0.99

Non-fine-tuned BioBERT represents the lowest performance, while Flan-T5 and vicuna excel in all other models. BioBERT (fine-tuned) indicates the performance of our previous study in terms of identifying PK-DDI [8]. The performance of the LLMs was assessed by repeatedly sampling random sentences from training and/or validation datasets (Undersampling) for zero and few-shot learning, as detailed in the Materials and methods section. The row for the fine-tuned BioBERT model was copied from Zirkle et al. [8] and does not include standard deviation, as their approach did not involve undersampling. This row is included to facilitate a comparison with our zero-shot and few-shot learning methods.

We also evaluated SetFit, a relatively small model designed to achieve a few-shot learning performance comparable to that of LLMs. The performance with an increasing number of training examples is shown in Fig. 2. The graph suggests that using four samples (two samples per class) for SetFit, which is the default size of few-shot training samples when we evaluated LLMs (Table 2), yielded low evaluation metrics with high variability. Nonetheless, running the SetFit model with 32 samples per class (64 total) yielded a performance comparable to LLMs evaluated before (Fig. 2).

Figure 2.

SetFit Model’s performance metrics: The x-axis represents the total count of training samples provided to the model, evenly divided into positive and negative classes (PK-DDI and non-PK-DDI).

Apply LLM to identify drug exposure changes due to patients intrinsic factors

Having identified Flan-T5 as having the highest overall performance for zero-shot learning for our PK-DDI evaluation task, we wanted to evaluate its performance on a new task, which was identifying sentences related to intrinsic factors affecting the drug’s clinical exposure. We constructed a manually annotated dataset that contains 198 positive and 302 negative sentences (see section Materials and methods) to assess if the high performance of Flan-T5 on the evaluation task (identifying PK-DDI sentences) would be carried over to the new task (identifying intrinsic factors affecting drug’s clinical exposure). As illustrated in Table 3, Flan-T5 achieved roughly 80% across all performance metrics without any training sentences (zero-shot learning). In comparison, we also applied the other top performing LLMs (T0pp, Tk-instruct, and Vicuna) as evaluated in the previous PK-DDI task to this new task of identifying intrinsic factors. Overall, these three models have lower performance compared to Flan-T5.

Table 3.

Performance of sentences relates to the patient’s intrinsic factors altering drugs’ clinical exposure.

Zero-shot learning	Precision	Sensitivity	F1 score	Specificity
Tk_Instruct	0.54	0.98	0.70	0.46
Flan-T5	0.82	0.81	0.81	0.88
Vicuna	0.45	0.89	0.59	0.28
T0pp	0.69	0.69	0.69	0.80
Few-shot learning	Precision	Sensitivity	F1 score	Specificity
Tk_Instruct	0.66 ± 0.27	0.12 ± 0.01	0.02 ± 0.02	0.99 ± 0.01
Flan-T5	0.85 ± 0.02	0.65 ± 0.05	0.74 ± 0.03	0.93 ± 0.01
Vicuna	0.60 ± 0.08	0.88 ± 0.07	0.70 ± 0.03	0.60 ± 0.14
T0pp	0.42 ± 0.05	0.98 ± 0.08	0.59 ± 0.02	0.12 ± 0.12

In the context of few-shot learning, all LLMs shown in this table fall short of demonstrating effective performance. However, it is notable that Flan-T5 exhibits considerable capabilities in zero-shot learning.

Now that we have confidence in Flan-T5’s capability of performing different tasks without training data (zero-shot learning), we decided to apply it directly to the task of scanning the entire FDA drug label documents to identify intrinsic factors that could affect the clinical exposure of a selected list of drugs of interest. In this task, as explained in the Material and methods section, the list of drugs of interest is related to the Comprehensive in vitro Proarrhythmia (CiPA) Initiative [15, 44–46]. We processed 708 024 sentences related to CiPA drugs through the aforesaid LLM to identify intrinsic factors affecting drug’s clinical exposure. The model predicted 32 205 sentences as positive, and after taking out identicals, it determined 553 of these sentences to be uniquely positive. Upon a comprehensive examination, we found that 434 of these sentences were accurately identified (true positives), whereas 119 were incorrectly predicted (false positives). As a result, the precision rate for the positively predicted sentences was 78.5%. The results for each of the 28 CiPA drugs, derived from scanning all FDA drug labels, are presented in Table 4.

Table 4.

Results from applying Flan-T5 to the real-world problem within the framework of zero-shot learning.

Drug	Total sentences (FDA label)	Predicted positively (true positive)
Astemizole	0	0
Azimilide	0	0
Bepridil	0	0
Chlorpromazine	20 013	8 (7)
Example sentence: Chlorpromazine should be administered cautiously to persons with cardiovascular, liver, or renal diseases.
Cisapride	0	0
Clarithromycin	53 959	56 (33)
Example sentence: In patients with severe renal impairment (CL CR 30 mL/min), the dose of clarithromycin should be reduced by 50%.
Clozapine	17 608	19 (15)
Example sentence: A subset (3%–10%) of the population has reduced activity of CYP2D6 (CYP2D6 poor metabolizers).
Diltiazem	45 653	23 (16)
Example sentence: Individual patients, particularly 60 years of age, may respond to a lower dose of 120 mg.
Disopyramide	1503	13 (11)
Example sentence: Hepatic impairment also causes an increase in the plasma half-life of disopyramide.
Dofetilide	8556	24 (22)
Example sentence: A population pharmacokinetic analysis showed that women have ~12%–18% lower dofetilide oral clearances than men (14%–22% greater plasma dofetilide levels), after correction for weight and creatinine clearance.
Droperidol	310	2 (2)
Example sentence: The initial dose of droperidol should be appropriately reduced in elderly, debilitated, and other poor-risk patients.
Ibutilide	286	4 (2)
Example sentence: In 285 patients with atrial fibrillation or atrial flutter who were treated with ibutilide fumarate injection, the clearance of ibutilide was independent of renal function, as assessed by creatinine clearance (range 21–140 mL/min).
Loratadine	13	0
Metoprolol	161 244	63 (44)
Example sentence: The elimination half-life of metoprolol is about 7.5 hours in poor metabolizers and 2.8 hours in extensive metabolizers.
Mexiletine	5371	9 (8)
Example sentence: However, the metabolic clearance of mexiletine in the extensive metabolizer phenotype decreased by about 70% making the poor and extensive metabolizer groups indistinguishable.
Nifedipine	45 339	22 (21)
Example sentence: After administration of nifedipine extended-release tablets to healthy elderly men and women (age > 60 years), the mean Cmax is 36% higher and the average plasma concentration is 70% greater than in younger patients.
Nitrendipine	0	0
Ondansetron	130 523	88 (69)
Example sentence: The clearance of ondansetron in pediatric patients 1 month to 4 months of age is slower and the half-life is 2.5 fold longer than patients who are > 4–24 months of age.
Pimozide	288	7 (7)
Example sentence: The time to achieve steady state Pimozide concentrations is expected to be longer (~2 weeks) in poor CYP 2D6 metabolizers because of the prolonged half-life.
Quinidine	3325	15 (13)
Example sentence: Quinidine clearance typically proceeds at 3–5 mL/min/kg in adults, but clearance in children may be twice or three times as rapid.
Ranolazine	11 494	14 (9)
Example sentence: More marked and progressive increases in serum creatinine, associated with increases in BUN or potassium, indicating acute renal failure, have been reported after initiation of ranolazine in patients with severe renal impairment.
Risperidone	103 653	81 (73)
Example sentence: The low dose was 0.125 mg/day for patients for patients weighing 20–45 kg, and it was 0.175 mg/day for patients weighing > 45 kg.
Sotalol	19 911	59 (46)
Example sentence: The half-life of sotalol is prolonged (up to 69 hours) in anuric patients.
Tamoxifen	20 451	7 (2)
Example sentence: In the youngest cohort of female pediatric patients (2–6 year olds), CL/F was 2.6-fold higher; in the oldest cohort (7–10.9 year olds) CL/F was approximately 1.9-fold higher.
Terfenadine	0	0
Vandetanib	0	0
Verapamil	58 524	39 (34)
Example sentence: In patients with hepatic insufficiency, metabolism of immediate-release verapamil is delayed, and elimination half-life prolonged up to 14–16 hours; the volume of distribution is increased, and plasma clearance reduced to about 30% of normal.

The first column lists the CiPA drugs. The second column, titled ‘Total Sentences’, displays the number of sentences in FDA labels where each drug is mentioned. We used the difflib Python package [42] to discard sentences with a similarity of 95% or more. The third column counts the sentences that were positively predicted and also includes those predicted correctly. Certain CiPA drugs, such as azimilide, may not be mentioned at all, likely due to lack of approval in the USA. In contrast, drugs like Loratadine are referenced in their own or other drugs’ labels, but none of these mentions relate to clinical exposure from (PK) DDI. The second row for each drug provides a model-identified example sentence. These sentences often contain quantitative data useful for assessing the reference drug’s high clinical exposure.

Discussion and conclusion

Our study presents a comprehensive evaluation of various open-source large language models (LLMs) including Flan-T5 [13], Tk-instruct [20], T0pp [23], Vicuna [27], with a separate analysis of BioBERT [7], and SetFit [32], based on their capability to detect sentences related to specific pharmacokinetic (PK) aspects, such as drug–drug interactions (DDI) and the potential for a drug’s clinical exposure to be changed by patients’ intrinsic factors (e.g. liver injury, age, etc.). The goal of our research is to evaluate the feasibility of implementing to implement open-source LLMs in a local network and quickly adapt them to perform various tasks through few-shot or zero-shot training, without the need to build extensive training data for each specific task.

Performance of LLMs in few-shot and zero-shot learning

In both scenarios of few and zero-shot learning we have noticed variations in the performance of the models. Regarding our first task, which involves identifying PK-DDI sentences, we have found that Flan-T5, Tk-instruct, and Vicuna consistently deliver results with high sensitivity and specificity. T0pp shows a promising result in zero-shot learning but falls short in few shot scenarios. This may be expected, as T0pp has been extensively trained for zero-shot learning [23]. Lastly, the baseline (non-fine-tuned) BioBERT model exhibited lower performance across the first task of this study, indicating its unsuitability for few-shot and zero-shot learning. Of note, our previous study that fine-tuned BioBERT model on extensive training sentences had evaluation metrics exceeding 80%, showing how subsequent training (fine-tuning) using large quantities of data can significantly enhance its performance [8]. Nevertheless, in terms of identifying PK-DDI sentences, a direct comparison (Table 2) shows that our LLMs, without any subsequent training (zero-shot learning), deliver superior performance and outperform the BioBERT that was fine-tuned on large quantities of training data [8].

Likewise, regarding the second task (intrinsic factors affecting drug’s clinical exposure), Flan-T5 in the context of both few-shot and zero-shot learning demonstrates satisfactory performance. Consequently, Flan-T5 was applied, without any training examples (zero-shot learning), to the real-world task: scanning all FDA prescription drug labels to identify intrinsic factors that can affect the clinical exposure of a selected list of drugs of interest. Even though it is difficult to fully evaluate the performance of Flan-T5 on such a huge unlabeled dataset, a thorough manual examination suggests it has a precision rate of 78.5%.

About prompt selection

We evaluated LLMs using a series of technical questions phrased in different ways for the purpose of our prompt selections. We observed that, for zero-shot learning, because a model was not given any training examples, a prompt that provides detailed definitions for some technical terms in the question is crucial for the models to perform efficiently. For instance, in our first task (identifying PK-DDI sentences), a basic prompt like ‘Identify the drug-drug interaction type in this sentence’ followed by two options (pharmacokinetic DDI versus non-pharmacokinetic DDI) for the Flan-T5 model to choose from, yielded satisfactory results. However, when the prompt was enhanced with more detailed explanation of pharmacokinetic DDI, the model’s performance was further improved. A more striking trend was observed in our second task. A vague prompt like ‘Does this statement relate to the pharmacokinetic properties influenced by intrinsic factors?’ was insufficient and led to poor performance. A more detailed prompt giving the definition of ‘intrinsic factors’ improved the model's accuracy substantially (Table 1: Task 2 concerning the Flan-T5 prompt). A similar pattern was observed for other LLMs. This suggests that these LLMs achieved different levels of understanding for different technical terms (i.e. ‘understand’ the term ‘pharmacokinetics’ better than the term ‘intrinsic factors’) through the pre-training procedures using large quantities of general corpora. When applied to a more domain-specific task, a more detailed explanation of some technical terms needs to be incorporated into the prompt so that the model can ‘understand’ the question through in-context learning [47–49]. For few-shot learning, we adopted a prompt style similar to zero-shot learning, with randomly selected example sentences added to facilitate in-context learning. There are reports that a more carefully selected and dynamically adjusted exemplar set might improve the performance [50–52], which can be future evaluated in our future studies.

Real-world application and error analysis

Our study extended beyond theoretical evaluations, applying the Flan-T5 model to a real-world task of identifying intrinsic factors affecting drug’s clinical exposure to CiPA drugs employing 708 024 sentences from all available FDA labels for prescription drugs (As of March 2023). This choice of data source allowed us to test the model's efficacy in a real-world, data-sensitive context, assessing its ability to identify complex relations in regulatory documents. The Flan-T5 model's precision rate of 78.5% in predicting true positives underscores its practical utility and reliability. This application demonstrates the feasibility of using locally implemented open-source LLMs in real-world biomedical contexts and opening the way for future research and development in this area.

The performance analysis revealed that certain sentences are written in a neutral, informative style. As an example, consider the following sentence: ‘Accumulation of ketorolac tromethamine has not been studied in special populations (elderly patients, renal failure patients, or hepatic disease patients)’. Here, the sentence describes the lack of research about the patient's intrinsic factors affecting the drug's pharmacokinetics. In instances like these, we manually annotated them as a negative, signifying their irrelevance to the patient's intrinsic factors that could modify the drug’s clinical exposure. However, unlike our annotation, the Flan-T5 model wrongly marked them as positive, which indicates that even though the model recognizes the patient's intrinsic factor, there is an insufficiency in the model's understanding of context and content relevance.

Occasionally, a sentence may not provide sufficient information to be classified as positive. For example, the sentence: ‘In healthy men, orally administered verapamil hydrochloride undergoes extensive metabolism in the liver’. It is true that a drug's extensive metabolism in the liver in some individuals causes a drug's pharmacokinetics to change [53]. However, the sentence itself does not indicate altering the drug's pharmacokinetics directly. We manually annotated as false, since it is irrelevant to intrinsic factors affecting drug’s clinical exposure. However, the model misclassified this sentence.

Practical implications and limitations

Implementation of open source LLMs within a secure local network, such as within a drug regulatory agency, could substantially enhance the efficiency and effectiveness of drug review process. These models could help regulators to extract critical information or identify missing information from the submission documents, compare new information with existing datasets and historical documents, and identify potential safety signals from adverse event reports and post-market surveillance data. As illustrated by our initial evaluation, it is possible to harness LLMs’ few-shot or even zero-shot learning capabilities to quickly adapt to different tasks and evolving regulatory needs.

However, it is crucial to address several potential limitations and challenges when deploying open-source LLMs within a regulatory agency. Some of the limitations are technical challenges. For example, LLMs require substantial computational resources, which may not be readily available for all organizations. At the time of writing, we were only able to implement LLMs with maximally 20 billion parameters with the available computing resources in the local network, while the biggest open source LLMs have three to five times more parameters. As a result, we were unable to compare the performance across LLMs with the biggest sizes, although error analysis of representative misclassified sentences (see previous section as well as reference [8]) suggests these sentences may contain incomplete or ambiguous information, and are less likely to be rectified simply by employing larger models. In addition, implementing and applying these open LLMs to specific tasks demands specialized technical skills in data science and IT infrastructure. Making these tools available to regular reviewers might necessitate investment in extensive personnel training or hiring experts to establish an internal graphical user interface, adding to the technical and engineering burden on the agency.

Some other limitations are inherent to LLMs, which are usually pre-trained on large amounts of public data. These data often include noises and biases, potentially degrading model performance if the training data are not sufficiently comprehensive. The robustness of these models can also be compromised by their sensitivity to minor input variations, leading to inconsistent predictions. Lastly, performance metrics derived from controlled test conditions might not accurately reflect the model's behavior in more diverse and complex real-world scenarios, especially when dealing with imbalanced datasets or when the model is applied across different disciplines.

Our study demonstrates the performance of various open-source LLMs on selected datasets but lacks the extensive validation needed to generalize the results to other regulatory contexts. The inherent limitations of few-shot and zero-shot learning could conceivably result in suboptimal outcomes for certain highly specialized tasks. Future validations to test the models' capability to handle other regulatory tasks, including efficacy assessment and risk management, using zero- or few-shot learning are planned. This subsequent work will aim to explore the broader applicability of these models in different regulatory environments, further assessing their robustness and adaptability in real-world scenarios. Additionally, for future work, it is possible to implement a Mixture of Experts strategy (MoE) [54, 55] and combine the strengths of different LLMs. These strategies are expected to improve performance and offer new insights into the best practices for model integration in regulatory settings. This expansion will help assess the versatility and effectiveness of LLMs across various facets of the regulatory decision-making process, providing a more comprehensive view of their potential impacts and benefits.

Conclusion

In conclusion, our research evaluated the feasibility of implementing open-source LLMs in a local, secure network and quickly adapting them to cover various NLP tasks. While a special emphasis of this study was on identifying pharmacokinetic interactions (DDI or intrinsic factors affecting drug exposure) from FDA drug labels, it is likely that the findings can provide insights into generally applying these LLMs to various NLP tasks without large quantities of task-specific training data in a regulatory setting.

Key Points

• Despite the availability of commercially hosted LLMs like ChatGPT, using such services may inadvertently expose sensitive information and risk data breaches, especially for sensitive sectors like healthcare and regulatory agencies.

• As a feasibility evaluation, we implemented a series of open-source LLMs within a regulatory agency’s local network and assessed their performance on specific tasks involving extracting relevant clinical pharmacology information from regulatory drug labels.

• Some open-source LLMs demonstrated a zero-shot or few-shot performance that is comparable to, or even better than, neural network models that needed thousands of training samples, pointing to the possibility of implementing LLMs within a secure local network and using these models to perform various natural language processing tasks when large numbers of training examples are unavailable.

Funding

This study was supported by the Center for Drug Evaluation and Research (CDER) and administered by the Oak Ridge Institute for Science and Education (ORISE) under an interagency agreement with the US Department of Energy (DOE) and the Food and Drug Administration (FDA). Also, this project used Betsy High-Performance computational resources at FDA, Center for Devices and Radiological Health (CDRH).

Conflict of interest: The authors declare that there are no conflicts of interest regarding the publication of this paper.

Author contributions

Zhihua Li (Conceptualization), Hamed Meshkin and Zhihua Li (Methodology), Hamed Meshkin, Joel Zirkle, and Zhihua Li (Dataset Validation and Verification), Hamed Meshkin and Zhihua Li (Data Analysis and Research Investigation), Hamed Meshkin. and Zhihua Li (Writing—Original Draft), Hamed Meshkin, Joel Zirkle, Anik Chaturbedi, Shilpa Chakravartula, Ghazal Arabidarrehdor, John Mann, Bradlee Thrasher, and Zhihua Li (Writing—Review and Editing), and Zhihua Li (Supervision).

Disclaimer

The project and conclusions in this study reflect the views of the authors and should not be interpreted as representing the views or policies of the Food and Drug Administration, the Department of Health and Human Services, or the United States government.

Data Availability

The large language models used in this study are open source and can be accessed through the following links:

Flan-T5-xxl: https://huggingface.co/google/flan-t5-xxl

Tk-Instruct: https://huggingface.co/allenai/tk-instruct-3b-def-pos-neg-expl

T0pp: https://huggingface.co/bigscience/T0pp

Vicuna: https://huggingface.co/lmsys/vicuna-13b-v1.5

SetFit: https://huggingface.co/docs/setfit/en/installation