We define “knowledge” as entities of interest to researchers working with PDCMs in the cancer research field. For example, the patient’s diagnosis provides a reference point to confirm that a PDCM faithfully recapitulates the biology of the original tumor and is essential for ensuring the model’s relevance and reliability in studies of cancer progression or treatment response. Thus, “diagnosis” is important to understand the model’s characteristics in the context of patient’s disease. The patient’s age can significantly affect the molecular and genetic characteristics of the tumor. For example, pediatric cancers often have distinct genetic drivers and tumor microenvironments compared to cancers in older adults. In addition, age-related biological factors, such as immune system, metabolism, and hormone levels, influence how a tumor responds to treatments. Thus, knowing the patient’s age is imperative for predictive accuracy of the model in preclinical testing and relevance of research findings. Therefore, we selected 15 most commonly used CancerModels.Org data model attributes (Table 1), which include the attributes defined in the minimal information standard for patient-derived xenograft models [19] and the draft minimal information standard for in vitro models [20].

We used 100 abstracts to develop the gold-standard corpus annotated for the 15 entities (Table 1). The abstracts were chosen from publications linked to the PDCMs submitted to CancerModels.Org platform. They were selected to cover all 3 types of models in the resource-PDXs, organoids, and cell lines. The final corpus is available on GitHub (see Data and Code Availability section).

Three annotators (ZP, TM, and EL) independently labeled entities in all 100 abstracts for a total of 40 hours. The annotation quality was tracked through interannotator agreement (IAA), a measure of agreement between each annotation produced by different annotators working on the same dataset. The IAA is an indication of how difficult the task is for humans and it becomes the target for system development. We used pairwise F₁-score as the IAA metric [21] in the exact match setting that requires the spans of the annotators to match exactly. We computed the agreement between each pair of annotators and averaged across the 3 sets of scores. The final IAA for each entity type is reported in Table 1. The IAA range is 42‐100 indicating moderate agreement. Note that the lowest agreement is for low occurrence entity types, for example, cancer_grade has only 8 instances with 42 IAA. These low-frequency entity types are more likely to be overlooked by the human experts as annotation is a cognitively demanding task. Thus, to ensure a high-quality gold-standard dataset, we overlayed the single annotations with an adjudication step, where the annotators discussed annotation disagreements and potential missed annotations to come to final joint decisions. The resulting gold dataset spans 15 entity types for a total 3313 entity mentions (refer Table 2 for distributions) was split into training, development, and test sets in the standard 60:20:20 ratio. The train set was used for creating entity extraction algorithms, the development set for refining the algorithms, and the test set for the final evaluation.

Various prompting techniques have been proposed since the emergence of LLMs [22-25]. At a high level, these prompting techniques can be divided into 2 categories, direct prompting [4] and soft prompting [18,24,26] . The main difference between the two methods is the prompt representation, that is whether the prompt consists of human language words or vectors (Figure 1). Direct prompting (or discrete prompting) is the most intuitive and now classic prompting method where users directly interact with LLMs using natural language. For example, a user may ask ChatGPT to “Write a thank you note to an old friend of my parents”; in this case, the text within the quotation marks is a discrete prompt. Soft prompting (or continuous prompting) uses a machine learning approach to train a sequence of continuous vectors, which are the “virtual tokens” of the prompt. It is worth noting that soft prompting differs from fine tuning. With soft prompting, the LLM parameters are not updated, only the soft prompt parameters are adjusted. In contrast, finetuning requires to update the parameters of the entire LLM, and therefore needs more computation resources. Both prompting techniques have their advantages and disadvantages. Compared to direct prompting, soft prompting does not require the tedious process of manually creating prompts; however, it requires some labeled data to train the prompt. In this work, we explore both direct and soft prompting as we aim to explore the latest developments in LLMs and prompting techniques for the task of extracting PDCM entities from abstracts of academic papers.

When asking LLMs to extract entities such as diagnoses or biomarkers, the most intuitive way is to ask LLMs to output the entities directly. In example 1 below, “ALK” is a biomarker entity. One may expect the model to output {“biomarker” [ALK]}. However, we note that the string “ALK” is mentioned multiple times in this example text, therefore it is not clear which “ALK” the model refers to. To get the most precise extraction to facilitate a more fine-grained analysis, we instruct the model to output the offsets of the specific mentions in the text (ie, the spans). For instance, if the model gives us [(48, 51, “ALK,” biomarker), (323, 326, “ALK,” biomarker), …], we know that from character 48 to character 51, there is a biomarker entity, “ALK.” Similarly, we can find another biomarker entity “ALK” at position 323‐326.

We started our exploration by designing prompts with an explicit instruction to specify the character offsets of each entity along with the entity text and type (eg, 48, 51, “ALK”, biomarker). However, our experiments show that it was challenging for the LLM to output the correct character offsets, a seemingly straightforward task (all the model needs to do is to count the number of characters); however, the complexity of this seemingly straightforward task is likely due to the LLM’s way of breaking words outside its vocabulary into so-called word pieces, for example, “organoid” is broken down into 2 word pieces “organ” and “-oid.” Considering that LLMs were trained as generative models [3,4], we subsequently cast the entity extraction task as a generation task, where we instructed the model to mark the entities with XML tags. For instance, if the model outputs “Oncogenic fusion of anaplastic lymphoma kinase (<biomarker>ALK</biomarker>) with echinoderm microtubule …,” then postprocessing the output with regular expressions would find the exact position of “ALK” in the text. Specifically, we asked the LLMs to mark the start and end of an entity with <entity_type> and </entity_type> tags, where entity_type is a placeholder for the specific entity type, such as biomarker or treatment (refer Table 1 for the full list).

Designing the direct prompts manually could be time-consuming and minor changes in the prompt language could lead to drastic changes in the model performance [24,27]. On the other hand, soft prompting requires some amount of gold data for its training and annotating gold data by domain experts could also be time-consuming. Fortunately, only a small set of labeled data are needed to train soft prompts. As described above, we created a gold dataset, which we used for training and evaluating our soft prompting approach.

There are a few soft prompting methods, the difference usually lies in how the prompt vectors are initialized and learned. Prompt-tuning [18] is a technique that learns the prompt by adding a list of virtual tokens (ie, vectors) in front of the input, where the virtual tokens can be randomly initialized, or drawn from a pretrained word embedding [28] set. Another method is P-tuning [24], which uses small neural networks such as feedforward neural networks [29] (multilayer perceptron) or recurrent neural networks [30] (eg, long-short term memory) as the prompt encoder to learn the prompt. Only the parameters in the prompt encoder are updated during training, while the weights in the LLMs remain frozen. In our experiments, we found P-tuning did not always converge to an optimal solution for our task perhaps due to the random initialization of the vectors rather than using carefully pretrained word embeddings. Therefore, we focused on prompt-tuning in this work. Following Lester et al [18], we initialized the vectors in the prompt with the embeddings of the label words in the entity type set (Table 1).

The standard approach for entity extraction in natural language processing is via token classification [31]. Concretely, a classifier is trained to predict the label for each token in a sentence according to a predefined label set. Additionally, each label is prepended with a B or I prefix to indicate the entity’s Beginning or Inside mention, respectively. An example is provided in Figure 2. “Ewing sarcoma” is an entity mention of the diagnosis type. Thus “Ewing” and “sarcoma” are labeled as “Diagnosis,” while all other tokens are labeled as “O,” meaning they are Outside of an entity. To be more precise, “Ewing” is at the beginning of the diagnosis entity, and “sarcoma” is inside of the entity, so they are labeled as “B-Diagnosis” and “I-Diagnosis,” respectively.

To summarize, we trained a multiclass classifier for the soft-prompting training step. There are 15 entity types in our dataset, therefore there are 15×2+1=31 labels for token classification, with one extra label for “O.”

For efficiency purposes, we used Apache cTAKES [32] to split an abstract into sentences which were then passed to the LLMs to extract entities one sentence at a time. Our direct prompt included the instruction, the definition of each entity type, 5 examples (few-shot in-context learning) and the query (the sentence). The in-context learning [4] is a common practice in LLM prompting and has consistently shown improved results as the examples guide the LLM onto an optimal path [33,34]. Figure 3 presents our prompt template, and examples are in Multimedia Appendix 1.

When choosing the LLMs, we used GPT-4o [35], one of the most powerful proprietary LLMs at the time of this study, and SOTA open LLMs from the LLaMA3 family [36], including LLaMA3.1 70B, LLaMA3.1 8B, LLaMA3.2 1B, and LLaMA3.2 3B. We did not use GPT-4o or LLaMA3.1 70B to train the soft prompts due to computational resource limitations; thus, our work here is representative of the computational environment in the vast majority of academic medical centers and research labs. We set the soft prompt length to 30. We trained the soft prompt on the training set for 50 epochs with a learning rate of 0.001. Hyperparameters were tuned on the development set using the LLaMA3.1 8B model.

We report the evaluation results on the test set in the next section. In addition, we apply 5-fold cross-validation and report the average scores with SDs. For the 5-fold cross-validation, we excluded the 3 abstracts used to sample the gold examples for direct prompting and split the remaining 97 abstracts into 5 folds with a 20:20:20:20:17 ratio. For direct prompting, we ran the model on each fold and reported the average scores. For soft prompting, we set aside one fold as the test set and trained the soft prompts on the remaining 4 folds.

Our experiments demonstrate that soft prompting, a relatively underexplored aspect of LLM prompting, can significantly enhance the performance of smaller LLMs. The 3 LLaMA models exhibit comparable performance under soft prompting (an F₁-score of 46 in the exact match setting, and 70 in the overlapping match setting). These results are particularly promising results given the limited training data, consisting of 60 abstracts with 2089 entity mentions. Please note that all F₁-scores mentioned in this section refer to the F₁-scores on the test set.

How much data is needed to train the soft prompt? To answer this question, we trained LLaMA3.2 1B model, the smallest model used in this work, with different amounts of training data. Figure 4 shows the relation between the proportion of training data and the F₁-scores on the test set (overlapping match). Solid performance was achieved with only 5% of the training data (26 sentences from 3 abstracts). With 25% of the training data (129 sentences from 15 abstracts), the model achieved an F₁-score of 68.21, only 2 points lower than using the entire training set, and only 3 points lower than GPT4-o with direct prompting. Despite the impressive performance of GPT4-o direct prompting, one potential issue is that not all data used in biomedical research can be sent to proprietary models such as GPT or the Gemini family models [8] via public application programming interfaces. That is, for applications using real patient data that require Health Insurance Portability and Accountability Act–compliant platforms, our findings demonstrate that achieving performance comparable to proprietary LLMs such as GPT4-o remains feasible through soft prompting. However, this approach necessitates a tradeoff, requiring a small set of labeled data for optimal effectiveness.

Some entities appear more frequently than other entities in our dataset. For example, diagnosis and treatment mentions are more frequent than mentions of cancer_grade. In Table 7, we present the number of instances of each entity type in our dataset and the corresponding performance of GPT4-o direct prompting. We can see that GPT4-o performs the best for the entity types that have the most instances—diagnosis, model type, and treatment entities. Of these frequent entity types, biomarker is the one with the lowest performance. Our error analysis points to several factors that could have contributed to these results, including ambiguous and inconsistent mentions and contextual dependencies. In this task, we defined a biomarker as “gene, protein or biological molecule identified in or associated with patient’s/model’s tumor.” Thus, biomarker entities can be mentioned using their full names (eg, epidermal growth factor receptor, lnc-RP11-536 K7.3, echinoderm microtubule-associated protein-like 4), standardized gene or protein symbols (NPM1, KRAS, PTEN) or abbreviations of metabolites (NADPH, D2HG). Moreover, a biomarker entity (eg, “MEK”) often overlaps with a treatment entity (eg, “MEK inhibitor”). The ambiguity in biomarker entity mentions might interfere with the model’s ability to recognize them consistently. In addition, biomarker entities are often mentioned as lists (see Example 2) resulting in a different frequency within and across the abstracts and patterns of entity mentions, in comparison with other entities. Overall, ambiguity emerges as the primary source of error. More precise definitions, accompanied by examples illustrating the distinct meanings, might present a solution. Table S2 in Multimedia Appendix 1 provides the breakdown of errors per entity type along with examples.

The moderate performance of entity types such as genetic_effect, molecular_char, and response_to_treatment, and tumour_type is due to the number of training instances ranging from 61 to 128 as well as the IAA ranging from 49.67 to 57.67. The moderate IAA scores of those entity types underscore the need for refined annotation protocols and modeling strategies that better capture domain-specific knowledge. Furthermore, the lower performance observed for entity types with smaller sample sizes (eg, model_id) highlights the need for enhancing model performance on low-frequency labels. Future research could explore strategies such as data augmentation to improve the model’s generalizability for underrepresented entities.

The extraction of PDCM-relevant knowledge is not an easy task for the domain experts as indicated by the IAA (F₁-score below 65 for all entity types except for model_id). In 9 out of 15 entity types, the system performance in an overlapping match setting exceeds the IAA (last two columns of Table 7). This is the case for categories with plentiful training instances (eg, diagnosis, model_type) as well as for categories with fewer training instances (eg, sample_type, cancer_grade). For the exact match setting, in 6 out of 15 entity types, the system performance exceeds the IAA (last two columns in Table 8). Therefore, the LLM could be a viable assistant, with its outputs reviewed by a domain expert to ensure the accuracy of the finalextraction. We believe such human-in-the-loop approaches present a promising direction for future research and application.

We would like to note that the work presented in the paper was done in a computational environment representative of the vast majority of academic medical centers and nonindustry research labs. Although we have access to SOTA Graphics Processing Units, we still found ourselves constrained as to the extent to which we could use very large language models. The larger community needs to address the growing gap in computational resources between big tech and the rest of the research community.

As this is a feasibility study, we limited ourselves to the extraction of entity mentions of 15 entity types chosen from attributes in the descriptive standards for PDCMs. While these are recognized by the PDCM and oncology community, they do not cover all knowledge in the PDCM-relevant texts. Some refinement of the entity types will be beneficial to improve prompting results.

We limited our corpus to 100 abstracts from papers associated with PDCMs deposited in CancerModels.Org. We did not assess the abstracts for the presence and equal distribution of all the entities. Thus, there were very few mentions of some entities in the corpus (eg, cancer_stage), negatively affecting our overall F₁-score. We decided not to exclude those entities as these results could guide refinements of future studies. The computational methods discussed here are applicable to other studies requiring the extraction of textual information from scientific papers. Future work could involve extending this method to extract knowledge from the main body of the papers.

This study investigates the potential of LLMs as powerful tools for extracting PDCM-relevant knowledge from scientific literature—an essential task for advancing cancer research and precision medicine. By comparing direct and soft prompting across both proprietary and open LLMs, we provide valuable insights into the most effective strategies for PDCM-relevant knowledge extraction. Our findings indicate that GPT-4o, when used with direct prompting, maintains competitive performance, while soft prompting significantly enhances the effectiveness of smaller LLMs. In conclusion, our results demonstrate that training soft prompts on smaller open models can achieve performance levels comparable to those of proprietary LLMs.

To our knowledge, this is the first study to implement SOTA LLMs prompting for knowledge extraction in the PDCM domain and the first to explore the emerging topic of soft prompting in this context. Our findings demonstrate that LLMs can effectively streamline the extraction of complex cancer model metadata, potentially reducing the burden of manual curation and accelerating the integration of PDCMs into research and clinical workflows. Additionally, this study lays the foundation for future research aimed at optimizing LLMs for large-scale knowledge extraction tasks. Efficiently extracting and harmonizing PDCM-relevant knowledge will ultimately drive progress in cancer research and precision oncology, equipping researchers and clinicians with better tools to improve patient outcomes. More broadly, our study contributes to the ongoing discourse on the applicability of LLMs, acknowledging that while they offer transformative potential, they are not a universal solution for all tasks.