UNASSIGNED: Clinical note section identification helps locate relevant information and could be beneficial for downstream tasks such as named entity recognition. However, the traditional supervised methods suffer from transferability issues. This study proposes a new framework for using large language models (LLMs) for section identification to overcome the limitations.
UNASSIGNED: We framed section identification as question-answering and provided the section definitions in free-text. We evaluated multiple LLMs off-the-shelf without any training. We also fine-tune our LLMs to investigate how the size and the specificity of the fine-tuning dataset impacts model performance.
UNASSIGNED: GPT4 achieved the highest F1 score of 0.77. The best open-source model (Tulu2-70b) achieved 0.64 and is on par with GPT3.5 (ChatGPT). GPT4 is also found to obtain F1 scores greater than 0.9 for 9 out of the 27 (33%) section types and greater than 0.8 for 15 out of 27 (56%) section types. For our fine-tuned models, we found they plateaued with an increasing size of the general domain dataset. We also found that adding a reasonable amount of section identification examples is beneficial.
UNASSIGNED: These results indicate that GPT4 is nearly production-ready for section identification, and seemingly contains both knowledge of note structure and the ability to follow complex instructions, and the best current open-source LLM is catching up.
UNASSIGNED: Our study shows that LLMs are promising for generalizable clinical note section identification. They have the potential to be further improved by adding section identification examples to the fine-tuning dataset.

Molecular Property Prediction (MPP) is vital for drug discovery, crop protection, and environmental science. Over the last decades, diverse computational techniques have been developed, from using simple physical and chemical properties and molecular fingerprints in statistical models and classical machine learning to advanced deep learning approaches. In this review, we aim to distill insights from current research on employing transformer models for MPP. We analyze the currently available models and explore key questions that arise when training and fine-tuning a transformer model for MPP. These questions encompass the choice and scale of the pretraining data, optimal architecture selections, and promising pretraining objectives. Our analysis highlights areas not yet covered in current research, inviting further exploration to enhance the field\'s understanding. Additionally, we address the challenges in comparing different models, emphasizing the need for standardized data splitting and robust statistical analysis.

BACKGROUND: Decoding human genomic sequences requires comprehensive analysis of DNA sequence functionality. Through computational and experimental approaches, researchers have studied the genotype-phenotype relationship and generate important datasets that help unravel complicated genetic blueprints. Thus, the recently developed artificial intelligence methods can be used to interpret the functions of those DNA sequences.
METHODS: This study explores the use of deep learning, particularly pre-trained genomic models like DNA_bert_6 and human_gpt2-v1, in interpreting and representing human genome sequences. Initially, we meticulously constructed multiple datasets linking genotypes and phenotypes to fine-tune those models for precise DNA sequence classification. Additionally, we evaluate the influence of sequence length on classification results and analyze the impact of feature extraction in the hidden layers of our model using the HERV dataset. To enhance our understanding of phenotype-specific patterns recognized by the model, we perform enrichment, pathogenicity and conservation analyzes of specific motifs in the human endogenous retrovirus (HERV) sequence with high average local representation weight (ALRW) scores.
RESULTS: We have constructed multiple genotype-phenotype datasets displaying commendable classification performance in comparison with random genomic sequences, particularly in the HERV dataset, which achieved binary and multi-classification accuracies and F1 values exceeding 0.935 and 0.888, respectively. Notably, the fine-tuning of the HERV dataset not only improved our ability to identify and distinguish diverse information types within DNA sequences but also successfully identified specific motifs associated with neurological disorders and cancers in regions with high ALRW scores. Subsequent analysis of these motifs shed light on the adaptive responses of species to environmental pressures and their co-evolution with pathogens.
CONCLUSIONS: These findings highlight the potential of pre-trained genomic models in learning DNA sequence representations, particularly when utilizing the HERV dataset, and provide valuable insights for future research endeavors. This study represents an innovative strategy that combines pre-trained genomic model representations with classical methods for analyzing the functionality of genome sequences, thereby promoting cross-fertilization between genomics and artificial intelligence.

Fine-tuning is an important technique in transfer learning that has achieved significant success in tasks that lack training data. However, as it is difficult to extract effective features for single-source domain fine-tuning when the data distribution difference between the source and the target domain is large, we propose a transfer learning framework based on multi-source domain called adaptive multi-source domain collaborative fine-tuning (AMCF) to address this issue. AMCF utilizes multiple source domain models for collaborative fine-tuning, thereby improving the feature extraction capability of model in the target task. Specifically, AMCF employs an adaptive multi-source domain layer selection strategy to customize appropriate layer fine-tuning schemes for the target task among multiple source domain models, aiming to extract more efficient features. Furthermore, a novel multi-source domain collaborative loss function is designed to facilitate the precise extraction of target data features by each source domain model. Simultaneously, it works towards minimizing the output difference among various source domain models, thereby enhancing the adaptability of the source domain model to the target data. In order to validate the effectiveness of AMCF, it is applied to seven public visual classification datasets commonly used in transfer learning, and compared with the most widely used single-source domain fine-tuning methods. Experimental results demonstrate that, in comparison with the existing fine-tuning methods, our method not only enhances the accuracy of feature extraction in the model but also provides precise layer fine-tuning schemes for the target task, thereby significantly improving the fine-tuning performance.

Small-molecule drug design aims to generate compounds that target specific proteins, playing a crucial role in the early stages of drug discovery. Recently, research has emerged that utilizes the GPT model, which has achieved significant success in various fields to generate molecular compounds. However, due to the persistent challenge of small datasets in the pharmaceutical field, there has been some degradation in the performance of generating target-specific compounds. To address this issue, we propose an enhanced target-specific drug generation model, Adapt-cMolGPT, which modifies molecular representation and optimizes the fine-tuning process. In particular, we introduce a new fine-tuning method that incorporates an adapter module into a pre-trained base model and alternates weight updates by sections. We evaluated the proposed model through multiple experiments and demonstrated performance improvements compared to previous models. In the experimental results, Adapt-cMolGPT generated a greater number of novel and valid compounds compared to other models, with these generated compounds exhibiting properties similar to those of real molecular data. These results indicate that our proposed method is highly effective in designing drugs targeting specific proteins.

Our best current science seems to suggest the laws of physics and the initial conditions of our universe are fine-tuned for the possibility of life. A significant number of scientists and philosophers believe that the fine-tuning is evidence for the multiverse hypothesis. This paper will focus on a much-discussed objection to the inference from the fine-tuning to the multiverse: the charge that this line of reasoning commits the inverse gambler\'s fallacy. Despite the existence of a literature going back decades, this philosophical debate has made little contact with scientific discussion of fine-tuning and the multiverse, which mainly revolves around a specific form of the multiverse hypothesis rooted in eternal inflation combined with string theory. Because of this, potentially important implications from science to philosophy, and vice versa, have been left underexplored. In this paper, I will take a first step at joining up these two discussions, by arguing that attention to the eternal inflation + string theory conception of the multiverse supports the inverse gambler\'s fallacy charge. It does this by supporting the idea that our universe is contingently fine-tuned, thus addressing the concern that proponents of the inverse gambler\'s fallacy charge have assumed this without argument.

BACKGROUND: Large language models (LLMs) have the potential to support promising new applications in health informatics. However, practical data on sample size considerations for fine-tuning LLMs to perform specific tasks in biomedical and health policy contexts are lacking.
OBJECTIVE: This study aims to evaluate sample size and sample selection techniques for fine-tuning LLMs to support improved named entity recognition (NER) for a custom data set of conflicts of interest disclosure statements.
METHODS: A random sample of 200 disclosure statements was prepared for annotation. All \"PERSON\" and \"ORG\" entities were identified by each of the 2 raters, and once appropriate agreement was established, the annotators independently annotated an additional 290 disclosure statements. From the 490 annotated documents, 2500 stratified random samples in different size ranges were drawn. The 2500 training set subsamples were used to fine-tune a selection of language models across 2 model architectures (Bidirectional Encoder Representations from Transformers [BERT] and Generative Pre-trained Transformer [GPT]) for improved NER, and multiple regression was used to assess the relationship between sample size (sentences), entity density (entities per sentence [EPS]), and trained model performance (F1-score). Additionally, single-predictor threshold regression models were used to evaluate the possibility of diminishing marginal returns from increased sample size or entity density.
RESULTS: Fine-tuned models ranged in topline NER performance from F1-score=0.79 to F1-score=0.96 across architectures. Two-predictor multiple linear regression models were statistically significant with multiple R2 ranging from 0.6057 to 0.7896 (all P<.001). EPS and the number of sentences were significant predictors of F1-scores in all cases ( P<.001), except for the GPT-2_large model, where EPS was not a significant predictor (P=.184). Model thresholds indicate points of diminishing marginal return from increased training data set sample size measured by the number of sentences, with point estimates ranging from 439 sentences for RoBERTa_large to 527 sentences for GPT-2_large. Likewise, the threshold regression models indicate a diminishing marginal return for EPS with point estimates between 1.36 and 1.38.
CONCLUSIONS: Relatively modest sample sizes can be used to fine-tune LLMs for NER tasks applied to biomedical text, and training data entity density should representatively approximate entity density in production data. Training data quality and a model architecture\'s intended use (text generation vs text processing or classification) may be as, or more, important as training data volume and model parameter size.

Peptide- and protein-based therapeutics are becoming a promising treatment regimen for myriad diseases. Toxicity of proteins is the primary hurdle for protein-based therapies. Thus, there is an urgent need for accurate in silico methods for determining toxic proteins to filter the pool of potential candidates. At the same time, it is imperative to precisely identify non-toxic proteins to expand the possibilities for protein-based biologics. To address this challenge, we proposed an ensemble framework, called VISH-Pred, comprising models built by fine-tuning ESM2 transformer models on a large, experimentally validated, curated dataset of protein and peptide toxicities. The primary steps in the VISH-Pred framework are to efficiently estimate protein toxicities taking just the protein sequence as input, employing an under sampling technique to handle the humongous class-imbalance in the data and learning representations from fine-tuned ESM2 protein language models which are then fed to machine learning techniques such as Lightgbm and XGBoost. The VISH-Pred framework is able to correctly identify both peptides/proteins with potential toxicity and non-toxic proteins, achieving a Matthews correlation coefficient of 0.737, 0.716 and 0.322 and F1-score of 0.759, 0.696 and 0.713 on three non-redundant blind tests, respectively, outperforming other methods by over $10\\%$ on these quality metrics. Moreover, VISH-Pred achieved the best accuracy and area under receiver operating curve scores on these independent test sets, highlighting the robustness and generalization capability of the framework. By making VISH-Pred available as an easy-to-use web server, we expect it to serve as a valuable asset for future endeavors aimed at discerning the toxicity of peptides and enabling efficient protein-based therapeutics.

OBJECTIVE: We aim to develop a novel method for rare disease concept normalization by fine-tuning Llama 2, an open-source large language model (LLM), using a domain-specific corpus sourced from the Human Phenotype Ontology (HPO).
METHODS: We developed an in-house template-based script to generate two corpora for fine-tuning. The first (NAME) contains standardized HPO names, sourced from the HPO vocabularies, along with their corresponding identifiers. The second (NAME+SYN) includes HPO names and half of the concept\'s synonyms as well as identifiers. Subsequently, we fine-tuned Llama 2 (Llama2-7B) for each sentence set and conducted an evaluation using a range of sentence prompts and various phenotype terms.
RESULTS: When the phenotype terms for normalization were included in the fine-tuning corpora, both models demonstrated nearly perfect performance, averaging over 99% accuracy. In comparison, ChatGPT-3.5 has only ∼20% accuracy in identifying HPO IDs for phenotype terms. When single-character typos were introduced in the phenotype terms, the accuracy of NAME and NAME+SYN is 10.2% and 36.1%, respectively, but increases to 61.8% (NAME+SYN) with additional typo-specific fine-tuning. For terms sourced from HPO vocabularies as unseen synonyms, the NAME model achieved 11.2% accuracy, while the NAME+SYN model achieved 92.7% accuracy.
CONCLUSIONS: Our fine-tuned models demonstrate ability to normalize phenotype terms unseen in the fine-tuning corpus, including misspellings, synonyms, terms from other ontologies, and laymen\'s terms. Our approach provides a solution for the use of LLMs to identify named medical entities from clinical narratives, while successfully normalizing them to standard concepts in a controlled vocabulary.

The newly released Segment Anything Model (SAM) is a popular tool used in image processing due to its superior segmentation accuracy, variety of input prompts, training capabilities, and efficient model design. However, its current model is trained on a diverse dataset not tailored to medical images, particularly ultrasound images. Ultrasound images tend to have a lot of noise, making it difficult to segment out important structures. In this project, we developed ClickSAM, which fine-tunes the Segment Anything Model using click prompts for ultrasound images. ClickSAM has two stages of training: the first stage is trained on single-click prompts centered in the ground-truth contours, and the second stage focuses on improving the model performance through additional positive and negative click prompts. By comparing the first stage\'s predictions to the ground-truth masks, true positive, false positive, and false negative segments are calculated. Positive clicks are generated using the true positive and false negative segments, and negative clicks are generated using the false positive segments. The Centroidal Voronoi Tessellation algorithm is then employed to collect positive and negative click prompts in each segment that are used to enhance the model performance during the second stage of training. With click-train methods, ClickSAM exhibits superior performance compared to other existing models for ultrasound image segmentation.