The scientific community may uncover the nature of life due to the great amount of biological data at its disposal. Implementing bioinformatics ensures converting this information into very valuable ones for decoding amino-acid sequences and the formation of new treatment strategies. However, the projects are big enough and need to be completed with modern techniques and outfits. Human assistants: Large Language Models (LLMs) are a potential ally in the biology field because of their great language processing competence. It is worth pointing out that the field of bioinformatics has just discovered the cutting-edge use of these models.

This blog is oriented toward the comprehensive examination of the ability of LLMs to solve many crucial bioinformatics issues, which include determining the encoding sites of the gene, identifying the gene sequence, detecting antibacterial and anticancer peptides, constructing molecules, and notifying the associated bioinformatics issues for the sake of education, are conducted. The investigation shows that the drafting of such tasks can be concluded with the support of the appropriate LLMs, including GPT versions. On top of this, researchers managed to uncover their weak spots with advanced bioinformatics applications. The goal is to facilitate further research on how AI can help in the fields of LLM and Science.

Powerhouses of Text and Code: LLMs

Large-scale neural network models, known as Large Language Models or GPT variants, are trained on enormous amounts of unlabeled data. They have recently drawn a lot of attention from several academic fields. Text-based input questions (prompts) may be used to train LLMs to generate natural-sounding contextual conversations through supervised fine-tuning, unsupervised pre-training, and human feedback. Beyond the capabilities of traditional natural language processing models, a natural language interface offers a versatile problem-solving platform for tasks ranging from text drafting to mathematical problem-solving. LLMs have demonstrated remarkable human-like reasoning and language-generating skills.

 These models, like GPT, excel in a variety of areas, including:

  • Zero-shot interaction:  Zero-shot interaction is defined as producing relevant text in response to stimuli without prior formal instruction in that activity.
  • Code comprehension and generation:  Understanding and even generating code might be considered code comprehension and generation, blurring the line between human and machine interaction.

These talents are quite promising for the discipline of bioinformatics, which requires data interpretation and processing.

Evaluating LLM Performance: A Rigorous Framework

The study used a comprehensive strategy to assess GPT-3.5 and GPT-4’s performance on a variety of bioinformatics tasks. The following is an overview of the primary strategies employed:

  1. Identifying Coding Regions (CDS) and Antimicrobials (AMPs):
  • Method: XML sequences were provided for both models; researchers aimed to highlight start and stop codons to estimate the range. For revealing AMPs, a list of known AMPs matched to a list of non-AMPs was compiled.
  • Evaluation: Topoint of the assessment standards like accuracy precision, recall, and F1-score were used for the assessment of CDS regions and AMPs identified. The model’s approach was analyzed regarding how good they were with asymmetrically distributed datasets.
  1. Optimizing Molecules for Drug Discovery:
  • Method: GPT-4, a current high-performance generation deep model, and existing compounds were given, and the participants were asked to create modifications to lift their characters, with special attention on SA and QED. The procedure of the model was made up of ways of creating SMILES strands that were meant for the modified molecules.
  • Evaluation: The accuracy of the generated SMILES was checked, and the impact of the suggested changes on SA and QED was assessed. Furthermore, it was studied if GPT-4 may explain the alterations.
  1. Gene and Protein Named Entity Recognition (NER):
  • Method: The models were trained using a corpus of biomedical information that had been carefully annotated for gene and protein mentions. They were then evaluated on their ability to employ strict and partial matching criteria to identify these objects from unseen text input.
  • Evaluation: The performance of GPT-3.5 and GPT-4 was compared to well-established approaches like BiLSTM and BioBERT versions, with typical NER assessment parameters including accuracy, recall, and F1 score.
  1. Educational Bioinformatics Problem Solving:
  • Method: A large selection of educational bioinformatics problems, including topics like phylogenetic analysis, protein structure prediction, and sequence analysis, were assembled. These challenges were presented to both LLMs, who were then asked to offer justifications and solutions.
  • Evaluation: The accuracy and comprehensiveness of the provided solutions were assessed using predetermined criteria. Additionally, performance was examined to see whether it was consistent across different topic subjects and difficulty levels.

Unlocking Potential: LLMs Take on Bioinformatics Challenges

This growing topic of research aims to better grasp the possibilities of LLMs in a variety of bioinformatics activities.

Here are some compelling examples:

  • Identifying coding regions: Understanding gene function requires locating DNA sequences that code for proteins, a task that LLMs can assist with.
  • Extracting information: Information retrieval from scientific literature is facilitated by LLMs’ ability to identify things in text, such as genes and proteins.
  • Peptide discovery: To expedite the creation of new medications, researchers are investigating the use of LLMs to find putative antibacterial or anticancer peptides.
  • Molecular optimization: LLMs can modify molecules to create new drugs or materials with specific properties.
  • Educational problem-solving: Through instructional venues, LLMs can offer solutions and support to students facing challenges in bioinformatics.

Challenges in Recognizing Entities and Solving Complex Problems

The performance of LLMs in named entity recognition (NER) for proteins and genes, which is essential for information extraction from biological literature, was examined in this work. Regretfully, GPT-3.5 and GPT-4 have shown their limits, failing to match known methods such as BioBERT variations and BiLSTMs. The fact that they were unable to correctly identify references to genes and proteins emphasizes the need for more research in this field.

Even though this study evaluates six critical bioinformatics activities, several bioinformatics sub-regions were not considered. In the future, researchers want to test and refine the appropriate GPT model implementations using more complex text contexts. This will take the form of designing and functionalizing huge biomolecules, predicting drug-receptor interactions, and generatively functionalizing massive biomolecule sequences.

However, there was one startling discovery: GPT-3.5 performed quite well on relatively simple problems, providing exact explanations and solutions. Although GPT-4 has shown potential in handling many problems, its ability to manage complex tasks is limited, suggesting that multiple-turn interactions may be necessary for full solutions.


The authors of the study show a promising prospect for the application of LLMs in bioinformatics, and this aspect is highly positive. These types of machines can detect altered base pairs (AMP recognition) and code areas and carry out molecular optimization processes, which makes them more adaptable. However, it is essential to recognize the present bottlenecks and take action to develop other areas, like complex problem-solving and NER (Neural Machine Translation). According to researchers, among these two AIs, Modof is the most logical for raising the logP, while GPT-4 appears to be more flexible because it can perform several tasks and give a remark for the reasons for the conclusions. To release the potential of LLMs as thought leaders in bioinformatics tools, we have to dig deeper and conduct more work and research.

It is important to take into consideration the prospect that bioinformatics may gain from significant developments in the domain of interesting and odd processes via LLMs. This benefits bioinformatics researchers. And we must try to figure out a way of having more AI researchers work in science and, in addition to that, give directions to scientists who use AI in their work.

Taking into account the limitations and striving very hard to get beyond them will result in the instruments that LLMs’ will become important tools in the bioinformatics toolbox, promoting the advancement of the discipline and developing scientific understanding. Bioinformatics researchers hold the promising prospect of leveraging advanced LLMs, thereby fostering the advancement of AI applications in scientific endeavors.

Article source: Reference Paper

Important Note: arXiv releases preprints that have not yet undergone peer review. As a result, it is important to note that these papers should not be considered conclusive evidence, nor should they be used to direct clinical practice or influence health-related behavior. It is also important to understand that the information presented in these papers is not yet considered established or confirmed.

Learn More:

Website | + posts

Anchal is a consulting scientific writing intern at CBIRT with a passion for bioinformatics and its miracles. She is pursuing an MTech in Bioinformatics from Delhi Technological University, Delhi. Through engaging prose, she invites readers to explore the captivating world of bioinformatics, showcasing its groundbreaking contributions to understanding the mysteries of life. Besides science, she enjoys reading and painting.


Please enter your comment!
Please enter your name here