Can AI Crack the Code of Life? Innovating Healthcare & Medicine through Large Language Models

MOLwise Biosciences Inc.

16th February 2025

Artificial Intelligence (AI), particularly Large Language Models (LLMs) have recently made significant strides, extending their capabilities beyond simple text generation, making them valuable tools for knowledge extraction and predictive modelling across multiple domains including Healthcare and Medicine. Globally used LLMs include the ChatGPT series, BERT (Bidirectional Encoder Representations from Transformer), PaLM, LaMDA, and Meta’s Llama series, DeepSeek, Baidu’s “Wenxin Yiyan” 360’s LLM, Alibaba’s “Tongyi Qianwen” and SenseTime’s LLM. 

A Quick Crash Course on Large Language Models (LLMs)

Large Language Models (LLMs) are deep learning-based AI systems designed to process and generate human-like text. These models enable analysis of vast amounts of text and recognize complex patterns in language. To understand how Large Language Models (LLMs) process language, imagine a detective solving a case with a team of assistants. 

Query (The Detective's Questions): The detective (LLM) needs to make sense of a crime scene (input text). To do this, detective asks relevant questions to understand the case - Where is the suspect mentioned in this report? Who does the word "he" refer to? Is there a connection between two events in the report?

Keys (The Filing System): The detective’s assistants (attention mechanism) have well-organized filing cabinets full of case files (words in the input text). Each document is labelled so the team can quickly locate the right information when needed.

Values (The Answers to the Questions): Once an assistant finds the right document based on the detective's questions (Query), they pull out the relevant details (Values) and hand them over. If the detective asks, “Who is ‘he’ in this statement?”, the assistant will check the Key (context of the sentence) and retrieve the Value (the person’s name). If the detective asks, “Where did the crime occur?”, the assistant will locate the Key (where location information is stored) and extract the Value (the actual location).

This process happens simultaneously for every word in a sentence, allowing the model to analyse relationships, track context, and understand meaning. Now, scale this up to billions of words and millions of detective-like interactions, and you get the power of an LLM which is able to predict, generate, and interpret human language with remarkable accuracy. Juang et al. (2021) have published the figure shown on the left, to visually represent how attention-mechanism works by highlighting how the model determines the meaning of the word "it" in the sentence: "The monkey  ate that banana because it was too hungry." In human language, resolving pronouns like "it" requires contextual understanding: does "it" refer to the monkey or the banana? Attention-mechanism enables the model to weigh the relevance of each word in relation to "it." In the image, stronger attention weights (darker lines) indicate that the model assigns higher importance to words like "monkey" and "banana", as they are the most likely antecedents. By dynamically adjusting attention scores across an entire text LLMs track relationships, resolve ambiguities, and improve natural language understanding.

How LLMs are revolutionizing Healthcare & Medicine

LLMs are reshaping healthcare & medicine through:

1) Healthcare Management at various levels by optimizing decision support, managing patient data, health awareness, etc. At the individual patient level, LLM-powered virtual assistants are improving patient engagement by explaining medical conditions, offering lifestyle recommendations, and even scheduling follow-ups, all while maintaining a conversational, patient-friendly approach. On a larger scale, LLMs are transforming hospital-wide operations by streamlining administrative workflows and operational logistics like documentation, billing, compliance reporting, processing insurance claims, staff scheduling, optimizing supply chain, resource allocation, managing patient records, thus reducing the workload on healthcare staff and improving efficiency. LLMs are being employed as powerful tools for aggregating & analysing patient data, hospital infection rates, epidemiological patterns, helping hospital systems identify trends and optimize care delivery. At the community-wide level, LLMs are addressing public health challenges, enabling large-scale data management and decision support. For e.g. managing vaccination drives, coordinating health initiatives and education efforts, predicting public health crisis, analyzing demographic health data, etc. As examples, Med-PaLM 2, HealthGPT, and CareGPT enhance patient interaction by summarizing medical knowledge and providing AI-powered consultation. BERT-based LLMs aid in clinical documentation and EHR management. Altogether, LLMs are paving the way for a more connected, data-driven, and responsive healthcare ecosystem.

2) Personalised Medicine and Clinical Research by accelerating drug discovery, precision diagnostics, and personalized medicine. LLMs are being used to model and analyse patient data, such as medical history, symptoms, and genomic information, to provide tailored treatment plans. Models trained with integrated-omics data (Genomics, Proteomics, Metabolomics, Lipidomics, etc.) can predict the most effective therapies for patients based on their genetic and metabolic makeup, a crucial step toward Personalised Medicine. LLMs are accelerating precision bio-diagnostics by early disease detection, medical image interpretation, rare disease identification, aiding development of portable diagnostic tools and biosensors for real-time analysis. LLMs are also accelerating drug discovery via De Novo Drug Design, molecular property prediction and forecasting of chemical interactions between molecules. ProGen, scGPT, ProtGPT-2, MegaMolBART, ChemBERTa  and AptaGPT assist in protein structure prediction and molecular generation expediting biopharmaceutical innovations. GeneFormer and scGPT analyse genetic and single-cell data, and can predict the roles of regulatory elements in DNA, which play crucial roles in gene expression. In addition, LLMs enable knowledge generation, synthesizing findings from thousands of research papers and clinical trials. BioGPT, PMC-LLaMA, and Bioformer specialize in biomedical literature mining, making research advancements accessible worldwide, and accelerating data-driven research.

The promise and the Pitfalls of LLMs

The promise of Large Language Models (LLMs) in medicine and biology is immense, as mentioned above. However, the integration of LLMs into medicine is not without its challenges. 

One of the major pitfalls lies in the accuracy and reliability of the generated information. LLMs can sometimes produce hallucinated outputs, which are plausible-sounding but incorrect. This can be dangerous in high-stakes environments like healthcare. Furthermore, these models are trained on datasets that may contain biases, leading to unequal performance across different patient demographics or even perpetuating harmful stereotypes. For example, if a model is trained on data that underrepresents certain populations, its recommendations might be less effective for those groups. Ethical concerns also abound, including issues around data privacy, as LLMs need to handle sensitive health information without compromising patient confidentiality. Lastly, the black-box nature of these models makes it difficult for clinicians to fully trust their recommendations, as the reasoning behind their outputs is often not transparent. 

While the potential of LLMs is undeniable, ensuring their safe, equitable, and ethical deployment will be key to maximizing their benefits.