New LLMs, Agents, and Graphs in Life Sciences

Generative AI in Healthcare

Generative AI models optimize objectives over large datasets (e.g., via gradient‑based training) and can synthesize new text, images, audio, or structures conditioned on prompts. Over the past decade, gen-AI has advanced through several architectural milestones:

• VAEs (2013): encoder–decoder generative models; early uses included anomaly detection in medical imaging and basic text generation.
• GANs (2014): generator vs discriminator to synthesize realistic samples; popularized image creation.
• Diffusion models (mid-2010s): learn to denoise from noise; basis of modern image generators, often paired with language/control modules (e.g., DALL-E 2/3).
• Transformers (2017): attention-based sequence models enabling long-range context and rich embeddings; foundation of today’s LLMs used in lab and clinic-adjacent tasks.

The Transformer model architecture. Courtesy of Yuening Jia. License CC BY-SA 3.0.

Transformers enable faster training, more coherent generation, and longer, context-aware outputs, excelling in natural language processing, they underpin current LLM models/products such as ChatGPT-5, Copilot, Gemini, DeepSeek.

LLMs are one branch of foundation models. Like LLMs, many FMs are built on the transformer architecture with self-attention, but they are typically designed to handle a wider range of modalities and tasks. Protein/molecule models (e.g., AlphaFold3, ESM3, RFdiffusion2) are often integrated with LLM‑driven tooling for planning, critique, and documentation

Healthcare is data-heavy and workflow-intensive. Some of that maps to generative models’ strengths in summarization, drafting, and retrieval, which somewhat reflects in figures:

• In US, Menlo Ventures reports healthcare deploying AI at twice the rate of the broader US economy in 2025.
• According to McKinsey’s Q4 2024 survey, 85% of healthcare organizations have already implemented or are developing generative AI use cases, with 64% reporting positive ROI.
• Same survey shows 61% of those healthcare organizations cited partnerships with third-party vendors as their primary AI adoption strategy (e.g., Anthropic’s partnership with major healthcare players such as Novo Nordisk, Genmab, and Sanofi).

LLMs Tailored for Life Sciences

LLMs can be broadly categorized as unimodal or multimodal. Unimodal models process a single type of data—typically text. In contrast, multimodal large language models (MLLMs) can understand and reason across multiple data types such as text, images, and audio. They can describe visuals, answer questions about videos, interpret charts, perform OCR, and engage in conversations that combine speech and vision.

General Models Adapted to Healthcare

🔹 Claude for Life Sciences (Anthropic) is a suite of updates by Anthropic aimed at enhancing Claude models for better biomedical research. The update adds connectors to scientific platforms like Benchling, BioRender, PubMed, Synapse, Wiley’s Scholar Gateway, and 10x Genomics to let researchers pull data, figures, and papers directly into Claude’s workspace. The new Agent Skills feature allows the model to follow detailed lab workflows—like quality control for single-cell RNA-seq—while a prompt library helps scientists get consistent results on tasks such as protocol generation, literature review, data analysis, and regulatory drafting. Claude Sonnet 4.5 scored 0.83 on the Protocol QA test (above the human baseline) compared to 0.74 by Sonnet 4; the similar improvement was shown in in BixBench (bioinformatics task benchmark).

🔹 GPT-5 (OpenAI) is OpenAI’s flagship multimodal system that answers with a fast default model or switches to a deeper “thinking” model for hard problems. It handles text, images, and tools (e.g., code execution, browsing) and includes agentic abilities to autonomously search and set up task workflows.

A new study by researchers from University of Minnesota tested GPT-5 and GPT-4o (previous best model) on five BioNLP tasks and nine biomedical QA datasets. GPT-5 scored reportedly about 0.76 average QA accuracy vs. 0.66 for GPT-4o, with jumps on reasoning sets like MedXpertQA (+0.29) and DiagnosisArena (+0.22).

🔹 LLaMA2 (Meta AI) is a family of foundation models (7B/13B/70B) available as base and chat-tuned variants for natural-language tasks like drafting, summarization, Q&A, and basic reasoning, plus Code Llama for code generation and editing. LLama’s base models are transformer, autoregressive LLMs pre-trained on ~2T tokens; chat models add supervised fine-tuning (SFT) + reinforcement learning from human feedback (RLHF) for instruction following and safer behavior.

A Nature Communications study compared LLaMA 2 models with GPT-3.5, GPT-4, and PMC-LLaMA (domain-tuned version) across 12 biomedical NLP tasks, including entity recognition, relation extraction, classification, QA, summarization, and simplification. The LLaMA 2 13B model scored about 20% below GPTs in zero- and few-shot settings but improved significantly after fine-tuning (from 0.28 to 0.51). Interestingly, PMC-LLaMA, despite additional biomedical training, didn’t show superior performance over LlaMA 2.

🔹 Med-PaLM (Google) is a line of LLMs tuned for clinical reasoning, built on Google PaLM. Functionally, Med-PaLM integrates medical comprehension, reasoning, and knowledge retrieval to handle professional exam questions, research interpretation, and consumer health inquiries. Its answers are reviewed against criteria such as scientific accuracy, consensus, reasoning quality, bias, and harm likelihood, and in studies were often preferred over physician-written answers.

in –style questions, while its successor achieved 86.5% accuracy on the benchmark and reached physician-level performance. Google is also developing , a multimodal version that combines text with medical images, health records, and genomic data. Despite reported high performance in medical question benchmarks, developers state the necessity of further research to assess the use of Med-PaLM2 in healthcare operations.