Building Agentic AI Applications for Biodata Exploration

Ana Claudia Sima, Vincent Emonet,

Tarcisio Mendes, Panayiotis Smeros, Ruijie Wang

kru@sib.swis

Schedule – Day 1

• The theory : LLM basics, RAG, model evolution

​

• Coffee break @ 10:30

​

• Introduction to coding agents
• Lunch @ 12:00

​

• Knowledge Representation and Data Access for Agentic AI
• Coffee break @ 15:00

​

• Hands-on : Using existing MCPs for biodata exploration in agentic workflows
• Apero @ 17:30 at the Irchel Bar (building Y00)

Schedule – Day 2

• Build your own MCP server for a biodata resource

​

• Coffee break @ 10:30

​

• Hands-on: Agentic application for biodata exploration using custom MCP

​

• Lunch @ 12:00

​

• Hands-on : continuation
• Coffee break @ 15:00

​

• Bonus (optional): Adapting the practicals to your own use case / using local LLMs

Who we are – Knowledge Representation Unit @ SIB

• Tarcisio Mendes, Ana Claudia Sima – KRU co-leads (Lausanne / Zurich)

​

• Vincent Emonet – Data Scientist and Software Engineer

​

• Panayiotis Smeros and Ruijie Wang - AI Research Scientists with broad experience in ML / LLMs for information retrieval, classification, NLP …

​

​

Tarcisio

Ana

Vincent

Panayiotis

Ruijie

https://www.sib.swiss/community/our-groups/vital-it/knowledge-representation-unit

Learning objectives?

Disclaimer: This is a rapidly advancing field!��What you learn today reflects a snapshot of this field as of May 2026….�

�We hope to also learn from you!

Do you already use Agentic AI?

What to expect from today

• An overview of LLMs, basics on how they are trained and potential uses

​

• A deep dive into using LLMs for question answering with context
• From unstructured and structured sources (inputs)
• Towards unstructured and structured answers (outputs)
• …while using the right tools for the right purpose (Agentic)

​

• Best practices for building and using AI Agents

​

• Agentic application for interacting with biodata leveraging a custom MCP
• Build your own mini data exploration agent!

Ask Questions Anytime!

Part I. The Theory

What is a Large Language Model?

A (deep learning) model trained to predict the next word in a sentence

​

How is it trained?

• Self-supervised learning on HUGE amounts of text
• Learning to predict 1 word at a time

​

12

Source: https://amitness.com

• Originally based on GPT3.5

​

• A “chatty” version of LLM

​

13

Note: ChatGPT is not an LLM per-se, but a service with many “bells and whistles”

leveraging an LLM

(but also including web search,

integration with external tools, etc)

The “art” of prompting

• Prompting = a guiding question / instruction given to the model to shape the generated response

14

McDuff, D., Schaekermann, M., Tu, T. et al. Towards accurate differential diagnosis with large language models. Nature (2025). https://doi.org/10.1038/s41586-025-08869-4

The “art” of prompting

• Prompting = a guiding question / instruction given to the model to shape the generated response

​

• Brittle!

​

15

Vinay, Rasita, et al. "Emotional Prompting Amplifies Disinformation Generation in AI Large Language Models." Frontiers in Artificial Intelligence 8: 1543603.

What can you use an LLM for?

• Summarisation
• Programming (Co-Pilot)
• Question Answering
• Reports from images
• …

​

​

​

• Recognise special pieces of information in a text
• “Named Entity Recognition”
• Sentiment analysis
• Fraud detection
• …

16

Generative tasks

Classification tasks

Key LLM training steps

1) Pre-train

2) Fine-tune

3) Retrieve & Augment

4) Validate / Evaluate

5) Publish / Deploy

Pre-training: Tokenization

Tokenization

• Breaking down sentences into tokens

​

• The totality of tokens = the vocabulary of an LLM
• Determined by the specific choice of tokenizer

​

• Tokens do not necessarily always correspond to words
• Why?

Try it for yourself!

• Use the live tokenizer playground to see how it works!

​

• https://huggingface.co/spaces/Xenova/the-tokenizer-playground

​

• Use the sentence: "Hello hello and welcome to this course! 🙂 "

20

https://huggingface.co/spaces/Xenova/the-tokenizer-playground

What is a Large Language Model? (Generalization)

• A (deep learning) model trained to predict the next word in a sentence
• Token = words / amino-acids / genes / ….
• What is a Language?
• A vocabulary +
• Sequences of tokens that represent information in that language

22

token

sequence

Bioinformatics applications

• Language models for biological research: a primer , Nature Methods, ‘24

​

​

23

Bioinformatics applications

• Large language models generate functional protein sequences across diverse families, Designing proteins with language models
• Train on massive set of protein sequences
• Idea: protein embeddings reflect protein structure and function

​

• Design of highly functional genome editors by modeling the universe of CRISPR-Cas sequences
• “using large language models (LLMs) trained on biological diversity at scale, we demonstrate the first successful precision editing of the human genome with a programmable gene editor designed with AI.”
• …and anything that can be modelled as a sequence
• Patient trajectories: Large Language Models forecast Patient Health Trajectories enabling Digital Twins

​

24

Pre-training: Embedding

Pre-training: embedding

• LLMs, like any machine learning system, work with numbers

​

• Embeddings = a way to transform words into numbers, while preserving their semantics

26

Source: https://dkharazi.github.io/notes/ml/nlp/embedding

Pre-training: Attention

Pre-training: Attention!

28

Source: A Multiscale Visualization of Attention in the Transformer Model, ACL 2019 Systems Demonstrations

​

“Self-Supervised” / “Unsupervised” – but training data matters

• The illusion of self-supervision – actually a lot of effort for curating good data
• Relies on work of thousands of human “data annotators”
• Ethical issues: breaches of copyrighted material
• Removing bias from training dataset
• etc

29

Result: Foundation Model

30

Source: https://viso.ai/deep-learning/foundation-models/

Limitations?

Drawbacks?

Non-exhaustive list of pitfalls and limitations…

• Hallucinations
• When it’s wrong, it’s confidently wrong!
• Cutoff date
• An LLM has no knowledge past its training date
• Example: GPT 4.1 (April ’25) has knowledge up to June ‘24
• COST!
• 100 Million $$$ to develop!
• Environmental considerations…
• PRIVACY!
• Please don’t put sensitive information in ChatGPT…
• Lack of interpretability
• Lack of provenance / attribution
• Bias
• ….

​

32

ANY SOLUTIONS?

Open-source LLMs: Llama, Mistral…

• Usually come in various size, the smaller the cheaper to run, but less “smart”

👶 ~7B: inference can run on your laptop in acceptable time if you have a GPU

🧒 ~70B: requires a larger server for decent inference time (1 A100/H100)

🧠 ~400B: largest, requires custom supercomputers (~4 H100s or more)

​

• Mixture-Of-Experts approaches such as Mixtral 8x7B or 8x22B combine multiple smaller models for better results with same hardware requirements

​

​

• HuggingFace: the most comprehensive resource to find open-source models
• e.g. https://huggingface.co/meta-llama/Meta-Llama-3-8B �

33

Augmenting LLMs…

• LLMs are much more powerful when augmented with tools, memory, datasets…
• Tools = services (e.g. accessible via APIs)
• How can LLMs discover these?
• MCP (Model Context Protocol)
• E.g. a PubMed MCP can enable searching PubMed literature (API) via LLM agents
• …but more on this later!

34

LLM (Agent)

Tool (API)

MCP Server

Fine-tuning

• Initially based on GPT3.5

​

• A “chatty” version of LLM
• Based on Instruction DataSet (small, e.g. 50K etc) – Alpaca Style Prompt Template

​

• “Secret sauce”: Reinforcement Learning through Human Feedback (RLHF)

36

We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides—the user and an AI assistant. We gave the trainers access to model-written suggestions to help them compose their responses. We mixed this new dialogue dataset with the InstructGPT dataset, which we transformed into a dialogue format.

​

Source: https://openai.com

Reinforcement Learning through Human Feedback

37

RLHF – an ongoing journey…

38

Crowdsourcing expertise?

39

40

Retrieval Augmented Generation (RAG)

• Complement the base model with an external knowledge base

​

• This can be:
• PDFs
• Structured Data
• CSVs
• Relational DBs
• …
• Images
• ….

Source: https://blog.gopenai.com

Why RAG?

Why RAG?

• Always up-to-date information
• Domain-specific / use case-specific data
• Relevant data for a given query
• ….

43

Second option: Fine-tuning a foundation model

• Requires a dataset of examples in the style of “instructions”, input and output

​

​

​

​

​

​

​

​

​

​

​

​

​

• Note: usually expensive to do

44

Part II. From Theory (Closer To) Practice

How can you use LLMs to explore biodata?

• Directly ask questions (“zero-shot”)

​

• Ask the LLM to formulate code (e.g. API calls) to answer questions

​

• Provide relevant context & tools, then ask LLM to generate the answer
• Structured and unstructured data (CSV files, PDFs)
• scripts
• API calls
• etc

Ask LLM questions over biodata directly

“Cheating!”

Disable and

try again!

Asking chatGPT question directly (“zero-shot”)

“Back in the day”:

when it was wrong…it was confidently wrong

Nope!!

Warning: in general, asking LLMs specialized questions directly will likely lead to

hallucinated answers!

Recommended: Ask Model To Ask an API

Recommended: Ask Model To Ask an API

Problem: not all APIs are as well known as Uniprot’s

• Solution: MCP servers as an interfaces (“adapters”) to APIs!
• E.g. a PubMed MCP can enable searching PubMed literature (API) via LLM agents

Up next: Agentic Workflows

The future: The AI Scientist?

57

The future: Agentic Workflows?

58

Source: https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

The future: Agentic Research?

59

The future?

60

Source: The Semantic Web, Scientific American, May 2001

Question Answering over Structured Data

Expasy

Some Intermediate Conclusions…

• The LLM space is constantly evolving…
• Tasks that seemed hard 6 months ago are now solved
• Powerful interfaces to both structured and unstructured data…
• …but need to be adapted to your use case

​

• Careful when using (especially) closed-source LLMs
• Don’t use them for sensitive data!

​

• Infrastructure is a problem
• Self-hosted (requires GPUs) or hosting platforms (paid)
• Small models are (clearly) less performant
• But big models are expensive to host

64

References

• EMBL-EBI webinars on LLMs: https://www.ebi.ac.uk/training/events/large-language-models-and-their-applications-bioinformatics/
• Developing an LLM: Building, Training, Fine-tuning (relatively technical video): https://www.youtube.com/watch?v=kPGTx4wcm_w - Sebastian Raschka , author of “Build a Large Language Model (from scratch)”
• Detailed notebook for running LLMs for various text analyses (includes analysis of next word probability, embeddings generated etc): https://colab.research.google.com/github/michael-franke/npNLG/blob/main/neural_pragmatic_nlg/07-LLMs/07b-pretrained-LLMs.ipynb Interactive overview of existing Language Models (from BERT to GPT3): https://informationisbeautiful.net/visualizations/the-rise-of-generative-ai-large-language-models-llms-like-chatgpt/?trk=article-ssr-frontend-pulse_little-text-block
• Will we run out of data? Limits of LLM scaling based on human-generated data: https://arxiv.org/pdf/2211.04325, interesting preprint analysing the limits of scaling beyond current model sizes due to the lack of sufficient additional training data in the near future
• Other LLMs to interact with: Gemini, claude, …
• Language models for biological research: a primer , Nature Methods, ’24
• Extended LLM course with multiple hands-on notebook examples: https://github.com/mlabonne/llm-course
• https://github.com/ncbi-nlp/LLM-Medicine-Primer
• McDuff, D., Schaekermann, M., Tu, T. et al. Towards accurate differential diagnosis with large language models. Nature (2025). https://doi.org/10.1038/s41586-025-08869-4
• Vinay, Rasita, et al. "Emotional Prompting Amplifies Disinformation Generation in AI Large Language Models." Frontiers in Artificial Intelligence 8: 1543603.

​

65

Thank you!

DATA SCIENTISTS FOR LIFE

Ana-claudia.sima@sib.swiss

Part III. Hands-on: Setup

Install: uv, get (Mistral) API key