Welcome!

Welcome to the Basel BioData Community meetup on LLMs in Research!

My name is Flavio Lombardo, I am a computational biologist at the University Hospital in Basel. I am happy to share some ideas on how LLMs are changing research (and other fields) and how they could help you.

Link to the meeting Click here for the Google Meet

• All opinions are my own and do not reflect my employer
• I do not have any conflict of interest with any of the topics here presented

Setup to Follow Along (1/2)

Required Packages

Install these R packages to follow along:

# Install required packages
install.packages("ellmer")
install.packages("pdftools")
install.packages("jsonlite")
install.packages("dplyr")
install.packages("purrr")

Get Your DeepSeek API Key (Free!)

Step 1: Visit platform.deepseek.com

Step 2: Sign up and get your API key (free $5 credit = ~35M tokens!)

Step 3: Add to your .Renviron file:

DEEPSEEK_API_KEY="your-key-here"

Don’t have an API key yet? No problem! Follow along and try later.

Setup to Follow Along (2/2)

Optional: Local Models with Ollama

For privacy-sensitive work or offline usage:

# Install Ollama helper package
install.packages("ollamar")

Then install Ollama app: - Visit: ollama.com - Download for your operating system - Pull a model (we’ll use qwen3:1.7b)

Need Help?

• Ask questions anytime!
• All code examples will be shared
• We’ll go step-by-step through examples
• No pressure to set up everything now

What are LLMs?

Large Language Models are AI systems trained on vast amounts of text data.

• Generate human-like text
• Understand and respond to questions
• Assist with coding, writing, and analysis
• Examples: GPT-4, Claude, Llama, Mistral, Deepseek

LLMs in Research Workflows (1/2)

LLMs in Research Workflows (2/2)

Why Cloud LLMs for Research?

• No setup - Start coding immediately
• State-of-the-art models - GPT-4, Claude, DeepSeek
• Structured extraction - Get typed data (integers, dates, strings)
• Cost-effective - DeepSeek: $0.14 per 1M tokens
• Easy scaling - Process hundreds of papers
• Regular updates - Models improve automatically

Cloud LLMs in R: ellmer Package

# Install ellmer (on CRAN)
install.packages("ellmer")

# Or development version
pak::pak("tidyverse/ellmer")

Setup: ellmer with DeepSeek API

Step 1: Get a DeepSeek API key from platform.deepseek.com

Step 2: Set up your API key

# In your .Renviron file (recommended) - use this for DeepSeek
DEEPSEEK_API_KEY="sk-your-deepseek-key-here"

# Or in R session (temporary)
Sys.setenv(DEEPSEEK_API_KEY = "sk-your-deepseek-key-here")

Step 3: Create a chat session with DeepSeek

library(ellmer)

# Create chat with DeepSeek
chat <- chat_deepseek(
  system_prompt = "You are a research assistant."
)

# Use it!
chat$chat("Explain CRISPR in simple terms")

Why DeepSeek? - Much cheaper: $0.14 per 1M tokens vs OpenAI’s $2.50 - Strong coding & research capabilities - Your $5 credit = ~35M tokens!

Structured Data Extraction with DeepSeek

DeepSeek requires forcing JSON output via api_args:

library(ellmer)
library(jsonlite)

# Create chat with JSON output forced
chat <- chat_deepseek(
  system_prompt = "You are a research assistant.
  Always respond with valid JSON only.",
  api_args = list(response_format = list(type = "json_object"))
)

# Ask for structured data
abstract <- "In our 2023 study, we analyzed 150 patients.
The treatment showed significant improvement (p = 0.001)..."

prompt <- 'Extract this data as JSON:
{
  "year": <integer>,
  "sample_size": <integer>,
  "p_value": <number>,
  "conclusion": "<string>"
}'

response <- chat$chat(paste(abstract, prompt))

# Parse JSON response
data <- fromJSON(response)
data$year          # 2023
data$sample_size   # 150
data$p_value       # 0.001

Example: Asking Specific Questions to a PDF

Go beyond structured extraction and ask direct questions.

library(pdftools)
library(ellmer)

# Extract text from a specific paper
pdf_path <- "Multimodal learning enables chat-based exploration of single-cell data.pdf"
text <- pdf_text(pdf_path) |> paste(collapse = "\n")

# Create a chat session
chat <- chat_deepseek(
  system_prompt = "You are a helpful research assistant.
  Answer questions based *only* on the provided text.",
  api_args = list(),
  seed = 1234
)

# Ask a specific question about the paper's content
question <- "What was the primary conclusion regarding this paper?"

# Combine context and question
prompt <- paste(
  "Here is the text from a research paper:",
  text,
  "---",
  "Please answer the following question based on the text:",
  question,
  sep = "\n"
)

# Get the answer
response <- chat$chat(prompt)

Local Models with Ollama (1/2)

For sensitive data or offline work, run models locally with Ollama:

Setup

# Install Ollama from ollama.com
# Then pull a model (in terminal):
ollama pull qwen3:1.7b

# Or from R:
install.packages("ollamar")
ollamar::pull("qwen3:1.7b")

Basic Chat Example

library(ellmer)

# Create a chat with local model
chat <- chat_ollama(
  model = "qwen3:1.7b",
  system_prompt = "You are a research assistant."
)

# Use it like any other chat
chat$chat("Explain p-values in simple terms")

Local Models with Ollama (2/2)

Structured Data Extraction

library(ellmer)
library(jsonlite)

# Define JSON schema for extraction
schema <- list(
  type = "object",
  properties = list(
    year = list(type = "integer"),
    sample_size = list(type = "integer")
  ),
  required = c("year", "sample_size")
)

# Create chat with JSON mode and clear instructions
chat <- chat_ollama(
  model = "qwen3:1.7b",
  system_prompt = "You must respond with valid JSON only.
  No explanations, no markdown, just JSON.",
  api_args = list(format = schema)
)

# Extract data with explicit JSON request
text <- "In 2023, we studied 150 patients..."
prompt <- paste0(
  "Extract as JSON: ", text,
  "\nReturn format: {\"year\": <int>, \"sample_size\": <int>}"
)

response <- chat$chat(prompt)
data <- fromJSON(response)

# Access results
data$year          # 2023
data$sample_size   # 150

Recommended Local Models

• qwen3:1.7b - Fast, good quality (recommended)
• llama3.2:3b - Balanced performance
• phi4:latest - Microsoft’s efficient model

Best Practices

Resources & Getting Started (1/2)

R Packages

• ellmer: ellmer.tidyverse.org - Main LLM interface
• pdftools: Extract text from PDF files
• jsonlite: Parse JSON responses
• purrr: Batch processing multiple files
• dplyr: Data manipulation and analysis

API Setup for Cloud Models

DeepSeek (Recommended - Cheapest): - Visit: platform.deepseek.com - Free $5 credit (~35M tokens!) - $0.14 per 1M tokens

Alternative Providers: - OpenAI: platform.openai.com - GPT-5, GPT-4o

• Anthropic: console.anthropic.com - Claude 4.5

Resources & Getting Started (2/2)

Learning Resources

Documentation: - ellmer Documentation - Main package docs - Structured Data Guide - Type-safe extraction - DeepSeek API Docs - API reference - Prompt Engineering Guide - Writing better prompts

Community & Support

Get Help: - Posit Community - Ask questions - R for Data Science - Learn R fundamentals - ellmer GitHub Issues - Report bugs

Stay Updated: - Follow Posit blog for ellmer updates - Join R communities on social media

Questions & Discussion

Let’s discuss:

• What research tasks could benefit from LLMs?
• Privacy concerns in your field?
• Experiences with AI tools?
• Ideas for future sessions?

Stay Connected

Thank You!

See you at the next meetup!

Questions? Let’s chat!