Andrew White |   EdisonScientific

Automating Scientific Discovery

Andrew White & Michaela Hinks






EdisonScientific
MIT NLP Seminar
December 2025

EdisonScientific

  • Spinout from FutureHouse formed in 2025
  • VC-backed AI Research Institute
  • Based in San Francisco
  • 35 employees

Science is changing independent of AI


Arxiv.org,10.6084/m9.figshare.17064419.v3

Number of Researchers are Growing

International R&D spending

PhD Researchers


NSF - https://ncses.nsf.gov/pubs/nsf24332; UNESCO UISI SDG9

Intellectual bottlenecks are growing


đź“ť Increasing paper count ($\approx$10M per year)

🧬 Larger data sets from cheaper experiments (genome at $200 per person, $1 / GB of sequencing)

🔍95% decline in disruptive papers since 1980

Park, M. et al. Nature 613, 138-144 (2023); Scannell, J.W. et al. Nat. Rev. Drug Discov. 11, 191–200 (2012); Deloitte 2025: Pharma innovation returns.

Edison Scientific Mission


Accelerate Scientific Discovery

Can large language models already do science?


LAB-Bench: Measuring Capabilities of Language Models for Biology Research

Jon M. Laurent, Joseph D. Janizek, Michael Ruzo, Michaela M. Hinks, Michael J. Hammerling, Siddharth Narayanan, Manvitha Ponnapati, Andrew D. White, Samuel G. Rodriques arXiv:2407.10362, 2024

Existing benchmarks

MMLU-Pro

  1. (Medicine) As of 2017, how many of the world's 1-year-old children today have been vaccinated against some disease?
  2. (History) On which continent are most of the Venus figurines found?
  3. (Chemistry) Find the logarithm of 3^2

Lab-Bench Questions

Not textbook knowledge

Lab-Bench Questions

Multimodal

Human baselines exceed LLMs

What about bioinformatics?

Bixbench Question:
How many peripheral immune cell types show significant differential expression (adjusted p-value < 0.05) of SOCS3?

Human baselines exceed LLMs

What is an agent?

Agent: trained, makes decisions

Environment: untrained, has tools, state

Protein Design Environment

  • Protein design with 5 existing deep learning models
  • Molecular dynamics, bioinformatics, literature research agent
  • Input: "design 92 binders for PD-L1"

Wet lab validation

Learning vs Frontier Models

Crows

Name Environment Key Tools
Literature Literature Research Search, Citation Traversal
Proteins Designing novel proteins AlphaFold2, Molecular Dynamics
Molecules Designing new molecules Retrosynthesis, self-driving robotic lab
Analysis Generating discoveries from data bioinformatics tools, code, file system

Agent vs ML Model

Modify surface residues of IL-10 to increase expression and solubility in E. coli without disrupting dimerization or receptor interaction.

Automating research of scientific literature


Language agents achieve superhuman synthesis of scientific knowledge

Michael D. Skarlinski, Sam Cox, Jon M. Laurent, James D. Braza, Michaela Hinks, Michael J. Hammerling, Manvitha Ponnapati, Samuel G. Rodriques, Andrew D. White arXiv:2409.13740, 2024

Evaluating

Overexpression studies of PRMT4 in SW480 UPF1 knockout cells show that which arginine residue in PRMT4 is important for asymmetric di-methylation of UPF1 R433?

Better at answering questions than PhD biology experts

Improving over time

Better than human written Wikipedia articles

Can be used to check for precedent and disagreement in literature

Demo

Edison Platform

  • Some amount of free usage, more for edu
  • Grants for academic projects
  • API - can be incorporated into your pipeline/agents
  • Majority of agent code is open source

API

Literature Research Agent Scale
  • Tasks per minute: 300
  • Research Papers 175,000,000
  • Wiki page for all diseases every 7 hours
  • All arxiv papers per week 30,000 papers / month
  • Check for contradictions 6.3M papers / year
  • All Wikipedia every 3 weeks

Model intelligence will continue to increase

Complete cycle of disease to mechanism to target to drug

ROBIN: A Multi-Agent System for Automating Scientific Discovery

Ali Essam Ghareeb*, Benjamin Chang*, Ludovico Mitchener, Angela Yiu, Caralyn J. Szostkiewicz, Jon M. Laurent, Muhammed T. Razzak, Andrew D. White†, Michaela M. Hinks‡, Samuel G. Rodriques

Kosmos: An AI Scientist for Autonomous Discovery


Ludovico Mitchener, Angela Yiu, Benjamin Chang, Mathieu Bourdenx, Tyler Nadolski, Arvis Sulovari, Eric C Landsness, Daniel L Barabasi, Siddharth Narayanan, Nicky Evans, Shriya Reddy, Martha Foiani, Aizad Kamal, Leah P Shriver, Fang Cao, Asmamaw T Wassie, Jon M Laurent, Edwin Melville-Green, Mayk Caldas, Albert Bou, Kaleigh F Roberts, Sladjana Zagorac, Timothy C Orr, Miranda E Orr, Kevin J Zwezdaryk, Ali E Ghareeb, Laurie McCoy, Bruna Gomes, Euan A Ashley, Karen E Duff, Tonio Buonassisi, Tom Rainforth, Randall J Bateman, Michael Skarlinski, Samuel G Rodriques, Michaela M Hinks, Andrew D White
arXiv:2511.02824, 2025

Kosmos overview

How do you validate systems like this?

Work with external groups. Input is their experimental data. Three discoveries reproduced in unpublished work. Four novel discoveries.


independent expert annotation of task difficulty and correctness

Example Discovery: What kosmos found

Human validation

Kosmos Scale

  • 120 sandboxed envs with 32GB RAM/8 CPUs
  • 3,000 papers parsed and considered
  • 24-48 hours of run time
  • Generates up to 4TB of data

Try Kosmos Yourself


  • 6 free runs with .edu email address
  • Standard cost: $200 per run
  • Available at platform.edisonscientific.com
questions