After a few years of procrastination, I've updated my textbook. Changes: 1. Tensorflow -> PyTorch 2. Darkmode 3. Added scaffold split section 4. Fixed many typos

I've been writing research articles for over 10 years now and one of the hardest parts is writing consistently and efficiently without procrastinating. I'm going to share some of my tips here 🧵 1/10

[placeholder for when I've finished]

In academia, we don't say "I love you." We say "minor revisions." And I think that's really beautiful.

After two years of work, we’ve made an AI Scientist that runs for days and makes genuine discoveries. Working with external collaborators, we report seven externally validated discoveries across multiple fields. It is available right now for anyone to use. 1/5

I packed-up a full-text paper scraper, vector database, and LLM into a CLI to answer questions from only highly-cited peer-reviewed papers. Feels unreal to be able instantly get answers by an LLM "reading" dozens of papers. 1/2

Sometimes people ask me advice about academia. I have no idea what I'm doing 80% of the time. All my PI proposals in last 2 years have been rejected. In my entire academic career, I've never been able to get a paper reviewed at JACS, Nat Chem, PNAS, etc. 1/2

About 75% of doing a PhD is getting your python environment set-up

Apple's AI researchers have embraced a kind of anti-LLM cynic ethos, publishing multiple papers trying to argue that reasoning LLMs are somehow limited/cannot generalize. Apple also has the worst AI products (Siri, Apple Intlligence). No idea what their "strategy" is here

Packaged our algorithm to design binding peptides into a simple colab. Enter a protein sequence, indicate where you want the peptide to bind, and in a few minutes you'll have a peptide predicted to tightly bind via AlphaFold + Bayesian Optimization. 1/2 colab.research.google.com/gi…

I just paid $60 to embed the text of the entire lord of the rings trilogy so I could have GPT answer a question I've wondered my whole evening: Do the people of Middle Earth poop? 1/6

We report a model that can go from natural language instructions, to robot actions, to synthesized molecule with an LLM. We synthesized catalysts, a novel dye, and insect repellent from 1-2 sentence instructions. This has been a seemingly unreachable goal for years! 1/3

This is the worst day ever to launch our reasoning model. We're just gonna announce it fresh next week. Everyone please act surprised.

The average PhD student spends 8 months of their academic career learning about CUDA drivers, conda, pip, setup_tools, and brew. Feels solved now with GPT-4

I try to prevent the rejections from interfering with my enthusiasm and curiosity for science. But holy shit, it's hard. And we all keep these failures private, which fucks with your perspective on academia and your brain chemistry. Anyway, happy Thursday #AcademicTwitter 2/2

Just got the official response from unicode about the protein emoji. It was REJECTED 🚫 This is obviously very difficult news for me and the protein emoji team. We're evaluating all our options right now.

We've completed our first draft of a protein emoji that we'll be submitting at end of month. Please let me know if you have feedback and/or show your support for a protein emoji! Art by Michael Osadciw

There are many skills you have to develop as a PI that are not taught in graduate school. The skill that I use most often as a PI is downloading a PDF, inserting my signature, and uploading it. Literally hours of my life are uploading and downloading PDFs.

I believe this rumor, since we've published two papers showing agents doing tasks better than PhD students in literature research, writing Wikipedia articles, and manipulating DNA. Wish we had better PR though😅

Can GPT-4 do drug discovery? No, but it can help. Let's walk through GPT-4 proposing new drugs. This is called knowledge-based screening. We're trying to fill a list of plausible compounds that could lead to new drugs based on research papers. 1/n

Finishing 2024 with one more research result! We’ve trained small language agents to do hard sci tasks: engineering proteins, manipulating DNA, and working with sci literature in a new library called Aviary. We beat humans and frontier LLMs on these tasks!

We’ve just finished writing the missing 15,616 Wikipedia articles to get complete coverage of all 19,255 human genes. We used PaperQA2, which has higher accuracy than existing human-written Wikipedia articles, as judged by blinded biology PhD students and postdocs. 1/5

I finished my estimate on required compute to make an atomic-resolution virtual cell: 10^38 FLOPs to simulate a human cell for 1 day. We should be able to do this simulation in 2074 using 200 TW of power. 1/3

My mom texted to tell me about ChatGPT. Huge milestone in AI hype cycle

How did I get GPT-4 to design new drugs? I gave GPT-4 access to small programs and it proposed a new molecule that could be used for leukemia. I did this with @langchain's agent lib. See chain of thought at bottom of GPT-4 paper cdn.openai.com/papers/gpt-4.…

Working hard on the most important project of my life

Here's an embedding of all drugs that entered clinical trials for last 60 years. You can see that molecules have been getting bigger and becoming more similar.

HLE has recently become the benchmark to beat for frontier agents. We @FutureHouseSF took a closer look at the chem and bio questions and found about 30% of them are likely invalid based on our analysis and third-party PhD evaluations. 1/7

Forget DALLE, PALM, M1 Pytorch. This is the biggest advancement in computer science this year: LaTeX math in github github.blog/2022-05-19-math-…

Molecular dynamics requires a lot of expert knowledge to set-up and analyze simulations. We set out to automate it with LLM agents: MDCrow!

Here's an open source 1b (1 byte) language model that gets 25% on MMLU. Very powerful relative to model size. I'm raising $2B on a $20B valuation to scale up to 500b model size.

Remember, don't let anyone find out I write proposals by playing video games all night with you and then panic at 1am followed by enormous amounts of coffee

Our preprint on using GPT-4 as an agent with tools for chemistry is out! We call it ChemCrow. Working with @SamCox822, @drecmb @pschwllr, we developed a set of tools for synthesis/cond, safety, commercial availability, patents, paper-qa arxiv.org/abs/2304.05376 1/5

Normal Friday night @FutureHouseSF. It wants to go in the lab so bad to get to work.

I've been exploring if GPT-4 and other models (please give me a key @AnthropicAI!!) can do "algebra" of molecules. Let's see a few examples 1/4 demo: whitead.github.io/svelte-che…

The last few months, I’ve been helping develop the mission for @FutureHouseSF - a non-profit AI Biology institute. Building on our work with autonomous agent driven chemistry and robotics, we’re building an entire org around the mission to build autonomous scientific agents 1/5

I'm excited to announce alphafold-inverse🎉! This AI model takes in the output protein structure from alphafold and gives the input amino acid sequence.

This paper is amazing. Their architecture comparison with AlphaFold2 really tells the whole story. Just flow matching with a transformer...

Here's an old render I did in Blender. Just showing relative sizes of some interesting biomacromolecules. Never enough free time now to do stuff like this. Miss it.

R01s are 13 pages of science. If you're a competent faculty member you should type at 100 words per minute - so you should be authoring an R01 every 7 minutes. Add in 10-15% margin for meeting/admin and you should be submitting about 50 R01's everyday.

ChemCrow is out today in @NatMachIntell! ChemCrow is an agent that uses chem tools and a cloud-based robotic lab for open-ended chem tasks. It’s been a journey to get to publication and I’d like to share some history about it. It started back in 2022. 1/8

The tweet: 🚨New AI model predicts if you have HIV infection from video of you breathing with 99% accuracy🚨 The model:

Finished with molcloud - now you can generate your own cloud of molecules 1. pip install molcloud 2. molcloud [smiles_file] github.com/whitead/molcloud

We used PaperQA2 to extract claims from papers and then see if they're contradicted anywhere in literature. This task is time consuming for humans, but we were able to use this for hundreds of papers to look for trends in disagreement in fields, decades, and journals.

Most of the code (except the interface) is already open source - github.com/whitead/paper-qa

Some more fun with machine generated molecules. Here is a molecule with perfectly satisfied valencies, but I don't think will be synthesized.

I have written up a 3.5k word/10 figure essay on how to write a reward function while avoiding reward hacking for chemistry. It covers all the ridiculous ways we had to avoid reward hacking for training ether0, our scientific reasoning model.

Read a paper that casually mentioned they trained on 330,000 synthesized molecules at a pharma company. They won't share this data, so it will be replicated over and over. We could reduce animal, economic, and environmental burden so much by sharing data from completed projects.

I've put together a few of my favorite discussions on the details of doing molecular dynamics. I'll add more as they come. Hopefully they're useful to you! 🧵1/n

The plan at FutureHouse has been to build scientific agents and use them to make novel discoveries. We’ve spent the last year researching the best way to make agents. We’ve made a ton of progress and now we’ve engineered them to be used at scale, by anyone. Today, we’re launching the FutureHouse Platform: an API and website to use our AI agents for scientific discovery. It’s been a bit of a journey! June 2024: we released a benchmark of what we believe is required of scientific agents to make an impact in biology, Lab-Bench. September 2024: we built one agent, PaperQA2, that could beat biology experts on literature research tasks by a few points. October 2024: we proved-out scaling by writing 17,000 missing Wikipedia articles for coding genes in humans. December 2024: we released a framework and training method to train agents across multiple tasks - beating biology experts in molecular cloning and literature research by >20 points of accuracy. May 2025: we’re releasing the FutureHouse Platform for anyone to deploy, visualize, and call on multiple agents. I’m so excited for this, because it’s the moment that we can see agents impacting people broadly. I’m so impressed with the team at FutureHouse for us to execute our plan in less than 1 year. From benchmark to wide deployment of agents that can exceed human performance on those benchmarks! So what exactly is the FutureHouse Platform? We’re starting with four agents: precedent search in literature (Owl), literature review (Falcon), chemical design (Phoenix), and concise literature search (Crow). The ethos of FutureHouse is to create tools for experts. Each agent’s individual actions, observations, and reasoning is displayed on the platform. Each scientific source is considered from retraction status, citation count, record of publisher, and citation graph. A complete description of the tools and how the LLM sees them is visible. I think you’ll find it very refreshing to have complete visibility into what the agents are doing. We’re scientific developers at heart at FutureHouse, so we built this platform API-first. For example, you can call Owl to determine if a hypothesis is novel. So - if you’re thinking about an agent that proposes new ideas, use our API to check them for novelty. Or checkout Z. Wei’s Fleming paper that uses Crow to check ADMET properties against literature by breaking a molecule into functional groups. We’ve open sourced almost everything already - including agents, the framework, the evals, and more. We have more benchmarking and head-to-head comparisons available in our blog post. See the complete run-down there on everything. You will notice our agents are slow! They do dozens of LLM queries, consider 100s of research papers (agents ONLY consider full-text papers), make calls to Open Targets, Clinical Trials APIs, and ponder citations. Please do not expect this to be like other LLMs/agents you’ve tried: the tradeoff in speed is made up for in accuracy, thoroughness and completeness. I hope, with patience, you find the output as exciting as we do! This truly represents a culmination of a ton of effort. Here are some things that kept me up at night: we wrote special tools for querying clinical trials. We found how to source open access papers and preprints at a scale to get to over 100 PDFs per question. We tested dozens of LLMs and permutations of them. We trained our own agents with Llama 3.1. We wrote a theoretical grounding on what an agent even is! We had to find a way to host ~50 tools, including many that require GPUs (not including the LLMs). Obviously this was a huge team effort: @m_skarlinski is the captain of the platform and has taught me and everyone at FutureHouse how to be part of a serious technology org. @SGRodriques is the indefatigable leader of FutureHouse and keeps us focused on the goal. Our entire front-end team is just half of @tylernadolsk time. And big thanks to James Braza for leading the fight against CI failures and teaching me so much about Python. @SidN137 and @Ryan__Rhys , for helping us define what an agent actually is. And @maykc for responding to my deranged slack DMs for more tools at all times. Everyone at FutureHouse contributed to this in some way, so thanks to them all! This is not the end, but it feels like the conclusion of the first chapter of FutureHouse’s mission to automate scientific discovery. DM me anything cool you find!

Having trouble keeping up with language models in science? We just prerinted a 497 reference review introducing and reviewing language models and agents. We narrowed our focus to chemistry/biochem on language models (80!), but cover 24 language model agents across science. 1/2

I'm going to industry on Jan 1. Will be on academic leave from @UofR and stay affiliated. Excited to try something new and will have more details to share soon!

Is it time for the protein emoji to return? Can we overcome the past unicode rejection now with the tailwinds of the Nobel Prize? Can proteins be recognized as important as 🫘 or 🙎or 🥀?

We're launching an agent that can do bioinformatics analysis, including repeating analysis from research papers. It is multimodal and results in a complete jupyter notebook (python or R) that ends in a concrete conclusion. Starting with closed-beta now

At FutureHouse, we’ve noticed scientific agents are good at applying average intelligence across tasks. They always seem to make the obvious choices, which is good, but discovery sometimes requires more intuition and insight than average. We’ve made the first step today towards superhuman insight by training a reasoning model for a specific domain of science: designing drug-like molecules. We’re releasing a 24B open-weights reasoning model called 𝚎𝚝𝚑𝚎𝚛𝟶. 𝚎𝚝𝚑𝚎𝚛𝟶 has been trained with reinforcement learning to exceed frontier and human experts across a range of molecular design tasks. 𝚎𝚝𝚑𝚎𝚛𝟶 takes in natural language, reasons in English, and outputs a new molecule. 𝚎𝚝𝚑𝚎𝚛𝟶 is now a tool for our chemistry design agent, Phoenix, which can call upon it to design molecules. Training a reasoning model for a scientific domain like chemistry, rather than math or programming, required a number of small technical advances. For example, we developed an iterative method of split specialist models and aggregation of reasoning traces. Another example is we used LLMs to rewrite questions that were partially solved. A major finding from this work is that we can train with >10x efficiency per experimental measurement when using a reasoning model, rather than fine-tuning. We also found that reasoning models can learn new tasks, developed specifically for this paper and not in pretraining corpora. We even saw a task have 0% performance until 100 steps into RL, at which it randomly solved once. This, along with our change in modality from natural language to molecules, bodes well for applying reasoning models far from natural language. Reasoning models in science are the future. Scientific tasks are naturally verifiable rewards: the physical world is the ultimate arbiter of accuracy, rather than human contractors. The data efficiency gain and ability to exceed frontier models with relatively few parameters/compute mean that we should expect more scientific reasoning models soon. Congrats to team @SidN137, James, @Ryan__Rhys, Albert, @GWellawatte , @maykcaldas , @ludomitch , and @SGRodriques. Thanks to @VoltagePark @nvidia and @huggingface for supporting us, and huge thanks to @ericschmidt for funding @FutureHouseSF The model weights, reward model, and new benchmark are open source. You can also read more about scientific reasoning models in our exclusive with Nature.

Working on a new blog post to estimate when we can simulate cells atom-by-atom. I analyzed 500 papers published in last 35 years to look at the duration of all-atom simulations and how many atoms can be modeled.

New preprint on pre-trained models for Bayesian optimization (BO) of sequences! We show LLMs trained on protein seqs can replace Gaussian processes in BO. Examples: BO of peptide inhibitors with AlphaFold and iterative design of proteins. 1/6 biorxiv.org/content/10.1101/…

ChatGPT API for molecular properties: "Why is this molecule soluble?" Trying out our XAI code with new ChatGPT API.

Found this ASCII molecule viewer. One of the most creative ideas I've seen and it's sitting at 5 stars on GitHub. Mesmerizing stuff github.com/dewberryants/asci…

Here's a picture showing which amino acids interact based on crystallography data. Blue = strong interaction. This is just a heuristic, but it's a nice reference for thinking beyond "hydrophobic doesn't like hydrophilic". Description: crowsandcats.blogspot.com/20…

Imagine if instead of donating $500M to Harvard or $400M to Stanford, the most well-funded PIs and institutions, donors funded 1,000 high risk $500,000 grants. Or launched 50 institutes at state universities. Harvard could withdraw 1% of their endowment to make-up the diff.

I believe we are underestimating the pace of AI over the next 12 months. The compound effect of people+LLMs will be huge - products will come faster, science will go faster, literature will be mined better. AI hype maybe too high, but we're missing hype on smart people + AI 1/3

Nature has a policy that all code must be released from papers - but apparently Deep Mind has released the "pseudocode" for the algorithm instead of reproducible code. Interesting how AI companies are can balance competitive advantage while still getting high impact pubs 1/2

Want to add to this that I'm grateful to @ChemicalScience, @MLSTjournal, and @digital_rsc for giving us a great space to publish some of our most radical papers.

9:00am @generate_biomed announces protein diffusion model making seemingly incredible folds, people note that it's unknown if these structures can experimentally form 3:45pm @_JosephWatson announces Baker Lab experimentally validated protein diffusion models Gotta keep up!

People in tech are freaking out about LLMs, but not in science yet. GPT-4 can just go from jira ticket to completed code in ~20-40% of the time. It doesn't matter if you ask follow-up questions - you need to completely rethink tech. 1/2

I got tenure!! 🎉Thank you students, mentors, and the wonderful computational chemistry community! It's been a long journey and I'm thrilled to have made it.

According to my machine learning model, these 9 molecules are optimal. Please synthesize and test them. 😂

New publishing concept: every 2-3 years you release a batch of ~10 papers around a theme. Call it an album.

My wife said my whiteboard has reached psychopath-level vibes and it's time to erase.

Starting next week in my lab, we're going to build an agent that can set-up and execute molecular dynamics simulations. Go from intent to system set-up. Follow along here: github.com/ur-whitelab/md-ag… Example: "Set-up a solvated simulation of GWGYKRR and KRAS"

I wrote a short opinion piece about the impact of large language models in chemistry that came out today in @NatRevChem! I cover agents, hallucinations, XAI, and what I think is next - would love to hear what others think! nature.com/articles/s41570-0…

I received tenure in July 2021. I then resolved to only work on projects that are truly exciting for me and my group. We went from 2-3 papers per year to 16 in the last 12 months. And they are exciting papers that were a joy to work on. 1/3

I really love equivariant neural networks. There is no field like it - such a blend of graphs, math, quantum mechanics, chemistry, and deep learning. This (very good paper) is citing math papers over 100 years old and deep learning papers from 2022. arxiv.org/abs/2302.03655

Why make one mistake, when you can automate your workflow to make 10,000 mistakes at once

What are the hardest problems right now in drug discovery? Has deep learning had any effect?

I was lucky enough to be a red teamer for GPT-4 and had a great time exploring it for chemistry. It's been tough to keep quiet! Want to thank the excellent team (especially @_lamaahmad!) for taking my feedback seriously about risks and mitigations for LLMs in chemistry. 1/2

2022 was the best year of my career: 18 papers, each of which I'm proud of, and we started more interesting research. I began learning 3 languages and potty-trained my son (very hard). Many fun side-projects too, like protein emoji and molcloud. The biggest change though 1/3

we made this: hasanyone.com/

Why do molecules smell? We create explanations for why molecules have one of 112 scents. The explanations are distilled into natural language. We're excited about data -> model -> explanation of mols. Work by Aditi (looking for PhD program!), @gandhi_heta, @GWellawatte 1/3

I have a new deaf student in my lab and she’s interested to find specific signs for deep learning and computational chemistry words. Anyone know of a community or resource for this?

I have a postdoc opening on deep learning and simulation of protein structure. If you have expertise in simulation only, I will help you develop your skills in deep learning. Will involve work on AlphaFold. Deadline 2/15/2022. Please RT🙏 #postdocjobs thewhitelab.org/#postdoc

However, they are not seen relieving themselves, further suggesting that the people of Middle Earth do not poop. END AI Answer. That's good enough for me, now I can finally get some sleep. 6/6

New paper giving our perspective for how to explain deep learning models in chemistry. We hope the simplicity and utility is clear, and that more people will start providing explanations for predictions! By @GWellawatte, @gandhi_heta, Aditi chemrxiv.org/engage/chemrxiv…

In-context learning (ICL) with LLMs can do regression w/uncertainty! We show Bayesian opt of molecules and catalyst/rxn conditions purely from text. ICL reqs no training, could do in a browser! Work by @marcporosoff @Kyam888 @MichtavyShane 1/6 arxiv.org/abs/2304.05341

Here are my unscientific power rankings for LLMs working with SMILES: 1. GPT-4 2. text-davinci-003 3. Claude-v1.2 4. gpt-3.5-turbo 5. claude-instant 6. text-curie-001 7. cohere-command-xlarge You can compare them here: whitead.github.io/svelte-che…

Our ether0 paper was accepted at NeurIPS 2025! Very proud of the @FutureHouseSF team!

I teach machine learning for chemistry & materials. I think this is the future of chem, chem eng, mat sci. 5 students enrolled. Am I crazy, at the wrong school, have a bad reputation? I feel like Prof. Rentz from The Chair. Guess I gotta change course name to sex, drugs, & ML

FutureHouse's goal has been to automate scientific discovery. Now we used our agents to make a genuine discovery – a new treatment for one kind of blindness (dAMD). We had multiple cycles of hypotheses, experiments, and data analysis – including identify the mechanism of action.

Here's why I'm staying in Academia and not going to industry 🧵 I applied to industry at many stages in my academic career and I got rejected literally every time😂So congratulations to those of you that are getting industry jobs🥳

My group is pleased to announce a new project: Emoji Chemistry😀⚗️🎉. We're working on a new periodic table of elements where tired letters 😴 are replaced with ✨emojis✨Here's an early screenshot of the future of chemistry 😎

Every AI chemistry talk ever: "We need more data!" never does anyone say "I'm so grateful for the data we have" or "here are 10 more data points." Just complaint after complaint from these chemists

My book on deep learning for molecules and materials has a new domain: dmol.pub shortest url I could get

I just solved one of the most difficult open problems in science: how to insert SVGs into @matplotlib figures. Now you can have SVG molecules from @RDKit_org, arbitrarily nested figures, recursive figures, etc. All in beautiful SVG format. I'm retiring now.

Read the tech report: arxiv.org/abs/2511.02824 Use it: platform.edisonscientific.co… See our new website with much more info: edisonscientific.com/

1/6 For the last few months @glenhocky and I have been asking what large language models (LLM) can do for chemistry. In our new preprint, we show LLMs know a bit of chemistry and can do a lot: like compute the dissociation curve of H2. arxiv.org/abs/2108.13360

Learning to write in LaTeX can be a difficult step in academia. I've put together a guide on how to go from novice to expert. It seems hard, but if you follow these steps you'll immediately earn the respect of your friends and colleagues.

Here's a demo showing scraping a Github repo and then asking complex questions about code. Advantages are it's specific to the repo, the LLM can convert between language & code, and it cites specific line numbers. My first @streamlit app :)

Congrats to @bradyajohnston to really solving the problem of going from molecular dynamics trajectory to Blender geometry. I's going to enable some cool new visualizations/art in computational chemistry. @bradyajohnston's code: github.com/BradyAJohnston/Mo…

Happy International Students' Day 🎉They overcome many extra challenges to become scholars and scientists, and their success at facing them does not appear on CVs. I appreciate their contributions to science and hope we can do more to support them!

Our first result at @FutureHouseSF! So much to talk about: 1. We quantified hallucinations major LLMs 2. Benchmarked RAGs on biology questions 3. Beat all models, tied meatbags 4. Wrote a complete human genome Wikipedia Will tweet more details later... futurehouse.org/wikicrow