Happy to share our new paper, out now in @J_A_C_S, in collaboration with @abbvie: https://t.co/HguEUjBy0A. We used 15 years of AbbVie Suzuki parallel library data to build yield prediction ML models and evaluate their potential use as tools in the drug discovery cycle:

🦜SPARROW is an algorithmic framework developed by @JennaFromer & #JameelClinic PI @cwcoley published last year in @NatComputSci. Now the two co-authors along w/ Alexandra Volkova have released new improvements for improving molecular design efficiency https://t.co/RLD0FcSmJr

Excited to share our latest work using contrastive learning to study publication bias: https://t.co/x0ncn0aVWH Instead of supervised learning, we design a novel strategy leveraging chemical data structure, enabling a deeper investigation of patterns in published substrate tables.

Happy to share our latest paper @WenhaoGao1 @RShprints , out now in @J_A_C_S! We use contrastive learning to train on substrate scopes from @CASChemistry data, showing despite publication bias in chemical reaction reporting, learned embeddings can still reflect reactivity!

ML can now predict the outcomes of chemical reactions, but often perform poorly in real-life scenarios. A team of researchers including #JameelClinic PI @cwcoley propose a stricter approach to evaluation to ensure model performance isn't overstated. https://t.co/8OfrwQsdHU

Many drugs, like vaccines & antibodies, rely on carbohydrates to function. Carbohydrates are made through glycosylations producing 2 different anomers. Researchers from the lab of #JClinic PI @cwcoley & @CHEMUCPH propose using ML to predict which anomer will be made.

Open to all- I will be giving a presentation on my work with AbbVie today at 12 PM EST! See below for details and the link to register so you can get the zoom info. There’s a great lineup of speakers, so do consider joining!

I *never* want to download a text file containing a single bibtex entry. Just let me copy it.

We are excited to announce the Symposium on Generative Modeling for Chemistry, Biology, and Material Discovery at the ACS Spring 2025 meeting. This symposium will explore the latest advancements and applications of generative models in molecular science research. (1/2)

A 2-year postdoc position in computational chemistry & ML is available in our group at @psl_univ (more info about the project below). Applications can be submitted through the CNRS job portal: https://t.co/hvaAxWAz6r. Deadline: Sept 24th. Retweets are appreciated! #chempostdoc

Incorporating Synthetic Accessibility in Drug Design: Predicting Yields of Suzuki Cross-Couplings by Leveraging AbbVie’s 15-Year Parallel Library Data Set JACS @cwcoley @MITChemE @MITEECS @MIT @techreview @TotalSyntheses #DrugDesign #SyntheticAccesibility

This work demonstrates a practical workflow for the evaluation and application of predictive ML models in small molecule drug discovery. Code and one-hot encoded versions of the datasets are available on our GitHub, linked in the paper. Feel free to reach out to talk more!

We deployed the models prospectively to suggest similar monomer replacements for new and follow-up libraries. In a case study, the models increased success rates by 11% for library design, and 50% for library rescue, showing their efficacy in increasing DMTA cycle success rates.

Recognizing that the chemist is the de facto expert in library design, we compared extrapolative model predictions to those of 11 expert AbbVie medicinal chemists. Across 6 libraries, the models often outperformed the chemists, at worst still performing in range of the chemists.

We evaluated models across various tasks, splits that reflected use-case scenarios, and common model types and representations. While differences between models were slight and extrapolation tasks were challenging, the RF model with fp + DFT features performed best overall.

The library dataset consisted of over 24,000 reactions and nearly 3,500 organohalides and organoborons. The majority of diversity was from substrates rather than conditions, with a wide yield distribution and overall 67% success rate (>0 isolated yield).

Was glad to be able to contribute to this work - we present a contrastive learning-based pretraining strategy for downstream reactivity prediction tasks, leveraging substrate scope data from @CASChemistry! Check out the arXiv link below in @WenhaoGao1's post to read the paper:

Thanks to @OPRD_ACS for featuring our recent Outlook article, "Dataset Design for Building Models of Chemical Reactivity", in this month's Highlights feature: https://t.co/H7cy5sO3TC Here's the link to our original Outlook for those who missed it!

Excited to be at NOLA for #ACSSpring2024! I’ll be presenting my work with @abbvie at the MEDI Poster Session, Wednesday 3/20 from 7-9 PM in Hall C! See attached pic for the abstract :) Feel free to reach out if you want to talk ML for chemistry or catch up!

🚨 Attention aspiring researchers! 🚨 Today is the LAST day to apply for our Summer Undergraduate Research Programs 🔬 Don't miss out on this opportunity to dive into cutting-edge projects and gain hands-on experience. 🌐 https://t.co/IBRGL8969T #Applynow #ResearchOpportunity