iiinova — where your digital self keeps evolving

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

9/10 Huge thanks to my co-first authors @igashov.bsky.social, @rne.bsky.social, Irina Morozova, our amazing collaborators (across EPFL, Merck, UCSF, ETH Zürich, Università del Piemonte Orientale, and AITHYRA), and @mmbronstein.bsky.social & Bruno Correia for their guidance and connections throughout

120

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

8/10 Using synthezisable design, we found de novo hits for BRD4 BD1 and SARS-CoV-2 Nsp3 Mac1, with binding confirmed by NMR and Xray.

110

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

7/10 This strategy also optimizes local properties like hydrogen bonding more efficiently than SMILES-based methods such as REINVENT (w @rebeccaneeser.bsky.social).

100

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

6/10 For the remaining targets, we paired LDDM with a new inference-time sampling strategy that keeps designs inside a combinatorial chemical space like Enamine REAL, so designs can be synthesized with little effort.

110

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

5/10 We tested LDDM in the wet lab in different modes. As a redesign and optimization: replacing R-groups of a VHL–CDO1 molecular glue yielded analogs with nanomolar affinity; swapping in non-canonical amino acids on a cathepsin S peptide inhibitor improved its potency 3.5-fold.

100

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

4/10 In in-silico benchmarks, we demonstrate good geometric validity, broad coverage of chemical space, accurate pharmacophore recovery, and strong local and covalent docking performance with less memorization than co-folding methods.

100

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

3/10 LDDM generates molecules in 3D conditioned on a protein pocket, and can be flexibly conditioned on arbitrary fragments or substructures. Therefore, one model handles de novo design, fragment linking/growing, and constrained or covalent docking.

100

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

2/10 Most generative models for small-molecule design are judged only on in-silico benchmarks. Far fewer are tested for whether the molecules they propose can be synthesized and survive the wet lab. We built LDDM to close that gap.

100

Adrian Dobbelstein @adobbelstein.bsky.social · 20/09/2026

Happy to share our preprint: "A Unified 3D Generative Model for Synthesizable Structure-Based Drug Design". Highlights, preprint, and code below: 🧵1/10

1154

Reposted by Adrian Dobbelstein

Tomáš Pluskal @pluskal-lab.org · 21/08/2025

The first day of our bioML symposium at @iocbprague.bsky.social went very well. It is amazing to see how fast this field is moving forward. Thanks to everyone for the fantastic talks! 🤖🧬

1156

Reposted by Adrian Dobbelstein

Arne Schneuing @rne.bsky.social · 07/03/2025

The code & camera-ready version of our #ICLR2025 paper on "Multi-domain Distribution Learning for De Novo Drug Design" are now available 📚 Paper: openreview.net/forum?id=g3V... 💻 Code: github.com/LPDI-EPFL/Dr... (1/4)

2257

Reposted by Adrian Dobbelstein

Ilia Igashov @igashov.bsky.social · 23/01/2025

Really excited: 2/2 our papers got accepted at #ICLR2025! Thanks to all amazing co-authors and hope to see you in Singapore! 🇸🇬 If you're interested in generative models and RL agents for drug design check our papers in the thread 🧵

141