Thomas Litfin

Isomorphic Labs Drug Design Engine (IsoDDE), a unified computational drug-design system

Announcement:
www.isomorphiclabs.com/articles/the...

Report:
storage.googleapis.com/isomorphicla...

10.02.2026 15:16 — 👍 30    🔁 10    💬 0    📌 3

Multiple protein structure alignment at scale with FoldMason Protein structure is conserved beyond sequence, making multiple structural alignment (MSTA) essential for analyzing distantly related proteins. Computational prediction methods have vastly extended ou...

FoldMason is out now in @science.org. It generates accurate multiple structure alignments for thousands of protein structures in seconds. Great work by Cameron L. M. Gilchrist and @milot.bsky.social.
📄 www.science.org/doi/10.1126/...
🌐 search.foldseek.com/foldmason
💾 github.com/steineggerla...

30.01.2026 06:11 — 👍 297    🔁 147    💬 4    📌 3

Here are the success rates of de novo pipelines based on which designs I could actually identify the methods for.

22.01.2026 15:47 — 👍 23    🔁 12    💬 2    📌 1

Building antibodies blindfolded: the paradox of de novo design By Natasha Murakowska and Joseph Harman

Loved this post from A-Alpha: aalphabio.substack.com/p/building-a.... If anything I think the IPSAE (or any other post-hoc metric) picture is even worse than they show: after optimization the fraction of false positives would (probably) be even higher than in this dataset

09.01.2026 22:46 — 👍 6    🔁 3    💬 1    📌 0

New preprint🚨
Imagine (re)designing a protein via inverse folding. AF2 predicts the designed sequence to a structure with pLDDT 94 & you get 1.8 Å RMSD to the input. Perfect design?
What if I told u that the structure has 4 solvent-exposed Trp and 3 Pro where a Gly should be?

Why to be wary🧵👇

16.12.2025 15:15 — 👍 57    🔁 21    💬 4    📌 1

Know when to co-fold'em This is the official web page for the James Fraser Lab at UCSF.

I'm really excited to break up the holiday relaxation time with a new preprint that benchmarks AlphaFold3 (AF3)/“co-folding” methods with 2 new stringent performance tests.

Thread below - but first some links:
A longer take:
fraserlab.com/2025/12/29/k...

Preprint:
www.biorxiv.org/content/10.6...

29.12.2025 22:25 — 👍 72    🔁 30    💬 5    📌 2

Figure comparing runtime and VRAM utilization between Boltz-2 (baseline), and LMI4Boltz (+memory, and +chunk). Main plot: runtime (y-axis) versus token count (x-axis), showing that all methods scale similarly, with +memory and +chunk handling larger token counts. Inset scatterplot: PDB test lDDT scores from Boltz-2 versus LMI4Boltz, showing a strong linear correlation (values near y = x). Right panels: – Top bar chart: maximum tokens processable on a 24 GB GPU increase from 1596 (Boltz-2) to 2356 (+memory) and 2660 (+chunk). – Bottom bar chart: H200 runtime for 1596 tokens remains comparable across methods.

🧶🧬 We present LMi4Boltz:
www.biorxiv.org/content/10.1...

Boltz-2 is an excellent open source alternative to AlphaFold3. However, high VRAM use restricts modeling large complexes. Using careful memory management, we increase the Boltz-2 size limit by >60% while maintaining execution speed.

31.10.2025 20:54 — 👍 24    🔁 7    💬 0    📌 0

I don't believe that is the consensus. Novel folds can be well predicted if there is strong support from evolutionary information.

29.04.2025 23:04 — 👍 1    🔁 0    💬 1    📌 0

I guess any holo structure template implicitly biases the model towards a known ligand-binding site. It would be interesting to know the ceiling of what can be achieved with this implicit information by providing the native ligand holo structure (kind of cross-docking vs re-docking).

12.02.2025 22:22 — 👍 1    🔁 0    💬 0    📌 0

When using the ground-truth template AF3 did not achieve perfect fidelity or the template issue only affected a subset of systems?

12.02.2025 03:33 — 👍 0    🔁 0    💬 1    📌 0