"In November 2024, preliminary results of CASP16 showed AlphaFold 3-based models did not significantly outperform older methods for predicting protein-ligand interactions. The top performing models in the CASP16 Pose Prediction for Pharma Targets section were ClusPro[13] and CoDock[14] utilizing AlphaFold 2 based predictions, human visual inspection, and manual adjustments.[15]
In April 2025 Isomorphic Labs raised $600 Million in its first ever external funding round, led by Thrive Capital.[16][17]"
Would like more details technically on their approach. Is it a neuro-symbolic approach, do they do something beyond just scaling, or some other architecture breakthrough, but alas drug discovery must remain proprietary. Thank gods for capitalism!
Alphafold's weights are public (1, 2, and multimer, all except Alphafold 3) and there are several techniques to use it to do drug design, from Alphafold 1's "32 blank spaces" technique to the ML technique that should receive the Nobel prize for worst named algorithm in history "hallucination".
The popular method was this. You see because Alphafold goes from DNA code -> 3d structure, and you know parts of the 3d layout of a successful medicine (if it interacts, it will get really close. So you just put very low values in the alphafold output and "lock" them), then you reverse the normal learning process. You lock the outcome in place and you use backpropagation to learn the inputs instead of the weights. The inputs ... are the DNA code to produce the medicine. Hence "hallucination". You pretend to already see the medicine's effect and work back from that rather than trying to determine the effect of a given medicine. So generating 100 high-quality candidates for a medicine becomes letting a python script run for a day or two on a single 3090 (it was a while ago) rather than 5 years of research.
Oh and it comes with nice confidence matrices, so you can rank the 100 candidates. You can postprocess also using Alphafold to check many things that disqualify medicines (toxicity is essentially interaction with non-target molecules, which you can simply test for).
You can even, I mean I didn't get it working, but it is possible to "detoxify" a known-toxic protein using hallucination. Figure out which small changes make it non-toxic, and combine that with also making sure it maintains its therapeutic effects. That works. It's absolutely incredible and open source. But you need someone who is comfortable with high-level AI code to make it work.
And this works, with serious and very difficult code changes, with more modern models. Oh and pharma companies are unwilling to hire people capable of doing this (at least in Europe) at wages they'd need to pay to get them. As per usual.
This is from February 2026 and notably there’s no way for the public to try out the model, nor is it published in a peer-reviewed journal
"In November 2024, preliminary results of CASP16 showed AlphaFold 3-based models did not significantly outperform older methods for predicting protein-ligand interactions. The top performing models in the CASP16 Pose Prediction for Pharma Targets section were ClusPro[13] and CoDock[14] utilizing AlphaFold 2 based predictions, human visual inspection, and manual adjustments.[15]
In April 2025 Isomorphic Labs raised $600 Million in its first ever external funding round, led by Thrive Capital.[16][17]"
https://en.wikipedia.org/wiki/Isomorphic_Labs
Shows the importance of a good harness.
What technical advancement made this possible? The text is vague about this.
there is a link to the white paper in the article.
Would like more details technically on their approach. Is it a neuro-symbolic approach, do they do something beyond just scaling, or some other architecture breakthrough, but alas drug discovery must remain proprietary. Thank gods for capitalism!
Alphafold's weights are public (1, 2, and multimer, all except Alphafold 3) and there are several techniques to use it to do drug design, from Alphafold 1's "32 blank spaces" technique to the ML technique that should receive the Nobel prize for worst named algorithm in history "hallucination".
The popular method was this. You see because Alphafold goes from DNA code -> 3d structure, and you know parts of the 3d layout of a successful medicine (if it interacts, it will get really close. So you just put very low values in the alphafold output and "lock" them), then you reverse the normal learning process. You lock the outcome in place and you use backpropagation to learn the inputs instead of the weights. The inputs ... are the DNA code to produce the medicine. Hence "hallucination". You pretend to already see the medicine's effect and work back from that rather than trying to determine the effect of a given medicine. So generating 100 high-quality candidates for a medicine becomes letting a python script run for a day or two on a single 3090 (it was a while ago) rather than 5 years of research.
Oh and it comes with nice confidence matrices, so you can rank the 100 candidates. You can postprocess also using Alphafold to check many things that disqualify medicines (toxicity is essentially interaction with non-target molecules, which you can simply test for).
You can even, I mean I didn't get it working, but it is possible to "detoxify" a known-toxic protein using hallucination. Figure out which small changes make it non-toxic, and combine that with also making sure it maintains its therapeutic effects. That works. It's absolutely incredible and open source. But you need someone who is comfortable with high-level AI code to make it work.
And this works, with serious and very difficult code changes, with more modern models. Oh and pharma companies are unwilling to hire people capable of doing this (at least in Europe) at wages they'd need to pay to get them. As per usual.
This is old news.