Naail Kashif-Khan
@naailkhan.bsky.social
Understanding proteins using computers and AI 🔬💻
Early-stage drug discovery scientist at a startup
Love heavy metal and the Arsenal 🎸🔴
Early-stage drug discovery scientist at a startup
Love heavy metal and the Arsenal 🎸🔴
Our results highlight the need for a high-throughput ML method for predicting binding affinity - Boltz-2 is cheap and fast but performs poorly. FEP works very well but is expensive and impractical to run on thousands or even millions of compounds.
September 10, 2025 at 12:41 PM
Our results highlight the need for a high-throughput ML method for predicting binding affinity - Boltz-2 is cheap and fast but performs poorly. FEP works very well but is expensive and impractical to run on thousands or even millions of compounds.
The Boltz-2 paper claims comparable performance to FEP methods, but our results show worse performance across all targets. Why? We see that accuracy is directly related to the similarity of the molecular glue compound to compounds in the training set:
September 10, 2025 at 12:41 PM
The Boltz-2 paper claims comparable performance to FEP methods, but our results show worse performance across all targets. Why? We see that accuracy is directly related to the similarity of the molecular glue compound to compounds in the training set:
We also investigate selectivity prediction, with experimental data from 18 compounds across four different CRBN targets. We see good predictive performance from FEP but poor performance from Boltz-2
September 10, 2025 at 12:41 PM
We also investigate selectivity prediction, with experimental data from 18 compounds across four different CRBN targets. We see good predictive performance from FEP but poor performance from Boltz-2
We present the most systematic evaluation of FEP and Boltz-2 on molecular glues to date (93 compounds across 6 different protein complexes). We find that FEP shows very good performance overall with errors between 0.3-1.25 kcal/mol to the experiment. Boltz-2 shows comparatively poor performance:
September 10, 2025 at 12:41 PM
We present the most systematic evaluation of FEP and Boltz-2 on molecular glues to date (93 compounds across 6 different protein complexes). We find that FEP shows very good performance overall with errors between 0.3-1.25 kcal/mol to the experiment. Boltz-2 shows comparatively poor performance: