FoldScape
A deep learning framework conditioned on sequence and controlled by the folding score Q as a control variable. It can generate ensembles of Cα conformations of proteins at any folding stage, combining hierarchical distance regression with a rectified flow decoder.
Timeline (last 90 days)
FoldScape is a sequence-conditioned deep learning framework that can generate the corresponding Cα conformational ensemble given a folding fraction Q.
FoldScape's generation strategy is divided into two steps: first, hierarchical, sliding-window one-dimensional inter-residue distance regression reconstructs the three-dimensional chain, and then, through chain assembly and sampling, a rectified flow decoder outputs the conformational ensemble.
FoldScape's training data covers two complementary trajectory sources: multi-scale coarse-grained folding trajectories and all-atom mechanical unfolding trajectories.
FoldScape designs a stride-one crop-superposition scheme that extends inference capability to chains several times longer than a single chain-level forward pass.
In comparative tests on 20 human proteins (36 to 857 residues in length), under matched sample sizes, the range of folding coordinates covered by FoldScape is about an order of magnitude greater than that of AlphaFold 3 ensembles with seed variation.
AlphaFold 3 anchors better in the local region immediately adjacent to the native state, whereas FoldScape does not reach that region, making the two complementary.
FoldScape's reconstruction accuracy remains roughly constant along Q, with no obvious degradation from collapsed coils to near-native molten globule states.
FoldScape cannot cover the very-near-native region immediately adjacent to the native state, which still needs to rely on models such as AlphaFold 3.
FoldScape provides a way to obtain conformational ensembles without per-protein molecular simulation, suitable for studying folding intermediates, mechanical responses, or partial unfolding behavior.
Combining FoldScape with AlphaFold 3 may be a more reasonable practical path.
All Facts (10)
FoldScape is a sequence-conditioned deep learning framework that can generate the corresponding Cα conformational ensemble given a folding fraction Q.
50%UnverifiedFoldScape's generation strategy is divided into two steps: first, hierarchical, sliding-window one-dimensional inter-residue distance regression reconstructs the three-dimensional chain, and then, through chain assembly and sampling, a rectified flow decoder outputs the conformational ensemble.
50%UnverifiedFoldScape's training data covers two complementary trajectory sources: multi-scale coarse-grained folding trajectories and all-atom mechanical unfolding trajectories.
50%UnverifiedFoldScape designs a stride-one crop-superposition scheme that extends inference capability to chains several times longer than a single chain-level forward pass.
50%UnverifiedIn comparative tests on 20 human proteins (36 to 857 residues in length), under matched sample sizes, the range of folding coordinates covered by FoldScape is about an order of magnitude greater than that of AlphaFold 3 ensembles with seed variation.
50%UnverifiedAlphaFold 3 anchors better in the local region immediately adjacent to the native state, whereas FoldScape does not reach that region, making the two complementary.
50%UnverifiedFoldScape's reconstruction accuracy remains roughly constant along Q, with no obvious degradation from collapsed coils to near-native molten globule states.
50%UnverifiedFoldScape cannot cover the very-near-native region immediately adjacent to the native state, which still needs to rely on models such as AlphaFold 3.
50%UnverifiedFoldScape provides a way to obtain conformational ensembles without per-protein molecular simulation, suitable for studying folding intermediates, mechanical responses, or partial unfolding behavior.
50%UnverifiedCombining FoldScape with AlphaFold 3 may be a more reasonable practical path.
50%