Victor Gertner

Results

Preprocessing speed

Data: PINDER monomers
Unit: ms / protein, one process
HW: H100 + Xeon 8468

AlphaSurf meshes in 27 ms; MSMS needs 258 ms to reach the same size. Spectral operators now dominate the pipeline.

Method	Vertices	Surface	Mesh proc.	Spectral ops	Features	Total	Fwd+bwd
AlphaSurf	1,320	27.1	12.4	184.6	28.5	252.6	32.2
EDTSurf gs 0.3	891	22.9	8.9	127.6	16.7	176.1	32.0
EDTSurf gs 0.4	1,549	29.0	12.9	185.0	24.8	251.7	34.0
EDTSurf gs 0.5	2,340	48.3	23.6	343.2	47.9	463.0	35.0
NanoShaper gs 0.4	2,595	70.9	26.3	375.5	56.8	529.5	35.0
NanoShaper gs 0.5	4,208	78.0	39.4	552.8	81.1	751.3	39.0
MSMS + coarsening	1,306	122.0	136.0	192.7	28.5	479.2	32.0

Fwd+bwd is the GPU training step. Default grids: EDTSurf 1,028 ms / 31K vertices, NanoShaper 739 ms / 70K, raw MSMS 122 ms / 13K. At gs 0.3, NanoShaper yields usable meshes for under 70% of proteins; AlphaSurf for close to 100%.

MaSIF-Ligand

Task: 7-class pocket classification
Data: 1,972 / 238 / 489 pockets
Metric: balanced accuracy, 3 seeds

Same accuracy as MSMS and NanoShaper, and the only top method faster than the training step.

Method	ms*	Bal. acc.
AlphaSurf	23.7	84.2
EDTSurf gs 0.3	23.9	81.7
EDTSurf gs 0.4	25.7	81.2
EDTSurf gs 0.5	30.0	79.4
NanoShaper gs 0.4	41.1	82.3
NanoShaper gs 0.5	45.7	84.3
MSMS + coars.	62.6	83.2

* input construction per protein, amortised over 20 CPU workers. Grey rows are slower than the 31 ms training step.

PINDER-Pair

Task: residue pairs closer than 5 Å
Data: 42K train dimers; 1,835 / 309 / 1,582 holo / apo / AF2
Metric: mean per-system AUROC, 3 seeds

NanoShaper and MSMS lead on Holo but lose up to 13 points on Apo and AF2. AlphaSurf is best on unbound heterodimers.

Method	ms*	Holo	Apo	AF2
EDTSurf gs 0.3	18.3	88.2	83.3	77.1
EDTSurf gs 0.4	24.9	89.6	84.3	78.3
AlphaSurf	25.7	89.5	85.5	81.7
EDTSurf gs 0.5	38.1	90.5	84.0	80.4
NanoShaper gs 0.4	49.7	93.1	80.2	73.5
MSMS + coarsening	56.4	92.3	82.3	71.7
NanoShaper gs 0.5	78.3	93.1	81.3	73.7

Method	ms*	Holo	Apo	AF2
EDTSurf gs 0.3	18.3	92.5	86.4	85.4
EDTSurf gs 0.4	24.9	93.5	87.2	86.4
AlphaSurf	25.7	92.9	86.9	86.8
EDTSurf gs 0.5	38.1	94.0	87.3	87.3
NanoShaper gs 0.4	49.7	95.6	74.6	71.7
MSMS + coarsening	56.4	94.9	73.3	71.5
NanoShaper gs 0.5	78.3	95.5	75.1	72.1

* input construction per protein, amortised over 20 CPU workers. Grey rows are slower than the 34 ms training step, so not usable on the fly.

MISATO, learning on dynamics

Task: residues within 10 Å of the ligand
Data: 16K MD runs × 100 frames
Metric: tested on holo, 3 seeds

Sampling a random MD frame per epoch means 1.5M meshes, too many to store. Built on the fly, it beats holo-only training.

Training	Surface	F1	AUROC	AUPRC
DistoDyn†	–	46.4	–	–
Holo	EDTSurf	42.7	82.2	40.0
Holo	AlphaSurf	48.9	85.7	47.7
MD frames	EDTSurf	43.1	82.8	41.0
MD frames	AlphaSurf	50.3	86.8	50.4

std ≤ 0.9. † best model of Portal et al. (2026).

ProteinGym, S3F at scale

Task: zero-shot mutation effect
Data: 217 DMS assays, 2.4M mutations
Pretrain: 31K CATH domains

Swapping dMaSIF point clouds for AlphaSurf points, all else unchanged: same Spearman, 80× faster surfaces.

Method	Spearman	Vertices	Surface	Fwd+bwd
S3F	0.470	4,136	424	56.5
α-S3F	0.474	563	5.2	23.8

ms per protein; 20 workers, one H100.

Tufting and augmentation

Data: PINDER-Pair
Metric: AUROC, 3 seeds

Tufting gives a small but consistent gain, largest on AF2.

AUROC	Holo	Apo	AF2
No tufting	92.12	86.40	85.46
Tufting	92.36	86.67	85.96

On MISATO, tufting adds +1.4 F1 and +2.4 AUPRC. On-the-fly coordinate noise (σ ≤ 0.5 Å) adds +0.5–0.9 AUROC on PINDER-Pair, mostly on Apo and AF2.

Surface viewer

Data: 1A27_AB

Simplex → SAS boundary

alpha 5.0

Repro details

Backbone: AtomSurf setup, ProNet residue graph + DiffusionNet surface encoder with bidirectional communication; only the surface mesh changes.

PINDER-Pair: 42,220 train dimers; test holo/apo/AF2 1,835/309/1,582 (valid meshes for all methods); AUROC from the graph head; 20 workers.

MISATO: 13,765/1,595/1,612 complexes; random MD frame at train time, frame 0 at test.

Seeds: 2024, 2025, 2026 (S3F: 2024). α-complexes with CGAL 6.2.