Corrupted frames in mdCATH coordinate data

[pietronvll]

Dear mdCATH Authors,
First of all, thanks for this amazing dataset!

I'm opening this issue to report that 48 out of 5,398 HDF5 files (0.89%) contain corrupted coordinate frames at the tail end of specific trajectories. I noticed this problem while using mdCATH for an ML project. My training loop randomly returned NaN values, and after some inspections (and thanks to Claude :) ), I pinpointed the issue to corrupted data. Let me also tag @g-turri to keep him posted on the updates on this.

Two types of corruption are present:

1. Integer overflow in coordinates (47 trajectories, 289 frames): One coordinate dimension contains values ~2.1 x 10^7 Angstroms, possibly consistent with an int32 overflow artifact (2^31 / 100 = 21,474,836.48). The non-overflowed dimensions retain plausible values.

2. All-zero coordinate frames (1 trajectory, 1 frame): All atomic coordinates are exactly 0.0.

In both cases the corruption always appears at the tail of a trajectory -- from some frame onward until the last frame. The frame immediately preceding the first bad frame is always physically valid.

The corruption was verified against the upstream HuggingFace repository by downloading fresh copies of 5 affected files and confirming byte-identical bad values, ruling out any downstream processing artifact.

Summary

Metric	Value
Files scanned	5,398
Trajectories scanned	134,950
Total frames scanned	62,581,026
Affected PDB domains	48 (0.89%)
Affected trajectories	48 (0.036%)
Corrupted frames	290 (0.00046%)

Each affected PDB has exactly one bad trajectory (one specific temperature/replica combination). The remaining trajectories for the same PDB are unaffected.

Affected trajectories

PDB ID	Temp	Replica	Total frames	First bad frame	Bad frames	Overflow dim	Max |coord|
1ca1A01	379	4	60	58	2	x	17,833,816
1fhoA00	348	2	10	3	7	y	21,474,838
1gkgA02	320	4	150	140	10	z	21,474,838
1hm7B00	379	0	270	260	10	z	21,474,838
1mjwA00	413	1	250	246	4	z	21,474,838
1pqwA00	450	2	430	420	10	z	21,474,838
1u0mA01	348	4	180	175	5	z	21,474,838
1y66A00	450	0	140	135	5	y	21,474,838
1zhhB01	450	0	310	307	3	y	21,474,838
1zxqA02	348	2	161	160	1	y	21,474,838
2akcA00	320	1	40	31	9	z	21,474,838
2dnxA00	348	3	440	438	2	y,z	21,474,838
2exrA02	320	4	470	468	2	z	21,474,838
2jbrA03	413	3	460	454	6	z	21,474,838
2jysA00	320	4	470	466	4	z	21,474,838
2k88A00	450	4	410	409	1	z	21,474,838
2lmkA00	450	3	40	30	10	z	21,474,838
2m5hA00	320	4	430	420	10	y	21,474,838
2mutA00	379	0	418	411	7	z	21,474,838
2p6wA00	348	0	476	470	6	z	21,474,838
2qenA03	348	3	448	447	1	(all-zero)	0
2qfzA02	348	0	360	352	8	z	21,474,838
2qyzA01	348	4	480	471	9	z	21,474,838
2rnnA00	348	2	30	22	8	z	21,474,838
2wdcA03	348	3	475	470	5	z	21,474,838
2wn3A02	379	1	405	404	1	z	21,474,838
2y94A03	348	3	320	318	2	y	21,474,838
2z90A01	450	0	50	46	4	z	21,474,838
2zuvA03	379	1	260	250	10	z	21,474,838
3b0dC00	320	3	100	91	9	z	21,474,838
3d8lA00	348	0	480	478	2	z	21,474,838
3gk0A00	348	0	356	352	4	z	21,474,838
3hn2A02	348	4	50	42	8	z	21,474,838
3j7aN00	413	1	357	350	7	z	21,474,838
3ogdA01	450	4	205	202	3	y	21,474,838
3t7aA02	450	3	460	450	10	z	21,474,838
3uoaB01	413	1	70	65	5	z	21,474,838
3w0lB03	413	2	110	104	6	z	21,474,838
3wutG00	413	1	120	111	9	z	21,474,838
4ekuA01	320	2	376	371	5	z	21,474,838
4fchA02	413	2	440	430	10	y	21,474,838
4j11D00	413	1	410	402	8	z	21,474,838
4lsxC00	450	1	450	447	3	y	21,474,838
4o30B00	379	3	220	211	9	y	21,474,838
4qpiC00	348	1	357	350	7	z	21,474,838
4tmpA00	348	4	220	216	4	y	21,474,838
4xxhA02	413	4	90	80	10	z	21,474,838
5d5pA02	413	3	450	441	9	z	21,474,838

[giadefa]

Hi Pietro, Thanks for the detailed report. It's in the nature of GPUGRID to have some bad trajectories. We should see what to do, but at least there is already your table. gianni

…

On Thu, Feb 19, 2026 at 12:11 PM Pietro Novelli ***@***.***> wrote: *pietronvll* created an issue (compsciencelab/mdCATH#7) <#7> Dear mdCATH Authors, First of all, thanks for this amazing dataset! I'm opening this issue to report that *48 out of 5,398 HDF5 files* (0.89%) contain corrupted coordinate frames at the tail end of specific trajectories. I noticed this problem while using mdCATH for an ML project. My training loop randomly returned NaN values, and after some inspections (and thanks to Claude :) ), I pinpointed the issue to corrupted data. Let me also tag @g-turri <https://github.com/g-turri> to keep him posted on the updates on this. Two types of corruption are present: 1. *Integer overflow in coordinates* (47 trajectories, 289 frames): One coordinate dimension contains values ~2.1 x 10^7 Angstroms, possibly consistent with an int32 overflow artifact (2^31 / 100 = 21,474,836.48). The non-overflowed dimensions retain plausible values. 2. *All-zero coordinate frames* (1 trajectory, 1 frame): All atomic coordinates are exactly 0.0. In both cases the corruption *always* appears at the tail of a trajectory -- from some frame onward until the last frame. The frame immediately preceding the first bad frame is always physically valid. The corruption was verified against the upstream HuggingFace repository by downloading fresh copies of 5 affected files and confirming byte-identical bad values, ruling out any downstream processing artifact. Summary Metric Value Files scanned 5,398 Trajectories scanned 134,950 Total frames scanned 62,581,026 Affected PDB domains 48 (0.89%) Affected trajectories 48 (0.036%) Corrupted frames 290 (0.00046%) Each affected PDB has exactly *one* bad trajectory (one specific temperature/replica combination). The remaining trajectories for the same PDB are unaffected. Affected trajectories PDB ID Temp Replica Total frames First bad frame Bad frames Overflow dim Max |coord| 1ca1A01 379 4 60 58 2 x 17,833,816 1fhoA00 348 2 10 3 7 y 21,474,838 1gkgA02 320 4 150 140 10 z 21,474,838 1hm7B00 379 0 270 260 10 z 21,474,838 1mjwA00 413 1 250 246 4 z 21,474,838 1pqwA00 450 2 430 420 10 z 21,474,838 1u0mA01 348 4 180 175 5 z 21,474,838 1y66A00 450 0 140 135 5 y 21,474,838 1zhhB01 450 0 310 307 3 y 21,474,838 1zxqA02 348 2 161 160 1 y 21,474,838 2akcA00 320 1 40 31 9 z 21,474,838 2dnxA00 348 3 440 438 2 y,z 21,474,838 2exrA02 320 4 470 468 2 z 21,474,838 2jbrA03 413 3 460 454 6 z 21,474,838 2jysA00 320 4 470 466 4 z 21,474,838 2k88A00 450 4 410 409 1 z 21,474,838 2lmkA00 450 3 40 30 10 z 21,474,838 2m5hA00 320 4 430 420 10 y 21,474,838 2mutA00 379 0 418 411 7 z 21,474,838 2p6wA00 348 0 476 470 6 z 21,474,838 2qenA03 348 3 448 447 1 (all-zero) 0 2qfzA02 348 0 360 352 8 z 21,474,838 2qyzA01 348 4 480 471 9 z 21,474,838 2rnnA00 348 2 30 22 8 z 21,474,838 2wdcA03 348 3 475 470 5 z 21,474,838 2wn3A02 379 1 405 404 1 z 21,474,838 2y94A03 348 3 320 318 2 y 21,474,838 2z90A01 450 0 50 46 4 z 21,474,838 2zuvA03 379 1 260 250 10 z 21,474,838 3b0dC00 320 3 100 91 9 z 21,474,838 3d8lA00 348 0 480 478 2 z 21,474,838 3gk0A00 348 0 356 352 4 z 21,474,838 3hn2A02 348 4 50 42 8 z 21,474,838 3j7aN00 413 1 357 350 7 z 21,474,838 3ogdA01 450 4 205 202 3 y 21,474,838 3t7aA02 450 3 460 450 10 z 21,474,838 3uoaB01 413 1 70 65 5 z 21,474,838 3w0lB03 413 2 110 104 6 z 21,474,838 3wutG00 413 1 120 111 9 z 21,474,838 4ekuA01 320 2 376 371 5 z 21,474,838 4fchA02 413 2 440 430 10 y 21,474,838 4j11D00 413 1 410 402 8 z 21,474,838 4lsxC00 450 1 450 447 3 y 21,474,838 4o30B00 379 3 220 211 9 y 21,474,838 4qpiC00 348 1 357 350 7 z 21,474,838 4tmpA00 348 4 220 216 4 y 21,474,838 4xxhA02 413 4 90 80 10 z 21,474,838 5d5pA02 413 3 450 441 9 z 21,474,838 — Reply to this email directly, view it on GitHub <#7>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3KUOVNML35M4UYQE5A3AD4MXVFNAVCNFSM6AAAAACVW5T7OWVHI2DSMVQWIX3LMV43ASLTON2WKOZTHE3DIMBVGU2TAMY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>