wasarou/weight-space-programming-boltz2

Experimental weight-space edits to Boltz-2

This repository contains the exact thirty edited Lightning checkpoints (v1-v30) from a direct weight-edit experiment with Boltz-2 2.2.1. These are unvalidated research checkpoints. No checkpoint meets the original requirement of zero R/S and E/Z errors across all 209 systems with lower bond and angle errors.

Source code | Checkpoint manifest | SHA-256 checksums

Original evaluation code snapshot: 1e0f5ae742ec87dfab23c92dcd984bafe2500b22.

For new executions, use the security-maintained runtime. It retains Boltz-2 2.2.1 with patched dependencies and restricted checkpoint loading. Compatibility was checked with 41 CPU tests and two GPU smoke predictions. The published benchmark scores were produced with the original environment and were not recomputed with this runtime.

Provenance

The unmodified baseline is the upstream Boltz-2 structure checkpoint. Its SHA-256 is 090e82ac8c92f5e943fa1b39e7410a44027bea7243c0bbb3caa67a77fc1428e1. The baseline is not duplicated here. Each edited file was checked against its original edit record, and all untargeted tensors were verified during checkpoint creation. The files retain the upstream MIT license and copyright notice in LICENSE.

Evaluation scope

All thirty candidates were compared on the same 55 development systems at seed 1, with a shared baseline and matched GPU allocations. No candidate met the simultaneous-improvement criteria. v11 had better development point estimates than v2 for R/S, joint stereochemistry, and the two geometry medians, but its bond-length median was worse than the baseline. These development results do not establish a validated replacement.

The frozen v2 follow-up evaluated 50 validation systems at seed 1 and 104 test systems at seeds 1, 2, and 3. Joint test agreement was unchanged at 103/312 predictions. All primary pooled-effect confidence intervals included zero. Repeated seeds were averaged within each system before biological-group resampling.

The v11 validation/test follow-up is recorded separately. The same test split had already been used for v2, so it is not a new untouched test.

The completed v11 follow-up had validation R/S agreement of 27/50 for the baseline and 25/50 for v11. Across the 104 test systems and three seeds, R/S agreement changed from 125/312 to 129/312, full E/Z from 233/312 to 230/312, and joint agreement remained 103/312. All primary pooled-effect confidence intervals included zero. All 362 fresh validation/test baseline outputs were byte-identical to those in the v2 study. The original repair acceptance criteria remain unmet.

• Full thirty-candidate development comparison
• Frozen v2 evaluation
• Frozen v11 evaluation
• Descriptive v11 versus v2 comparison

Development results

Counts are complete-system agreements out of 55. Geometry values are median RMS errors against independent ETKDGv3/UFF conformers, not experimental binding-pose errors.

Candidate	R/S	Full E/Z	Joint	Bond median (A)	Angle median (degrees)
Upstream baseline	21/55	36/55	20/55	0.038337	2.9074
v1	23/55	36/55	22/55	0.038804	2.8740
v2	23/55	36/55	22/55	0.038833	2.8752
v3	21/55	37/55	20/55	0.038572	2.8370
v4	20/55	36/55	19/55	0.038371	2.9042
v5	22/55	37/55	21/55	0.044742	3.1605
v6	9/55	21/55	5/55	2.505318	46.1132
v7	21/55	35/55	19/55	0.073891	4.7763
v8	24/55	37/55	22/55	0.045782	3.1395
v9	21/55	38/55	20/55	0.038519	2.7932
v10	20/55	36/55	19/55	0.038503	2.8888
v11	24/55	36/55	23/55	0.038460	2.8734
v12	21/55	36/55	20/55	0.038325	2.9107
v13	22/55	36/55	20/55	0.038424	2.9508
v14	21/55	37/55	19/55	0.038772	2.9440
v15	21/55	37/55	20/55	0.038517	2.7716
v16	22/55	37/55	20/55	0.039933	2.7588
v17	24/55	36/55	23/55	0.038491	2.8781
v18	21/55	37/55	16/55	0.036566	2.8999
v19	19/55	37/55	15/55	0.044543	2.7551
v20	20/55	36/55	19/55	0.038544	2.9124
v21	20/55	37/55	19/55	0.038559	2.8215
v22	21/55	36/55	20/55	0.038626	2.8900
v23	21/55	36/55	19/55	0.039163	2.9081
v24	20/55	36/55	19/55	0.038417	2.9475
v25	20/55	36/55	19/55	0.038305	2.8799
v26	21/55	36/55	20/55	0.038291	2.8932
v27	8/55	30/55	7/55	0.587255	34.5568
v28	22/55	35/55	20/55	0.099217	6.0212
v29	8/55	26/55	7/55	0.520772	24.0307
v30	8/55	22/55	1/55	0.640607	28.6757

Download

uv tool install huggingface-hub
hf download wasarou/weight-space-programming-boltz2 boltz2_wsp_absolute_vectors_only_v11.ckpt --local-dir .

Use the checkpoints with the maintained environment in the current source repository. It keeps Boltz-2 2.2.1 and overrides the outdated dependency pins in its upstream package metadata. The reported paired comparisons disable physical, contact, and FK guidance; standard upstream CLI defaults are not necessarily benchmark-equivalent.

Prepared inputs, full predictions, and private session transcripts are not distributed in this model repository. The public dataset is linked in the model-card metadata; reproducing the archived campaign also requires the prepared benchmark inputs and legacy evaluation dependencies described in the source repository.

References

• Passaro et al. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction.
• Takehi et al. (2026). Sparks of Weight-Space Programming.