IC-311On siltuximab GRAVY reduction, Lambo-2 achieves the best concept shift while ESM2 produces the most naturalness-disrupted designs

Aya Abdelsalam Ismail, Tuomas Oikarinen, Amy Wang, Julius Adebayo, Samuel Don Stanton, Hector Corrada Bravo, Kyunghyun Cho, Nathan C. Frey

SourceConcept Bottleneck Language Models For Protein Design

The paper evaluates several released protein design models on the task of reducing the GRAVY (hydropathy) index of the FDA-approved antibody siltuximab, constrained to edit distance at most 5 from the original sequence. Lambo-2, a classifier-guided discrete diffusion model, achieves the largest shift in the GRAVY distribution among all models tested. WJS and the authors' CB-PLM produce designs with the least disruption to protein naturalness (as measured by TAP scores), whereas ESM2 and a non-deep-learning hydrophilic resample baseline produce the most disrupted designs. The authors note that their general-purpose CB-PLM, trained on over 700 concepts, delivers results comparable to single-concept-optimized state-of-the-art models.

Evidence
correlational
Key metric
Edit distance constraint of 5 from siltuximab; qualitative ranking: Lambo-2 best GRAVY shift, CB-PLM second; WJS and CB-PLM least TAP disruption, ESM2 and hydrophilic resample most TAP disruption
Caveat
Results are for a single protein (siltuximab) and a single property (GRAVY); specific GRAVY values and TAP scores are reported only in figures, not in the text. The comparison involves models trained on different datasets and with different objectives.
Model
Lambo-2, ESM-2, WJS
Methods
Therapeutic Antibody Profiler [eval]
Related work
Lambo-2 [compared-to], PROPEN [compared-to], WJS [compared-to], ESM-2 [compared-to]
Extraction
automatic-extraction