كتابة تعليق

The authors address an important and highly relevant problem in computational drug discovery: how to efficiently estimate small-molecule promiscuity across the human proteome. Moving beyond conventional one-target or limited-panel screening toward proteome-scale assessment of off-target liability is an interesting direction with clear potential utility for early-stage drug discovery. The proposed OmniBind framework is computationally attractive, particularly in its attempt to compress proteome-wide affinity information into a rapidly accessible molecular-level promiscuity score. The manuscript is generally well motivated and presents several interesting analyses. Nevertheless, we have several concerns regarding the construction of the training labels, validation of the resulting promiscuity predictions, and assessment of model generalizability that we believe should be addressed.

1. The final OmniBind model is trained predominantly on labels derived from Ligand-Transformer predictions, with an additional set of pseudo-labels generated by a random-forest model trained on a small subset of protein affinities. This raises the question of how errors or systematic biases in these teacher models propagate into OmniBind. In particular, the low MSE reported in Figure S2 evaluates OmniBind against Ligand-Transformer-derived CPP scores rather than experimental measurements and therefore primarily demonstrates successful distillation of the teacher model. The comparisons with Boltz-2 provide useful cross-model validation, but do not establish experimental accuracy. Can the authors quantify the performance of the intermediate random-forest model and Ligand-Transformer against experimental data, or provide an experimentally grounded evaluation of OmniBind that is independent of its training labels?

2. The “Experimental promiscuity estimation” section is difficult to interpret conceptually. The manuscript defines computational promiscuity as the mean predicted affinity across 15000 proteins, whereas the experimental validation uses proxy quantities derived from a much smaller and non-random subset of experimentally assayed proteins. For ChEMBL and PDSP, the proxy is the mean affinity across available measurements among the most frequently assayed proteins, whereas SPD uses the fraction of tested proteins with AC_{50} ≤ 10 μM. These quantities differ both from the proteome-wide definition and from one another. More importantly, could the authors assess how representative these sparse experimental panels are of the proteome-wide quantity—for example, by subsampling the Ligand-Transformer predictions using the same protein panels and comparing the resulting panel-level estimates with the complete-proteome promiscuity score?

3. Figure 3 appears to use different numbers of compounds across panels A and B, although the text and caption indicate that the molecular-weight analyses were performed on the same drug sets as the corresponding Cmax analyses. Please clarify the sample selection for each panel and, if different subsets were used, explain why. If the intention is to compare correlations directly, it would be preferable to compute the molecular-weight correlations on exactly the same compounds used for the corresponding OPS–Cmax analyses.

4. The statement that an MSE of 0.41 indicates insufficient training data is not fully supported by the reported metric alone. Because MSE is scale-dependent, please provide additional evidence such as training/validation performance, learning curves, or a comparison with an appropriate baseline to justify this interpretation.

5. For the final held-out set of 1,000 TargetMol molecules, it would be helpful to report the chemical similarity of test compounds to the training set, for example the maximum Tanimoto similarity to any training molecule and/or the distribution of nearest-neighbour similarities. A scaffold-based split or analysis would further clarify whether OmniBind generalizes to novel chemotypes rather than primarily interpolating among structurally related compounds.

6. Please define P_i explicitly in the transformed expression for \pi_k. It appears to correspond to the initial concentration/abundance of protein P_i, but the relationship to [P0i][P_{0i}] in the preceding equation is not stated.

7. Figure 2 refers to the final predictor as “OmniPred” and Figure S1 refers to it as “PromPt,” whereas the manuscript refers to this model as OmniBind. Is PromPt an earlier name or a separate model? Please define the term or replace it with OmniBind for consistency.

8. The stated range of pairwise correlations between experimental promiscuity proxies (r=0.25–0.49) does not appear to match Figure 2C. Please verify the reported correlation range and clarify which pairwise comparisons are included in this statement. It would be helpful to include the full evaluation (where r=0.25) in the SI.

9. The Cmax changes are interesting. There are likely close analogs (perhaps in ChEMBL, perhaps in the patent literature) where Cmax varies quite a bit. It would be interesting to see how this method predicts in narrow chemical spaces and whether the predicted binders gained/lost can rationalize such changes in PK.

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.