Affinity values with different trends

[MiaJY-Yang]

Excellent work on Boltz2 - it's truly impressive and highly valuable for glycan-protein interaction studies!

I'm currently using the model to predict glycan-protein affinities and have a question about output interpretation. In my results, I've noticed two distinct output categories with three values each:

1. affinity_pred_value, affinity_pred_value1, affinity_pred_value2
2. affinity_probability_binary with three analogous outputs The challenge I'm facing is that the three affinity_pred_value outputs show completely divergent trends in my experiments. Could you please clarify:
3. What is the difference between these three values?
4. Which value should be prioritized for interpreting actual binding affinities? This clarification would greatly help ensure I'm interpreting results correctly.

[BenedictIrwinApheris]

From what I've seen:

Value1 is the low and Value2 is the high. Seems that the average of value1 and value2 gives value, so it is a symmetric interval.

There are then 2x predictions

affinity_pred_value which is like pIC50 as 6-log_10(IC50 in uM) affinity_probability_binary which is a probability in [0,1]

[qiuzy]

you can read boltz2 manuscript published in arxiv (https://doi.org/10.1101/2025.06.14.659707). In Page 26, it tells more details of the affinity model ensemble. This two predicted affinities come from two trained models. Also, you can find more explanation in repo doc: output part in https://github.com/jwohlwend/boltz/blob/main/docs/prediction.md