Absci Investor Presentation Deck
CASE STUDY: DESIGNING BETTER HER2 BINDERS
Al quantitatively predicts antibody affinity
Predicted ACE score
9
8-
7
6
LO
5
4
3
2
Pearson R= 0.97
Spearman p= 0.84
RMSE= 0.42
3 4 5 6
Measure ACE score
7
8 9
HIGH PREDICTIVE PERFORMANCE
Pearson R correlation of 0.93
- Trained on 90% of dataset
- Results shown for 10% of dataset not
seen by model
Pearson R
Spearman p
RMSE
0.4
0.6
0.8
Measured replicate 1 vs Measured replicate 2
Measured vs Predicted same replicate
Measured vs Predicted different replicates
0.0
0.2
HIGH QUALITY DATA
Models trained on one replicate can
predict unseen data from a
different replicate
1.0
Count
20
15
10
LO
5
0
7.50
Bachas, S., Rakocevic, G. et al., "Antibody optimization enabled by artificial intelligence predictions of binding affinity and naturalness," 2022 pre-print in bioRxiv.
ABSCI CORPORATION 2023 ALL RIGHTS RESERVED
Affinity-optimized
Trastuzumab
7.75 8.00
8.25 8.50 8.75 9.00 9.25 9.50
- log10 (KD)
HIGH AFFINITY PREDICTIONS
Models can find variants with higher
affinity than seen in training data
- 92 of top 100 predicted high-affinity
variants bind tighter than trastuzumab
absci. 23View entire presentation