A Slice of Infinity: Building Robust, Interpretable AI for Open-ended Biodesign

Transcript

1. AI, Engineering Biology & Beyond 15 January 2026 michiel.stock@ugent.be A Slice of In f inity: Building Robust, Interpretable AI for Open- ended Biodesign Michiel Stock 1
2. Presentation outline 2 How can we use computational intelligence for biodesign? 1. searching 2. representing
3. 1. Searching How can we discover if we don’t know what to look for? 3
4. Open-endedness 4 Hughes, E., Dennis, M., Parker-Holder, J., Behbahani, F., Mavalankar, A., Shi, Y., Schaul, T., Rocktaschel, T., 2024. Open-Endedness is Essential for Arti fi cial Superhuman Intelligence. https://doi.org/10.48550/ arXiv.2406.04268 A system is open-ended if the sequence of entities it produces is novel and learnable from the observer's point of view. New entities past future present novelty: the further in the future, the more surprising the entities learnable: taking more historical entities into account makes the future more predictable
5. Open-ended systems are compositional 5
6. Novelty and diversity over f itness 6 Objective-driven Novelty-driven
7. 7 Using the MAP-elites algorithm to design a multiplex assembly of genetic circuits every cell is the optimal multiplex
 for a given price/# slots combination Ongoing work with Simeon Castle and Thomas Gorochowski (Bristol University). Designing multiplexer with QD
8. AI Phenotype realization Entity itself Selection Adaptable GP map Genetic variation Other entities Bioengineer System Interaction with phenotype (e.g., via stress response) Predation, Competition, Symbiosis Growth control Kill switch Artificial selection Changing environment Synthetic genetic circuits Developmental reprogramming Engineered mutagenesis and recombination Dynamic genetic operators Horizontal gene transfer Induced mutagenesis Realizing open-endedness in synbio 8 Stock, M., Gorochowski, T., 2023. Open-endedness in synthetic biology: a route to continual innovation for biological design. Science Advances 10. https://doi.org/10.31219/osf.io/wv5ac
9. 2. Representing How do we represent complex constructs computationally? 9
10. Representation building blocks (in the hyperdimensional space) 10
11. The hyperdimensional space homogeen all hypervectors look the same homogeneous holografisch information is contained in all elements of the hypervector holographic robust every hypervector is surrounded by an enormous number of similar vectors robust
12. HDC in one slide 12
13. Learning and reasoning with HDC memory-based learning 13 data atoms memory output learning reasoning encoding decoding
14. Pretrained embeddings into the hyperspace HDC as a glue for combining embeddings 14 images proteins text molecules pretrained deep neural network embedding vector hypervector random projection Sutor, P., Yuan, D., Summers-Stay, D., Fermuller, C., Aloimonos, Y., 2022. Gluing neural networks symbolically through hyperdimensional computing. https://doi.org/10.48550/arXiv.2205.15534
15. HDC for computational biology 15 fast and efficient HDC traditional extremely fast and energy-ef fi cient multimodal combining several data sources explainable interpretable and explainable! neuro-symbolic compositional ...ATCAAC... can represent complex compositions Stock M, Van Criekinge W, Boeckaerts D, Taelman S, Van Haeverbeke M, Dewulf P, et al. (2024) Hyperdimensional computing: A fast, robust, and interpretable paradigm for biological data. PLoS Comput Biol 20(9): e1012426. https://doi.org/10.1371/journal.pcbi.1012426
16. Deep learning vs HDC (according to me) 16 Deep learning learning a complex mapping from one space to the other using a dense dataset • learning arbitrarily complex maps from huge amounts of data • fi xed, “continuous” space HDC representing arbitrary compositions and structures in a homogeneous space • ef fi ciently characterize sparse combinations and structures • “open-ended” space
17. Phage-based synbio the enemy of our enemy 17 f( , ) safe phage with a speci fi c bacterial host Modifying host speci fi city by changing the tail fi bers 17 naughty bacterial cell heroic phages
18. PhageHostLearn 18 Boeckaerts, D., Stock, M., Ferriol-González, C., Oteo-Iglesias, J., Sanjuán, R., Domingo-Calap, P., De Baets, B., Briers, Y., 2024. Prediction of Klebsiella phage-host speci fi city at the strain level. Nat Commun 15, 4355. https:// doi.org/10.1038/s41467-024-48675-6 Dimi Boeckaerts
19. Predicting phage host-speci f icity messy data 19 many targets possible complex glycan structures to overcome mining for bacterial genomes for interacting phages The HDC operators can fuse all this information into “interaction vectors” Victor Németh
20. Predicting phage host-speci f icity results 20
21. Design rules phage lytic proteins 21 Criel, B., Taelman, S., Criekinge, W.V., Stock, M., Briers, Y., 2021. PhaLP: A Database for the Study of Phage Lytic Proteins and Their Evolution. Viruses 2021, Vol. 13, Page 1240 13, 1240. https://doi.org/10.3390/ V13071240 Ste f f Taelman
22. In summary Synthetic biology can be open- ended: • in f inite arrangement of building blocks; • focus on novelty and diversity of constructs; • learning requires representations that are highly compositional and f lexible. 22
23. Acknowledgements and further reading Michiel Stock KERMIT, Ghent University https://kermit.ugent.be/
 michiel.stock@ugent.be Tutorial: https://github.com/Kermit-UGent/ HDC_tutorial Special thanks to Biocompute lab, LAB and the MEMO research group for collaborations.