$ cat projects/crispr-quantum.md
Quantum-classical neural networks for CRISPR
research 2023 — published 23 Oct 2024 Author
at a glance
- Venue
- Symposium on High Performance Computing Systems (SSCAD 2024)
- Published
- 23 October 2024
- Language
- Portuguese
- Area
- HPC · quantum computing · bioinformatics
context
Written with two co-authors at the Department of Computer Science of PUC Minas, Belo Horizonte, and published in the proceedings of SSCAD 2024, pages 13–24, by the Brazilian Computer Society (SBC).
the problem
The DepMap project publishes, for hundreds of cancer cell lines, how strongly each line depends on each gene in CRISPR knock-out experiments, and how many copies of each gene it carries. The paper asks whether a network with quantum layers can predict the first from the second, and whether hybrid quantum neural networks are viable for regression at all.
Its stated hypothesis is about performance: that quantum models can improve the analysis and prediction of gene-edit effects by exploring the data faster and more efficiently. That makes the classical baseline, and the run times, the numbers that matter.
approach
What was tried — and, where it applies, what it taught. The second half is the part that usually gets edited out, and the part that is actually useful.
-
- Built three networks of almost the same size with TensorFlow, Keras and PennyLane: a classical one, one with a single quantum layer between classical layers (H-NNW), and one whose intermediate layers are all quantum (HQNNW). Each was trained for 20 epochs, with the quantum circuits simulated on classical hardware.
learned Simulation dominates the cost. The classical network used 1.12 s of CPU time; the hybrid took 20 min 1 s and the fully quantum network 1 h 17 min 22 s.
-
- Compared the three on the same measures, from loss and accuracy to R², CPU time and wall time.
learned No network wins outright. The hybrids end with higher accuracy (0.9750 against 0.9352) and lower final MSE and MAE, while the classical network ends training with the lowest loss (0.0059) and the lowest epoch-20 MAE, MSE, MAPE, MSLE and log-cosh. All three R² values are negative, so none predicts better than a constant equal to the mean.
outcome
what I would do differently
State the result the way the table shows it. The paper concludes that the hybrids offer advantages in prediction and could beat classical models with the right optimisation. With every R² negative and the classical network holding the lowest training loss, the claim the data supports is a comparison under simulation, not an advantage.
Put the baseline that a negative R² is measured against, a constant prediction of the mean, into the table, and say in the abstract that no quantum hardware was used. Both would have made the result easier to read correctly.