Beyond Silicon: Could DNA Support Future Forms of Molecular AI? | Prof. Tarik A. Rashid | Dr. Tarik Ahmed Rashid
Public Research / 8 Min Read
Beyond Silicon: Could DNA Support Future Forms of Molecular AI?
Introduction
Disregard Hardware or Graphical Processing Units (GPUs), because AI may shortly run on DNA. What if artificial intelligence didn't depend on silicon chips but on strands of life itself? In DNA computing, millions of solutions can be explored simultaneously inside a single drop of biological material. In a traditional computer, information is carried by electronic circuits and in binary form. A different physical substrate is used in DNA computing: sequences of adenine (A), thymine (T), guanine (G) and cytosine (C). The information can be encoded in DNA strands, while molecular operations such as hybridisation, ligation, amplification, separation, and sequencing can be utilised as computational steps. The primary goal of DNA computing isn't to process common programs quickly. Rather, it promises that many candidate answers can be represented in parallel by many DNA molecules. This is an attractive conceptual possibility for DNA computing because combinatorial search problems have a very rapid growth rate in the number of solutions. The original experiment performed by Adleman demonstrated that a slight problem in a directed Hamiltonian path could be solved using molecular biology operations. Later, Lipton extended the theoretical discussion, suggesting approaches to solving hard computational problems like Boolean satisfiability using DNA. Practically, however, the field has always been constrained by a lack of perfect control of biochemical reactions. These DNA strands can misalign themselves, fail to align, amplify unequally, form secondary structures or be lost in the filtering and readout process during the amplification. This distinction is important for modern research on DNA computing, where the theoretical possibility does not imply that anything can be done, and that what can be done in simulation should not be done in experiments.
Scientific Background: How DNA Can Represent Computation
Designed strands can recognise complementary strands of DNA because base pairing is predictable in the molecule; A pairs with T, and G pairs with C, making DNA useful for molecular computation. This recognition can be applied to produce molecules, to pick out target strands, or to initiate downstream reactions. Each node or edge in a graph can have a unique DNA sequence for path-based DNA computing. If the fragments can be joined together, they can create longer strands, which can be interpreted as possible routes. Several laboratory operations are commonly associated with DNA-computing models:
- Hybridisation: Complementary strands of DNA anneal to each other and allow designed fragments to identify each other.
- Ligation: DNA ligase may be able to ligate DNA together to form longer strands that can be whole or partial paths.
- PCR amplification: Polymerase Chain Reaction can amplify strands that have certain regions at which primers bind to them, such as source and destination markers.
- Single-strand conformation polymorphism (SSCP) and gel separation by length can be used to help distinguish strands by structure or length (both are imperfect filters).
- Readout: Remaining candidates may be interpreted using length-based detection, sequencing or other molecular/electronic measurements.
Recent literature reviews highlight the advancement of DNA computation beyond the initial graph search applications to molecular circuits, strand-displacement systems, diagnostics, and ideas for DNA-based data-processing. The same reviews point out that scalability, error control, cost, speed and readout are still significant challenges to practical deployment.
The Shortest Path Problem
The Shortest Path Problem is to find a path from a source node to a destination node that has the smallest sum of costs. It can be anything from distance, time, energy, money, risk, or something else measurable. Dijkstra's algorithm and Bellman-Ford are some popular algorithms used in conventional computing for such problems. In DNA computing, the problem is solved in another way: molecules are employed to create and to filter candidate paths. Designed oligonucleotides can be used to encode each node and edge in a small graph. The candidate paths are generated by molecular assembly, while invalid paths are discarded in the filtering steps. Then, the shortest valid path can be determined by length, encoded weight or analysis with a sequence. Some experimental and semi-experimental DNA strategies have been used for the weighted graph problem and shortest-path computation, such as length-based and concentration-based encoding strategies.
Standard DNA-Computing Workflow for Path OPTIMIZATION
A simplified procedure for the shortest-path optimization based on DNA is as follows:
- Encoding: is assigning unique DNA sequences to nodes, edges or edge weights.
- Pool generation: Mix designed strands to allow for many possible molecular paths in parallel.
- Source-destination selection: Amplify the strands that begin and end at the desired nodes using PCR or other selection techniques.
- Validity filter: Discard paths that repeat nodes, do not contain required markers, have wrong fragments, or do not have a proper structure.
- Cost evaluation: Estimate the cost of a path based on length, concentration, melting properties, labels or sequencing data.
- Readout: Determine the optimal remaining candidate path and unravel the path into nodes and edges of the graph.
This process is a scientifically valid process, but it is fragile. Requires the production of the right candidate molecules in the initial pool, survival through amplification and filtering, and the ability of the read-out method to differentiate the best path from the molecular noise.
Why Standard DNA Computing Needs Improvement?
A classical DNA-computing model is frequently similar to a one-shot search. A vast library of candidate molecules is created, and the system eliminates unwanted molecules one by one. If the path is lost at an early stage, it is possible that it will not be recovered. This contrasts with many modern optimisation algorithms that continually modify a set of candidate solutions. The drawbacks of a typical molecular workflow are:
Incomplete initial pool generation: Not all valid paths may be generated to a useful quantity at the start.
- PCR bias: certain strands of DNA will grow better than other strands, adjusting the apparent population distribution.
- Filtering may retain invalid molecules or discard valid molecules: false positives/false negatives.
- Secondary structures: strands can fold or bind to themselves, which can affect amplification and separation.
- Readout uncertainty: When dealing with complex graphs, the resolution offered by gel-based or structure-based separation might be inadequate.
- Scalability: larger graphs will need a lot more molecules, reactions, controls, and validation steps.
Evolutionary DNA Computing: Core Idea
Evolutionary algorithms search by keeping a group of candidate solutions. Better candidates are chosen, mixed, modified, and passed on to the next generation. This approach works well when the search area is too big for a complete search or when the system gains from slowly improving imperfect solutions. The evolutionary DNA-computing model proposed by Godar Ibrahim, Tarik Rashid, and Ahmed Sadiq is significant because it sees DNA strands as an evolving group instead of a single fixed set of molecules. Their work outlines a simulated model where strands are repeatedly recombined, wrong solutions are removed, and better or nearly optimal solutions are chosen. This makes their method more flexible than a standard DNA algorithm. In simple terms, the improvement is this: instead of discarding all failed strands, the algorithm tries to reuse useful partial information. It does not rely only on the initial random molecular generation but creates new generations of candidate paths. This is the scientific reason the method is intriguing.
Proposed Evolutionary Enhancements
The dropped strands will be rescued and reused
Dropped strands will be retrieved and used again. Partial paths that are discarded can still be of benefit and may be mended and reinserted with the correction of missing segments. In simulation, it's easy, but in the wet lab, it needs tools such as barcoding and sequencing or microfluidics.
Semi-Crossover and Mutation
Crossover merges nodes and mutation fixes nodes or edges to diversify. For DNA systems, they can be obtained by strand displacement or by overlap-based assembly, which can be difficult to control.
Fitness Selection
In simulation, fitness can be derived directly, while in the wet-lab, it must be deduced from other measurable signals, including concentration, length or the sequencing output. The biggest problem is finding a good physical fitness surrogate.
Adaptive Mutation Strategy
If progress comes to a standstill to avoid local optima, mutation strength can become greater. In experiments, this will be setting the reaction conditions and/or the libraries, but it must be carefully managed to ensure that the strands are not invalid.
Practical Feasibility: What Is Realistic and What Is Not?
DNA computing is experimentally real, and evolutionary algorithms are well-established in computer science. However, combining both in a practical wet-lab system is still very challenging. Unlike simulation, where every candidate solution can be tracked and controlled, DNA experiments depend on physical molecules, noisy reactions, imperfect filtering, and indirect measurements. For this reason, the proposed method should first be tested on small, well-defined graph problems. Its performance can then be compared with a standard DNA-computing workflow using clear measures such as best path cost, average path cost, recovery rate, error rate, molecular diversity, resource cost, and reproducibility. In this form, evolutionary DNA computing is best understood as a promising research idea, not yet as a scalable replacement for conventional computing.
References
1. L. M. Adleman, "Molecular computation of solutions to combinatorial problems," Science, vol. 266, no. 5187, pp. 1021-1024, 1994. DOI: 10.1126/science.7973651.
2. R. J. Lipton, "DNA solution of hard computational problems," Science, vol. 268, no. 5210, pp. 542-545, 1995. DOI: 10.1126/science.7725098.
3. G. J. Ibrahim, T. A. Rashid, and A. T. Sadiq, "Evolutionary DNA computing algorithm for job scheduling problem," IETE Journal of Research, pp. 514–527, May 2018, doi: 10.1080/03772063.2017.1362964.
4. G. J. Ibrahim, T. A. Rashid, and A. T. Sadiq, "Improving DNA Computing Using Evolutionary Techniques," International Journal of Advanced Computer Science and Applications, vol. 7, no. 3, 2016. DOI: 10.14569/IJACSA.2016.070316.
5. Z. Ibrahim, Y. Tsuboi, O. Ono, and M. Khalid, "Hybrid Concentration-Controlled Direct-Proportional Length-Based DNA Computing for Numerical Optimization of the Shortest Path Problem," in Biologically Inspired Approaches to Advanced Information Technology, Lecture Notes in Computer Science, vol. 3853, pp. 206-221, Springer, 2006. DOI: 10.1007/11613022_18.
6. R. E. Polak and A. J. Keung, "A molecular assessment of the practical potential of DNA-based computation," Current Opinion in Biotechnology, vol. 81, article 102940, 2023. DOI: 10.1016/j.copbio.2023.102940.
7. M. Zhang and D. Han, "Concept, development and applications of DNA computation," Fundamental Research, 2023. DOI: 10.1016/j.fmre.2023.06.015.
8. D. Y. Zhang and G. Seelig, "Dynamic DNA nanotechnology using strand-displacement reactions," Nature Chemistry, vol. 3, pp. 103-113, 2011. DOI: 10.1038/nchem.957.
9. T. A. Rashid, "Computer Science & Artificial Intelligence," 2026. Available: tarikrashid.com.