Over the past decades, we have witnessed a staggering number of advancements in genetic engineering. Some popular events include the discovery of DNA's double helical structure, the development of the polymerase chain reaction, the cloning of Dolly the sheep, and the advent of CRISPR-Cas9 gene editing, just to name a few [1]. Parallelling these developments is the monumental growth of artificial intelligence (AI), namely generative models that form the basis of popular applications such as ChatGPT. Today, generative models greatly enhance biology research and clinical discovery, with tools such as AlphaFold even allowing scientists to predict protein structure [2].

This year at iGEM Toronto, we are developing Plasmid.AI – a platform that uses generative AI to create new, functional DNA sequences. Our ultimate goal for this platform is to write entirely new genomes to create novel organisms, namely new bacteria and new phages, which have fine-tuned phenotypes. These designer organisms can have success in various domains, such as in agriculture, in gut probiotics, in bioremediation, in bioproduction, and in therapeutics. Given the complexity of this task, our project this year focuses on the simplest proof-of-concept: the generation and validation of new plasmid parts/components.

In addition to developing DNA-writing technology, our team seeks to have the Plasmid.AI platform positively impact communities. To this end, our human practices and entrepreneurship teams have honed in on solving antibiotic resistance as the most crucial domain for promoting social good and commercialization. By focusing on this issue in particular, Plasmid.AI can have the greatest positive impact locally and globally.

Thus, the overarching aim of our project this year is twofold:

1. Develop the foundational technology to write new life – starting with new plasmid components
2. Lay the groundwork for this foundational technology to positively impact communities – starting with antibiotic resistance

These goals will serve as the building blocks to our future project goals of writing new genomes.

Plasmids are highly important biotechnological tools with numerous applications. Some of them are listed below:

Being able to develop plasmids and plasmid components with fine-tuned functionality would greatly enhance their impact in the above applications and their utility as biotechnological tools. However, despite recent advances in DNA sequence models, no tools have demonstrated the capacity to generate functionally relevant plasmids/plasmid components de novo. This is a gap that part (1) of our project this year addresses.

Dry lab generates whole plasmids using machine learning (ML). Through an in silico validation pipeline, we then filter the sampled sequences by predicted viability and deliver promising plasmid components to the wet lab.

There are three main components of our model development for plasmid generation: sequence tokenization, model architecture and training, and model inference.

Tokenization: We used a sophisticated tokenization strategy called Byte Pair Encoding (BPE) trained on our dataset of wildtype plasmid sequences. This tokenizer identifies and encodes frequently occurring subsequences, effectively compressing repetitive DNA patterns common in long plasmids.

Model Architecture and Training: For our model architecture, we use the Mamba2 model, which has demonstrated success in long-sequence modelling tasks. This choice was driven by Mamba's ability to handle long plasmid sequences efficiently, making it well-suited for plasmid data. Its various features allow it to overcome the limitations of traditional attention-based transformers in handling extended sequences. Our training objective follows the next-token prediction paradigm, where the model learns to predict the next token in the sequence given the preceding tokens. This method enables the model to learn the underlying patterns and structures of plasmid sequences, capturing both local and global features of plasmid organization.

Model Inference: During inference, we generate novel plasmid sequences through a carefully designed process. We initiate sequence generation with the CLS token, signaling the start of a new sequence. We then utilize nucleus sampling [3] to predict the next tokens within the sequence, ultimately generating many plasmid sequences in parallel.

Overall Batch Summary Stats: Our in-silico pipeline validates batches of generated plasmids, identifying promising plasmid origins of replication (ORIs) and selecting antimicrobial resistance (AMR) proteins with potential for experimental success. Before sequence-specific validation, we assessed the overall quality of each batch using clustering tools like MOBsuite [4]. We examined sequence clusters, alignment to known motifs, and checked for circular contigs, which are indicative of novel plasmids.

ORI Alignment and Annotation: Generated sequences were aligned to the oriC database using MMseqs2 [5] and BLAST+ [5]. High-scoring sequences were passed to the annotation phase, where key replication motifs were identified using MEME Suite. Class B plasmids were further prioritized based on alignment and manual review.

Antimicrobial Protein Validation: The generative model, trained on whole-plasmid sequences, assumes generated proteins should be functional. However, due to the unannotated training data, the model needed to generalize beyond known wild-type proteins. Using AMRFinderPlus [6], we aligned generated plasmid sequences against a diverse protein set, including antimicrobial resistance factors.

Structure-Based Protein Validation: To detect protein misfolding or expression issues early, we used the ESMFold [7] deep-learning model for structure prediction. Validation metrics such as plDDT, PAE, and pTM-scores were applied to ensure protein accuracy. Regions with high pLDDT/pTM scores and low PAE values can be considered potentially viable for downstream applications.

The wet lab team is tasked with validating dry lab's AI-generated sequences in vitro. To validate the efficacy of AI-generated plasmids, the wet lab team synthesizes these plasmids and test them in bacterial cultures. This process starts with first confirming the functionality of these generated plasmids. The generated plasmids' origin of replication (oriV) are isolated, synthesized, and subsequently cloned into a testing backbone and transformed for characterization.

To facilitate high-throughput validation, the testing backbone is engineered with Golden Gate type IIS-restriction enzyme cut sites for transgene insertion. The functionality of the AI-generated oriVs will be assessed through growth assays, where the number of colony-forming units (CFUs) in experimental populations will be compared to negative control populations. A significant difference in CFUs in the experimental group would indicate successful plasmid replication and functionality.

Further validation of plasmid viability will be conducted by extracting plasmids from transformed bacteria followed by sequencing, hence confirming the propagation of the oriV and other essential components such as antibiotic resistance genes (ARGs).

The dry lab-wet lab pipeline consists of exhaustively generating and testing candidate AI-generated sequences for purpose and performance. As numerous and diverse plasmid components are being generated and evaluated at any time, high throughput experimentation workflows, tools, and best practices must be developed to support team operations. This has the added benefit of expediting use of our dry lab-wet lab pipeline by other researchers or in industry. The hardware team's mandate is to develop such workflows and tools.

Sample storage and management is a perennially present problem across all kinds of labs. Laboratories at the University of Toronto typically use HECHMET for laboratory inventory management [8], which is a barcode based check-in/check-out system. Although this solution is optimised for generic chemicals, it creates high overhead for small samples with short object life cycles. To enable high throughput operations, we are developing a solution aimed at tackling managing large numbers of samples, potentially across different locations.

The sample manager product system will consist of one or more holders for samples with sensors and wireless communication capabilities, a server for data retention and information transmission, and an application endpoint for data display and check-in/check-out functionality. Wet lab team input and solutions used in the manufacturing industry were incorporated. Cost and ease of use were considered as part of the design process.

We based the product system based on a self-driving lab architecture described in [9]. Although initially developed for a different purpose, we are adapting the architecture mentioned to form a readily extendable internet of things (IoT) framework for developing future networked laboratory automation projects. Furthermore, data collected from usage trends identified here may be fed into the Lean and Six Sigma process design strategies, improving wet lab team parallelism while improving quality and decreasing variability.

Most assays done by the wet lab team will require some form of growth quantification, in particular for in vitro oriV validation. Colony counting and OD600 assays are two such approaches. However, the former requires significant repetitive effort and time, while the latter requires specialised equipment in the form of a spectrophotometer/nephelometer [10]. Advancements in computer vision, however, have enabled object detection models to succeed in the former task [11].

We intend to implement and finetune a vision transformer model to perform colony detection on petri dish plates. Furthermore, we plan to deploy the vision transformer to a single chip computer, which will be housed on a benchtop stand with a camera and related accessories. The whole assembly will be internet of things (IoT) accessible. Considerations made by the team include cost, throughput, ease of construction, and eventual open source availability.

A laboratory and a factory share a similar goal in a sense that they are both production facilities for material objects. In our case, our wet lab floor can be likened to a factory floor in a sense that we are trying to optimize the wet lab team's throughput by as much as possible. As such, why not apply industrial engineering techniques to the lab? The hardware team applied process engineering techniques, such as Lean, Six Sigma, and DMAIC, to the wet lab floor, with the aim of reducing variability and increasing throughput [12].

Three major batches of sequences were generated and validated. Other batches (such as attempts with diffusion models) were produced but did not show sufficient promise to make it to the validation pipeline.

Batch 1: The first batch of plasmids was generated using a Mamba model trained on a smaller dataset of less than 50,000 plasmids. This initial model used nucleotide-level tokenization, which limited its ability to capture longer-range dependencies in the plasmid sequences. An initial set of 6 oris (a mix of Class A and Class B) was selected. However, after attempting to design these plasmids in the wet lab, Class A plasmids were deemed to be too difficult to test at this stage, as their machinery is more complex and difficult to detect. This wet Lab - dry Lab iteration meant that instead 5 promising Class B oris were delivered to wet lab for testing. Finally, one promising protein was also selected. None of the oris in this batch worked.

Batch 2: For the second batch, we transitioned to the Mamba2 architecture and implemented a Byte Pair Encoding (BPE) tokenizer. We also significantly expanded our training dataset. 8 Class B oris and 5 proteins were selected from the validation pipeline. At least 6 oris from this batch have shown promising results as of the time of writing.

Batch 3: For this batch, we focused on exploring the model's capability to generate plasmids with specific, user-defined characteristics. We also further refined our training dataset and fine-tuning stage using results of the wet lab experiments. 9 oris were selected for experimental validation. These oris have not been fully tested in vitro yet.

In addition to developing the foundational technology of writing new DNA sequences, our team seeks to lay the groundwork for this project to have a positive social impact. This allows our work to benefit the communities around us, which is an important aspect of synthetic biology and engineering. As such, our Human Practices (HP) and Entrepreneurship teams have undertaken extensive work to determine the most suitable project application for Plasmid.AI – at this stage, we have honed in on antibiotic resistance.

HPs objective is to integrate the innovative capabilities of Plasmid AI with human practices to effectively address the pressing challenge of antibiotic resistance (ABR). Through careful background research, we've identified ABR as a critical issue that demands not just attention but innovative and thoughtful solutions (see Integrated Human Practices section of wiki). We understand that this challenge goes beyond the laboratory; it deeply impacts the lives of individuals and communities. Thus, we are committed to a holistic approach that embraces insights from a diverse range of stakeholders.

As we navigate the rapidly evolving landscape of AI, we are committed to establishing a regulatory framework that draws on existing models to ensure responsible and ethical use of our technology. In tandem with this, we remain aware of our ecological footprint. Collaborating with an Associate Engineer from the Ministry of Environment, we are dedicated to developing strategies that minimize our environmental impact while advancing our mission.

We believe education is a powerful catalyst for change. This year, we welcomed incoming freshmen through seminars that illustrate the exciting possibilities within synthetic biology and highlight our work with Plasmid AI. In partnership with the University of Toronto's Science Communication Club, we're crafting articles that not only address antibiotic resistance but also encourage broader discussions about its implications. Through these initiatives, we aim to spark meaningful conversations, engage communities, and create well-rounded solutions to this pressing global challenge.

Antibiotics are crucial in treating bacterial infections, and have prevented millions of deaths since the discovery of penicillin in 1928 [13]. However, misuse of these drugs in medicine and agriculture, combined with a dry antibiotic pipeline [14], has led to the generation of superbugs – namely, bacteria that can withstand antibiotic treatment [15]. This, coupled with broader drug resistance from other microbes (viruses, protists, etc.), constitutes a phenomenon called antimicrobial resistance (AMR), which represents one of the most pressing global health challenges humanity currently faces. According to the World Health Organization (WHO), AMR is projected to cause 10 million deaths per year by 2050, cost more than $100 trillion in lost economic output, and drastically increase poverty rates [16].

Many organizations have developed watchlists for the most concerning superbug bacteria, notably the ESKAPE pathogens and the WHO's critical priority pathogens list [17]. In particular, pathogens like Carbapenem-resistant Enterobacteriaceae (CRE) and extensively drug-resistant tuberculosis (XDR-TB) are at the forefront of antibacterial research.

Multiple treatment methods currently exist to tackle the pathogens above, but all with limited efficacy:

A strategy to counter superbugs involves creating AI-generated plasmids that can outcompete the natural plasmids carrying antibiotic resistance genes. While all plasmids impose a metabolic burden on bacteria, those harbouring genes conferring antibiotic resistance are strongly selected for due to their survival advantage in the presence of antibiotics. Designing AI-generated plasmids that provide an even stronger selective advantage will cause bacteria to replace the resistance-carrying plasmids with our engineered versions.

When antibiotic pressure is removed, bacteria no longer have a selective reason to retain plasmids carrying resistance genes, making them once again susceptible to antibiotics. Our AI-generated plasmids function to aid the bacteria in accelerating this process of discarding the plasmids carrying antibiotic resistance by providing an even more favourable alternative.

Specifically, more favourable plasmids are less metabolically burdensome, easier to replicate, and capable of directly competing with existing resistance plasmids. By generating AI-plasmid components with these characteristics, we can resensitize superbugs to antibiotic resistance via removal of its plasmid-based antibiotic resistance.

Our objective is to create a go-to market stage company for our AI-powered discovery engine that integrates our proprietary foundational model (dry lab) with rapid in vitro screening and clinical validation (wet lab). Our company will markedly enhance the preclinical screening of novel phages to create superbug therapies faster and more cost-effectively than competitors.

We're co-incubating with 2 independent incubators located in Canada and the UK. These incubators have allowed us to refine our problem statement and understand our value proposition. We've developed a rapid fire 3-minute pitch, a 6-minute investor ready pitch, a robust business plan, and other deliverables necessary to raise seed-stage capital and beyond.