Industrial insights

Since establishing contact with VJTbio in July, we have engaged in multiple friendly exchanges and discussions, establishing a close contact. Our communication with VJTbio has been continuous throughout the entire project, and at each stage of the project, our collaboration with the company has provided us with significant inspiration. It has helped us clarify our direction and insist on our original intention of creating an iGEM project that is meaningful to the industry and the world.

online communication

We established WeChat groups with key members of the company, frequently seeking advice, sharing insights, and learning from one another within the group. Online communication is highly convenient, and we have also received warm and friendly treatment from the company. We are deeply grateful for the patient guidance and numerous inspirations provided by company.

We have had discussions regarding the reliability of antibody de novo design and the recognition of antibody de novo design within the industry. We have come to realize that the application of antibody de novo design is not yet mature, and there is also a lack of recognition for de novo design within the industry. AI companies often obtain a large number of positive antibody sequences as a training dataset and then use this as a basis for prediction to design antibodies with higher affinity. So we also chose extending the spieces derived from murine antibodies with AI methods in the project instead of de novo design.

Regarding the verification of our project, we also consulted the company about the verification method in the industry. The company introduced a mainstream method to us, which is to express and purify the full-length antibody sequence including the designed FR and CDR, and then perform Protein Protein Interaction Experiments with antigen protein. Through such communication, we decided to add structure score of the antibody sequence to our project, which is divided into two parts: on the one hand, whether the antibody sequence can be folded into a normal structure; on the other hand, we tend to maintain the original structure to retain antigen-antibody binding.

We are very thankful for the patient guidance and numerous inspirations provided by company.

Other biotech company

As students with an information background, we believe that it's necessary for us to enhance our understanding of synthetic biology and gain insights into the industry by conducting on-site visits. This will further improve our iGEM project.

In addition to visiting VJTbio, we have also visited numerous biotechnology-related companies throughout the season, gaining valuable experiences from these field trips.

During the visit and exchange at QiTan Tech on April 6th, we not only toured many of the company's bioinstrumentation, but also gained a deep understanding of the structure, properties, and meaningful applications of nanopore proteins. The company's years of efforts in optimizing nanopore proteins are precisely what have led to the company's current level of accuracy in genetic sequencing.

We engaged in thorough discussions concerning protein optimization and design. During these discussions, we conveyed our team's interests and the directions we intend to explore. Also, we learned that the company's algorithms in designing and optimizing nanopore proteins have room for improvement, and our team believed that there is potential for collaboration in this area with the company. However, through further dialogue, a consensus was reached between our team and the company. Namely, the company was focusing on enhancing sequencing accuracy, and even incremental progress from 99% to 99.5% holds significant importance within the industry. Nonetheless, this endeavor demands substantial time for accumulation and rigorous validation. As undergraduate students, rather than graduate students, we probably should embark on tasks that leverage the creativity and imagination more inherent to undergraduates.

Furthermore, we also engaged in discussions with the company about the breakthroughs in the field of de novo protein design in 2023. During the exchange, we learned that "de novo protein design" is highly challenging, currently achievable mainly by the David Baker research group. This realization subsequently led us to consider focusing on antibodies with more predictable sequences and structures, and continuing our efforts in the design and optimization of antibodies. It was after this that we began to pay attention to learning many aspects of antibodies, including the antibody structure prediction tool IgFold, the progress of de novo design of antibody CDR-H3 loops, antibody language models, etc.

Zhongguancun Life Science Park in Beijing

We specifically visited the Zhongguancun Life Science Park, which focuses on cutting-edge tracks such as cell and gene therapy, innovative pharmaceuticals, AI-driven drug development, AI-assisted diagnostics, and more. We visited some companies within the park, with a particular emphasis on holding in discussions with Beijing Syngentech and BeiCell Biotechnology.

Through engagements with both companies on May 28th, we not only toured many of the company's biotech instruments, enriching our biological background knowledge, but also gained a deeper understanding of the role of synthetic biology in biomedicine and life health. We learned that AI has already played a significant role in new drug development. What's more, both companies have utilized AI to optimize wet lab processes, making substantial contributions to clinical diagnosis, treatment, and drug development.

After this visit, we have all developed a strong interest in the role of AI in facilitating new drug development. Through our further research, we found that antibody drugs constitute a significant portion of novel pharmaceuticals. We hope our project can contribute to the optimization and design of antibody drugs and ultimately established a close collaboration with VJTbio as previously described.

Additionally, during the exchange, we also acknowledged the critical significance of "safety" in the new drug development process. In our project, we placed significant emphasis on incorporating measures to ensure the security of data sources and the integrity of algorithmic output sequences. We hold the view that comprehending how AI models predict and optimize antibody sequences and characteristics is crucial for ensuring their reliability and safety, so we pay more attention to develop more interpretable codes and algorithms of AI models in the process of designing and optimizing antibodies.

Scholarly exchanges

On April 19th, we had the privilege of inviting Yinglu Cui from the research group of Professor Bian Wu at the Chinese Academy of Sciences to visit Tsinghua University. We held a group seminar where we engaged in extensive discussions regarding protein design and enzyme engineering.

Professor Cui introduced to us the research undertaken by their group in the field of "computational-enzyme design." We reviewed the history of protein optimization and design, highlighting areas such as rational design, semi-rational design, directed evolution, etc. She outlined the various breakthroughs and ongoing challenges in the transition to the current phase of computational design. Further, she provided insights into the prospects of applying computer-designed non-natural enzymes in the field of biomedicine.

Our team presented our research and algorithm designs in protein optimization, sharing some of the algorithmic models we've experimented with. Professor Cui shared valuable insights from her lab's experience in this field, pointing out that the one-hot encoding we've been using lacks universality. Instead, she suggested we should include pre-encoding work and emphasized the importance of assessing the generalization capability when researching model algorithms in the protein design realm. She noted that some current models, such as DLKcat, suffer from weak generalization abilities. In response to Professor Cui's suggestions, we systematically worked on improvements, addressing each point. Especially in terms of pre-encoding work, we have done a lot of pre-encoding research and practical attempts. Through constant comparison and trade-offs, we finally chose BLOSUM62 in our project, which is the most suitable for our project, taking into account a large number of rare species.

After communicating with several protein design and optimization companies, we have transitioned our focus from protein optimization design to antibody optimization design. However, we haven't yet identified a specific task. Before establishing a concrete plan, we learned about recent breakthroughs in using artificial intelligence for de novo antibody design and extensively reviewed relevant literature. Then, we discussed the details with Professor Xie. After listening to our existing research and algorithm attempts, as well as our interest in antibody drugs, he pointed out a meaningful background for us: humanization of murine antibodies. We involved in thorough discussions about the context, significance, and implications of such theme.

We further investigated antibody humanization and found that traditional humanization techniques primarily involve a series of biological tools and simulation computations implemented on computers, whereas machine learning deep learning methods are less used in the field, and are mainly employed for binary classification and humanization scoring of humanized antibodies. We aim to delve deeper into using machine learning and deep learning tools to address challenges in this domain, incorporating considerations for antibody structures. When we propose a preliminary technical route, we once again communicated with Professor Xie, who acknowledged our idea and expressed anticipation for us to construct a comprehensive automated tool and add our innovative ideas.

Ziting Zhang also pursued her undergraduate studies in the Department of Automation at Tsinghua University. Currently, her research interests converge at the intersection of bioinformatics. As a scholar in immunology, she believes that our team's project this year holds considerable meaning. Through exchanges with her, we gained a deeper understanding of immunological knowledge related to antibodies and how to leverage the advantages of our team's background in information. Through this exchange, we spent more energy in the project to improve our algorithm. On the one hand, due to our proficiency in algorithms and rich experience in tuning parameters, we have extensively tested many machine learning and deep learning models, and finally selected the most reasonable one by comparing the performance and mathematical principles of each algorithm. On the other hand, through our background knowledge and algorithm experience, we have done a lot of fine-tuning to make the whole model more perfect.

Antibodies are divided into constant and variable regions, within the variable region are the high-frequency mutated CDR regions (where the antigen comes into contact with the antibody), and the relatively conserved and species-specific FR regions. As the antigen-antibody binding region, CDR has garnered significant attention in the academic field, and there are a lot of difficulties in the design of CDR region, especially CDR-H3 loop.

We shared the details of our project with Kong, and he also expressed interests in developing similar software in the past. We also inquired about the databases currently used for CDR design, sought guidance on the essential aspects of designing CDR regions from sequence to structure, and learned about some details of Equivariant Graph Neural Networks (EGNs). Through this exchange, we gained insights into the unique characteristics of CDR regions and came to a full realization of the high variability of the CDR H3 loop. Based on this understanding, we made improvements to the structural scoring aspect of our project.

A large proportion of our data comes from rare species, which makes it difficult for conventional algorithms to avoid biasing species with a large amount of data, and cannot weigh these rare species. We discovered a method called OCLR (One Class Logistic Regression) while reading the paper and found the author, Dr. Guo. Since we want to apply One Class Logistic Regression to scoring antibody sequences from rare species, we introduced our project and consulted with Dr. Guo about the key advantages of the one-class approach. Dr. Guo noted that the approach exhibits excellent scalability, making it highly suitable for our application scenario.

Dr. Guo also provided us with a lot of advice on the application of OCLR, and walks us through the details of how he uses OCLR: by training the existing cell type data, the weight vector corresponding to each cell type can be obtained, and then for each test cell, the correlation coefficient of the cell type weight vector can be used to predict which cell type it belongs to. Inspired by this approach, we also want to train a feature vector for each species, which will greatly help us score rare species.

In the use of the oneclass model, we also sought advice from Zeyu Chen, who from the same research group with Dr. Guo. He recommended that we first experiment with methods such as SVM one class and IsolationForest, which are more commonly used. During practical experimentation, we observed that the performance of these traditional machine learning algorithms was not satisfactory.

After discussing the results we obtained with Chen again, we realized that the aforementioned traditional machine learning algorithms could be repurposed for outlier detection and integrated into our project to optimize it, using in filtering refenrence dataset for example. Following several rounds of comparative validation, we selected Local Outlier Factor as the method for outlier detection. Additionally, we embarked on constructing our own deep one class algorithm to score antibody sequences, aiming to establish a more suitable algorithmic model for our project.

Team collaborations

CCiC (10th Conference of China iGEMer Community)

We listened to presentations from dozens of teams, both online and offline, as they shared their projects. We also engaged in separate discussions with over ten teams. Through interactions with other teams, we gained clearer insights into certain aspects and unique contributions of our own project, further identifying some oversights along with our distinctive features. We also made new acquaintances with many software teams in China, and discussed the follow-up Human Practice activities together, such as Synbio Plus AI Conference.

It is worth mentioning that in our exchange programs with other teams, SJTU-software team raised the question, "With such progress in de novo antibody design, why should we still use murine antibody humanization to design human antibodies?" This prompted us to reflect on whether our project has limitations. Through further communication with VJTbio Company, we learned that the industry currently considers de novo AI design to be unreliable, while techniques like humanization and canineization of antibodies are mainstream approaches. Nevertheless, we realized that our previous focus solely on humanization indeed had constrictions. After many visits and exchange activities mentioned above, we have synthesized the suggestions of all parties, carried out careful consideration and technical feasibility validation, and finally decided to extend our Topic towards antibody species-specification.

Exchange with PekingHSC

When we were at CCiC, we had a preliminary understanding of PekingHSC's projects via poster and project sharing. Due to the geographical advantage, we visited Peking University's main campus on July 24th and engaged in in-depth discussions with students from the School of Medicine at Peking University. We once again presented the details of our project and engaged in comprehensive and thorough discussions covering various aspects such as Human Practice, Education, Safety, and even the process of obtaining visas for travel to France during 2023 Grand Jamboree.

One of our team's mentors has deep-rooted experience in the field of oncolytic virus combating tumors, which is PekingHSC's topic, and possesses substantial industry knowledge. Therefor, we introduced Professor Xie Zhen to them, and they conducted discussions focusing on virus safety and other related topics. Members of PekingHSC also introduced us to some laboratories at Peking University dedicated to antibody research, which is very helpful for us to understand the antibody knowledge related to the project. Both of our teams assisted each other, gained significant insights, and forged lasting friendships.

Synbio Plus AI Conference

We communicated with other teams about the applications of AI in synthetic biology, with a focus on introducing the algorithmic models and considerations used in our project. Many teams utilized pre-trained models, and we shared some insights regarding generalization capabilities: emphasizing the alignment between the data used during pre-training and the data employed in actual project is crucial for achieving better results.

In addition, we placed particular emphasis on our team's concern regarding the safety aspects of utilizing AI in new drug development, along with a series of measures. We shared our insights derived from discussions with multiple companies. We believe that the current limitations of AI encompass: dataset quality, interpretability, and security. Similar to the perspective of a member of SYSU-Software, we concur that sometimes data holds more significance than the algorithms themselves. When applying AI technology to antibody design, limitations in data quality and quantity might constrain the potential risks associated with the designed antibodies. Moreover, comprehending how AI models predict and optimize antibody sequences and characteristics is crucial for ensuring their reliability and safety. However, prevailing deep learning models often possess intricate structures and parameters, making it difficult to elucidate the rationales behind their decisions, which raises doubts about their reliability and trustworthiness. Hence, our team is also devoted to developing more interpretable models and algorithms, aiming to provide explanations for the decision-making processes and outcomes of AI models in antibody design, hoping that the recognition of AI-designed drugs in the industry will be slightly improved.

During the conference, we also delved into the ethical challenges AI faces, as well as the relevant legal regulations and academic standards. This prompted us to engage in further reading about the legal regulations and academic standards related to AI after the discussions concluded. This, in turn, facilitated the refinement of our project.

Responsibilities and Benefits

In the process of our Human Practice, we had in-depth exchanges with biopharmaceutical companies such as VJTbio to understand the actual needs of the industry, and also discussed the direction of the project with researchers in the field of immunology.

Our project hopes to help those groups who are committed to pet care and pet rescue. We hope to provide faster, AI-assisted antibody drug development auxiliary technology for cats, dogs and other pets.

Our project hopes to help those groups who are committed to developing animal husbandry and developing animal husbandry drugs. We hope to accelerate the development of antibody drugs for various livestock and poultry such as cattle, sheep, pigs and chickens, and hope to use antibody drugs with fewer side effects and mild and long-lasting effects.

Our project hopes to help all groups who care about animal drug experiments. We hope to use deep learning tools to accelerate the development of multi-species antibody drugs. We hope to maximize the role of existing antibody drugs and computer-assisted capabilities, and accelerate the development of antibody drugs as much as possible, and reduce inefficient animal experiments in antibody drug development.

We have evaluated various types of risks, reflecting the responsible nature of our design. We have also fully validated the project, with a strong emphasis on the safety of training data and the reliability of output sequences. Antibody drugs generally do not pose a risk of drug-drug interactions with common small-molecule drugs. Due to their high specificity, they are less likely to have adverse effects on normal tissues, resulting in a lower risk of adverse reactions. Moreover, their lower immunogenicity reduces the risk of allergic reactions and immune-related adverse events. For example, in the case of antibiotics, which are widely used for animal treatments, their extensive use may not only kill beneficial bacteria but also lead to the development of antibiotic resistance in bacteria, making subsequent treatments more challenging. Antibody drugs, on the other hand, do not face these challenges and can also be used to treat diseases caused by other pathogens such as viral infections. Our project is also trying to reduce the immunogenicity of the output sequence to the target species as much as possible, making the drug safer.

Public Survey: Animal Antibody Drug Awareness Survey

We published questionnaires in Chinese and English and conducted extensive surveys. Of questionnaires collected 36% had some knowledge of how animals are treated and what happens to them, and 39% of the respondents had sought medical assistance for their animals.

Yet most of them have used antibiotic drugs in the process.

At the same time, respondents using antibiotic drugs were not well informed about the replacement of antibiotic drugs by antibody drugs.

But one does not exclude the use of antibody drugs as an alternative to antibiotic drugs for treatment, which are more harmful to both animals and the environment.

and are willing to pay a higher price for it.

It can be seen from the survey that most people are willing to use antibody drugs to treat animals, and are even willing to pay more expensive prices for it. However, antibody drugs are still scarce. This reflects the significance of our project. In the questionnaires we received, there were no concerns about our project. Regarding possible hidden dangers, the drugs we design play a computer-assisted role for researchers, and drugs need to be fully verified before the drugs are actually approved for marketing.

Team activity: Close contact with animals

We also organized activities to create opportunities for team members to have close contact with animals. In early September we headed to a cat cafe and enjoyed a morning with the adorable cats. We also communicated with the cat cafe owner about the cats’ daily health. Cute cats inspire compassion and affection in us, and we feel even more deeply our responsibility and the benefits of our projects.